{"manifest":{"format":"knogra-workspace","version":"2.0","appVersion":"1.6.0","createdAt":"2026-08-21T11:05:24.304Z","name":"graph-modern-ai-technology-2026-08-21"},"graph":{"nodes":[{"id":"n-1784-815-828-256","title":"Sequential vs. Parallel Processing","tags":[],"properties":{},"createdAt":"2026-07-23T14:10:28.256Z","updatedAt":"2026-07-23T14:19:10.078Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-815-855-132","title":"The Transformer Block","tags":[],"properties":{},"createdAt":"2026-07-23T14:10:55.132Z","updatedAt":"2026-07-23T14:19:10.079Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-815-862-176","title":"Multi-Head Attention","tags":[],"properties":{},"createdAt":"2026-07-23T14:11:02.176Z","updatedAt":"2026-07-23T17:42:53.251Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-816-030-709","title":"Scaled Dot-Product Attention","tags":[],"properties":{"equation":"\\text{Attention}(Q, K, V) = \\text{softmax}\\left(\\frac{QK^T}{\\sqrt{d_k}}\\right)V"},"createdAt":"2026-07-23T14:13:50.709Z","updatedAt":"2026-07-23T17:53:38.429Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-816-091-547","title":"Sparse Attention","tags":[],"properties":{},"createdAt":"2026-07-23T14:14:51.547Z","updatedAt":"2026-07-23T14:16:03.896Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-816-097-653","title":"Cross-Attention","tags":[],"properties":{},"createdAt":"2026-07-23T14:14:57.653Z","updatedAt":"2026-07-23T14:16:03.897Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-816-099-231","title":"Softmax Bottleneck","tags":[],"properties":{},"createdAt":"2026-07-23T14:14:59.231Z","updatedAt":"2026-07-23T17:53:43.564Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-816-107-389","title":"Grouped-Query Attention (GQA)","tags":[],"properties":{},"createdAt":"2026-07-23T14:15:07.389Z","updatedAt":"2026-07-23T14:18:50.176Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-816-112-838","title":"Query-Key-Value (QKV) Projection","tags":[],"properties":{},"createdAt":"2026-07-23T14:15:12.838Z","updatedAt":"2026-07-23T14:18:37.024Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-816-384-703","title":"The Turing Trap","tags":[],"properties":{},"createdAt":"2026-07-23T14:19:44.703Z","updatedAt":"2026-07-23T17:30:16.883Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-823-059-854","title":"Universal Turing Machine","tags":[],"properties":{},"createdAt":"2026-07-23T16:10:59.854Z","updatedAt":"2026-07-23T16:46:10.704Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-823-066-853","title":"Recursive Self-Improvement","tags":[],"properties":{},"createdAt":"2026-07-23T16:11:06.853Z","updatedAt":"2026-07-23T16:46:10.705Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-823-072-505","title":"Neuro-Symbolic Integration","tags":[],"properties":{},"createdAt":"2026-07-23T16:11:12.505Z","updatedAt":"2026-07-23T17:32:45.894Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-823-083-572","title":"Moravec’s Paradox","tags":[],"properties":{},"createdAt":"2026-07-23T16:11:23.572Z","updatedAt":"2026-07-23T16:51:45.369Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-823-094-285","title":"Agency & Long-Horizon Planning","tags":[],"properties":{},"createdAt":"2026-07-23T16:11:34.285Z","updatedAt":"2026-07-23T18:34:40.261Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-823-267-623","title":"Phase Transitions","tags":[],"properties":{},"createdAt":"2026-07-23T16:14:27.623Z","updatedAt":"2026-07-23T18:39:11.074Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-823-344-722","title":"Induction Heads","tags":[],"properties":{},"createdAt":"2026-07-23T16:15:44.722Z","updatedAt":"2026-07-23T17:44:47.975Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-824-073-339","title":"ImageNet Challenge","tags":[],"properties":{},"createdAt":"2026-07-23T16:27:53.339Z","updatedAt":"2026-07-23T16:28:08.694Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-824-132-453","title":"The XOR Problem","tags":[],"properties":{},"createdAt":"2026-07-23T16:28:52.453Z","updatedAt":"2026-07-23T16:36:03.767Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-824-173-383","title":"David Rumelhart","tags":[],"properties":{},"createdAt":"2026-07-23T16:29:33.383Z","updatedAt":"2026-07-23T16:36:03.763Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-824-174-809","title":"James McClelland","tags":[],"properties":{},"createdAt":"2026-07-23T16:29:34.809Z","updatedAt":"2026-07-23T16:36:03.762Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-824-181-993","title":"Parallel Distributed Processing (PDP)","tags":[],"properties":{},"createdAt":"2026-07-23T16:29:41.993Z","updatedAt":"2026-07-23T16:39:39.162Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-824-369-059","title":"NETtalk","tags":[],"properties":{},"createdAt":"2026-07-23T16:32:49.059Z","updatedAt":"2026-07-23T16:36:03.762Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-824-503-335","title":"Graceful Degradation","tags":[],"properties":{},"createdAt":"2026-07-23T16:35:03.335Z","updatedAt":"2026-07-23T16:35:34.299Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-824-668-566","title":"Hebbian Theory","tags":[],"properties":{},"createdAt":"2026-07-23T16:37:48.566Z","updatedAt":"2026-07-23T16:41:41.307Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-824-758-795","title":"Universal Approximation Theorem","tags":[],"properties":{},"createdAt":"2026-07-23T16:39:18.795Z","updatedAt":"2026-07-23T16:41:41.307Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-824-876-198","title":"Synaptic Plasticity","tags":[],"properties":{"comment":"The biological process by which specific patterns of synaptic activity result in changes in synaptic strength, serving as the neurochemical basis for learning and memory."},"createdAt":"2026-07-23T16:41:16.198Z","updatedAt":"2026-07-23T16:48:21.593Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-825-281-004","title":"Differentiable Logic","tags":[],"properties":{},"createdAt":"2026-07-23T16:48:01.004Z","updatedAt":"2026-07-23T16:50:43.484Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-825-283-108","title":"Systematicity","tags":[],"properties":{},"createdAt":"2026-07-23T16:48:03.108Z","updatedAt":"2026-07-23T16:50:43.484Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-825-371-425","title":"Symbol grounding problem","tags":[],"properties":{},"createdAt":"2026-07-23T16:49:31.425Z","updatedAt":"2026-07-23T17:05:32.580Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-825-477-825","title":"Intrinsic vs. Borrowed Grounding","tags":[],"properties":{},"createdAt":"2026-07-23T16:51:17.825Z","updatedAt":"2026-07-23T16:51:45.370Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-825-570-102","title":"Functional vs. Phenomological Understanding","tags":[],"properties":{},"createdAt":"2026-07-23T16:52:50.102Z","updatedAt":"2026-07-23T17:00:36.039Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-826-096-398","title":"Tree-of-Thought Reasoning","tags":[],"properties":{},"createdAt":"2026-07-23T17:01:36.398Z","updatedAt":"2026-07-23T17:05:32.581Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-826-102-465","title":"Simulators Hypothesis","tags":[],"properties":{},"createdAt":"2026-07-23T17:01:42.465Z","updatedAt":"2026-07-23T17:05:32.581Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-826-107-744","title":"Linear Representation of Features","tags":[],"properties":{},"createdAt":"2026-07-23T17:01:47.744Z","updatedAt":"2026-07-23T17:05:32.582Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-826-169-536","title":"World Models","tags":[],"properties":{},"createdAt":"2026-07-23T17:02:49.536Z","updatedAt":"2026-07-23T17:05:32.582Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-826-493-490","title":"Liability & Accountability","tags":[],"properties":{},"createdAt":"2026-07-23T17:08:13.490Z","updatedAt":"2026-07-23T18:36:40.898Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-827-044-159","title":"CUDA Moat","tags":[],"properties":{},"createdAt":"2026-07-23T17:17:24.159Z","updatedAt":"2026-07-23T17:18:10.140Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-827-051-233","title":"CoWoS","tags":[],"properties":{},"createdAt":"2026-07-23T17:17:31.233Z","updatedAt":"2026-07-23T17:18:10.141Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-827-060-218","title":"Inference-Compute Trade-off","tags":[],"properties":{},"createdAt":"2026-07-23T17:17:40.218Z","updatedAt":"2026-07-23T17:18:10.141Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-827-071-779","title":"Sovereign AI","tags":[],"properties":{},"createdAt":"2026-07-23T17:17:51.779Z","updatedAt":"2026-07-23T17:18:10.142Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-827-122-695","title":"Inference vs. Training","tags":[],"properties":{},"createdAt":"2026-07-23T17:18:42.695Z","updatedAt":"2026-07-23T17:19:12.351Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-827-369-389","title":"Unit Economics of Intelligence","tags":[],"properties":{},"createdAt":"2026-07-23T17:22:49.389Z","updatedAt":"2026-07-23T17:23:16.295Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-827-452-679","title":"First AI Winter (1974–1980)","tags":[],"properties":{},"createdAt":"2026-07-23T17:24:12.679Z","updatedAt":"2026-07-23T17:29:28.571Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-827-454-845","title":"The Lighthill Report","tags":[],"properties":{},"createdAt":"2026-07-23T17:24:14.845Z","updatedAt":"2026-07-23T17:29:28.572Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-827-458-834","title":"Second AI Winter (1987–1993)","tags":[],"properties":{},"createdAt":"2026-07-23T17:24:18.834Z","updatedAt":"2026-07-23T17:29:28.572Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-827-474-296","title":"Expert Systems","tags":[],"properties":{},"createdAt":"2026-07-23T17:24:34.296Z","updatedAt":"2026-07-23T17:29:28.573Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-827-480-696","title":"Expert Systems Collapse","tags":[],"properties":{},"createdAt":"2026-07-23T17:24:40.696Z","updatedAt":"2026-07-23T17:29:28.574Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-827-552-241","title":"Marvin Minsky","tags":[],"properties":{},"createdAt":"2026-07-23T17:25:52.241Z","updatedAt":"2026-07-23T17:29:28.574Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-827-564-110","title":"Perceptrons (1969)","tags":[],"properties":{},"createdAt":"2026-07-23T17:26:04.110Z","updatedAt":"2026-07-23T17:29:28.575Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-827-930-093","title":"Inference-Time Search","tags":[],"properties":{},"createdAt":"2026-07-23T17:32:10.093Z","updatedAt":"2026-07-23T18:34:40.261Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-828-107-362","title":"PPO","tags":[],"properties":{},"createdAt":"2026-07-23T17:35:07.362Z","updatedAt":"2026-07-23T17:38:44.335Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-828-173-143","title":"Richard Sutton","tags":[],"properties":{},"createdAt":"2026-07-23T17:36:13.143Z","updatedAt":"2026-07-23T17:38:44.333Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-828-178-544","title":"Andrew Barto","tags":[],"properties":{},"createdAt":"2026-07-23T17:36:18.544Z","updatedAt":"2026-07-23T17:38:44.335Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-828-196-325","title":"Temporal Difference (TD) Learning","tags":[],"properties":{"equation":"V(s_t) \\leftarrow V(s_t) + \\alpha [r_{t+1} + \\gamma V(s_{t+1}) - V(s_t)]"},"createdAt":"2026-07-23T17:36:36.325Z","updatedAt":"2026-07-23T17:38:44.336Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-828-609-869","title":"Superposition","tags":[],"properties":{},"createdAt":"2026-07-23T17:43:29.869Z","updatedAt":"2026-07-23T17:44:47.974Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-828-743-395","title":"Cosine Similarity","tags":[],"properties":{"equation":"\\frac{\\mathbf{u} \\cdot \\mathbf{v}}{\\|\\mathbf{u}\\| \\|\\mathbf{v}\\|}"},"createdAt":"2026-07-23T17:45:43.395Z","updatedAt":"2026-07-23T17:47:26.273Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-828-782-607","title":"One-Hot Encoding","tags":[],"properties":{},"createdAt":"2026-07-23T17:46:22.607Z","updatedAt":"2026-07-23T17:47:26.274Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-828-813-272","title":"Embedding Layer","tags":[],"properties":{"equation":"\\mathbf{e} = \\mathbf{x}W_E","comment":"The first layer in a transformer that maps one-hot encoded token indices into dense, trainable vectors."},"createdAt":"2026-07-23T17:46:53.272Z","updatedAt":"2026-07-23T17:47:26.275Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-828-881-856","title":"Approximate Nearest Neighbor (ANN)","tags":[],"properties":{},"createdAt":"2026-07-23T17:48:01.856Z","updatedAt":"2026-07-23T17:49:31.095Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-828-956-557","title":"HNSW (Hierarchical Navigable Small World)","tags":[],"properties":{},"createdAt":"2026-07-23T17:49:16.557Z","updatedAt":"2026-07-23T17:49:31.096Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-828-965-523","title":"Hybrid Search","tags":[],"properties":{},"createdAt":"2026-07-23T17:49:25.523Z","updatedAt":"2026-07-23T17:49:31.097Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-829-052-575","title":"Confabulation","tags":[],"properties":{},"createdAt":"2026-07-23T17:50:52.575Z","updatedAt":"2026-07-23T17:53:43.569Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-829-158-925","title":"Sycophancy","tags":[],"properties":{},"createdAt":"2026-07-23T17:52:38.925Z","updatedAt":"2026-07-23T17:53:43.570Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-829-870-984","title":"KV Cache","tags":[],"properties":{},"createdAt":"2026-07-23T18:04:30.984Z","updatedAt":"2026-07-23T18:04:40.474Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-829-904-051","title":"Few-Shot Prompting","tags":[],"properties":{},"createdAt":"2026-07-23T18:05:04.051Z","updatedAt":"2026-07-23T18:07:19.710Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-829-951-047","title":"Prompt Optimization","tags":[],"properties":{},"createdAt":"2026-07-23T18:05:51.047Z","updatedAt":"2026-07-23T18:06:28.055Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-830-009-038","title":"Persona Adoption","tags":[],"properties":{},"createdAt":"2026-07-23T18:06:49.038Z","updatedAt":"2026-07-23T18:07:12.959Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-830-080-875","title":"Lost in the Middle","tags":[],"properties":{},"createdAt":"2026-07-23T18:08:00.875Z","updatedAt":"2026-07-23T18:19:52.415Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-830-838-752","title":"In-Context Optimization","tags":[],"properties":{"comment":"This addresses the hypothesis that the forward pass of a transformer is actually performing an internal version of an optimization algorithm."},"createdAt":"2026-07-23T18:20:38.752Z","updatedAt":"2026-07-23T18:20:54.728Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-831-088-029","title":"Chinchilla Scaling","tags":[],"properties":{"equation":"C \\approx 6ND","comment":"The specific empirical finding from DeepMind that redirected the industry to balance parameter count and data volume more efficiently."},"createdAt":"2026-07-23T18:24:48.029Z","updatedAt":"2026-07-23T18:25:45.195Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-831-335-145","title":"MCP Servers","tags":[],"properties":{"comment":"The 'server' is the specific architectural component in the protocol that hosts the tools and data, acting as the bridge between the AI and the local or remote resource."},"createdAt":"2026-07-23T18:28:55.145Z","updatedAt":"2026-07-23T18:34:40.262Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-831-395-200","title":"Function Calling","tags":[],"properties":{"comment":"The specific technical implementation where a model outputs structured JSON to signal a tool request, which is the foundation of most current tool-use implementations."},"createdAt":"2026-07-23T18:29:55.200Z","updatedAt":"2026-07-23T18:34:40.261Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-831-536-167","title":"Action-Perception Loop","tags":[],"properties":{"comment":"The fundamental iterative cycle where an agent acts, observes the environment's response, and updates its internal state."},"createdAt":"2026-07-23T18:32:16.167Z","updatedAt":"2026-07-23T18:34:01.565Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-202-073","title":"GPT Family","tags":[],"properties":{"comment":"The foundational proprietary family from OpenAI, establishing the tiered model (4o, 4o-mini, o1)."},"createdAt":"2026-07-23T18:43:22.073Z","updatedAt":"2026-07-23T18:52:23.364Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-203-225","title":"Gemini Family","tags":[],"properties":{"comment":"Google's multimodal-first family, spanning from high-speed 'Flash' to high-capability 'Ultra' tiers."},"createdAt":"2026-07-23T18:43:23.225Z","updatedAt":"2026-07-23T18:52:23.372Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-210-084","title":"Claude Family","tags":[],"properties":{"comment":"Anthropic's family (Haiku, Sonnet, Opus), emphasizing safety, nuance, and constitutional alignment."},"createdAt":"2026-07-23T18:43:30.084Z","updatedAt":"2026-07-23T18:52:23.368Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-216-210","title":"DeepSeek Family","tags":[],"properties":{"comment":"An influential family focused on high-efficiency reasoning and coding, representing a major bridge between open-weight access and SOTA performance."},"createdAt":"2026-07-23T18:43:36.210Z","updatedAt":"2026-07-23T18:52:23.373Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-218-654","title":"Llama Family","tags":[],"properties":{"comment":"Meta's open-weight series that serves as the primary foundation for the global open-source ecosystem."},"createdAt":"2026-07-23T18:43:38.654Z","updatedAt":"2026-07-23T18:52:23.374Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-224-828","title":"Mistral Family","tags":[],"properties":{"comment":"The family that popularized Mixture of Experts (MoE) as a viable strategy for high-efficiency model tiers."},"createdAt":"2026-07-23T18:43:44.828Z","updatedAt":"2026-07-23T18:52:23.370Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-269-284","title":"Grok Family","tags":[],"properties":{"comment":"The model family from xAI, emphasizing real-time data integration and high-compute reasoning tiers."},"createdAt":"2026-07-23T18:44:29.284Z","updatedAt":"2026-07-23T18:52:23.375Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-271-126","title":"Kimi Family","tags":[],"properties":{"comment":"The family from Moonshot AI that pioneered ultra-long context windows as a primary competitive advantage."},"createdAt":"2026-07-23T18:44:31.126Z","updatedAt":"2026-07-23T18:52:23.375Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-277-664","title":"GLM Family","tags":[],"properties":{"comment":"A leading bilingual family from Zhiping AI, utilizing a unique hybrid architecture for high-performance Chinese/English tasks."},"createdAt":"2026-07-23T18:44:37.664Z","updatedAt":"2026-07-23T18:52:23.376Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-439-257","title":"Claude Haiku","tags":[],"properties":{"comment":"The fastest and most cost-effective model in the family, optimized for near-instant responses and efficiency."},"createdAt":"2026-07-23T18:47:19.257Z","updatedAt":"2026-07-23T18:50:13.005Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-441-578","title":"Claude Sonnet","tags":[],"properties":{"comment":"The middle-tier model that currently serves as Anthropic's state-of-the-art for most use cases, balancing high speed with frontier-level performance."},"createdAt":"2026-07-23T18:47:21.578Z","updatedAt":"2026-07-23T18:50:13.008Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-448-042","title":"Claude Opus","tags":[],"properties":{"comment":"The flagship high-intelligence model designed for complex analysis and high-order reasoning."},"createdAt":"2026-07-23T18:47:28.042Z","updatedAt":"2026-07-23T18:50:13.009Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-454-967","title":"Claude Fable","tags":[],"properties":{},"createdAt":"2026-07-23T18:47:34.967Z","updatedAt":"2026-07-23T18:50:13.011Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-565-259","title":"Claude Mythos 5","tags":[],"properties":{"comment":"The restricted sibling of Fable 5, sharing the same architecture but without the safety classifiers, used for defensive security research."},"createdAt":"2026-07-23T18:49:25.259Z","updatedAt":"2026-07-23T18:50:13.011Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-570-962","title":"Project Glasswing","tags":[],"properties":{"comment":"The collaborative program between Anthropic and the US government that governed the early release and safety testing of Mythos-class models."},"createdAt":"2026-07-23T18:49:30.962Z","updatedAt":"2026-07-23T18:50:13.012Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-704-799","title":"GPT-5.6 Sol","tags":[],"properties":{"comment":"The flagship tier of the 5.6 generation, optimized for maximum reasoning and complex coding."},"createdAt":"2026-07-23T18:51:44.799Z","updatedAt":"2026-07-23T18:52:04.172Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-706-148","title":"GPT-5.6 Luna","tags":[],"properties":{"comment":"The fast, cost-efficient tier designed for high-volume tasks and low-latency responses."},"createdAt":"2026-07-23T18:51:46.148Z","updatedAt":"2026-07-23T18:52:04.172Z","attachments":[],"aiArtifacts":[]},{"id":"n-1784-832-711-608","title":"GPT-5.6 Terra","tags":[],"properties":{"comment":"The balanced mid-tier model, offering GPT-5.5 level intelligence at a significantly lower cost."},"createdAt":"2026-07-23T18:51:51.608Z","updatedAt":"2026-07-23T18:52:04.173Z","attachments":[],"aiArtifacts":[]},{"id":"n-mermaid-mrxl4f8r-1","title":"Modern AI Technology","tags":["root","core","branch"],"properties":{"comment":"The field of modern artificial intelligence: its scientific foundations, engineering methods, system capabilities, industry, economics, societal consequences, history, and open debates."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.416Z","attachments":[],"aiArtifacts":[],"isAnchor":true},{"id":"n-mermaid-mrxl4f8r-10","title":"Machine Learning","tags":["science","core","branch"],"properties":{"comment":"The discipline of building computer systems that learn from data rather than following explicitly programmed rules."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.064Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-100","title":"Benchmark Saturation","tags":["capabilities","frontier","leaf"],"properties":{"comment":"The predicament of capability tests losing usefulness as frontier models max them out or absorb their answers from training data."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-101","title":"LLM-as-Judge","tags":["capabilities","frontier","leaf"],"properties":{"comment":"The evaluation method of using a strong AI model to grade other models' outputs."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-102","title":"Open-Weight Models","tags":["industry","core","branch"],"properties":{"comment":"AI models whose trained weights are published for anyone to download, inspect, and run."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.420Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-103","title":"Model APIs","tags":["industry","core","branch"],"properties":{"comment":"The commercial form of AI in which hosted model access is sold and metered per token."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.421Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-104","title":"Self-Hosting","tags":["industry","core","branch"],"properties":{"comment":"The practice of running AI models on infrastructure the user controls rather than through hosted services."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.419Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-105","title":"Model Families","tags":["industry","core","branch"],"properties":{"comment":"The product strategy of offering models in capability tiers from flagship to small."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.423Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-106","title":"Frontier Labs","tags":["industry","core","branch"],"properties":{"comment":"The small set of organizations with the capital, talent, and computing resources to train AI models at the highest capability level."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.423Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-107","title":"OpenAI","tags":["industry","frontier","org","branch"],"properties":{"comment":"The AI laboratory behind the GPT models, ChatGPT, and the o-series reasoning models, founded in 2015."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:52:04.171Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-108","title":"Anthropic","tags":["industry","frontier","org","branch"],"properties":{"comment":"The AI safety-focused laboratory behind the Claude models, founded in 2021."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:50:13.013Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-109","title":"Google DeepMind","tags":["industry","frontier","org","branch"],"properties":{"comment":"Google's consolidated AI research organization, producer of the Gemini models and earlier of AlphaGo and AlphaFold."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:36:29.640Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-11","title":"Supervised Learning","tags":["science","core","branch"],"properties":{"comment":"A machine learning paradigm in which models learn from examples paired with correct answers or labels."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.817Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-110","title":"Meta AI","tags":["industry","frontier","org","branch"],"properties":{"comment":"Meta's AI division, known for the openly released Llama model family."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:25.156Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-111","title":"xAI","tags":["industry","frontier","org","branch"],"properties":{"comment":"The AI laboratory founded by Elon Musk in 2023, producer of the Grok models and the Colossus training cluster."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:20:48.836Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-112","title":"Mistral AI","tags":["industry","frontier","org","branch"],"properties":{"comment":"The Paris-based AI laboratory known for efficient open-weight models and European AI sovereignty positioning."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.025Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-113","title":"DeepSeek","tags":["industry","frontier","org","branch"],"properties":{"comment":"The Chinese AI laboratory whose efficient, openly released reasoning models in early 2025 triggered a global market repricing."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.023Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-114","title":"Alibaba Qwen","tags":["industry","frontier","org","branch"],"properties":{"comment":"Alibaba's AI model family and laboratory, a major anchor of the open-weight ecosystem."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:52:23.371Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-115","title":"Executives","tags":["industry","frontier","branch"],"properties":{"comment":"The executives and founders leading the artificial intelligence industry's major companies and laboratories."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.422Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-116","title":"Sam Altman","tags":["industry","frontier","person","branch"],"properties":{"comment":"The chief executive of OpenAI and the most publicly visible leader of the AI industry."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:30:16.884Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-117","title":"Dario Amodei","tags":["industry","frontier","person","branch"],"properties":{"comment":"The chief executive and co-founder of Anthropic, a former physicist and scaling-law researcher."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.113Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-118","title":"Demis Hassabis","tags":["industry","frontier","person","branch"],"properties":{"comment":"The chief executive and co-founder of DeepMind, Nobel laureate in chemistry, leading Google's consolidated AI effort."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:36:30.663Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-119","title":"Elon Musk","tags":["industry","frontier","person","branch"],"properties":{"comment":"The entrepreneur who co-founded OpenAI, founded xAI in 2023, and runs Tesla, SpaceX, and the X platform."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:20:48.836Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-12","title":"Unsupervised Learning","tags":["science","core","branch"],"properties":{"comment":"A machine learning paradigm in which models find structure in data without any labels or correct answers provided."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.326Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-120","title":"Jensen Huang","tags":["industry","frontier","person","branch"],"properties":{"comment":"The founder and chief executive of Nvidia."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:18:10.132Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-121","title":"Satya Nadella","tags":["industry","frontier","person","branch"],"properties":{"comment":"The chief executive of Microsoft, architect of its OpenAI partnership and company-wide AI product strategy."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.105Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-122","title":"Mark Zuckerberg","tags":["industry","frontier","person","branch"],"properties":{"comment":"The founder and chief executive of Meta, backer of the open-weight Llama strategy."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.105Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-123","title":"Liang Wenfeng","tags":["industry","frontier","person","branch"],"properties":{"comment":"The founder of the AI laboratory DeepSeek and of the quantitative fund High-Flyer that incubated it."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.106Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-124","title":"Mustafa Suleyman","tags":["industry","frontier","person","branch"],"properties":{"comment":"The co-founder of DeepMind and Inflection AI who became chief executive of Microsoft AI in 2024."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.107Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-125","title":"Arthur Mensch","tags":["industry","frontier","person","branch"],"properties":{"comment":"The co-founder and chief executive of the French AI laboratory Mistral."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.107Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-126","title":"Compute & Silicon","tags":["industry","frontier","branch"],"properties":{"comment":"The companies that design and manufacture the computing hardware on which AI runs: chip designers, the foundry layer, and specialized compute clouds."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.424Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-127","title":"Nvidia","tags":["industry","frontier","org","branch"],"properties":{"comment":"The chip company whose GPUs and CUDA software ecosystem dominate AI training and inference hardware."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:20:48.837Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-128","title":"AMD","tags":["industry","frontier","org","branch"],"properties":{"comment":"The chip company challenging the GPU incumbent in AI accelerators with its Instinct line."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:19:12.348Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-129","title":"TSMC","tags":["industry","frontier","org","branch"],"properties":{"comment":"The Taiwanese semiconductor foundry that fabricates nearly all advanced AI chips."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:19:12.350Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-13","title":"Reinforcement Learning","tags":["science","core","branch"],"properties":{"comment":"A machine learning paradigm in which an agent learns behavior through trial, error, and reward signals."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:53:51.980Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-130","title":"CoreWeave","tags":["industry","frontier","org","branch"],"properties":{"comment":"The specialized GPU cloud provider that became a major AI-compute supplier and a prominent AI-era public listing."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:20:48.837Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-131","title":"Groq","tags":["industry","frontier","org","branch"],"properties":{"comment":"The chipmaker whose language processing units specialize in very fast AI inference serving."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:19:12.351Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-132","title":"Platforms & Infrastructure","tags":["industry","frontier","branch"],"properties":{"comment":"The platform and infrastructure companies through which AI models reach developers and enterprises: hyperscale clouds, model hubs, API aggregators, and data suppliers."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.425Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-133","title":"Microsoft","tags":["industry","frontier","org","branch"],"properties":{"comment":"The technology corporation whose OpenAI partnership, Copilot products, and Azure cloud made it the leading AI incumbent."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.111Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-134","title":"Google","tags":["industry","frontier","org","branch"],"properties":{"comment":"The technology corporation that invented the transformer, produces the Gemini models and TPU chips, and integrates AI into its search products."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:19:10.078Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-135","title":"Amazon Web Services","tags":["industry","frontier","org","branch"],"properties":{"comment":"Amazon's cloud division, the largest cloud provider, offering multi-vendor model access and custom AI chips."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:05.474Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-136","title":"Hugging Face","tags":["industry","frontier","org","branch"],"properties":{"comment":"The platform where open-weight AI models, datasets, and machine learning libraries are published and shared."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:25.159Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-137","title":"OpenRouter","tags":["industry","frontier","org","branch"],"properties":{"comment":"The API aggregation service providing unified access and billing across AI models from many providers."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:05.476Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-138","title":"Together AI","tags":["industry","frontier","org","branch"],"properties":{"comment":"The cloud platform specialized in serving and fine-tuning open-weight AI models."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:05.477Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-139","title":"Scale AI","tags":["industry","frontier","org","branch"],"properties":{"comment":"The data-labeling and evaluation company that supplied human feedback infrastructure for AI training."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.425Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-14","title":"Neural Networks","tags":["science","core","branch"],"properties":{"comment":"A class of machine learning models built from layers of simple numerical units with adjustable connection weights."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.820Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-140","title":"Application Layer","tags":["industry","frontier","branch"],"properties":{"comment":"The companies building consumer and business products on top of foundation models, including search, coding, image, video, voice, and companionship applications."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.421Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-141","title":"Perplexity","tags":["industry","frontier","org","branch"],"properties":{"comment":"The AI search company known for answering questions directly with citations."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:40:17.538Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-142","title":"Cursor","tags":["industry","frontier","org","branch"],"properties":{"comment":"The company behind the AI-native code editor of the same name."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:40:17.539Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-143","title":"Midjourney","tags":["industry","frontier","org","branch"],"properties":{"comment":"The independent AI image-generation laboratory known for its distinctive visual aesthetic."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:40:17.535Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-144","title":"Runway","tags":["industry","frontier","org","branch"],"properties":{"comment":"The generative-video company whose tools are used in film and advertising production."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:40:17.536Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-145","title":"ElevenLabs","tags":["industry","frontier","org","branch"],"properties":{"comment":"The voice-AI company known for high-quality speech synthesis and voice cloning."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:40:17.537Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-146","title":"Character.AI","tags":["industry","frontier","org","branch"],"properties":{"comment":"The AI companion chatbot platform known for high user engagement and prominence in AI companionship safety debates."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:40:17.539Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-147","title":"AI Datacenters","tags":["economy","frontier","branch"],"properties":{"comment":"The gigawatt-scale computing facilities where frontier AI models are trained and served."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:35:33.527Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-148","title":"Inference Cost","tags":["economy","frontier","branch"],"properties":{"comment":"The per-token economics of serving AI models to users."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.026Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-149","title":"Investment Boom","tags":["economy","frontier","branch"],"properties":{"comment":"The wave of capital investment into AI infrastructure, laboratories, and companies during the current era."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:35:33.528Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-15","title":"Deep Learning","tags":["science","core","branch"],"properties":{"comment":"The use of neural networks with many layers, enabling learned hierarchies of features from raw data."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:33.320Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-150","title":"Energy Demand","tags":["economy","frontier","branch"],"properties":{"comment":"The electricity requirements of AI datacenters as a constraint on and driver of the industry's growth."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:35:33.529Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-151","title":"Export Controls","tags":["economy","frontier","branch"],"properties":{"comment":"The government restrictions, primarily American, on exporting advanced AI chips to China."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:40.898Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-152","title":"AI Governance","tags":["society","frontier","safety","branch"],"properties":{"comment":"The developing body of laws, regulations, and international agreements governing AI development and deployment."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:44.726Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-153","title":"EU AI Act","tags":["society","frontier","safety","leaf"],"properties":{"comment":"The European Union's comprehensive risk-tiered law regulating artificial intelligence, phasing in from 2025."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:40.892Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-154","title":"Labor & Automation","tags":["society","frontier","leaf"],"properties":{"comment":"The societal question of how AI automation affects employment and the nature of work."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:36.663Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-155","title":"Misinformation & Deepfakes","tags":["society","frontier","safety","branch"],"properties":{"comment":"The societal problem of cheap, convincing AI-generated fake text, images, audio, and video."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:36.665Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-156","title":"AI Companionship","tags":["society","frontier","branch"],"properties":{"comment":"The social phenomenon of humans forming personal relationships with AI systems as companions, confidants, and tutors."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:36.667Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-157","title":"Symbolic AI","tags":["history","core","event","branch"],"properties":{"comment":"The historical AI paradigm treating intelligence as explicit rules, logic, and symbol manipulation, dominant from the 1950s through the 1980s."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:33.316Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-158","title":"Connectionism","tags":["history","core","event","branch"],"properties":{"comment":"The historical AI school holding that intelligence should be learned by neural networks rather than programmed as rules."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:33.318Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-159","title":"AI Winters","tags":["history","core","event","branch"],"properties":{"comment":"The historical periods of collapsed AI funding and credibility following eras of overpromising, principally the mid-1970s and late 1980s."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:33.314Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-16","title":"Backpropagation","tags":["science","core","branch"],"properties":{"comment":"The algorithm that trains neural networks by propagating error signals backward through layers to compute weight adjustments."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.820Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-160","title":"Deep Learning Revolution","tags":["history","core","event","branch"],"properties":{"comment":"The 2012 turning point when the AlexNet neural network won the ImageNet competition and vindicated deep learning."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:33.319Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-161","title":"AlphaGo","tags":["history","core","event","branch"],"properties":{"comment":"The 2016 event in which DeepMind's Go-playing system defeated world champion Lee Sedol."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:38:44.332Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-162","title":"Transformer Breakthrough","tags":["history","core","event","branch"],"properties":{"comment":"The 2017 publication of the research paper Attention Is All You Need, which introduced the transformer architecture."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:53.252Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-163","title":"ChatGPT Moment","tags":["history","core","event","branch"],"properties":{"comment":"The November 2022 release of ChatGPT and its explosive public adoption, which brought generative AI into mainstream awareness."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:30:16.878Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-164","title":"Reasoning Era","tags":["history","frontier","event","branch"],"properties":{"comment":"The period from late 2024 onward defined by reasoning models and capability gains from inference-time computation."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:32:45.894Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-165","title":"AI Safety","tags":["debates","core","debate","safety","branch"],"properties":{"comment":"The debate over whether AI systems can be prevented from causing harm, spanning present-day misuse and long-horizon risk from highly capable systems."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.277Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-166","title":"Capability Progress","tags":["debates","core","debate","branch"],"properties":{"comment":"The debate over why AI capabilities keep improving and whether the improvement will continue."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.282Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-167","title":"Algorithmic Progress","tags":["debates","core","debate","branch"],"properties":{"comment":"The thesis that improvements in methods, architectures, and training efficiency drive AI progress as much as increases in raw scale."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.072Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-168","title":"LLM Understanding Debate","tags":["debates","core","debate","branch"],"properties":{"comment":"The debate over whether large language models genuinely understand the world or merely imitate patterns in their training data."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.284Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-169","title":"Stochastic Parrot View","tags":["debates","core","debate","branch"],"properties":{"comment":"The view that large language models recombine patterns from their training data without genuine meaning or comprehension."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:21:00.323Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-17","title":"Loss Function","tags":["science","core","branch"],"properties":{"comment":"The function that quantifies a model's prediction error as a single number to be minimized during training."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.821Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-170","title":"World-Model View","tags":["debates","core","debate","branch"],"properties":{"comment":"The view that large language models acquire genuine internal representations of the world through learning to predict text."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:21:00.324Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-171","title":"Artificial General Intelligence","tags":["debates","core","debate","branch"],"properties":{"comment":"The contested concept of artificial intelligence matching human breadth across cognitive work, serving simultaneously as research goal, marketing term, and governance trigger."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.071Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-172","title":"Superintelligence","tags":["debates","core","debate","leaf"],"properties":{"comment":"The concept of AI capability decisively beyond human level and the discourse around its possibility and risks."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.286Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-173","title":"Data Wall","tags":["debates","frontier","debate","branch"],"properties":{"comment":"The open question of whether high-quality human-written training text is running out and what that would mean for AI progress."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.286Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-174","title":"Bubble Question","tags":["debates","frontier","debate","branch"],"properties":{"comment":"The open question of whether the AI investment boom reflects rational anticipation of value or a speculative bubble."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.287Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-18","title":"Gradient Descent","tags":["science","core","branch"],"properties":{"comment":"The optimization method that trains models by iteratively stepping model weights in the direction that reduces the loss."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.822Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-19","title":"Generalization","tags":["science","core","branch"],"properties":{"comment":"The property of a trained model performing well on inputs it never saw during training."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.822Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-2","title":"The Big Ideas","tags":["science","core","perspective","branch"],"properties":{"comment":"The scientific foundations of modern artificial intelligence: machine learning principles, neural network architectures, empirical scaling phenomena, and the researchers who established them."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:33.238Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-20","title":"Overfitting","tags":["science","core","leaf"],"properties":{"comment":"The failure mode in which a model memorizes training data instead of learning transferable patterns."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.823Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-21","title":"Transfer Learning","tags":["science","core","branch"],"properties":{"comment":"The technique of reusing representations learned on one task as the starting point for a different task."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.823Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-22","title":"Transformer Architecture","tags":["science","core","branch"],"properties":{"comment":"The neural network architecture introduced by Google researchers in 2017 that underlies essentially all modern frontier AI models."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.066Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-23","title":"Attention Mechanism","tags":["science","core","leaf"],"properties":{"comment":"The transformer operation in which each token dynamically weighs the relevance of all other tokens in context."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:53:38.427Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-24","title":"Parameters","tags":["science","core","leaf"],"properties":{"comment":"The learned numerical weights of a neural network, whose count serves as the standard measure of model size."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:53.253Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-25","title":"Positional Encoding","tags":["science","core","leaf"],"properties":{"comment":"The transformer component that encodes each token's position in the sequence, since attention itself is order-blind."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:53.254Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-26","title":"Decoder-Only Architecture","tags":["science","core","leaf"],"properties":{"comment":"The transformer variant using only the generation half of the original architecture, the standard layout for large language models."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:53.254Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-27","title":"Mixture of Experts","tags":["science","core","branch"],"properties":{"comment":"A neural architecture that divides the network into specialist sub-networks and activates only a few per token to reduce compute cost."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.024Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-28","title":"Tokenization","tags":["science","core","branch"],"properties":{"comment":"The process of splitting text into subword units called tokens, the basic unit AI models process and price."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:19:52.412Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-29","title":"Embeddings","tags":["science","core","branch"],"properties":{"comment":"The technique of representing meaning as vectors in high-dimensional space where semantic similarity corresponds to geometric proximity."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:49:31.094Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-3","title":"The Model Factory","tags":["engineering","core","perspective","branch"],"properties":{"comment":"The engineering process by which AI models are built and operated: large-scale training, behavioral refinement, serving infrastructure, specialized hardware, and deployment safety practice."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:33.234Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-30","title":"Latent Space","tags":["science","core","leaf"],"properties":{"comment":"The geometric space of a model's internal learned representations, where directions and distances carry semantic relationships."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:53:38.429Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-31","title":"Scaling Laws","tags":["science","core","branch"],"properties":{"comment":"The empirical finding that AI model capability improves predictably as compute, data, and parameter count increase."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.067Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-32","title":"Scaling Hypothesis","tags":["science","core","debate","branch"],"properties":{"comment":"The thesis that increasing compute, data, and model size is the primary driver of artificial intelligence progress."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.068Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-33","title":"Emergent Capabilities","tags":["science","core","debate","branch"],"properties":{"comment":"The phenomenon of AI models developing abilities at scale that they were never explicitly trained for."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.068Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-34","title":"Alignment","tags":["science","core","safety","branch"],"properties":{"comment":"The research problem of making AI models reliably pursue the intentions of their operators and users."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.069Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-35","title":"Interpretability","tags":["science","core","safety","branch"],"properties":{"comment":"The research field devoted to understanding the internal workings of trained neural networks."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.070Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-36","title":"Mechanistic Interpretability","tags":["science","frontier","safety","leaf"],"properties":{"comment":"The interpretability research program that reverse-engineers the specific computational circuits and features inside neural networks."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.071Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-37","title":"Research Pioneers","tags":["science","core","branch"],"properties":{"comment":"The scientists whose research created the foundations of modern artificial intelligence, several of whom are now its most prominent public voices."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.066Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-38","title":"Geoffrey Hinton","tags":["science","core","person","branch"],"properties":{"comment":"A British-Canadian computer scientist, Nobel laureate, and deep learning pioneer known for backpropagation, AlexNet, and later public warnings about AI risk."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:44.730Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-39","title":"Yann LeCun","tags":["science","core","person","branch"],"properties":{"comment":"A French-American computer scientist, Turing Award laureate, deep learning pioneer, and Meta's chief AI scientist known for convolutional networks and skepticism of near-term AI risk claims."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:50.499Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-4","title":"The Toolbox","tags":["capabilities","core","perspective","branch"],"properties":{"comment":"The capabilities of modern AI systems and the product forms in which people use them, from conversation and coding to image generation and autonomous task execution."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:33.235Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-40","title":"Yoshua Bengio","tags":["science","core","person","branch"],"properties":{"comment":"A Canadian computer scientist, Turing Award laureate, and deep learning pioneer at Mila who now leads international AI safety assessment work."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:40.897Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-41","title":"Fei-Fei Li","tags":["science","core","person","branch"],"properties":{"comment":"A Stanford computer scientist who created the ImageNet dataset and advocates human-centered AI."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:25:45.193Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-42","title":"Ilya Sutskever","tags":["science","core","person","branch"],"properties":{"comment":"A computer scientist who co-authored AlexNet, co-founded OpenAI as its chief scientist, championed the scaling hypothesis, and later founded Safe Superintelligence Inc."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:24:35.574Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-43","title":"Andrew Ng","tags":["science","core","person","branch"],"properties":{"comment":"A computer scientist and educator who co-founded Google Brain and Coursera and taught machine learning to a global audience of practitioners."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.824Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-44","title":"Jürgen Schmidhuber","tags":["science","core","person","branch"],"properties":{"comment":"A Swiss-based computer scientist whose laboratory produced LSTM networks and early versions of several foundational deep learning ideas."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:50.497Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-45","title":"Andrej Karpathy","tags":["science","frontier","person","branch"],"properties":{"comment":"A computer scientist and educator, OpenAI founding member and former Tesla AI director, known for influential AI teaching and coining the term vibe coding."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:50.501Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-46","title":"Pre-training","tags":["engineering","core","branch"],"properties":{"comment":"The first phase of building a language model: training on web-scale text corpora with the next-token prediction objective."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.017Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-47","title":"Next-Token Prediction","tags":["engineering","core","leaf"],"properties":{"comment":"The training objective of predicting the next token of text given all preceding tokens."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.321Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-48","title":"Self-Supervised Learning","tags":["engineering","core","branch"],"properties":{"comment":"The training technique of deriving supervision signals from the data itself rather than from human-provided labels."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.322Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-49","title":"Training Data","tags":["engineering","core","branch"],"properties":{"comment":"The web-scale corpora of text, code, images, and transcripts on which AI models are trained."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.323Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-5","title":"The Arena","tags":["industry","core","perspective","branch"],"properties":{"comment":"The competitive landscape of the artificial intelligence industry: research laboratories, executives, hardware suppliers, platform companies, application builders, and their business models."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.418Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-50","title":"Data Curation","tags":["engineering","core","branch"],"properties":{"comment":"The engineering practice of filtering, deduplicating, and mixing training corpora to improve model quality."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.324Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-51","title":"Synthetic Data","tags":["engineering","frontier","branch"],"properties":{"comment":"Training data generated by AI models themselves rather than collected from human sources."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.324Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-52","title":"Training Compute","tags":["engineering","core","branch"],"properties":{"comment":"The total computation consumed by a model training run, measured in floating-point operations."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.325Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-53","title":"Base Model","tags":["engineering","core","leaf"],"properties":{"comment":"A model directly produced by pre-training, possessing broad knowledge but no assistant behavior."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.319Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-54","title":"Foundation Model","tags":["engineering","core","branch"],"properties":{"comment":"A large model trained once at great expense to serve as the shared basis for many downstream applications."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.316Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-55","title":"Post-training","tags":["engineering","core","branch"],"properties":{"comment":"The stage of model production that turns a pre-trained base model into a usable assistant through fine-tuning and reinforcement learning."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.018Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-56","title":"Fine-Tuning","tags":["engineering","core","branch"],"properties":{"comment":"The technique of continuing a trained model's training on narrower data to adapt or specialize it."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-57","title":"Supervised Fine-Tuning","tags":["engineering","core","branch"],"properties":{"comment":"The fine-tuning of a model on curated demonstrations of ideal behavior, typically high-quality prompt-response pairs."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.023Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-58","title":"RLHF","tags":["engineering","core","branch"],"properties":{"comment":"The post-training method in which human preference comparisons train a reward model that then guides reinforcement learning of the language model."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:53:51.975Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-59","title":"Reward Model","tags":["engineering","core","leaf"],"properties":{"comment":"A model trained to predict human preferences between responses, used as an automated judge during preference-based training."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:53:51.979Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-6","title":"The Gold Rush","tags":["economy","frontier","perspective","branch"],"properties":{"comment":"The economics of the artificial intelligence boom: infrastructure investment, energy consumption, the cost of serving models, capital flows, and semiconductor geopolitics."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:33.238Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-60","title":"RLVR","tags":["engineering","frontier","branch"],"properties":{"comment":"The post-training method using reinforcement learning on tasks with objectively checkable answers such as mathematics and code tests."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:38:44.334Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-61","title":"RLAIF","tags":["engineering","frontier","leaf"],"properties":{"comment":"The post-training method in which AI models rather than humans provide the feedback signal for preference training."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-62","title":"Inference","tags":["engineering","core","branch"],"properties":{"comment":"The operational phase in which a trained model processes prompts and generates outputs for users."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.019Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-63","title":"Context Window","tags":["engineering","core","branch"],"properties":{"comment":"The maximum amount of text, measured in tokens, that a model can consider at once."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:20:54.725Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-64","title":"Prompt","tags":["engineering","core","branch"],"properties":{"comment":"The input text that steers a model's output, comprising questions, instructions, documents, or examples."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:19:52.414Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-65","title":"Prompt Engineering","tags":["engineering","core","leaf"],"properties":{"comment":"The craft of improving model outputs by structuring inputs effectively."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:20:54.727Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-66","title":"System Prompt","tags":["engineering","core","leaf"],"properties":{"comment":"The standing instructions an operator places before all user input to set a model's persona, rules, and boundaries."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:07:19.707Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-67","title":"Sampling","tags":["engineering","core","leaf"],"properties":{"comment":"The procedure by which a model selects each next token from its output probability distribution, tuned by parameters such as temperature."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:04:40.472Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-68","title":"Quantization","tags":["engineering","core","branch"],"properties":{"comment":"The technique of compressing model weights into lower-precision numbers to reduce memory and serving cost."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.024Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-69","title":"Test-Time Compute","tags":["engineering","frontier","branch"],"properties":{"comment":"The approach of improving answer quality by spending more computation at inference time rather than only during training."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:04:40.473Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-7","title":"Promise & Peril","tags":["society","core","perspective","branch"],"properties":{"comment":"The societal consequences of artificial intelligence: government regulation, effects on employment, synthetic media, and personal relationships between humans and AI systems."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:33.237Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-70","title":"Distillation","tags":["engineering","core","branch"],"properties":{"comment":"The technique of training a small model to imitate a larger one, capturing much of its capability at lower serving cost."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.020Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-71","title":"AI Accelerators","tags":["engineering","core","branch"],"properties":{"comment":"The category of chips specialized for the dense matrix arithmetic of neural network computation."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.020Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-72","title":"GPU","tags":["engineering","core","branch"],"properties":{"comment":"The graphics processing unit, the parallel processor class that serves as the dominant hardware for AI training and inference."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:18:10.136Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-73","title":"Custom AI Chips","tags":["engineering","core","branch"],"properties":{"comment":"AI accelerator chips designed in-house by cloud companies, such as tensor processing units, as alternatives to merchant GPUs."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-74","title":"Red-Teaming","tags":["engineering","core","safety","branch"],"properties":{"comment":"The practice of systematically attacking an AI model before release to discover harmful behaviors."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.021Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-75","title":"Jailbreaking","tags":["engineering","core","safety","branch"],"properties":{"comment":"The practice of crafting inputs that bypass an AI model's safety training to elicit prohibited outputs."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.022Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-76","title":"Guardrails","tags":["engineering","core","safety","branch"],"properties":{"comment":"The deployed safety layer of filters, classifiers, and rules wrapped around a model in production."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.022Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-77","title":"Large Language Models","tags":["capabilities","core","branch"],"properties":{"comment":"The class of general-purpose AI systems trained on web-scale text that can draft, summarize, translate, code, and converse."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.889Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-78","title":"In-Context Learning","tags":["capabilities","core","branch"],"properties":{"comment":"The ability of a language model to learn a new task from examples provided in the prompt, without any retraining."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.073Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-79","title":"Hallucination","tags":["capabilities","core","safety","branch"],"properties":{"comment":"The failure mode in which a language model generates confident but false information such as fabricated facts or citations."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:44.733Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-8","title":"The Long Road","tags":["history","core","perspective","branch"],"properties":{"comment":"The history of artificial intelligence from its founding schools of thought in the 1950s through its funding collapses and breakthroughs to the present day."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:33.240Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-80","title":"Multimodality","tags":["capabilities","core","branch"],"properties":{"comment":"The extension of AI models beyond text to understanding and generating images, audio, and video."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.890Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-81","title":"Vision-Language Models","tags":["capabilities","core","branch"],"properties":{"comment":"AI models that process images and text jointly, enabling description, chart reading, and screenshot reasoning."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:21:00.323Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-82","title":"Speech Models","tags":["capabilities","core","branch"],"properties":{"comment":"AI models for speech recognition and speech synthesis."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-83","title":"Image Generation","tags":["capabilities","core","product","branch"],"properties":{"comment":"The capability of producing images from text descriptions."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-84","title":"Video Generation","tags":["capabilities","frontier","product","branch"],"properties":{"comment":"The capability of producing video from text or image prompts."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-85","title":"Diffusion Models","tags":["capabilities","core","branch"],"properties":{"comment":"The generative technique that creates images and video by learning to reverse a noising process."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-86","title":"Reasoning Models","tags":["capabilities","frontier","branch"],"properties":{"comment":"The class of models trained to generate extended internal reasoning before producing an answer."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.891Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-87","title":"Chain-of-Thought","tags":["capabilities","core","leaf"],"properties":{"comment":"The technique of solving problems through explicit step-by-step reasoning rather than direct answers."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.073Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-88","title":"AI Agents","tags":["capabilities","frontier","branch"],"properties":{"comment":"AI systems that pursue goals through sequences of actions such as searching, calling tools, and iterating, rather than producing single responses."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:34:40.258Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-89","title":"Tool Use","tags":["capabilities","core","branch"],"properties":{"comment":"The mechanism by which AI models call external functions and services and incorporate their results."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:34:40.259Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-9","title":"Burning Questions","tags":["debates","core","perspective","branch"],"properties":{"comment":"The major unresolved questions of artificial intelligence, on which leading experts publicly disagree: safety, machine understanding, the sources and limits of progress, and the sustainability of the investment boom."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:33.236Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-90","title":"Computer Use","tags":["capabilities","frontier","branch"],"properties":{"comment":"The agent capability of operating real user interfaces by reading screens and issuing clicks and keystrokes."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:34:40.260Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-91","title":"Multi-Agent Systems","tags":["capabilities","frontier","leaf"],"properties":{"comment":"Architectures in which multiple AI model instances cooperate on a task in distinct roles."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:34:01.565Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-92","title":"Model Context Protocol","tags":["capabilities","frontier","branch"],"properties":{"comment":"An open standard for connecting AI systems to external tools and data sources, introduced by Anthropic."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:40:44.996Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-93","title":"Retrieval-Augmented Generation","tags":["capabilities","core","branch"],"properties":{"comment":"The technique of grounding model answers by retrieving relevant documents first and generating from them."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.893Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-94","title":"Vector Database","tags":["capabilities","core","branch"],"properties":{"comment":"A database that indexes content by embedding vectors and retrieves by semantic similarity."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.894Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-95","title":"AI Coding Assistants","tags":["capabilities","core","product","branch"],"properties":{"comment":"The product category of AI tools that write, complete, and debug software code."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.894Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-96","title":"Chat Assistants","tags":["capabilities","core","product","branch"],"properties":{"comment":"The conversational AI product form through which most of the public interacts with language models."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.895Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-97","title":"AI Search","tags":["capabilities","frontier","product","branch"],"properties":{"comment":"The product category of search systems that answer questions directly with citations rather than returning links."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.895Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-98","title":"Evaluation","tags":["capabilities","core","branch"],"properties":{"comment":"The discipline of measuring AI model capabilities."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.896Z","attachments":[],"aiArtifacts":[],"isAnchor":false},{"id":"n-mermaid-mrxl4f8r-99","title":"Benchmarks","tags":["capabilities","core","branch"],"properties":{"comment":"Standardized test sets used to compare AI model capabilities."},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z","attachments":[],"aiArtifacts":[],"isAnchor":false}],"edges":[{"id":"e-mermaid-mrxl4f8r-1","title":"has_perspective","sourceId":"n-mermaid-mrxl4f8r-1","targetId":"n-mermaid-mrxl4f8r-2","typeId":"has-perspective","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:33.243Z"},{"id":"e-mermaid-mrxl4f8r-10","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-10","targetId":"n-mermaid-mrxl4f8r-11","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.826Z"},{"id":"e-mermaid-mrxl4f8r-100","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-98","targetId":"n-mermaid-mrxl4f8r-101","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-101","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-5","targetId":"n-mermaid-mrxl4f8r-102","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.428Z"},{"id":"e-mermaid-mrxl4f8r-102","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-5","targetId":"n-mermaid-mrxl4f8r-103","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.428Z"},{"id":"e-mermaid-mrxl4f8r-103","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-5","targetId":"n-mermaid-mrxl4f8r-104","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.427Z"},{"id":"e-mermaid-mrxl4f8r-104","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-5","targetId":"n-mermaid-mrxl4f8r-105","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.430Z"},{"id":"e-mermaid-mrxl4f8r-105","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-5","targetId":"n-mermaid-mrxl4f8r-106","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.430Z"},{"id":"e-mermaid-mrxl4f8r-106","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-106","targetId":"n-mermaid-mrxl4f8r-107","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:30:16.889Z"},{"id":"e-mermaid-mrxl4f8r-107","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-106","targetId":"n-mermaid-mrxl4f8r-108","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:10:26.655Z"},{"id":"e-mermaid-mrxl4f8r-108","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-106","targetId":"n-mermaid-mrxl4f8r-109","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:10:26.653Z"},{"id":"e-mermaid-mrxl4f8r-109","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-106","targetId":"n-mermaid-mrxl4f8r-110","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:10:26.656Z"},{"id":"e-mermaid-mrxl4f8r-11","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-10","targetId":"n-mermaid-mrxl4f8r-12","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.826Z"},{"id":"e-mermaid-mrxl4f8r-110","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-106","targetId":"n-mermaid-mrxl4f8r-111","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:10:26.654Z"},{"id":"e-mermaid-mrxl4f8r-111","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-106","targetId":"n-mermaid-mrxl4f8r-112","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:10:26.656Z"},{"id":"e-mermaid-mrxl4f8r-112","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-106","targetId":"n-mermaid-mrxl4f8r-113","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:10:26.657Z"},{"id":"e-mermaid-mrxl4f8r-113","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-106","targetId":"n-mermaid-mrxl4f8r-114","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:10:26.657Z"},{"id":"e-mermaid-mrxl4f8r-114","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-5","targetId":"n-mermaid-mrxl4f8r-115","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.429Z"},{"id":"e-mermaid-mrxl4f8r-115","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-115","targetId":"n-mermaid-mrxl4f8r-116","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.115Z"},{"id":"e-mermaid-mrxl4f8r-116","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-115","targetId":"n-mermaid-mrxl4f8r-117","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.119Z"},{"id":"e-mermaid-mrxl4f8r-117","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-115","targetId":"n-mermaid-mrxl4f8r-118","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.115Z"},{"id":"e-mermaid-mrxl4f8r-118","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-115","targetId":"n-mermaid-mrxl4f8r-119","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.116Z"},{"id":"e-mermaid-mrxl4f8r-119","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-115","targetId":"n-mermaid-mrxl4f8r-120","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:17:15.475Z"},{"id":"e-mermaid-mrxl4f8r-12","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-10","targetId":"n-mermaid-mrxl4f8r-13","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.825Z"},{"id":"e-mermaid-mrxl4f8r-120","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-115","targetId":"n-mermaid-mrxl4f8r-121","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.117Z"},{"id":"e-mermaid-mrxl4f8r-121","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-115","targetId":"n-mermaid-mrxl4f8r-122","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.117Z"},{"id":"e-mermaid-mrxl4f8r-122","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-115","targetId":"n-mermaid-mrxl4f8r-123","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.118Z"},{"id":"e-mermaid-mrxl4f8r-123","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-115","targetId":"n-mermaid-mrxl4f8r-124","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.119Z"},{"id":"e-mermaid-mrxl4f8r-124","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-115","targetId":"n-mermaid-mrxl4f8r-125","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.119Z"},{"id":"e-mermaid-mrxl4f8r-125","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-5","targetId":"n-mermaid-mrxl4f8r-126","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.431Z"},{"id":"e-mermaid-mrxl4f8r-126","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-126","targetId":"n-mermaid-mrxl4f8r-127","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:20:48.845Z"},{"id":"e-mermaid-mrxl4f8r-127","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-126","targetId":"n-mermaid-mrxl4f8r-128","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:19:12.354Z"},{"id":"e-mermaid-mrxl4f8r-128","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-126","targetId":"n-mermaid-mrxl4f8r-129","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:19:12.356Z"},{"id":"e-mermaid-mrxl4f8r-129","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-126","targetId":"n-mermaid-mrxl4f8r-130","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:20:48.846Z"},{"id":"e-mermaid-mrxl4f8r-13","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-10","targetId":"n-mermaid-mrxl4f8r-14","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.827Z"},{"id":"e-mermaid-mrxl4f8r-130","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-126","targetId":"n-mermaid-mrxl4f8r-131","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:19:12.356Z"},{"id":"e-mermaid-mrxl4f8r-131","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-5","targetId":"n-mermaid-mrxl4f8r-132","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.431Z"},{"id":"e-mermaid-mrxl4f8r-132","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-132","targetId":"n-mermaid-mrxl4f8r-133","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-133","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-132","targetId":"n-mermaid-mrxl4f8r-134","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-134","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-132","targetId":"n-mermaid-mrxl4f8r-135","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-135","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-132","targetId":"n-mermaid-mrxl4f8r-136","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-136","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-132","targetId":"n-mermaid-mrxl4f8r-137","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-137","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-132","targetId":"n-mermaid-mrxl4f8r-138","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-138","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-132","targetId":"n-mermaid-mrxl4f8r-139","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.432Z"},{"id":"e-mermaid-mrxl4f8r-139","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-5","targetId":"n-mermaid-mrxl4f8r-140","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.429Z"},{"id":"e-mermaid-mrxl4f8r-14","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-10","targetId":"n-mermaid-mrxl4f8r-17","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.827Z"},{"id":"e-mermaid-mrxl4f8r-140","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-140","targetId":"n-mermaid-mrxl4f8r-141","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:40:17.542Z"},{"id":"e-mermaid-mrxl4f8r-141","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-140","targetId":"n-mermaid-mrxl4f8r-142","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:40:17.543Z"},{"id":"e-mermaid-mrxl4f8r-142","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-140","targetId":"n-mermaid-mrxl4f8r-143","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:40:17.541Z"},{"id":"e-mermaid-mrxl4f8r-143","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-140","targetId":"n-mermaid-mrxl4f8r-144","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:40:17.541Z"},{"id":"e-mermaid-mrxl4f8r-144","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-140","targetId":"n-mermaid-mrxl4f8r-145","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:40:17.542Z"},{"id":"e-mermaid-mrxl4f8r-145","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-140","targetId":"n-mermaid-mrxl4f8r-146","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:40:17.544Z"},{"id":"e-mermaid-mrxl4f8r-146","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-6","targetId":"n-mermaid-mrxl4f8r-147","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:35:33.532Z"},{"id":"e-mermaid-mrxl4f8r-147","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-6","targetId":"n-mermaid-mrxl4f8r-148","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:35:33.532Z"},{"id":"e-mermaid-mrxl4f8r-148","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-6","targetId":"n-mermaid-mrxl4f8r-149","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:35:33.533Z"},{"id":"e-mermaid-mrxl4f8r-149","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-6","targetId":"n-mermaid-mrxl4f8r-150","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:35:33.533Z"},{"id":"e-mermaid-mrxl4f8r-15","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-10","targetId":"n-mermaid-mrxl4f8r-18","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.828Z"},{"id":"e-mermaid-mrxl4f8r-150","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-6","targetId":"n-mermaid-mrxl4f8r-151","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:35:33.534Z"},{"id":"e-mermaid-mrxl4f8r-151","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-7","targetId":"n-mermaid-mrxl4f8r-152","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:40.901Z"},{"id":"e-mermaid-mrxl4f8r-152","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-152","targetId":"n-mermaid-mrxl4f8r-153","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:40.900Z"},{"id":"e-mermaid-mrxl4f8r-153","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-7","targetId":"n-mermaid-mrxl4f8r-154","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:36.670Z"},{"id":"e-mermaid-mrxl4f8r-154","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-7","targetId":"n-mermaid-mrxl4f8r-155","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:36.671Z"},{"id":"e-mermaid-mrxl4f8r-155","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-7","targetId":"n-mermaid-mrxl4f8r-156","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:36.672Z"},{"id":"e-mermaid-mrxl4f8r-156","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-8","targetId":"n-mermaid-mrxl4f8r-157","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:33.325Z"},{"id":"e-mermaid-mrxl4f8r-157","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-8","targetId":"n-mermaid-mrxl4f8r-158","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:33.326Z"},{"id":"e-mermaid-mrxl4f8r-158","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-8","targetId":"n-mermaid-mrxl4f8r-159","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:33.324Z"},{"id":"e-mermaid-mrxl4f8r-159","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-8","targetId":"n-mermaid-mrxl4f8r-160","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:33.326Z"},{"id":"e-mermaid-mrxl4f8r-16","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-10","targetId":"n-mermaid-mrxl4f8r-19","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.829Z"},{"id":"e-mermaid-mrxl4f8r-160","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-8","targetId":"n-mermaid-mrxl4f8r-161","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:33.327Z"},{"id":"e-mermaid-mrxl4f8r-161","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-8","targetId":"n-mermaid-mrxl4f8r-162","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:33.328Z"},{"id":"e-mermaid-mrxl4f8r-162","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-8","targetId":"n-mermaid-mrxl4f8r-163","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:30:16.885Z"},{"id":"e-mermaid-mrxl4f8r-163","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-8","targetId":"n-mermaid-mrxl4f8r-164","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:31:36.460Z"},{"id":"e-mermaid-mrxl4f8r-164","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-9","targetId":"n-mermaid-mrxl4f8r-165","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.287Z"},{"id":"e-mermaid-mrxl4f8r-165","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-9","targetId":"n-mermaid-mrxl4f8r-166","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.289Z"},{"id":"e-mermaid-mrxl4f8r-166","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-166","targetId":"n-mermaid-mrxl4f8r-167","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.293Z"},{"id":"e-mermaid-mrxl4f8r-167","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-9","targetId":"n-mermaid-mrxl4f8r-168","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.289Z"},{"id":"e-mermaid-mrxl4f8r-168","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-168","targetId":"n-mermaid-mrxl4f8r-169","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:00:36.040Z"},{"id":"e-mermaid-mrxl4f8r-169","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-168","targetId":"n-mermaid-mrxl4f8r-170","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:05:32.584Z"},{"id":"e-mermaid-mrxl4f8r-17","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-10","targetId":"n-mermaid-mrxl4f8r-21","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.829Z"},{"id":"e-mermaid-mrxl4f8r-170","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-9","targetId":"n-mermaid-mrxl4f8r-171","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.290Z"},{"id":"e-mermaid-mrxl4f8r-171","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-171","targetId":"n-mermaid-mrxl4f8r-172","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.293Z"},{"id":"e-mermaid-mrxl4f8r-172","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-9","targetId":"n-mermaid-mrxl4f8r-173","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.290Z"},{"id":"e-mermaid-mrxl4f8r-173","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-9","targetId":"n-mermaid-mrxl4f8r-174","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.291Z"},{"id":"e-mermaid-mrxl4f8r-174","title":"enables","sourceId":"n-mermaid-mrxl4f8r-11","targetId":"n-mermaid-mrxl4f8r-57","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-175","title":"enables","sourceId":"n-mermaid-mrxl4f8r-12","targetId":"n-mermaid-mrxl4f8r-48","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.334Z"},{"id":"e-mermaid-mrxl4f8r-176","title":"enables","sourceId":"n-mermaid-mrxl4f8r-13","targetId":"n-mermaid-mrxl4f8r-58","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:53:51.984Z"},{"id":"e-mermaid-mrxl4f8r-177","title":"enables","sourceId":"n-mermaid-mrxl4f8r-13","targetId":"n-mermaid-mrxl4f8r-60","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:38:44.339Z"},{"id":"e-mermaid-mrxl4f8r-178","title":"enables","sourceId":"n-mermaid-mrxl4f8r-17","targetId":"n-mermaid-mrxl4f8r-18","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.830Z"},{"id":"e-mermaid-mrxl4f8r-179","title":"enables","sourceId":"n-mermaid-mrxl4f8r-21","targetId":"n-mermaid-mrxl4f8r-56","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-18","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-14","targetId":"n-mermaid-mrxl4f8r-15","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:11.018Z"},{"id":"e-mermaid-mrxl4f8r-180","title":"enables","sourceId":"n-mermaid-mrxl4f8r-173","targetId":"n-mermaid-mrxl4f8r-51","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.336Z"},{"id":"e-mermaid-mrxl4f8r-181","title":"enables","sourceId":"n-mermaid-mrxl4f8r-28","targetId":"n-mermaid-mrxl4f8r-63","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:19:52.417Z"},{"id":"e-mermaid-mrxl4f8r-182","title":"enables","sourceId":"n-mermaid-mrxl4f8r-29","targetId":"n-mermaid-mrxl4f8r-94","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:49:31.098Z"},{"id":"e-mermaid-mrxl4f8r-183","title":"enables","sourceId":"n-mermaid-mrxl4f8r-22","targetId":"n-mermaid-mrxl4f8r-77","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:21:08.564Z"},{"id":"e-mermaid-mrxl4f8r-184","title":"enables","sourceId":"n-mermaid-mrxl4f8r-27","targetId":"n-mermaid-mrxl4f8r-148","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.034Z"},{"id":"e-mermaid-mrxl4f8r-185","title":"enables","sourceId":"n-mermaid-mrxl4f8r-46","targetId":"n-mermaid-mrxl4f8r-55","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.030Z"},{"id":"e-mermaid-mrxl4f8r-186","title":"enables","sourceId":"n-mermaid-mrxl4f8r-31","targetId":"n-mermaid-mrxl4f8r-147","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:25:45.197Z"},{"id":"e-mermaid-mrxl4f8r-187","title":"enables","sourceId":"n-mermaid-mrxl4f8r-31","targetId":"n-mermaid-mrxl4f8r-33","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.079Z"},{"id":"e-mermaid-mrxl4f8r-188","title":"enables","sourceId":"n-mermaid-mrxl4f8r-52","targetId":"n-mermaid-mrxl4f8r-147","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:20:48.840Z"},{"id":"e-mermaid-mrxl4f8r-189","title":"enables","sourceId":"n-mermaid-mrxl4f8r-54","targetId":"n-mermaid-mrxl4f8r-77","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:21:08.565Z"},{"id":"e-mermaid-mrxl4f8r-19","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-14","targetId":"n-mermaid-mrxl4f8r-16","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.831Z"},{"id":"e-mermaid-mrxl4f8r-190","title":"enables","sourceId":"n-mermaid-mrxl4f8r-58","targetId":"n-mermaid-mrxl4f8r-34","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:53:51.983Z"},{"id":"e-mermaid-mrxl4f8r-191","title":"enables","sourceId":"n-mermaid-mrxl4f8r-63","targetId":"n-mermaid-mrxl4f8r-78","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:20:54.729Z"},{"id":"e-mermaid-mrxl4f8r-192","title":"enables","sourceId":"n-mermaid-mrxl4f8r-63","targetId":"n-mermaid-mrxl4f8r-93","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:19:52.418Z"},{"id":"e-mermaid-mrxl4f8r-193","title":"enables","sourceId":"n-mermaid-mrxl4f8r-68","targetId":"n-mermaid-mrxl4f8r-104","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:42:02.478Z"},{"id":"e-mermaid-mrxl4f8r-194","title":"enables","sourceId":"n-mermaid-mrxl4f8r-69","targetId":"n-mermaid-mrxl4f8r-86","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:32:45.898Z"},{"id":"e-mermaid-mrxl4f8r-195","title":"enables","sourceId":"n-mermaid-mrxl4f8r-77","targetId":"n-mermaid-mrxl4f8r-96","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.903Z"},{"id":"e-mermaid-mrxl4f8r-196","title":"enables","sourceId":"n-mermaid-mrxl4f8r-77","targetId":"n-mermaid-mrxl4f8r-88","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.904Z"},{"id":"e-mermaid-mrxl4f8r-197","title":"enables","sourceId":"n-mermaid-mrxl4f8r-77","targetId":"n-mermaid-mrxl4f8r-95","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.904Z"},{"id":"e-mermaid-mrxl4f8r-198","title":"enables","sourceId":"n-mermaid-mrxl4f8r-77","targetId":"n-mermaid-mrxl4f8r-81","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:21:00.330Z"},{"id":"e-mermaid-mrxl4f8r-199","title":"enables","sourceId":"n-mermaid-mrxl4f8r-85","targetId":"n-mermaid-mrxl4f8r-83","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-2","title":"has_perspective","sourceId":"n-mermaid-mrxl4f8r-1","targetId":"n-mermaid-mrxl4f8r-3","typeId":"has-perspective","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:33.240Z"},{"id":"e-mermaid-mrxl4f8r-20","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-19","targetId":"n-mermaid-mrxl4f8r-20","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.831Z"},{"id":"e-mermaid-mrxl4f8r-200","title":"enables","sourceId":"n-mermaid-mrxl4f8r-85","targetId":"n-mermaid-mrxl4f8r-84","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-201","title":"enables","sourceId":"n-mermaid-mrxl4f8r-89","targetId":"n-mermaid-mrxl4f8r-90","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:34:40.263Z"},{"id":"e-mermaid-mrxl4f8r-202","title":"enables","sourceId":"n-mermaid-mrxl4f8r-89","targetId":"n-mermaid-mrxl4f8r-92","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:34:40.263Z"},{"id":"e-mermaid-mrxl4f8r-203","title":"enables","sourceId":"n-mermaid-mrxl4f8r-93","targetId":"n-mermaid-mrxl4f8r-97","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.905Z"},{"id":"e-mermaid-mrxl4f8r-204","title":"enables","sourceId":"n-mermaid-mrxl4f8r-71","targetId":"n-mermaid-mrxl4f8r-151","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-205","title":"enables","sourceId":"n-mermaid-mrxl4f8r-72","targetId":"n-mermaid-mrxl4f8r-52","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:18:10.145Z"},{"id":"e-mermaid-mrxl4f8r-206","title":"enables","sourceId":"n-mermaid-mrxl4f8r-147","targetId":"n-mermaid-mrxl4f8r-150","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:35:33.534Z"},{"id":"e-mermaid-mrxl4f8r-207","title":"enables","sourceId":"n-mermaid-mrxl4f8r-102","targetId":"n-mermaid-mrxl4f8r-104","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.427Z"},{"id":"e-mermaid-mrxl4f8r-208","title":"enables","sourceId":"n-mermaid-mrxl4f8r-70","targetId":"n-mermaid-mrxl4f8r-105","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:44:21.462Z"},{"id":"e-mermaid-mrxl4f8r-209","title":"enables","sourceId":"n-mermaid-mrxl4f8r-149","targetId":"n-mermaid-mrxl4f8r-147","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:35:33.535Z"},{"id":"e-mermaid-mrxl4f8r-21","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-2","targetId":"n-mermaid-mrxl4f8r-22","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.076Z"},{"id":"e-mermaid-mrxl4f8r-210","title":"enables","sourceId":"n-mermaid-mrxl4f8r-75","targetId":"n-mermaid-mrxl4f8r-76","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.032Z"},{"id":"e-mermaid-mrxl4f8r-211","title":"enables","sourceId":"n-mermaid-mrxl4f8r-79","targetId":"n-mermaid-mrxl4f8r-93","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:53:43.573Z"},{"id":"e-mermaid-mrxl4f8r-212","title":"enables","sourceId":"n-mermaid-mrxl4f8r-99","targetId":"n-mermaid-mrxl4f8r-166","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-213","title":"enables","sourceId":"n-mermaid-mrxl4f8r-35","targetId":"n-mermaid-mrxl4f8r-34","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.079Z"},{"id":"e-mermaid-mrxl4f8r-214","title":"enables","sourceId":"n-mermaid-mrxl4f8r-33","targetId":"n-mermaid-mrxl4f8r-171","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.081Z"},{"id":"e-mermaid-mrxl4f8r-215","title":"enables","sourceId":"n-mermaid-mrxl4f8r-166","targetId":"n-mermaid-mrxl4f8r-171","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.292Z"},{"id":"e-mermaid-mrxl4f8r-216","title":"enables","sourceId":"n-mermaid-mrxl4f8r-157","targetId":"n-mermaid-mrxl4f8r-159","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:33.325Z"},{"id":"e-mermaid-mrxl4f8r-217","title":"enables","sourceId":"n-mermaid-mrxl4f8r-158","targetId":"n-mermaid-mrxl4f8r-14","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:41:41.310Z"},{"id":"e-mermaid-mrxl4f8r-218","title":"enables","sourceId":"n-mermaid-mrxl4f8r-160","targetId":"n-mermaid-mrxl4f8r-15","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:33.329Z"},{"id":"e-mermaid-mrxl4f8r-219","title":"enables","sourceId":"n-mermaid-mrxl4f8r-161","targetId":"n-mermaid-mrxl4f8r-13","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:38:44.339Z"},{"id":"e-mermaid-mrxl4f8r-22","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-22","targetId":"n-mermaid-mrxl4f8r-23","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:53.258Z"},{"id":"e-mermaid-mrxl4f8r-220","title":"enables","sourceId":"n-mermaid-mrxl4f8r-162","targetId":"n-mermaid-mrxl4f8r-22","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:53.259Z"},{"id":"e-mermaid-mrxl4f8r-221","title":"enables","sourceId":"n-mermaid-mrxl4f8r-163","targetId":"n-mermaid-mrxl4f8r-96","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:30:16.886Z"},{"id":"e-mermaid-mrxl4f8r-222","title":"enables","sourceId":"n-mermaid-mrxl4f8r-164","targetId":"n-mermaid-mrxl4f8r-86","typeId":"enables","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:32:45.899Z"},{"id":"e-mermaid-mrxl4f8r-223","title":"explains","sourceId":"n-mermaid-mrxl4f8r-32","targetId":"n-mermaid-mrxl4f8r-166","typeId":"explains","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:24:34.471Z"},{"id":"e-mermaid-mrxl4f8r-224","title":"explains","sourceId":"n-mermaid-mrxl4f8r-167","targetId":"n-mermaid-mrxl4f8r-166","typeId":"explains","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.294Z"},{"id":"e-mermaid-mrxl4f8r-225","title":"explains","sourceId":"n-mermaid-mrxl4f8r-169","targetId":"n-mermaid-mrxl4f8r-77","typeId":"explains","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:21:00.330Z"},{"id":"e-mermaid-mrxl4f8r-226","title":"explains","sourceId":"n-mermaid-mrxl4f8r-170","targetId":"n-mermaid-mrxl4f8r-77","typeId":"explains","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:21:00.331Z"},{"id":"e-mermaid-mrxl4f8r-227","title":"spotlights","sourceId":"n-mermaid-mrxl4f8r-165","targetId":"n-mermaid-mrxl4f8r-34","typeId":"spotlights","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:44.736Z"},{"id":"e-mermaid-mrxl4f8r-228","title":"spotlights","sourceId":"n-mermaid-mrxl4f8r-165","targetId":"n-mermaid-mrxl4f8r-74","typeId":"spotlights","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:44.742Z"},{"id":"e-mermaid-mrxl4f8r-229","title":"spotlights","sourceId":"n-mermaid-mrxl4f8r-165","targetId":"n-mermaid-mrxl4f8r-75","typeId":"spotlights","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:44.737Z"},{"id":"e-mermaid-mrxl4f8r-23","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-22","targetId":"n-mermaid-mrxl4f8r-24","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:53.260Z"},{"id":"e-mermaid-mrxl4f8r-230","title":"spotlights","sourceId":"n-mermaid-mrxl4f8r-165","targetId":"n-mermaid-mrxl4f8r-76","typeId":"spotlights","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:44.737Z"},{"id":"e-mermaid-mrxl4f8r-231","title":"spotlights","sourceId":"n-mermaid-mrxl4f8r-165","targetId":"n-mermaid-mrxl4f8r-79","typeId":"spotlights","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:44.738Z"},{"id":"e-mermaid-mrxl4f8r-232","title":"spotlights","sourceId":"n-mermaid-mrxl4f8r-165","targetId":"n-mermaid-mrxl4f8r-35","typeId":"spotlights","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:44.739Z"},{"id":"e-mermaid-mrxl4f8r-233","title":"spotlights","sourceId":"n-mermaid-mrxl4f8r-165","targetId":"n-mermaid-mrxl4f8r-152","typeId":"spotlights","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:44.735Z"},{"id":"e-mermaid-mrxl4f8r-234","title":"spotlights","sourceId":"n-mermaid-mrxl4f8r-173","targetId":"n-mermaid-mrxl4f8r-49","typeId":"spotlights","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.334Z"},{"id":"e-mermaid-mrxl4f8r-235","title":"spotlights","sourceId":"n-mermaid-mrxl4f8r-174","targetId":"n-mermaid-mrxl4f8r-149","typeId":"spotlights","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:23:16.297Z"},{"id":"e-mermaid-mrxl4f8r-236","title":"spotlights","sourceId":"n-mermaid-mrxl4f8r-174","targetId":"n-mermaid-mrxl4f8r-147","typeId":"spotlights","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:23:16.298Z"},{"id":"e-mermaid-mrxl4f8r-237","title":"spotlights","sourceId":"n-mermaid-mrxl4f8r-174","targetId":"n-mermaid-mrxl4f8r-159","typeId":"spotlights","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:28.577Z"},{"id":"e-mermaid-mrxl4f8r-238","title":"leads","sourceId":"n-mermaid-mrxl4f8r-116","targetId":"n-mermaid-mrxl4f8r-107","typeId":"leads","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:30:16.890Z"},{"id":"e-mermaid-mrxl4f8r-239","title":"leads","sourceId":"n-mermaid-mrxl4f8r-117","targetId":"n-mermaid-mrxl4f8r-108","typeId":"leads","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.126Z"},{"id":"e-mermaid-mrxl4f8r-24","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-22","targetId":"n-mermaid-mrxl4f8r-25","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:53.260Z"},{"id":"e-mermaid-mrxl4f8r-240","title":"leads","sourceId":"n-mermaid-mrxl4f8r-118","targetId":"n-mermaid-mrxl4f8r-109","typeId":"leads","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:36:29.648Z"},{"id":"e-mermaid-mrxl4f8r-241","title":"leads","sourceId":"n-mermaid-mrxl4f8r-119","targetId":"n-mermaid-mrxl4f8r-111","typeId":"leads","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:20:48.847Z"},{"id":"e-mermaid-mrxl4f8r-242","title":"leads","sourceId":"n-mermaid-mrxl4f8r-120","targetId":"n-mermaid-mrxl4f8r-127","typeId":"leads","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:18:10.143Z"},{"id":"e-mermaid-mrxl4f8r-243","title":"leads","sourceId":"n-mermaid-mrxl4f8r-121","targetId":"n-mermaid-mrxl4f8r-133","typeId":"leads","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.124Z"},{"id":"e-mermaid-mrxl4f8r-244","title":"leads","sourceId":"n-mermaid-mrxl4f8r-122","targetId":"n-mermaid-mrxl4f8r-110","typeId":"leads","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.124Z"},{"id":"e-mermaid-mrxl4f8r-245","title":"leads","sourceId":"n-mermaid-mrxl4f8r-123","targetId":"n-mermaid-mrxl4f8r-113","typeId":"leads","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.125Z"},{"id":"e-mermaid-mrxl4f8r-246","title":"leads","sourceId":"n-mermaid-mrxl4f8r-124","targetId":"n-mermaid-mrxl4f8r-133","typeId":"leads","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.124Z"},{"id":"e-mermaid-mrxl4f8r-247","title":"leads","sourceId":"n-mermaid-mrxl4f8r-125","targetId":"n-mermaid-mrxl4f8r-112","typeId":"leads","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.123Z"},{"id":"e-mermaid-mrxl4f8r-248","title":"drives","sourceId":"n-mermaid-mrxl4f8r-38","targetId":"n-mermaid-mrxl4f8r-16","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:36:03.773Z"},{"id":"e-mermaid-mrxl4f8r-249","title":"drives","sourceId":"n-mermaid-mrxl4f8r-38","targetId":"n-mermaid-mrxl4f8r-160","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:11.020Z"},{"id":"e-mermaid-mrxl4f8r-25","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-22","targetId":"n-mermaid-mrxl4f8r-26","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:53.261Z"},{"id":"e-mermaid-mrxl4f8r-250","title":"drives","sourceId":"n-mermaid-mrxl4f8r-38","targetId":"n-mermaid-mrxl4f8r-165","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:44.739Z"},{"id":"e-mermaid-mrxl4f8r-251","title":"drives","sourceId":"n-mermaid-mrxl4f8r-39","targetId":"n-mermaid-mrxl4f8r-15","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:11.017Z"},{"id":"e-mermaid-mrxl4f8r-252","title":"drives","sourceId":"n-mermaid-mrxl4f8r-39","targetId":"n-mermaid-mrxl4f8r-170","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:05.944Z"},{"id":"e-mermaid-mrxl4f8r-253","title":"drives","sourceId":"n-mermaid-mrxl4f8r-40","targetId":"n-mermaid-mrxl4f8r-15","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:07:49.522Z"},{"id":"e-mermaid-mrxl4f8r-254","title":"drives","sourceId":"n-mermaid-mrxl4f8r-40","targetId":"n-mermaid-mrxl4f8r-152","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:40.903Z"},{"id":"e-mermaid-mrxl4f8r-255","title":"drives","sourceId":"n-mermaid-mrxl4f8r-41","targetId":"n-mermaid-mrxl4f8r-160","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:11.021Z"},{"id":"e-mermaid-mrxl4f8r-256","title":"drives","sourceId":"n-mermaid-mrxl4f8r-41","targetId":"n-mermaid-mrxl4f8r-49","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.335Z"},{"id":"e-mermaid-mrxl4f8r-257","title":"drives","sourceId":"n-mermaid-mrxl4f8r-42","targetId":"n-mermaid-mrxl4f8r-32","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:24:35.579Z"},{"id":"e-mermaid-mrxl4f8r-258","title":"drives","sourceId":"n-mermaid-mrxl4f8r-42","targetId":"n-mermaid-mrxl4f8r-160","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:45.923Z"},{"id":"e-mermaid-mrxl4f8r-259","title":"drives","sourceId":"n-mermaid-mrxl4f8r-43","targetId":"n-mermaid-mrxl4f8r-10","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.830Z"},{"id":"e-mermaid-mrxl4f8r-26","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-22","targetId":"n-mermaid-mrxl4f8r-27","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:03:08.023Z"},{"id":"e-mermaid-mrxl4f8r-260","title":"drives","sourceId":"n-mermaid-mrxl4f8r-44","targetId":"n-mermaid-mrxl4f8r-14","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:41:41.311Z"},{"id":"e-mermaid-mrxl4f8r-261","title":"drives","sourceId":"n-mermaid-mrxl4f8r-45","targetId":"n-mermaid-mrxl4f8r-95","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-262","title":"drives","sourceId":"n-mermaid-mrxl4f8r-116","targetId":"n-mermaid-mrxl4f8r-171","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:11:40.831Z"},{"id":"e-mermaid-mrxl4f8r-263","title":"drives","sourceId":"n-mermaid-mrxl4f8r-117","targetId":"n-mermaid-mrxl4f8r-34","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:13:05.912Z"},{"id":"e-mermaid-mrxl4f8r-264","title":"drives","sourceId":"n-mermaid-mrxl4f8r-118","targetId":"n-mermaid-mrxl4f8r-161","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:36:30.671Z"},{"id":"e-mermaid-mrxl4f8r-265","title":"drives","sourceId":"n-mermaid-mrxl4f8r-118","targetId":"n-mermaid-mrxl4f8r-13","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:36:30.669Z"},{"id":"e-mermaid-mrxl4f8r-266","title":"drives","sourceId":"n-mermaid-mrxl4f8r-119","targetId":"n-mermaid-mrxl4f8r-147","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:20:48.843Z"},{"id":"e-mermaid-mrxl4f8r-267","title":"drives","sourceId":"n-mermaid-mrxl4f8r-106","targetId":"n-mermaid-mrxl4f8r-54","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:21:08.567Z"},{"id":"e-mermaid-mrxl4f8r-268","title":"drives","sourceId":"n-mermaid-mrxl4f8r-107","targetId":"n-mermaid-mrxl4f8r-163","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:30:16.887Z"},{"id":"e-mermaid-mrxl4f8r-269","title":"drives","sourceId":"n-mermaid-mrxl4f8r-107","targetId":"n-mermaid-mrxl4f8r-164","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:31:36.461Z"},{"id":"e-mermaid-mrxl4f8r-27","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-22","targetId":"n-mermaid-mrxl4f8r-28","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:53.262Z"},{"id":"e-mermaid-mrxl4f8r-270","title":"drives","sourceId":"n-mermaid-mrxl4f8r-107","targetId":"n-mermaid-mrxl4f8r-96","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:30:16.888Z"},{"id":"e-mermaid-mrxl4f8r-271","title":"drives","sourceId":"n-mermaid-mrxl4f8r-108","targetId":"n-mermaid-mrxl4f8r-34","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:13:05.910Z"},{"id":"e-mermaid-mrxl4f8r-272","title":"drives","sourceId":"n-mermaid-mrxl4f8r-108","targetId":"n-mermaid-mrxl4f8r-92","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:30:47.593Z"},{"id":"e-mermaid-mrxl4f8r-273","title":"drives","sourceId":"n-mermaid-mrxl4f8r-109","targetId":"n-mermaid-mrxl4f8r-161","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:36:29.649Z"},{"id":"e-mermaid-mrxl4f8r-274","title":"drives","sourceId":"n-mermaid-mrxl4f8r-109","targetId":"n-mermaid-mrxl4f8r-162","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:19:10.082Z"},{"id":"e-mermaid-mrxl4f8r-275","title":"drives","sourceId":"n-mermaid-mrxl4f8r-109","targetId":"n-mermaid-mrxl4f8r-13","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:36:29.648Z"},{"id":"e-mermaid-mrxl4f8r-276","title":"drives","sourceId":"n-mermaid-mrxl4f8r-110","targetId":"n-mermaid-mrxl4f8r-102","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:25.161Z"},{"id":"e-mermaid-mrxl4f8r-277","title":"drives","sourceId":"n-mermaid-mrxl4f8r-111","targetId":"n-mermaid-mrxl4f8r-147","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:20:48.843Z"},{"id":"e-mermaid-mrxl4f8r-278","title":"drives","sourceId":"n-mermaid-mrxl4f8r-112","targetId":"n-mermaid-mrxl4f8r-102","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:25.161Z"},{"id":"e-mermaid-mrxl4f8r-279","title":"drives","sourceId":"n-mermaid-mrxl4f8r-112","targetId":"n-mermaid-mrxl4f8r-27","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.035Z"},{"id":"e-mermaid-mrxl4f8r-28","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-22","targetId":"n-mermaid-mrxl4f8r-29","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:47:26.276Z"},{"id":"e-mermaid-mrxl4f8r-280","title":"drives","sourceId":"n-mermaid-mrxl4f8r-113","targetId":"n-mermaid-mrxl4f8r-70","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.033Z"},{"id":"e-mermaid-mrxl4f8r-281","title":"drives","sourceId":"n-mermaid-mrxl4f8r-113","targetId":"n-mermaid-mrxl4f8r-148","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.035Z"},{"id":"e-mermaid-mrxl4f8r-282","title":"drives","sourceId":"n-mermaid-mrxl4f8r-113","targetId":"n-mermaid-mrxl4f8r-151","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-283","title":"drives","sourceId":"n-mermaid-mrxl4f8r-113","targetId":"n-mermaid-mrxl4f8r-102","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-284","title":"drives","sourceId":"n-mermaid-mrxl4f8r-113","targetId":"n-mermaid-mrxl4f8r-27","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.034Z"},{"id":"e-mermaid-mrxl4f8r-285","title":"drives","sourceId":"n-mermaid-mrxl4f8r-114","targetId":"n-mermaid-mrxl4f8r-102","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:25.162Z"},{"id":"e-mermaid-mrxl4f8r-286","title":"drives","sourceId":"n-mermaid-mrxl4f8r-114","targetId":"n-mermaid-mrxl4f8r-105","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:52:23.379Z"},{"id":"e-mermaid-mrxl4f8r-287","title":"drives","sourceId":"n-mermaid-mrxl4f8r-127","targetId":"n-mermaid-mrxl4f8r-72","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:18:10.144Z"},{"id":"e-mermaid-mrxl4f8r-288","title":"drives","sourceId":"n-mermaid-mrxl4f8r-127","targetId":"n-mermaid-mrxl4f8r-52","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:20:48.848Z"},{"id":"e-mermaid-mrxl4f8r-289","title":"drives","sourceId":"n-mermaid-mrxl4f8r-128","targetId":"n-mermaid-mrxl4f8r-72","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-29","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-29","targetId":"n-mermaid-mrxl4f8r-30","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:47:26.275Z"},{"id":"e-mermaid-mrxl4f8r-290","title":"drives","sourceId":"n-mermaid-mrxl4f8r-129","targetId":"n-mermaid-mrxl4f8r-71","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-291","title":"drives","sourceId":"n-mermaid-mrxl4f8r-130","targetId":"n-mermaid-mrxl4f8r-147","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:20:48.844Z"},{"id":"e-mermaid-mrxl4f8r-292","title":"drives","sourceId":"n-mermaid-mrxl4f8r-131","targetId":"n-mermaid-mrxl4f8r-148","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-293","title":"drives","sourceId":"n-mermaid-mrxl4f8r-131","targetId":"n-mermaid-mrxl4f8r-71","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-294","title":"drives","sourceId":"n-mermaid-mrxl4f8r-133","targetId":"n-mermaid-mrxl4f8r-95","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-295","title":"drives","sourceId":"n-mermaid-mrxl4f8r-134","targetId":"n-mermaid-mrxl4f8r-162","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:19:10.083Z"},{"id":"e-mermaid-mrxl4f8r-296","title":"drives","sourceId":"n-mermaid-mrxl4f8r-134","targetId":"n-mermaid-mrxl4f8r-73","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-297","title":"drives","sourceId":"n-mermaid-mrxl4f8r-134","targetId":"n-mermaid-mrxl4f8r-97","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-298","title":"drives","sourceId":"n-mermaid-mrxl4f8r-135","targetId":"n-mermaid-mrxl4f8r-103","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:05.478Z"},{"id":"e-mermaid-mrxl4f8r-299","title":"drives","sourceId":"n-mermaid-mrxl4f8r-135","targetId":"n-mermaid-mrxl4f8r-73","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-3","title":"has_perspective","sourceId":"n-mermaid-mrxl4f8r-1","targetId":"n-mermaid-mrxl4f8r-4","typeId":"has-perspective","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:33.241Z"},{"id":"e-mermaid-mrxl4f8r-30","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-2","targetId":"n-mermaid-mrxl4f8r-31","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.077Z"},{"id":"e-mermaid-mrxl4f8r-300","title":"drives","sourceId":"n-mermaid-mrxl4f8r-136","targetId":"n-mermaid-mrxl4f8r-102","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:25.163Z"},{"id":"e-mermaid-mrxl4f8r-301","title":"drives","sourceId":"n-mermaid-mrxl4f8r-137","targetId":"n-mermaid-mrxl4f8r-103","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:05.479Z"},{"id":"e-mermaid-mrxl4f8r-302","title":"drives","sourceId":"n-mermaid-mrxl4f8r-138","targetId":"n-mermaid-mrxl4f8r-103","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:05.479Z"},{"id":"e-mermaid-mrxl4f8r-303","title":"drives","sourceId":"n-mermaid-mrxl4f8r-139","targetId":"n-mermaid-mrxl4f8r-50","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-304","title":"drives","sourceId":"n-mermaid-mrxl4f8r-141","targetId":"n-mermaid-mrxl4f8r-97","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-305","title":"drives","sourceId":"n-mermaid-mrxl4f8r-142","targetId":"n-mermaid-mrxl4f8r-95","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-306","title":"drives","sourceId":"n-mermaid-mrxl4f8r-143","targetId":"n-mermaid-mrxl4f8r-83","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-307","title":"drives","sourceId":"n-mermaid-mrxl4f8r-144","targetId":"n-mermaid-mrxl4f8r-84","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-308","title":"drives","sourceId":"n-mermaid-mrxl4f8r-145","targetId":"n-mermaid-mrxl4f8r-82","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-309","title":"drives","sourceId":"n-mermaid-mrxl4f8r-145","targetId":"n-mermaid-mrxl4f8r-155","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:24.266Z"},{"id":"e-mermaid-mrxl4f8r-31","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-31","targetId":"n-mermaid-mrxl4f8r-32","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.081Z"},{"id":"e-mermaid-mrxl4f8r-310","title":"drives","sourceId":"n-mermaid-mrxl4f8r-146","targetId":"n-mermaid-mrxl4f8r-96","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:21:26.340Z"},{"id":"e-mermaid-mrxl4f8r-311","title":"drives","sourceId":"n-mermaid-mrxl4f8r-146","targetId":"n-mermaid-mrxl4f8r-156","typeId":"drives","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:15.900Z"},{"id":"e-mermaid-mrxl4f8r-32","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-2","targetId":"n-mermaid-mrxl4f8r-33","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.077Z"},{"id":"e-mermaid-mrxl4f8r-33","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-2","targetId":"n-mermaid-mrxl4f8r-34","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.078Z"},{"id":"e-mermaid-mrxl4f8r-34","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-2","targetId":"n-mermaid-mrxl4f8r-35","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.078Z"},{"id":"e-mermaid-mrxl4f8r-35","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-35","targetId":"n-mermaid-mrxl4f8r-36","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.082Z"},{"id":"e-mermaid-mrxl4f8r-36","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-2","targetId":"n-mermaid-mrxl4f8r-37","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.076Z"},{"id":"e-mermaid-mrxl4f8r-37","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-37","targetId":"n-mermaid-mrxl4f8r-38","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:50.504Z"},{"id":"e-mermaid-mrxl4f8r-38","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-37","targetId":"n-mermaid-mrxl4f8r-39","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:50.505Z"},{"id":"e-mermaid-mrxl4f8r-39","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-37","targetId":"n-mermaid-mrxl4f8r-40","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:07:49.522Z"},{"id":"e-mermaid-mrxl4f8r-4","title":"has_perspective","sourceId":"n-mermaid-mrxl4f8r-1","targetId":"n-mermaid-mrxl4f8r-5","typeId":"has-perspective","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.426Z"},{"id":"e-mermaid-mrxl4f8r-40","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-37","targetId":"n-mermaid-mrxl4f8r-41","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:50.506Z"},{"id":"e-mermaid-mrxl4f8r-41","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-37","targetId":"n-mermaid-mrxl4f8r-42","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:50.502Z"},{"id":"e-mermaid-mrxl4f8r-42","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-37","targetId":"n-mermaid-mrxl4f8r-43","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:26:03.889Z"},{"id":"e-mermaid-mrxl4f8r-43","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-37","targetId":"n-mermaid-mrxl4f8r-44","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:50.502Z"},{"id":"e-mermaid-mrxl4f8r-44","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-37","targetId":"n-mermaid-mrxl4f8r-45","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:50.507Z"},{"id":"e-mermaid-mrxl4f8r-45","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-3","targetId":"n-mermaid-mrxl4f8r-46","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.027Z"},{"id":"e-mermaid-mrxl4f8r-46","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-46","targetId":"n-mermaid-mrxl4f8r-47","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.330Z"},{"id":"e-mermaid-mrxl4f8r-47","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-46","targetId":"n-mermaid-mrxl4f8r-48","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.330Z"},{"id":"e-mermaid-mrxl4f8r-48","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-46","targetId":"n-mermaid-mrxl4f8r-49","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.331Z"},{"id":"e-mermaid-mrxl4f8r-49","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-46","targetId":"n-mermaid-mrxl4f8r-50","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.331Z"},{"id":"e-mermaid-mrxl4f8r-5","title":"has_perspective","sourceId":"n-mermaid-mrxl4f8r-1","targetId":"n-mermaid-mrxl4f8r-6","typeId":"has-perspective","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:33.243Z"},{"id":"e-mermaid-mrxl4f8r-50","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-46","targetId":"n-mermaid-mrxl4f8r-51","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.332Z"},{"id":"e-mermaid-mrxl4f8r-51","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-46","targetId":"n-mermaid-mrxl4f8r-52","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.332Z"},{"id":"e-mermaid-mrxl4f8r-52","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-46","targetId":"n-mermaid-mrxl4f8r-53","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.333Z"},{"id":"e-mermaid-mrxl4f8r-53","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-46","targetId":"n-mermaid-mrxl4f8r-54","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.329Z"},{"id":"e-mermaid-mrxl4f8r-54","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-3","targetId":"n-mermaid-mrxl4f8r-55","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.027Z"},{"id":"e-mermaid-mrxl4f8r-55","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-55","targetId":"n-mermaid-mrxl4f8r-56","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-56","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-55","targetId":"n-mermaid-mrxl4f8r-58","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:53:51.984Z"},{"id":"e-mermaid-mrxl4f8r-57","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-55","targetId":"n-mermaid-mrxl4f8r-60","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:32:45.904Z"},{"id":"e-mermaid-mrxl4f8r-58","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-55","targetId":"n-mermaid-mrxl4f8r-61","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-59","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-56","targetId":"n-mermaid-mrxl4f8r-57","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-6","title":"has_perspective","sourceId":"n-mermaid-mrxl4f8r-1","targetId":"n-mermaid-mrxl4f8r-7","typeId":"has-perspective","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:33.242Z"},{"id":"e-mermaid-mrxl4f8r-60","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-58","targetId":"n-mermaid-mrxl4f8r-59","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:53:51.982Z"},{"id":"e-mermaid-mrxl4f8r-61","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-3","targetId":"n-mermaid-mrxl4f8r-62","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.028Z"},{"id":"e-mermaid-mrxl4f8r-62","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-62","targetId":"n-mermaid-mrxl4f8r-63","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:19:52.416Z"},{"id":"e-mermaid-mrxl4f8r-63","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-62","targetId":"n-mermaid-mrxl4f8r-64","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:19:52.420Z"},{"id":"e-mermaid-mrxl4f8r-64","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-62","targetId":"n-mermaid-mrxl4f8r-67","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:04:40.477Z"},{"id":"e-mermaid-mrxl4f8r-65","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-62","targetId":"n-mermaid-mrxl4f8r-68","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.033Z"},{"id":"e-mermaid-mrxl4f8r-66","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-62","targetId":"n-mermaid-mrxl4f8r-69","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:04:40.478Z"},{"id":"e-mermaid-mrxl4f8r-67","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-64","targetId":"n-mermaid-mrxl4f8r-65","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:07:19.715Z"},{"id":"e-mermaid-mrxl4f8r-68","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-64","targetId":"n-mermaid-mrxl4f8r-66","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:07:19.713Z"},{"id":"e-mermaid-mrxl4f8r-69","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-3","targetId":"n-mermaid-mrxl4f8r-70","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.028Z"},{"id":"e-mermaid-mrxl4f8r-7","title":"has_perspective","sourceId":"n-mermaid-mrxl4f8r-1","targetId":"n-mermaid-mrxl4f8r-8","typeId":"has-perspective","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:33.245Z"},{"id":"e-mermaid-mrxl4f8r-70","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-3","targetId":"n-mermaid-mrxl4f8r-71","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.029Z"},{"id":"e-mermaid-mrxl4f8r-71","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-71","targetId":"n-mermaid-mrxl4f8r-72","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-72","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-71","targetId":"n-mermaid-mrxl4f8r-73","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-73","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-3","targetId":"n-mermaid-mrxl4f8r-74","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.029Z"},{"id":"e-mermaid-mrxl4f8r-74","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-74","targetId":"n-mermaid-mrxl4f8r-75","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.031Z"},{"id":"e-mermaid-mrxl4f8r-75","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-3","targetId":"n-mermaid-mrxl4f8r-76","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.030Z"},{"id":"e-mermaid-mrxl4f8r-76","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-4","targetId":"n-mermaid-mrxl4f8r-77","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.898Z"},{"id":"e-mermaid-mrxl4f8r-77","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-77","targetId":"n-mermaid-mrxl4f8r-78","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:21:00.327Z"},{"id":"e-mermaid-mrxl4f8r-78","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-77","targetId":"n-mermaid-mrxl4f8r-79","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:21:00.328Z"},{"id":"e-mermaid-mrxl4f8r-79","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-4","targetId":"n-mermaid-mrxl4f8r-80","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.898Z"},{"id":"e-mermaid-mrxl4f8r-8","title":"has_perspective","sourceId":"n-mermaid-mrxl4f8r-1","targetId":"n-mermaid-mrxl4f8r-9","typeId":"has-perspective","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:33.242Z"},{"id":"e-mermaid-mrxl4f8r-80","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-80","targetId":"n-mermaid-mrxl4f8r-81","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-81","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-80","targetId":"n-mermaid-mrxl4f8r-82","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-82","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-80","targetId":"n-mermaid-mrxl4f8r-83","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-83","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-80","targetId":"n-mermaid-mrxl4f8r-84","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-84","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-80","targetId":"n-mermaid-mrxl4f8r-85","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-85","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-4","targetId":"n-mermaid-mrxl4f8r-86","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.899Z"},{"id":"e-mermaid-mrxl4f8r-86","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-86","targetId":"n-mermaid-mrxl4f8r-87","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.906Z"},{"id":"e-mermaid-mrxl4f8r-87","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-4","targetId":"n-mermaid-mrxl4f8r-88","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:34:01.569Z"},{"id":"e-mermaid-mrxl4f8r-88","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-88","targetId":"n-mermaid-mrxl4f8r-89","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:34:40.264Z"},{"id":"e-mermaid-mrxl4f8r-89","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-88","targetId":"n-mermaid-mrxl4f8r-90","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:34:40.264Z"},{"id":"e-mermaid-mrxl4f8r-9","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-2","targetId":"n-mermaid-mrxl4f8r-10","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.075Z"},{"id":"e-mermaid-mrxl4f8r-90","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-88","targetId":"n-mermaid-mrxl4f8r-91","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:34:01.570Z"},{"id":"e-mermaid-mrxl4f8r-91","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-88","targetId":"n-mermaid-mrxl4f8r-92","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:34:40.263Z"},{"id":"e-mermaid-mrxl4f8r-92","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-4","targetId":"n-mermaid-mrxl4f8r-93","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.900Z"},{"id":"e-mermaid-mrxl4f8r-93","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-93","targetId":"n-mermaid-mrxl4f8r-94","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.906Z"},{"id":"e-mermaid-mrxl4f8r-94","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-4","targetId":"n-mermaid-mrxl4f8r-95","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.901Z"},{"id":"e-mermaid-mrxl4f8r-95","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-4","targetId":"n-mermaid-mrxl4f8r-96","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.902Z"},{"id":"e-mermaid-mrxl4f8r-96","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-4","targetId":"n-mermaid-mrxl4f8r-97","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.902Z"},{"id":"e-mermaid-mrxl4f8r-97","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-4","targetId":"n-mermaid-mrxl4f8r-98","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.903Z"},{"id":"e-mermaid-mrxl4f8r-98","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-98","targetId":"n-mermaid-mrxl4f8r-99","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e-mermaid-mrxl4f8r-99","title":"consists_of","sourceId":"n-mermaid-mrxl4f8r-99","targetId":"n-mermaid-mrxl4f8r-100","typeId":"consists-of","tags":[],"properties":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"e1784815820916","title":"","sourceId":"n-mermaid-mrxl4f8r-162","targetId":"n-mermaid-mrxl4f8r-23","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:10:20.916Z","updatedAt":"2026-07-23T17:42:53.259Z"},{"id":"e1784815828260","title":"","sourceId":"n-mermaid-mrxl4f8r-162","targetId":"n-1784-815-828-256","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:10:28.260Z","updatedAt":"2026-07-23T14:19:10.083Z"},{"id":"e1784815855136","title":"","sourceId":"n-mermaid-mrxl4f8r-162","targetId":"n-1784-815-855-132","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:10:55.136Z","updatedAt":"2026-07-23T14:19:10.084Z"},{"id":"e1784815895320","title":"","sourceId":"n-mermaid-mrxl4f8r-23","targetId":"n-1784-815-862-176","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:11:35.320Z","updatedAt":"2026-07-23T17:42:53.263Z"},{"id":"e1784815957303","title":"","sourceId":"n-mermaid-mrxl4f8r-22","targetId":"n-1784-815-855-132","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:12:37.303Z","updatedAt":"2026-07-23T14:19:10.084Z"},{"id":"e1784816030720","title":"","sourceId":"n-mermaid-mrxl4f8r-23","targetId":"n-1784-816-030-709","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:13:50.720Z","updatedAt":"2026-07-23T17:53:38.434Z"},{"id":"e1784816091551","title":"","sourceId":"n-mermaid-mrxl4f8r-23","targetId":"n-1784-816-091-547","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:14:51.551Z","updatedAt":"2026-07-23T14:16:03.902Z"},{"id":"e1784816097657","title":"","sourceId":"n-mermaid-mrxl4f8r-23","targetId":"n-1784-816-097-653","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:14:57.657Z","updatedAt":"2026-07-23T14:16:03.902Z"},{"id":"e1784816099237","title":"","sourceId":"n-mermaid-mrxl4f8r-23","targetId":"n-1784-816-099-231","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:14:59.237Z","updatedAt":"2026-07-23T17:53:38.431Z"},{"id":"e1784816107393","title":"","sourceId":"n-mermaid-mrxl4f8r-23","targetId":"n-1784-816-107-389","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:15:07.393Z","updatedAt":"2026-07-23T14:16:03.904Z"},{"id":"e1784816112844","title":"","sourceId":"n-mermaid-mrxl4f8r-23","targetId":"n-1784-816-112-838","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:15:12.844Z","updatedAt":"2026-07-23T14:18:37.027Z"},{"id":"e1784816121310","title":"","sourceId":"n-1784-816-112-838","targetId":"n-1784-816-030-709","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:15:21.310Z","updatedAt":"2026-07-23T14:18:37.028Z"},{"id":"e1784816185121","title":"","sourceId":"n-1784-815-862-176","targetId":"n-1784-815-855-132","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:16:25.121Z","updatedAt":"2026-07-23T14:17:43.847Z"},{"id":"e1784816188655","title":"","sourceId":"n-1784-815-862-176","targetId":"n-1784-816-112-838","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:16:28.656Z","updatedAt":"2026-07-23T14:17:43.847Z"},{"id":"e1784816192903","title":"","sourceId":"n-1784-815-862-176","targetId":"n-1784-816-107-389","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:16:32.903Z","updatedAt":"2026-07-23T14:17:43.848Z"},{"id":"e1784816303584","title":"","sourceId":"n-1784-816-030-709","targetId":"n-1784-816-099-231","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:18:23.584Z","updatedAt":"2026-07-23T17:53:38.432Z"},{"id":"e1784816311221","title":"","sourceId":"n-1784-816-030-709","targetId":"n-1784-816-107-389","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:18:31.221Z","updatedAt":"2026-07-23T14:18:50.177Z"},{"id":"e1784816384708","title":"","sourceId":"n-mermaid-mrxl4f8r-163","targetId":"n-1784-816-384-703","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:19:44.708Z","updatedAt":"2026-07-23T17:30:16.888Z"},{"id":"e1784816434920","title":"","sourceId":"n-mermaid-mrxl4f8r-96","targetId":"n-mermaid-mrxl4f8r-88","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:20:34.920Z","updatedAt":"2026-07-23T18:21:00.335Z"},{"id":"e1784816445301","title":"","sourceId":"n-mermaid-mrxl4f8r-96","targetId":"n-mermaid-mrxl4f8r-106","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:20:45.301Z","updatedAt":"2026-07-23T17:30:16.891Z"},{"id":"e1784816478435","title":"","sourceId":"n-mermaid-mrxl4f8r-54","targetId":"n-mermaid-mrxl4f8r-80","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:21:18.435Z","updatedAt":"2026-07-23T18:21:08.567Z"},{"id":"e1784816507389","title":"","sourceId":"n-mermaid-mrxl4f8r-54","targetId":"n-mermaid-mrxl4f8r-53","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:21:47.389Z","updatedAt":"2026-07-23T18:21:08.566Z"},{"id":"e1784816617362","title":"","sourceId":"n-mermaid-mrxl4f8r-55","targetId":"n-mermaid-mrxl4f8r-54","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:23:37.363Z","updatedAt":"2026-07-23T14:24:12.935Z"},{"id":"e1784816637547","title":"","sourceId":"n-mermaid-mrxl4f8r-53","targetId":"n-mermaid-mrxl4f8r-52","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:23:57.547Z","updatedAt":"2026-07-23T14:24:12.936Z"},{"id":"e1784816639126","title":"","sourceId":"n-mermaid-mrxl4f8r-53","targetId":"n-mermaid-mrxl4f8r-49","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:23:59.126Z","updatedAt":"2026-07-23T18:22:47.350Z"},{"id":"e1784816979287","title":"","sourceId":"n-mermaid-mrxl4f8r-46","targetId":"n-mermaid-mrxl4f8r-22","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:29:39.287Z","updatedAt":"2026-07-23T18:21:08.569Z"},{"id":"e1784818168450","title":"","sourceId":"n-mermaid-mrxl4f8r-55","targetId":"n-mermaid-mrxl4f8r-54","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:49:28.450Z","updatedAt":"2026-07-23T18:21:08.569Z"},{"id":"e1784818628202","title":"","sourceId":"n-mermaid-mrxl4f8r-54","targetId":"n-mermaid-mrxl4f8r-126","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T14:57:08.202Z","updatedAt":"2026-07-23T17:19:12.357Z"},{"id":"e1784823059856","title":"","sourceId":"n-mermaid-mrxl4f8r-171","targetId":"n-1784-823-059-854","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:10:59.856Z","updatedAt":"2026-07-23T16:46:10.712Z"},{"id":"e1784823066856","title":"","sourceId":"n-mermaid-mrxl4f8r-171","targetId":"n-1784-823-066-853","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:11:06.856Z","updatedAt":"2026-07-23T16:46:10.713Z"},{"id":"e1784823072508","title":"","sourceId":"n-mermaid-mrxl4f8r-171","targetId":"n-1784-823-072-505","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:11:12.508Z","updatedAt":"2026-07-23T16:50:43.486Z"},{"id":"e1784823083576","title":"","sourceId":"n-mermaid-mrxl4f8r-171","targetId":"n-1784-823-083-572","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:11:23.576Z","updatedAt":"2026-07-23T16:46:10.714Z"},{"id":"e1784823094290","title":"","sourceId":"n-mermaid-mrxl4f8r-171","targetId":"n-1784-823-094-285","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:11:34.290Z","updatedAt":"2026-07-23T16:46:10.715Z"},{"id":"e1784823155966","title":"","sourceId":"n-mermaid-mrxl4f8r-172","targetId":"n-1784-823-094-285","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:12:35.966Z","updatedAt":"2026-07-23T16:13:34.387Z"},{"id":"e1784823161495","title":"","sourceId":"n-mermaid-mrxl4f8r-172","targetId":"n-mermaid-mrxl4f8r-31","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:12:41.495Z","updatedAt":"2026-07-23T18:25:45.198Z"},{"id":"e1784823238320","title":"","sourceId":"n-mermaid-mrxl4f8r-33","targetId":"n-mermaid-mrxl4f8r-78","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:13:58.320Z","updatedAt":"2026-07-23T18:39:11.082Z"},{"id":"e1784823244092","title":"","sourceId":"n-mermaid-mrxl4f8r-33","targetId":"n-mermaid-mrxl4f8r-87","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:14:04.092Z","updatedAt":"2026-07-23T18:39:11.083Z"},{"id":"e1784823249858","title":"","sourceId":"n-mermaid-mrxl4f8r-33","targetId":"n-mermaid-mrxl4f8r-36","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:14:09.858Z","updatedAt":"2026-07-23T18:39:11.083Z"},{"id":"e1784823267630","title":"","sourceId":"n-mermaid-mrxl4f8r-33","targetId":"n-1784-823-267-623","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:14:27.630Z","updatedAt":"2026-07-23T18:39:11.083Z"},{"id":"e1784823336507","title":"","sourceId":"n-mermaid-mrxl4f8r-36","targetId":"n-mermaid-mrxl4f8r-30","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:15:36.507Z","updatedAt":"2026-07-23T17:44:47.976Z"},{"id":"e1784823344727","title":"","sourceId":"n-mermaid-mrxl4f8r-36","targetId":"n-1784-823-344-722","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:15:44.727Z","updatedAt":"2026-07-23T16:15:55.840Z"},{"id":"e1784823631895","title":"","sourceId":"n-mermaid-mrxl4f8r-35","targetId":"n-mermaid-mrxl4f8r-168","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:20:31.895Z","updatedAt":"2026-07-23T17:00:36.041Z"},{"id":"e1784823977113","title":"","sourceId":"n-mermaid-mrxl4f8r-38","targetId":"n-mermaid-mrxl4f8r-158","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:26:17.113Z","updatedAt":"2026-07-23T16:36:03.771Z"},{"id":"e1784823983422","title":"","sourceId":"n-mermaid-mrxl4f8r-38","targetId":"n-mermaid-mrxl4f8r-170","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:26:23.422Z","updatedAt":"2026-07-23T17:05:32.585Z"},{"id":"e1784824073345","title":"","sourceId":"n-mermaid-mrxl4f8r-160","targetId":"n-1784-824-073-339","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:27:53.345Z","updatedAt":"2026-07-23T16:28:08.698Z"},{"id":"e1784824126361","title":"","sourceId":"n-mermaid-mrxl4f8r-158","targetId":"n-mermaid-mrxl4f8r-16","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:28:46.361Z","updatedAt":"2026-07-23T16:41:41.312Z"},{"id":"e1784824132458","title":"","sourceId":"n-mermaid-mrxl4f8r-158","targetId":"n-1784-824-132-453","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:28:52.458Z","updatedAt":"2026-07-23T16:36:03.772Z"},{"id":"e1784824173390","title":"","sourceId":"n-mermaid-mrxl4f8r-158","targetId":"n-1784-824-173-383","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:29:33.391Z","updatedAt":"2026-07-23T16:36:03.768Z"},{"id":"e1784824174812","title":"","sourceId":"n-mermaid-mrxl4f8r-158","targetId":"n-1784-824-174-809","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:29:34.812Z","updatedAt":"2026-07-23T16:36:03.768Z"},{"id":"e1784824181997","title":"","sourceId":"n-mermaid-mrxl4f8r-158","targetId":"n-1784-824-181-993","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:29:41.997Z","updatedAt":"2026-07-23T16:36:03.773Z"},{"id":"e1784824369064","title":"","sourceId":"n-mermaid-mrxl4f8r-158","targetId":"n-1784-824-369-059","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:32:49.064Z","updatedAt":"2026-07-23T16:36:03.767Z"},{"id":"e1784824385343","title":"","sourceId":"n-1784-824-173-383","targetId":"n-mermaid-mrxl4f8r-18","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:33:05.343Z","updatedAt":"2026-07-23T16:33:52.052Z"},{"id":"e1784824386573","title":"","sourceId":"n-1784-824-173-383","targetId":"n-1784-824-181-993","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:33:06.573Z","updatedAt":"2026-07-23T16:33:52.053Z"},{"id":"e1784824390380","title":"","sourceId":"n-1784-824-173-383","targetId":"n-1784-824-174-809","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:33:10.380Z","updatedAt":"2026-07-23T16:35:34.301Z"},{"id":"e1784824417017","title":"","sourceId":"n-mermaid-mrxl4f8r-18","targetId":"n-mermaid-mrxl4f8r-16","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:33:37.017Z","updatedAt":"2026-07-23T16:33:52.056Z"},{"id":"e1784824494141","title":"","sourceId":"n-1784-824-174-809","targetId":"n-1784-824-369-059","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:34:54.141Z","updatedAt":"2026-07-23T16:35:54.170Z"},{"id":"e1784824495258","title":"","sourceId":"n-1784-824-174-809","targetId":"n-1784-824-181-993","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:34:55.258Z","updatedAt":"2026-07-23T16:35:34.301Z"},{"id":"e1784824503339","title":"","sourceId":"n-1784-824-174-809","targetId":"n-1784-824-503-335","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:35:03.339Z","updatedAt":"2026-07-23T16:35:34.300Z"},{"id":"e1784824668571","title":"","sourceId":"n-mermaid-mrxl4f8r-14","targetId":"n-1784-824-668-566","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:37:48.571Z","updatedAt":"2026-07-23T16:41:41.312Z"},{"id":"e1784824758801","title":"","sourceId":"n-mermaid-mrxl4f8r-14","targetId":"n-1784-824-758-795","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:39:18.801Z","updatedAt":"2026-07-23T16:41:41.313Z"},{"id":"e1784824771892","title":"","sourceId":"n-mermaid-mrxl4f8r-14","targetId":"n-1784-824-181-993","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:39:31.892Z","updatedAt":"2026-07-23T16:39:39.168Z"},{"id":"e1784824876205","title":"","sourceId":"n-mermaid-mrxl4f8r-14","targetId":"n-1784-824-876-198","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:41:16.205Z","updatedAt":"2026-07-23T16:41:41.313Z"},{"id":"e1784825066829","title":"","sourceId":"n-mermaid-mrxl4f8r-2","targetId":"n-mermaid-mrxl4f8r-171","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:44:26.829Z","updatedAt":"2026-07-23T18:39:11.080Z"},{"id":"e1784825078224","title":"","sourceId":"n-mermaid-mrxl4f8r-2","targetId":"n-mermaid-mrxl4f8r-167","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:44:38.224Z","updatedAt":"2026-07-23T18:39:11.080Z"},{"id":"e1784825248613","title":"","sourceId":"n-1784-823-072-505","targetId":"n-mermaid-mrxl4f8r-157","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:47:28.613Z","updatedAt":"2026-07-23T16:50:43.487Z"},{"id":"e1784825250429","title":"","sourceId":"n-1784-823-072-505","targetId":"n-mermaid-mrxl4f8r-86","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:47:30.429Z","updatedAt":"2026-07-23T17:32:45.899Z"},{"id":"e1784825255085","title":"","sourceId":"n-1784-823-072-505","targetId":"n-mermaid-mrxl4f8r-158","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:47:35.085Z","updatedAt":"2026-07-23T16:50:43.489Z"},{"id":"e1784825281012","title":"","sourceId":"n-1784-823-072-505","targetId":"n-1784-825-281-004","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:48:01.012Z","updatedAt":"2026-07-23T16:50:43.489Z"},{"id":"e1784825283116","title":"","sourceId":"n-1784-823-072-505","targetId":"n-1784-825-283-108","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:48:03.116Z","updatedAt":"2026-07-23T16:50:43.490Z"},{"id":"e1784825289603","title":"","sourceId":"n-1784-823-072-505","targetId":"n-mermaid-mrxl4f8r-87","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:48:09.603Z","updatedAt":"2026-07-23T17:32:45.901Z"},{"id":"e1784825422658","title":"","sourceId":"n-1784-823-072-505","targetId":"n-1784-825-371-425","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:50:22.658Z","updatedAt":"2026-07-23T16:51:45.373Z"},{"id":"e1784825469864","title":"","sourceId":"n-1784-825-371-425","targetId":"n-1784-823-083-572","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:51:09.864Z","updatedAt":"2026-07-23T16:51:45.374Z"},{"id":"e1784825477833","title":"","sourceId":"n-1784-825-371-425","targetId":"n-1784-825-477-825","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:51:17.833Z","updatedAt":"2026-07-23T16:51:45.375Z"},{"id":"e1784825483287","title":"","sourceId":"n-1784-825-371-425","targetId":"n-mermaid-mrxl4f8r-170","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:51:23.287Z","updatedAt":"2026-07-23T17:05:32.586Z"},{"id":"e1784825486063","title":"","sourceId":"n-1784-825-371-425","targetId":"n-mermaid-mrxl4f8r-169","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:51:26.063Z","updatedAt":"2026-07-23T17:00:36.042Z"},{"id":"e1784825570106","title":"","sourceId":"n-mermaid-mrxl4f8r-168","targetId":"n-1784-825-570-102","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:52:50.106Z","updatedAt":"2026-07-23T17:00:36.041Z"},{"id":"e1784825577190","title":"","sourceId":"n-mermaid-mrxl4f8r-168","targetId":"n-1784-825-371-425","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T16:52:57.190Z","updatedAt":"2026-07-23T17:05:32.586Z"},{"id":"e1784826096404","title":"","sourceId":"n-mermaid-mrxl4f8r-170","targetId":"n-1784-826-096-398","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:01:36.404Z","updatedAt":"2026-07-23T17:05:32.587Z"},{"id":"e1784826102471","title":"","sourceId":"n-mermaid-mrxl4f8r-170","targetId":"n-1784-826-102-465","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:01:42.471Z","updatedAt":"2026-07-23T17:05:32.588Z"},{"id":"e1784826107749","title":"","sourceId":"n-mermaid-mrxl4f8r-170","targetId":"n-1784-826-107-744","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:01:47.749Z","updatedAt":"2026-07-23T17:05:32.588Z"},{"id":"e1784826169542","title":"","sourceId":"n-mermaid-mrxl4f8r-170","targetId":"n-1784-826-169-536","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:02:49.542Z","updatedAt":"2026-07-23T17:05:32.589Z"},{"id":"e1784826269857","title":"","sourceId":"n-1784-826-169-536","targetId":"n-mermaid-mrxl4f8r-44","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:04:29.857Z","updatedAt":"2026-07-23T17:05:32.589Z"},{"id":"e1784826307881","title":"","sourceId":"n-mermaid-mrxl4f8r-170","targetId":"n-mermaid-mrxl4f8r-44","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:05:07.881Z","updatedAt":"2026-07-23T17:05:32.590Z"},{"id":"e1784826493495","title":"","sourceId":"n-mermaid-mrxl4f8r-152","targetId":"n-1784-826-493-490","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:08:13.495Z","updatedAt":"2026-07-23T18:36:40.904Z"},{"id":"e1784826496815","title":"","sourceId":"n-mermaid-mrxl4f8r-152","targetId":"n-mermaid-mrxl4f8r-151","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:08:16.815Z","updatedAt":"2026-07-23T18:36:40.904Z"},{"id":"e1784826499636","title":"","sourceId":"n-mermaid-mrxl4f8r-152","targetId":"n-mermaid-mrxl4f8r-106","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:08:19.636Z","updatedAt":"2026-07-23T18:36:40.905Z"},{"id":"e1784826506063","title":"","sourceId":"n-mermaid-mrxl4f8r-152","targetId":"n-mermaid-mrxl4f8r-74","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:08:26.063Z","updatedAt":"2026-07-23T18:36:40.905Z"},{"id":"e1784826595366","title":"","sourceId":"n-mermaid-mrxl4f8r-106","targetId":"n-mermaid-mrxl4f8r-136","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:09:55.366Z","updatedAt":"2026-07-23T17:10:26.658Z"},{"id":"e1784826989787","title":"","sourceId":"n-mermaid-mrxl4f8r-127","targetId":"n-mermaid-mrxl4f8r-147","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:16:29.787Z","updatedAt":"2026-07-23T17:20:48.844Z"},{"id":"e1784827044169","title":"","sourceId":"n-mermaid-mrxl4f8r-127","targetId":"n-1784-827-044-159","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:17:24.169Z","updatedAt":"2026-07-23T17:18:10.146Z"},{"id":"e1784827051241","title":"","sourceId":"n-mermaid-mrxl4f8r-127","targetId":"n-1784-827-051-233","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:17:31.241Z","updatedAt":"2026-07-23T17:18:10.147Z"},{"id":"e1784827060228","title":"","sourceId":"n-mermaid-mrxl4f8r-127","targetId":"n-1784-827-060-218","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:17:40.228Z","updatedAt":"2026-07-23T17:18:10.147Z"},{"id":"e1784827071789","title":"","sourceId":"n-mermaid-mrxl4f8r-127","targetId":"n-1784-827-071-779","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:17:51.789Z","updatedAt":"2026-07-23T17:18:10.148Z"},{"id":"e1784827122706","title":"","sourceId":"n-mermaid-mrxl4f8r-126","targetId":"n-1784-827-122-695","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:18:42.706Z","updatedAt":"2026-07-23T17:19:12.357Z"},{"id":"e1784827129432","title":"","sourceId":"n-mermaid-mrxl4f8r-126","targetId":"n-mermaid-mrxl4f8r-150","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:18:49.432Z","updatedAt":"2026-07-23T17:22:22.877Z"},{"id":"e1784827134135","title":"","sourceId":"n-mermaid-mrxl4f8r-126","targetId":"n-mermaid-mrxl4f8r-147","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:18:54.135Z","updatedAt":"2026-07-23T17:22:22.874Z"},{"id":"e1784827139676","title":"","sourceId":"n-mermaid-mrxl4f8r-126","targetId":"n-mermaid-mrxl4f8r-151","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:18:59.676Z","updatedAt":"2026-07-23T17:22:22.878Z"},{"id":"e1784827369400","title":"","sourceId":"n-mermaid-mrxl4f8r-174","targetId":"n-1784-827-369-389","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:22:49.401Z","updatedAt":"2026-07-23T17:23:16.300Z"},{"id":"e1784827376720","title":"","sourceId":"n-1784-827-369-389","targetId":"n-mermaid-mrxl4f8r-148","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:22:56.720Z","updatedAt":"2026-07-23T17:23:16.300Z"},{"id":"e1784827452691","title":"","sourceId":"n-mermaid-mrxl4f8r-159","targetId":"n-1784-827-452-679","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:24:12.691Z","updatedAt":"2026-07-23T17:29:28.578Z"},{"id":"e1784827454854","title":"","sourceId":"n-mermaid-mrxl4f8r-159","targetId":"n-1784-827-454-845","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:24:14.854Z","updatedAt":"2026-07-23T17:29:28.578Z"},{"id":"e1784827458845","title":"","sourceId":"n-mermaid-mrxl4f8r-159","targetId":"n-1784-827-458-834","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:24:18.845Z","updatedAt":"2026-07-23T17:29:28.579Z"},{"id":"e1784827474305","title":"","sourceId":"n-mermaid-mrxl4f8r-159","targetId":"n-1784-827-474-296","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:24:34.305Z","updatedAt":"2026-07-23T17:29:28.579Z"},{"id":"e1784827480703","title":"","sourceId":"n-mermaid-mrxl4f8r-159","targetId":"n-1784-827-480-696","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:24:40.703Z","updatedAt":"2026-07-23T17:29:28.580Z"},{"id":"e1784827546517","title":"","sourceId":"n-1784-827-474-296","targetId":"n-1784-827-480-696","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:25:46.517Z","updatedAt":"2026-07-23T17:29:28.581Z"},{"id":"e1784827552248","title":"","sourceId":"n-mermaid-mrxl4f8r-159","targetId":"n-1784-827-552-241","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:25:52.248Z","updatedAt":"2026-07-23T17:29:28.581Z"},{"id":"e1784827564121","title":"","sourceId":"n-mermaid-mrxl4f8r-159","targetId":"n-1784-827-564-110","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:26:04.121Z","updatedAt":"2026-07-23T17:29:28.582Z"},{"id":"e1784827570892","title":"","sourceId":"n-1784-827-552-241","targetId":"n-1784-827-564-110","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:26:10.892Z","updatedAt":"2026-07-23T17:29:28.582Z"},{"id":"e1784827589847","title":"","sourceId":"n-1784-827-564-110","targetId":"n-1784-827-452-679","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:26:29.847Z","updatedAt":"2026-07-23T17:29:28.580Z"},{"id":"e1784827596685","title":"","sourceId":"n-1784-827-454-845","targetId":"n-1784-827-452-679","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:26:36.685Z","updatedAt":"2026-07-23T17:29:28.583Z"},{"id":"e1784827604739","title":"","sourceId":"n-1784-827-480-696","targetId":"n-1784-827-458-834","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:26:44.739Z","updatedAt":"2026-07-23T17:29:28.583Z"},{"id":"e1784827626475","title":"","sourceId":"n-mermaid-mrxl4f8r-157","targetId":"n-1784-827-452-679","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:27:06.475Z","updatedAt":"2026-07-23T17:29:28.584Z"},{"id":"e1784827857649","title":"","sourceId":"n-mermaid-mrxl4f8r-164","targetId":"n-mermaid-mrxl4f8r-87","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:30:57.649Z","updatedAt":"2026-07-23T17:32:45.901Z"},{"id":"e1784827867431","title":"","sourceId":"n-mermaid-mrxl4f8r-164","targetId":"n-1784-823-094-285","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:31:07.431Z","updatedAt":"2026-07-23T17:31:36.462Z"},{"id":"e1784827872691","title":"","sourceId":"n-mermaid-mrxl4f8r-164","targetId":"n-mermaid-mrxl4f8r-69","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:31:12.691Z","updatedAt":"2026-07-23T17:32:45.902Z"},{"id":"e1784827924586","title":"","sourceId":"n-mermaid-mrxl4f8r-86","targetId":"n-mermaid-mrxl4f8r-60","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:32:04.586Z","updatedAt":"2026-07-23T17:32:45.903Z"},{"id":"e1784827930103","title":"","sourceId":"n-mermaid-mrxl4f8r-86","targetId":"n-1784-827-930-093","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:32:10.103Z","updatedAt":"2026-07-23T17:32:45.903Z"},{"id":"e1784828077101","title":"","sourceId":"n-mermaid-mrxl4f8r-13","targetId":"n-mermaid-mrxl4f8r-59","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:34:37.101Z","updatedAt":"2026-07-23T17:53:51.985Z"},{"id":"e1784828107373","title":"","sourceId":"n-mermaid-mrxl4f8r-13","targetId":"n-1784-828-107-362","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:35:07.373Z","updatedAt":"2026-07-23T17:38:44.340Z"},{"id":"e1784828173156","title":"","sourceId":"n-mermaid-mrxl4f8r-13","targetId":"n-1784-828-173-143","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:36:13.156Z","updatedAt":"2026-07-23T17:38:44.340Z"},{"id":"e1784828178555","title":"","sourceId":"n-mermaid-mrxl4f8r-13","targetId":"n-1784-828-178-544","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:36:18.555Z","updatedAt":"2026-07-23T17:38:44.341Z"},{"id":"e1784828196344","title":"","sourceId":"n-mermaid-mrxl4f8r-13","targetId":"n-1784-828-196-325","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:36:36.344Z","updatedAt":"2026-07-23T17:38:44.341Z"},{"id":"e1784828265464","title":"","sourceId":"n-1784-828-173-143","targetId":"n-1784-828-196-325","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:37:45.464Z","updatedAt":"2026-07-23T17:38:44.342Z"},{"id":"e1784828279797","title":"","sourceId":"n-1784-828-178-544","targetId":"n-1784-828-196-325","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:37:59.797Z","updatedAt":"2026-07-23T17:38:44.343Z"},{"id":"e1784828527507","title":"","sourceId":"n-mermaid-mrxl4f8r-43","targetId":"n-mermaid-mrxl4f8r-38","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:42:07.507Z","updatedAt":"2026-07-23T17:42:23.829Z"},{"id":"e1784828596882","title":"","sourceId":"n-mermaid-mrxl4f8r-30","targetId":"n-1784-816-099-231","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:43:16.882Z","updatedAt":"2026-07-23T17:53:38.433Z"},{"id":"e1784828603512","title":"","sourceId":"n-mermaid-mrxl4f8r-30","targetId":"n-mermaid-mrxl4f8r-170","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:43:23.512Z","updatedAt":"2026-07-23T17:44:47.977Z"},{"id":"e1784828609878","title":"","sourceId":"n-mermaid-mrxl4f8r-30","targetId":"n-1784-828-609-869","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:43:29.878Z","updatedAt":"2026-07-23T17:44:47.978Z"},{"id":"e1784828612415","title":"","sourceId":"n-mermaid-mrxl4f8r-30","targetId":"n-1784-823-344-722","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:43:32.415Z","updatedAt":"2026-07-23T17:44:47.978Z"},{"id":"e1784828743412","title":"","sourceId":"n-mermaid-mrxl4f8r-29","targetId":"n-1784-828-743-395","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:45:43.412Z","updatedAt":"2026-07-23T17:47:26.277Z"},{"id":"e1784828782618","title":"","sourceId":"n-mermaid-mrxl4f8r-29","targetId":"n-1784-828-782-607","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:46:22.618Z","updatedAt":"2026-07-23T17:47:26.278Z"},{"id":"e1784828813293","title":"","sourceId":"n-mermaid-mrxl4f8r-29","targetId":"n-1784-828-813-272","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:46:53.293Z","updatedAt":"2026-07-23T17:47:26.279Z"},{"id":"e1784828873572","title":"","sourceId":"n-mermaid-mrxl4f8r-94","targetId":"n-mermaid-mrxl4f8r-63","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:47:53.572Z","updatedAt":"2026-07-23T18:19:52.419Z"},{"id":"e1784828881871","title":"","sourceId":"n-mermaid-mrxl4f8r-94","targetId":"n-1784-828-881-856","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:48:01.871Z","updatedAt":"2026-07-23T17:49:31.100Z"},{"id":"e1784828956572","title":"","sourceId":"n-mermaid-mrxl4f8r-94","targetId":"n-1784-828-956-557","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:49:16.572Z","updatedAt":"2026-07-23T17:49:31.100Z"},{"id":"e1784828965538","title":"","sourceId":"n-mermaid-mrxl4f8r-94","targetId":"n-1784-828-965-523","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:49:25.538Z","updatedAt":"2026-07-23T17:49:31.101Z"},{"id":"e1784829046922","title":"","sourceId":"n-mermaid-mrxl4f8r-79","targetId":"n-mermaid-mrxl4f8r-47","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:50:46.922Z","updatedAt":"2026-07-23T18:23:53.333Z"},{"id":"e1784829052589","title":"","sourceId":"n-mermaid-mrxl4f8r-79","targetId":"n-1784-829-052-575","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:50:52.589Z","updatedAt":"2026-07-23T17:53:43.574Z"},{"id":"e1784829104256","title":"","sourceId":"n-mermaid-mrxl4f8r-79","targetId":"n-mermaid-mrxl4f8r-58","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:51:44.256Z","updatedAt":"2026-07-23T17:53:51.985Z"},{"id":"e1784829158941","title":"","sourceId":"n-mermaid-mrxl4f8r-79","targetId":"n-1784-829-158-925","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:52:38.941Z","updatedAt":"2026-07-23T17:53:43.576Z"},{"id":"e1784829168765","title":"","sourceId":"n-mermaid-mrxl4f8r-79","targetId":"n-mermaid-mrxl4f8r-78","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:52:48.765Z","updatedAt":"2026-07-23T18:20:54.731Z"},{"id":"e1784829175379","title":"","sourceId":"n-mermaid-mrxl4f8r-79","targetId":"n-1784-816-099-231","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T17:52:55.379Z","updatedAt":"2026-07-23T17:53:43.571Z"},{"id":"e1784829684837","title":"","sourceId":"n-mermaid-mrxl4f8r-3","targetId":"n-mermaid-mrxl4f8r-57","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:01:24.837Z","updatedAt":"2026-07-24T17:21:28.031Z"},{"id":"e1784829702855","title":"","sourceId":"n-mermaid-mrxl4f8r-71","targetId":"n-mermaid-mrxl4f8r-68","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:01:42.855Z","updatedAt":"2026-07-24T17:21:28.033Z"},{"id":"e1784829870990","title":"","sourceId":"n-mermaid-mrxl4f8r-62","targetId":"n-1784-829-870-984","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:04:30.990Z","updatedAt":"2026-07-23T18:04:40.479Z"},{"id":"e1784829904056","title":"","sourceId":"n-mermaid-mrxl4f8r-64","targetId":"n-1784-829-904-051","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:05:04.056Z","updatedAt":"2026-07-23T18:07:19.716Z"},{"id":"e1784829908964","title":"","sourceId":"n-mermaid-mrxl4f8r-64","targetId":"n-mermaid-mrxl4f8r-87","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:05:08.964Z","updatedAt":"2026-07-23T18:07:19.716Z"},{"id":"e1784829913822","title":"","sourceId":"n-mermaid-mrxl4f8r-64","targetId":"n-mermaid-mrxl4f8r-63","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:05:13.822Z","updatedAt":"2026-07-23T18:19:52.419Z"},{"id":"e1784829951052","title":"","sourceId":"n-mermaid-mrxl4f8r-65","targetId":"n-1784-829-951-047","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:05:51.052Z","updatedAt":"2026-07-23T18:06:28.059Z"},{"id":"e1784829972469","title":"","sourceId":"n-mermaid-mrxl4f8r-65","targetId":"n-mermaid-mrxl4f8r-78","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:06:12.469Z","updatedAt":"2026-07-23T18:20:54.731Z"},{"id":"e1784830009044","title":"","sourceId":"n-mermaid-mrxl4f8r-66","targetId":"n-1784-830-009-038","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:06:49.044Z","updatedAt":"2026-07-23T18:07:12.961Z"},{"id":"e1784830021126","title":"","sourceId":"n-mermaid-mrxl4f8r-66","targetId":"n-mermaid-mrxl4f8r-76","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:07:01.126Z","updatedAt":"2026-07-23T18:07:12.962Z"},{"id":"e1784830080881","title":"","sourceId":"n-mermaid-mrxl4f8r-63","targetId":"n-1784-830-080-875","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:08:00.881Z","updatedAt":"2026-07-23T18:19:52.421Z"},{"id":"e1784830838756","title":"","sourceId":"n-mermaid-mrxl4f8r-78","targetId":"n-1784-830-838-752","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:20:38.756Z","updatedAt":"2026-07-23T18:20:54.732Z"},{"id":"e1784831017973","title":"","sourceId":"n-mermaid-mrxl4f8r-41","targetId":"n-mermaid-mrxl4f8r-31","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:23:37.973Z","updatedAt":"2026-07-23T18:25:45.199Z"},{"id":"e1784831088038","title":"","sourceId":"n-mermaid-mrxl4f8r-31","targetId":"n-1784-831-088-029","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:24:48.038Z","updatedAt":"2026-07-23T18:25:45.200Z"},{"id":"e1784831282539","title":"","sourceId":"n-mermaid-mrxl4f8r-4","targetId":"n-mermaid-mrxl4f8r-92","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:28:02.539Z","updatedAt":"2026-07-23T18:31:45.905Z"},{"id":"e1784831335150","title":"","sourceId":"n-mermaid-mrxl4f8r-92","targetId":"n-1784-831-335-145","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:28:55.150Z","updatedAt":"2026-07-23T18:34:40.267Z"},{"id":"e1784831342799","title":"","sourceId":"n-mermaid-mrxl4f8r-92","targetId":"n-mermaid-mrxl4f8r-103","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:29:02.799Z","updatedAt":"2026-07-23T18:40:45.000Z"},{"id":"e1784831352317","title":"","sourceId":"n-mermaid-mrxl4f8r-92","targetId":"n-1784-827-930-093","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:29:12.317Z","updatedAt":"2026-07-23T18:34:40.266Z"},{"id":"e1784831395206","title":"","sourceId":"n-mermaid-mrxl4f8r-89","targetId":"n-1784-831-395-200","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:29:55.206Z","updatedAt":"2026-07-23T18:34:40.265Z"},{"id":"e1784831398516","title":"","sourceId":"n-mermaid-mrxl4f8r-89","targetId":"n-1784-823-094-285","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:29:58.516Z","updatedAt":"2026-07-23T18:34:40.265Z"},{"id":"e1784831418377","title":"","sourceId":"n-mermaid-mrxl4f8r-89","targetId":"n-1784-827-930-093","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:30:18.377Z","updatedAt":"2026-07-23T18:34:40.266Z"},{"id":"e1784831536171","title":"","sourceId":"n-mermaid-mrxl4f8r-88","targetId":"n-1784-831-536-167","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:32:16.171Z","updatedAt":"2026-07-23T18:34:01.570Z"},{"id":"e1784831599833","title":"","sourceId":"n-mermaid-mrxl4f8r-88","targetId":"n-1784-823-094-285","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:33:19.833Z","updatedAt":"2026-07-23T18:34:01.571Z"},{"id":"e1784832202083","title":"","sourceId":"n-mermaid-mrxl4f8r-105","targetId":"n-1784-832-202-073","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:43:22.083Z","updatedAt":"2026-07-23T18:52:23.377Z"},{"id":"e1784832203236","title":"","sourceId":"n-mermaid-mrxl4f8r-105","targetId":"n-1784-832-203-225","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:43:23.236Z","updatedAt":"2026-07-23T18:52:23.380Z"},{"id":"e1784832210092","title":"","sourceId":"n-mermaid-mrxl4f8r-105","targetId":"n-1784-832-210-084","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:43:30.092Z","updatedAt":"2026-07-23T18:52:23.377Z"},{"id":"e1784832216220","title":"","sourceId":"n-mermaid-mrxl4f8r-105","targetId":"n-1784-832-216-210","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:43:36.220Z","updatedAt":"2026-07-23T18:52:23.380Z"},{"id":"e1784832218663","title":"","sourceId":"n-mermaid-mrxl4f8r-105","targetId":"n-1784-832-218-654","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:43:38.663Z","updatedAt":"2026-07-23T18:52:23.381Z"},{"id":"e1784832224838","title":"","sourceId":"n-mermaid-mrxl4f8r-105","targetId":"n-1784-832-224-828","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:43:44.838Z","updatedAt":"2026-07-23T18:52:23.379Z"},{"id":"e1784832269294","title":"","sourceId":"n-mermaid-mrxl4f8r-105","targetId":"n-1784-832-269-284","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:44:29.294Z","updatedAt":"2026-07-23T18:52:23.382Z"},{"id":"e1784832271132","title":"","sourceId":"n-mermaid-mrxl4f8r-105","targetId":"n-1784-832-271-126","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:44:31.132Z","updatedAt":"2026-07-23T18:52:23.382Z"},{"id":"e1784832277675","title":"","sourceId":"n-mermaid-mrxl4f8r-105","targetId":"n-1784-832-277-664","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:44:37.675Z","updatedAt":"2026-07-23T18:52:23.383Z"},{"id":"e1784832364299","title":"","sourceId":"n-1784-832-202-073","targetId":"n-mermaid-mrxl4f8r-107","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:46:04.299Z","updatedAt":"2026-07-23T18:52:04.174Z"},{"id":"e1784832439268","title":"","sourceId":"n-1784-832-210-084","targetId":"n-1784-832-439-257","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:47:19.268Z","updatedAt":"2026-07-23T18:50:13.014Z"},{"id":"e1784832441590","title":"","sourceId":"n-1784-832-210-084","targetId":"n-1784-832-441-578","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:47:21.590Z","updatedAt":"2026-07-23T18:50:13.015Z"},{"id":"e1784832448052","title":"","sourceId":"n-1784-832-210-084","targetId":"n-1784-832-448-042","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:47:28.052Z","updatedAt":"2026-07-23T18:50:13.016Z"},{"id":"e1784832454979","title":"","sourceId":"n-1784-832-210-084","targetId":"n-1784-832-454-967","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:47:34.979Z","updatedAt":"2026-07-23T18:50:13.016Z"},{"id":"e1784832565271","title":"","sourceId":"n-1784-832-210-084","targetId":"n-1784-832-565-259","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:49:25.271Z","updatedAt":"2026-07-23T18:50:13.017Z"},{"id":"e1784832578857","title":"","sourceId":"n-1784-832-570-962","targetId":"n-1784-832-565-259","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:49:38.857Z","updatedAt":"2026-07-23T18:50:13.018Z"},{"id":"e1784832596318","title":"","sourceId":"n-1784-832-210-084","targetId":"n-mermaid-mrxl4f8r-108","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:49:56.318Z","updatedAt":"2026-07-23T18:50:13.018Z"},{"id":"e1784832602253","title":"","sourceId":"n-mermaid-mrxl4f8r-108","targetId":"n-1784-832-570-962","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:50:02.254Z","updatedAt":"2026-07-23T18:50:13.019Z"},{"id":"e1784832704806","title":"","sourceId":"n-1784-832-202-073","targetId":"n-1784-832-704-799","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:51:44.806Z","updatedAt":"2026-07-23T18:52:04.175Z"},{"id":"e1784832706158","title":"","sourceId":"n-1784-832-202-073","targetId":"n-1784-832-706-148","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:51:46.158Z","updatedAt":"2026-07-23T18:52:04.176Z"},{"id":"e1784832711618","title":"","sourceId":"n-1784-832-202-073","targetId":"n-1784-832-711-608","typeId":"related","tags":[],"properties":{},"createdAt":"2026-07-23T18:51:51.618Z","updatedAt":"2026-07-23T18:52:04.177Z"}],"edgeTypes":[{"id":"consists-of","name":"Consists_of","thematicStyleSlotId":"edge-style-1","createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"drives","name":"Drives","thematicStyleSlotId":"edge-style-2","createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"enables","name":"Enables","thematicStyleSlotId":"edge-style-3","createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"explains","name":"Explains","thematicStyleSlotId":"edge-style-1","createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"has-perspective","name":"Has_perspective","thematicStyleSlotId":"edge-style-3","createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"leads","name":"Leads","thematicStyleSlotId":"edge-style-2","createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"related","name":"related","thematicStyleSlotId":"edge-style-1","createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"id":"spotlights","name":"Spotlights","thematicStyleSlotId":"edge-style-1","createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"}],"scenes":[{"id":"scene-attention-mechanism-n-mermaid-mrxl4f8r-23","title":"Transformer Architecture","centralNodeId":"n-mermaid-mrxl4f8r-23","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-22":{"position":{"x":246.11571428571432,"y":-155.7728571428571},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-23":{"position":{"x":60,"y":-50.49999999999997},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-815-862-176":{"position":{"x":-168.98470185607013,"y":-23.02977964779541},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-162":{"position":{"x":464.07090696824025,"y":-92.37063650129099},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-030-709":{"position":{"x":-254.64875775244508,"y":-281.6679764910107},"scale":1,"design":{"id":"equation-compact-node","params":{"equationScale":0.8,"effects":{"backgroundOpacity":0.5,"backgroundAltOpacity":0.5}}}},"n-1784-816-091-547":{"position":{"x":-276.78828155345474,"y":139.16895963566827},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-097-653":{"position":{"x":34.5599206412962,"y":86.76238509592812},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-099-231":{"position":{"x":187.1495190748677,"y":172.88881324298734},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-107-389":{"position":{"x":355.4455844801433,"y":171.3401521235779},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-112-838":{"position":{"x":-256.4801298055705,"y":-124.01127644785069},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-22":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-220":{"design":{"id":"default","params":{}}},"e1784815895320":{"design":{"id":"default","params":{}}},"e1784815820916":{"design":{"id":"default","params":{}}},"e1784816030720":{"design":{"id":"default","params":{}}},"e1784816091551":{"design":{"id":"default","params":{}}},"e1784816097657":{"design":{"id":"default","params":{}}},"e1784816099237":{"design":{"id":"default","params":{}}},"e1784816107393":{"design":{"id":"default","params":{}}},"e1784816112844":{"design":{"id":"default","params":{}}},"e1784816121310":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.1308562197092085,"pan":{"x":546.1486268174475,"y":457.108239095315},"focalPoint":{"x":59.999999999999986,"y":-50.499999999999986}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:13:33.989Z","updatedAt":"2026-07-23T16:22:50.027Z"},{"id":"scene-base-model-n-mermaid-mrxl4f8r-53","title":"Foundation Model","centralNodeId":"n-mermaid-mrxl4f8r-53","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-54":{"position":{"x":26.703478970277907,"y":158.84299975360688},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-46":{"position":{"x":-187.2965210297221,"y":209.84299975360688},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-53":{"position":{"x":-35.290118970116595,"y":28.870718807226936},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-55":{"position":{"x":221.68435350472313,"y":212.9584397164772},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-52":{"position":{"x":-30.747775329802778,"y":-108.86069216552191},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-49":{"position":{"x":-274.1839066824831,"y":81.01082250810038},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-53":{"design":{"id":"default","params":{}}},"e1784816507389":{"design":{"id":"default","params":{}}},"e1784816617362":{"design":{"id":"default","params":{}}},"e1784816637547":{"design":{"id":"default","params":{}}},"e1784816639126":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":666.9351784551749,"y":356.6939217891596},"focalPoint":{"x":-35.29011897011659,"y":28.870718807226922}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:22:58.776Z","updatedAt":"2026-07-23T16:22:50.032Z"},{"id":"scene-claude-family-n-1784-832-210-084","title":"Model Families","centralNodeId":"n-1784-832-210-084","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-105":{"position":{"x":-126.17781375418522,"y":160.04673515861938},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-210-084":{"position":{"x":-96.21355785116981,"y":-24.35887963631825},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-439-257":{"position":{"x":-323.7850954018852,"y":-115.53506670433457},"scale":1,"design":{"id":"default-node","params":{"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-1784-832-441-578":{"position":{"x":-176.92488817292343,"y":-207.2188341419537},"scale":1,"design":{"id":"default-node","params":{"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-1784-832-448-042":{"position":{"x":-11.795461270408904,"y":-208.98669284520466},"scale":1,"design":{"id":"default-node","params":{"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-1784-832-454-967":{"position":{"x":137.5482076025312,"y":-132.74815332684125},"scale":1,"design":{"id":"default-node","params":{"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-1784-832-565-259":{"position":{"x":140.4326998884634,"y":68.40224216886557},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-570-962":{"position":{"x":298.8109006348128,"y":-26.388443816866797},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-108":{"position":{"x":156.09980776469624,"y":232.38287518256143},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784832210092":{"design":{"id":"default","params":{}}},"e1784832439268":{"design":{"id":"default","params":{}}},"e1784832441590":{"design":{"id":"default","params":{}}},"e1784832448052":{"design":{"id":"default","params":{}}},"e1784832454979":{"design":{"id":"default","params":{}}},"e1784832565271":{"design":{"id":"default","params":{}}},"e1784832578857":{"design":{"id":"default","params":{}}},"e1784832596318":{"design":{"id":"default","params":{}}},"e1784832602253":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[75],"control-point-weights":[0.5]}}},"viewport":{"zoom":1.163189186359038,"pan":{"x":620.9145700736104,"y":384.3339853847868},"focalPoint":{"x":-5.944493083927394,"y":13.468157028049978}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T18:46:12.176Z","updatedAt":"2026-07-23T18:50:12.998Z"},{"id":"scene-david-rumelhart-n-1784-824-173-383","title":"Connectionism","centralNodeId":"n-1784-824-173-383","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-158":{"position":{"x":109.27443694443821,"y":-127.908993464277},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-173-383":{"position":{"x":46.83333333333334,"y":20.666666666666657},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-18":{"position":{"x":-8.911786502415957,"y":148.87166506570705},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-181-993":{"position":{"x":-197.06271208153387,"y":-64.13453853025524},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-174-809":{"position":{"x":287.53569505253665,"y":112.48108270359518},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-16":{"position":{"x":-12.62305127070993,"y":235.62731777760962},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784824173390":{"design":{"id":"default","params":{}}},"e1784824385343":{"design":{"id":"default","params":{}}},"e1784824386573":{"design":{"id":"default","params":{}}},"e1784824390380":{"design":{"id":"default","params":{}}},"e1784824174812":{"design":{"id":"default","params":{}}},"e1784824181997":{"design":{"id":"default","params":{}}},"e1784824417017":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":557.75,"y":354},"focalPoint":{"x":37.5,"y":30.666666666666668}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T16:29:52.209Z","updatedAt":"2026-07-23T16:33:52.042Z"},{"id":"scene-gpt-family-n-1784-832-202-073","title":"Model Families","centralNodeId":"n-1784-832-202-073","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-105":{"position":{"x":6.726456292839586,"y":161.67214354582518},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-202-073":{"position":{"x":-96.21355785116981,"y":-24.35887963631825},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-107":{"position":{"x":-331.08957588388546,"y":111.90602304997952},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-704-799":{"position":{"x":95.32227821342352,"y":-129.57827712308202},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-706-148":{"position":{"x":-267.62321659931746,"y":-127.62096938578549},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-711-608":{"position":{"x":-100.43212579099598,"y":-223.38618895887424},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784832202083":{"design":{"id":"default","params":{}}},"e1784832364299":{"design":{"id":"default","params":{}}},"e1784832704806":{"design":{"id":"default","params":{}}},"e1784832706158":{"design":{"id":"default","params":{}}},"e1784832711618":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.163189186359038,"pan":{"x":725.9145700736104,"y":428.3339853847868},"focalPoint":{"x":-96.2135578511698,"y":-24.358879636318267}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T18:45:52.012Z","updatedAt":"2026-07-23T18:52:04.165Z"},{"id":"scene-grouped-query-attention-gqa-n-1784-816-107-389","title":"Transformer Architecture","centralNodeId":"n-1784-816-107-389","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-23":{"position":{"x":-658.2345428255613,"y":-261.45262590013544},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-030-709":{"position":{"x":-586.8398703091519,"y":82.45410903231522},"scale":1,"design":{"id":"equation-compact-node","params":{"effects":{"backgroundOpacity":0.5,"backgroundAltOpacity":0.5},"colorOverrides":{"text":"#f287e9"}}}},"n-1784-816-107-389":{"position":{"x":-381.0624682232512,"y":-186.45671777549944},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784816030720":{"design":{"id":"default","params":{}}},"e1784816311221":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1,"pan":{"x":1034.5812065673567,"y":539.3172584339102},"focalPoint":{"x":-420.58120656735673,"y":-139.31725843391018}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:18:38.096Z","updatedAt":"2026-07-23T16:22:50.033Z"},{"id":"scene-james-mcclelland-n-1784-824-174-809","title":"Connectionism","centralNodeId":"n-1784-824-174-809","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-158":{"position":{"x":-140.76125810809845,"y":-209.7234095012055},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-173-383":{"position":{"x":-203.2023617192033,"y":-61.14774937026186},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-181-993":{"position":{"x":-126.45546964450992,"y":134.57568511647048},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-174-809":{"position":{"x":37.5,"y":30.666666666666657},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-369-059":{"position":{"x":170.73263464331,"y":-124.37877437762475},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-503-335":{"position":{"x":329.92805110651426,"y":-74.84445156448443},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784824173390":{"design":{"id":"default","params":{}}},"e1784824174812":{"design":{"id":"default","params":{}}},"e1784824503339":{"design":{"id":"default","params":{}}},"e1784824390380":{"design":{"id":"default","params":{}}},"e1784824494141":{"design":{"id":"default","params":{}}},"e1784824495258":{"design":{"id":"default","params":{}}},"e1784824369064":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":572.75,"y":403},"focalPoint":{"x":27.5,"y":-2}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T16:33:52.715Z","updatedAt":"2026-07-23T16:35:34.289Z"},{"id":"scene-labor-automation-n-mermaid-mrxl4f8r-154","title":"Promise & Peril","centralNodeId":"n-mermaid-mrxl4f8r-154","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-154":{"position":{"x":-12.771428571428544,"y":128.50857142857149},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-7":{"position":{"x":186.90285714285716,"y":-2.765714285714182},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-153":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.0479041916167664,"pan":{"x":627.3832335329341,"y":265.33532934131733},"focalPoint":{"x":-12.771428571428569,"y":128.50857142857146}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T18:36:27.900Z","updatedAt":"2026-07-23T18:36:33.866Z"},{"id":"scene-latent-space-n-mermaid-mrxl4f8r-30","title":"Transformer Architecture","centralNodeId":"n-mermaid-mrxl4f8r-30","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-30":{"position":{"x":60,"y":-50.49999999999997},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-29":{"position":{"x":-107.3471428571429,"y":-50.21714285714279},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-36":{"position":{"x":305.12262573303695,"y":-188.72239021570226},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-099-231":{"position":{"x":111.279579936922,"y":73.29985735651246},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-170":{"position":{"x":43.669507983710275,"y":-183.50118431993712},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-828-609-869":{"position":{"x":-142.46998590060312,"y":195.11573331476018},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-344-722":{"position":{"x":376.84445061003464,"y":-19.991349424087705},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-29":{"design":{"id":"default","params":{}}},"e1784823336507":{"design":{"id":"default","params":{}}},"e1784828596882":{"design":{"id":"default","params":{}}},"e1784828603512":{"design":{"id":"default","params":{}}},"e1784828609878":{"design":{"id":"default","params":{}}},"e1784828612415":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.1308562197092085,"pan":{"x":516.1486268174475,"y":379.108239095315},"focalPoint":{"x":86.52857142857141,"y":18.474285714285724}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T17:42:53.451Z","updatedAt":"2026-07-23T17:44:47.962Z"},{"id":"scene-mechanistic-interpretability-n-mermaid-mrxl4f8r-36","title":"Emergent Capabilities","centralNodeId":"n-mermaid-mrxl4f8r-36","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-36":{"position":{"x":-59.62857800193137,"y":-24.82728652440977},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-33":{"position":{"x":-310.035601091214,"y":113.29402021997569},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-35":{"position":{"x":174.6895816494254,"y":-154.07418226121757},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-30":{"position":{"x":-142.94664437143803,"y":-172.09147027744078},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-344-722":{"position":{"x":33.51458005478936,"y":150.34951589880131},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784823249858":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-35":{"design":{"id":"default","params":{}}},"e1784823336507":{"design":{"id":"default","params":{}}},"e1784823344727":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1,"pan":{"x":673.6285780019314,"y":424.82728652440977},"focalPoint":{"x":-59.62857800193137,"y":-24.82728652440977}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T16:14:59.538Z","updatedAt":"2026-07-23T16:22:50.102Z"},{"id":"scene-mermaid-mrxl4f8r","title":"Mermaid import","centralNodeId":"n-mermaid-mrxl4f8r-1","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-1":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-3":{"position":{"x":330.6294285714286,"y":-108.41},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-4":{"position":{"x":250.42514285714284,"y":-3.876000000000003},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-9":{"position":{"x":-162.11200000000002,"y":-184.86685714285713},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-7":{"position":{"x":-276.92742857142855,"y":115.38428571428571},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-6":{"position":{"x":-84,"y":203},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-2":{"position":{"x":97.25114285714288,"y":-169.52342857142855},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-5":{"position":{"x":252.77771428571427,"y":171.6157142857143},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-8":{"position":{"x":-299.2451428571428,"y":-68.65657142857143},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-2":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[-50],"control-point-weights":[0.5]}},"e-mermaid-mrxl4f8r-3":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-8":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[-25],"control-point-weights":[0.5]}},"e-mermaid-mrxl4f8r-6":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[25],"control-point-weights":[0.5]}},"e-mermaid-mrxl4f8r-5":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[-25],"control-point-weights":[0.5]}},"e-mermaid-mrxl4f8r-1":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[25],"control-point-weights":[0.5]}},"e-mermaid-mrxl4f8r-4":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[75],"control-point-weights":[0.5]}},"e-mermaid-mrxl4f8r-7":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[-25],"control-point-weights":[0.5]}}},"viewport":{"zoom":1.4338385907414994,"pan":{"x":700.9668168783286,"y":388},"focalPoint":{"x":6.299999999999977,"y":0}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:33.228Z"},{"id":"scene-mermaid-mrxl4f8r-N010","title":"The Big Ideas","centralNodeId":"n-mermaid-mrxl4f8r-2","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-1":{"position":{"x":-84,"y":203},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-2":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-10":{"position":{"x":-281.38857142857154,"y":-100.86571428571429},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-37":{"position":{"x":207,"y":74},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-22":{"position":{"x":-301.26428571428573,"y":-207.35857142857142},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-31":{"position":{"x":-19,"y":-188},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-32":{"position":{"x":-34.432857142857145,"y":-320.6714285714285},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-33":{"position":{"x":159.52285714285713,"y":-283.1},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-34":{"position":{"x":214.33857142857144,"y":-164.46285714285713},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-35":{"position":{"x":217,"y":-36},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-36":{"position":{"x":513,"y":-78},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-171":{"position":{"x":-233.90157953982458,"y":-406.51383479617374},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-167":{"position":{"x":-384.9015795398246,"y":108.48616520382627},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-78":{"position":{"x":340.17172801337756,"y":-403.6894914708825},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-87":{"position":{"x":491.6767601368425,"y":-343.6566706856724},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-267-623":{"position":{"x":503.7162177109401,"y":-261.8331129762089},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-1":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-9":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-36":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-21":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-30":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-32":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-33":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-34":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-187":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-213":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[175],"control-point-weights":[0.5]}},"e1784825066829":{"design":{"id":"default","params":{}}},"e1784825078224":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-214":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[75],"control-point-weights":[0.5]}},"e-mermaid-mrxl4f8r-31":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-35":{"design":{"id":"default","params":{}}},"e1784823238320":{"design":{"id":"default","params":{}}},"e1784823244092":{"design":{"id":"default","params":{}}},"e1784823249858":{"design":{"id":"default","params":{}}},"e1784823267630":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1,"pan":{"x":560.8942857142858,"y":505.33571428571423},"focalPoint":{"x":53.10571428571416,"y":-105.33571428571423}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:39:11.057Z"},{"id":"scene-mermaid-mrxl4f8r-N020","title":"Machine Learning","centralNodeId":"n-mermaid-mrxl4f8r-10","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-13":{"position":{"x":-291.9142857142857,"y":106.84},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-10":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-2":{"position":{"x":372.72248163265306,"y":140.69959183673467},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-11":{"position":{"x":47.62285714285715,"y":218.83428571428573},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-12":{"position":{"x":-110,"y":191},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-14":{"position":{"x":-240.98285714285714,"y":-0.8228571428571432},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-16":{"position":{"x":-428.34285714285716,"y":-2.8114285714285705},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-17":{"position":{"x":-347.04571428571427,"y":-156.94857142857143},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-18":{"position":{"x":-236.08,"y":-251.60571428571433},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-19":{"position":{"x":-38.16000000000001,"y":-226.15999999999997},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-20":{"position":{"x":-55.79348571428572,"y":-332.67226122448983},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-21":{"position":{"x":239.03428571428572,"y":-331.8971428571428},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-43":{"position":{"x":307.05142857142863,"y":-194.12571428571428},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-12":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-9":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-10":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-11":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-13":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-14":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-15":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-16":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-17":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-178":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-259":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-19":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-20":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.0706731693673852,"pan":{"x":635.2724491758001,"y":446.69533448703834},"focalPoint":{"x":-19.868293877551004,"y":-43.613061224489826}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:25.811Z"},{"id":"scene-mermaid-mrxl4f8r-N021","title":"Supervised Learning","description":"Imported from a Mermaid flowchart. Central: N021. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-11","nodes":{"n-mermaid-mrxl4f8r-10":{"position":{"x":6,"y":-220},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-11":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-57":{"position":{"x":-6,"y":220},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-10":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-174":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.1666666666666667,"pan":{"x":614,"y":400},"focalPoint":{"x":0,"y":0}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.037Z"},{"id":"scene-mermaid-mrxl4f8r-N022","title":"Unsupervised Learning","description":"Imported from a Mermaid flowchart. Central: N022. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-12","nodes":{"n-mermaid-mrxl4f8r-10":{"position":{"x":110,"y":-191},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-12":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-11":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":531.5,"y":543.25},"focalPoint":{"x":55,"y":-95.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.038Z"},{"id":"scene-mermaid-mrxl4f8r-N023","title":"Reinforcement Learning","centralNodeId":"n-mermaid-mrxl4f8r-13","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-13":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-10":{"position":{"x":345.38333333333327,"y":-235.74999999999997},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-161":{"position":{"x":-303.64626444996026,"y":-191.13744590212602},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-828-173-143":{"position":{"x":-170.26845970027676,"y":103.94265415510223},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-58":{"position":{"x":-66.8466264647782,"y":-205.84255804142728},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-60":{"position":{"x":75.27345279064387,"y":-200.40394817435845},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-828-107-362":{"position":{"x":-365.44060248610396,"y":-60.67132418274812},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-828-178-544":{"position":{"x":-344.85988054268955,"y":59.496051825641544},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-59":{"position":{"x":186.08165676106273,"y":-0.5928206142537968},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-828-196-325":{"position":{"x":166.90240259720656,"y":202.0177613480042},"scale":1,"design":{"id":"equation-compact-node","params":{"equationScale":0.8,"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-12":{"design":{"id":"default","params":{}}},"e1784828077101":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-176":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-177":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-219":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[-75],"control-point-weights":[0.5]}},"e1784828173156":{"design":{"id":"default","params":{}}},"e1784828107373":{"design":{"id":"default","params":{}}},"e1784828178555":{"design":{"id":"default","params":{}}},"e1784828196344":{"design":{"id":"default","params":{}}},"e1784828265464":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[25],"control-point-weights":[0.5]}},"e1784828279797":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[100],"control-point-weights":[0.5]}}},"viewport":{"zoom":1.3043478260869565,"pan":{"x":612.0801724673653,"y":414.3530830933595},"focalPoint":{"x":1.4718677750199465,"y":-11.004030371575606}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:38:44.322Z"},{"id":"scene-mermaid-mrxl4f8r-N024","title":"Neural Networks","centralNodeId":"n-mermaid-mrxl4f8r-14","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-14":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-16":{"position":{"x":-220,"y":12},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-10":{"position":{"x":220,"y":-12},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-158":{"position":{"x":-217.64972476299852,"y":-123.4883581272032},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-44":{"position":{"x":221.77218499283387,"y":94.57599280410851},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-15":{"position":{"x":190.0950993303204,"y":-130.63819797731628},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-668-566":{"position":{"x":-142.05926140687154,"y":94.42554793975015},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-758-795":{"position":{"x":-46.73856742232412,"y":-173.30774410575063},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-876-198":{"position":{"x":54.44971805211434,"y":180.0720032262872},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-19":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-13":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-18":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-217":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-260":{"design":{"id":"default","params":{}}},"e1784824126361":{"design":{"id":"default","params":{}}},"e1784824668571":{"design":{"id":"default","params":{}}},"e1784824758801":{"design":{"id":"default","params":{}}},"e1784824876205":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":614,"y":400},"focalPoint":{"x":0,"y":0}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:41:41.295Z"},{"id":"scene-mermaid-mrxl4f8r-N025","title":"Deep Learning","centralNodeId":"n-mermaid-mrxl4f8r-15","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-39":{"position":{"x":-164.92114008918415,"y":-24.965126710748024},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-15":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-160":{"position":{"x":218,"y":33},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-14":{"position":{"x":3.2708973541893527,"y":165.9443626123544},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-40":{"position":{"x":56.89384063059249,"y":-156.797101051967},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-8":{"position":{"x":233.88988975207926,"y":-158.3413478672782},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-38":{"position":{"x":366.7339225699205,"y":-88.41754517764265},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-41":{"position":{"x":408.6491940867687,"y":55.73510048505618},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-42":{"position":{"x":334.0107247904342,"y":185.98860001189018},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-251":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-218":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-18":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-253":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-159":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-249":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-255":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-258":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":450.5,"y":375.25},"focalPoint":{"x":109,"y":16.5}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:11.003Z"},{"id":"scene-mermaid-mrxl4f8r-N026","title":"Backpropagation","description":"Imported from a Mermaid flowchart. Central: N026. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-16","nodes":{"n-mermaid-mrxl4f8r-14":{"position":{"x":220,"y":-12},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-16":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-19":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":449,"y":409},"focalPoint":{"x":110,"y":-6}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.041Z"},{"id":"scene-mermaid-mrxl4f8r-N027","title":"Loss Function","description":"Imported from a Mermaid flowchart. Central: N027. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-17","nodes":{"n-mermaid-mrxl4f8r-10":{"position":{"x":199,"y":94},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-17":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-14":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":464.75,"y":329.5},"focalPoint":{"x":99.5,"y":47}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.041Z"},{"id":"scene-mermaid-mrxl4f8r-N028","title":"Gradient Descent","description":"Imported from a Mermaid flowchart. Central: N028. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-18","nodes":{"n-mermaid-mrxl4f8r-10":{"position":{"x":130,"y":177},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-18":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-15":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":516.5,"y":267.25},"focalPoint":{"x":65,"y":88.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.042Z"},{"id":"scene-mermaid-mrxl4f8r-N029","title":"Generalization","description":"Imported from a Mermaid flowchart. Central: N029. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-19","nodes":{"n-mermaid-mrxl4f8r-10":{"position":{"x":30,"y":218},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-19":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-20":{"position":{"x":-30,"y":-218},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-16":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-20":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.174496644295302,"pan":{"x":614,"y":400},"focalPoint":{"x":0,"y":0}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.042Z"},{"id":"scene-mermaid-mrxl4f8r-N031","title":"Transfer Learning","description":"Imported from a Mermaid flowchart. Central: N031. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-21","nodes":{"n-mermaid-mrxl4f8r-10":{"position":{"x":-77,"y":206},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-21":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-17":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":671.75,"y":245.5},"focalPoint":{"x":-38.5,"y":103}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.043Z"},{"id":"scene-mermaid-mrxl4f8r-N041","title":"Training Data","description":"Imported from a Mermaid flowchart. Central: N041. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-49","nodes":{"n-mermaid-mrxl4f8r-46":{"position":{"x":196,"y":100},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-49":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-48":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":467,"y":325},"focalPoint":{"x":98,"y":50}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.043Z"},{"id":"scene-mermaid-mrxl4f8r-N042","title":"Data Curation","description":"Imported from a Mermaid flowchart. Central: N042. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-50","nodes":{"n-mermaid-mrxl4f8r-46":{"position":{"x":126,"y":180},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-50":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-49":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":519.5,"y":265},"focalPoint":{"x":63,"y":90}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.043Z"},{"id":"scene-mermaid-mrxl4f8r-N043","title":"Synthetic Data","description":"Imported from a Mermaid flowchart. Central: N043. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-51","nodes":{"n-mermaid-mrxl4f8r-46":{"position":{"x":27,"y":218},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-51":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-50":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":593.75,"y":236.5},"focalPoint":{"x":13.5,"y":109}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.044Z"},{"id":"scene-mermaid-mrxl4f8r-N044","title":"Data Wall","description":"Imported from a Mermaid flowchart. Central: N044. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-173","nodes":{"n-mermaid-mrxl4f8r-9":{"position":{"x":-75,"y":207},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-173":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-172":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":670.25,"y":244.75},"focalPoint":{"x":-37.5,"y":103.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.044Z"},{"id":"scene-mermaid-mrxl4f8r-N045","title":"Tokenization","description":"Imported from a Mermaid flowchart. Central: N045. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-28","nodes":{"n-mermaid-mrxl4f8r-22":{"position":{"x":-82,"y":204},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-28":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-27":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":675.5,"y":247},"focalPoint":{"x":-41,"y":102}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.044Z"},{"id":"scene-mermaid-mrxl4f8r-N046","title":"Embeddings","centralNodeId":"n-mermaid-mrxl4f8r-29","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-30":{"position":{"x":290.5666666666666,"y":-211.6},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-29":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-22":{"position":{"x":-195.49999999999997,"y":161},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-94":{"position":{"x":195.79232651768257,"y":166.68790899162605},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-828-743-395":{"position":{"x":-293.7938464246089,"y":1.3028292407433675},"scale":1,"design":{"id":"equation-compact-node","params":{"equationScale":0.8,"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-1784-828-782-607":{"position":{"x":-161.5755148885293,"y":-179.10180006133967},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-828-813-272":{"position":{"x":-332.57617197806115,"y":-134.01657801730923},"scale":1,"design":{"id":"equation-compact-node","params":{"equationScale":0.8,"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-29":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-28":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-182":{"design":{"id":"default","params":{}}},"e1784828743412":{"design":{"id":"default","params":{}}},"e1784828782618":{"design":{"id":"default","params":{}}},"e1784828813293":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.3043478260869565,"pan":{"x":614,"y":400},"focalPoint":{"x":0,"y":0}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:47:26.263Z"},{"id":"scene-mermaid-mrxl4f8r-N050","title":"The Model Factory","centralNodeId":"n-mermaid-mrxl4f8r-3","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-1":{"position":{"x":-216.71428571428572,"y":111.42857142857144},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-3":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-46":{"position":{"x":-146.6571428571428,"y":-198.6},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-55":{"position":{"x":33.91428571428571,"y":-248.17142857142858},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-62":{"position":{"x":198.99999999999997,"y":-204.08571428571426},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-70":{"position":{"x":202.0857142857143,"y":-90.39999999999999},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-71":{"position":{"x":597.4285714285714,"y":5.971428571428567},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-74":{"position":{"x":135.85714285714286,"y":119.11428571428573},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-75":{"position":{"x":354,"y":262},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-76":{"position":{"x":89,"y":201},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-113":{"position":{"x":401.4285714285714,"y":-178.97142857142856},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-57":{"position":{"x":385.997834843134,"y":81.46312612823756},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-68":{"position":{"x":601.5869736851848,"y":-284.5567484421117},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-27":{"position":{"x":498.11443730527645,"y":-79.05714285714285},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-112":{"position":{"x":521.0631265308875,"y":-179.48229474419594},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-148":{"position":{"x":352.0519920056819,"y":-51.15546857795624},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-2":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-45":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-54":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-61":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-69":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-70":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-73":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-75":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-185":{"design":{"id":"default","params":{}}},"e1784829684837":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-74":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-210":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-280":{"design":{"id":"default","params":{}}},"e1784829702855":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-65":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-284":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-184":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-281":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-279":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.09375,"pan":{"x":509.98437500000034,"y":397.9375},"focalPoint":{"x":182.87142857142825,"y":-9.085714285714285}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:21:28.009Z"},{"id":"scene-mermaid-mrxl4f8r-N060","title":"Transformer Architecture","centralNodeId":"n-mermaid-mrxl4f8r-22","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-2":{"position":{"x":280.09714285714284,"y":229.40428571428572},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-22":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-23":{"position":{"x":-187,"y":115},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-815-862-176":{"position":{"x":-361.15898757035586,"y":116.82593463791885},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-162":{"position":{"x":-34.95052160318836,"y":210.1936492129947},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-24":{"position":{"x":-220,"y":2.1571428571428575},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-25":{"position":{"x":-198,"y":-96},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-26":{"position":{"x":-128,"y":-179},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-27":{"position":{"x":-30.42142857142857,"y":-302.89142857142855},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-28":{"position":{"x":82,"y":-204},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-29":{"position":{"x":170,"y":-140},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-30":{"position":{"x":337.3471428571429,"y":-140.28285714285718},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-21":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-22":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-220":{"design":{"id":"default","params":{}}},"e1784815820916":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-23":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-24":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-25":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-26":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-27":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-28":{"design":{"id":"default","params":{}}},"e1784815895320":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-29":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.1308562197092085,"pan":{"x":546.1486268174475,"y":457.108239095315},"focalPoint":{"x":59.999999999999986,"y":-50.499999999999986}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:42:53.243Z"},{"id":"scene-mermaid-mrxl4f8r-N065","title":"Mixture of Experts","description":"Imported from a Mermaid flowchart. Central: N065. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-27","nodes":{"n-mermaid-mrxl4f8r-22":{"position":{"x":26,"y":218},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-27":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-26":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":594.5,"y":236.5},"focalPoint":{"x":13,"y":109}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.046Z"},{"id":"scene-mermaid-mrxl4f8r-N090","title":"Pre-training","centralNodeId":"n-mermaid-mrxl4f8r-46","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-54":{"position":{"x":214,"y":-51},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-46":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-53":{"position":{"x":165,"y":-145},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-3":{"position":{"x":154.79,"y":238.3142857142857},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-47":{"position":{"x":-193,"y":106},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-48":{"position":{"x":-220,"y":3},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-49":{"position":{"x":-196,"y":-100},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-50":{"position":{"x":-144.74186563423123,"y":-184.68546640855783},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-51":{"position":{"x":-17.838571428571427,"y":-224.66285714285712},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-52":{"position":{"x":102.9857142857143,"y":-224.32285714285715},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-79":{"position":{"x":-347.43239982406294,"y":193.76885684590462},"scale":0.8,"design":{"id":"default-node","params":{"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-12":{"position":{"x":-385.577146762661,"y":5.414248938122851},"scale":0.75,"design":{"id":"default-node","params":{"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-173":{"position":{"x":-253.62345019015126,"y":-281.23400752403955},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-41":{"position":{"x":-328.56045998552537,"y":-102.54385152051363},"scale":0.8,"design":{"id":"default-node","params":{"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-31":{"position":{"x":366.1588574982427,"y":-205.74604487793252},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-53":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-45":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-46":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-47":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-48":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-49":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-50":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-51":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-52":{"design":{"id":"default","params":{}}},"e1784829046922":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-175":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-234":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-256":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-180":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.0671296225425286,"pan":{"x":588.900706515285,"y":397.828738752724},"focalPoint":{"x":23.520379300233206,"y":2.0346743276630024}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:23:53.313Z"},{"id":"scene-mermaid-mrxl4f8r-N092","title":"Self-Supervised Learning","description":"Imported from a Mermaid flowchart. Central: N092. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-48","nodes":{"n-mermaid-mrxl4f8r-46":{"position":{"x":220,"y":-3},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-48":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-47":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":449,"y":402.25},"focalPoint":{"x":110,"y":-1.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.047Z"},{"id":"scene-mermaid-mrxl4f8r-N093","title":"Scaling Laws","centralNodeId":"n-mermaid-mrxl4f8r-31","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-31":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-2":{"position":{"x":25,"y":219},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-32":{"position":{"x":-19.874285714285712,"y":-172.01428571428573},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-172":{"position":{"x":227.25177836985685,"y":-66.2201043092817},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-41":{"position":{"x":245.5308461275686,"y":39.097852344541444},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-147":{"position":{"x":221.44925791033845,"y":-206.21487674345693},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-33":{"position":{"x":-259.7415504049254,"y":-136.2612085827666},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-831-088-029":{"position":{"x":341.3454426529868,"y":184.6705878742012},"scale":1,"design":{"id":"equation-compact-node","params":{"equationScaling":0.8,"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-30":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-31":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-186":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-187":{"design":{"id":"default","params":{}}},"e1784823161495":{"design":{"id":"default","params":{}}},"e1784831017973":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-32":{"design":{"id":"default","params":{}}},"e1784831088038":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.1705685618729098,"pan":{"x":594,"y":386},"focalPoint":{"x":17.085714285714285,"y":11.959999999999999}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:25:45.184Z"},{"id":"scene-mermaid-mrxl4f8r-N094","title":"Training Compute","description":"Imported from a Mermaid flowchart. Central: N094. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-52","nodes":{"n-mermaid-mrxl4f8r-46":{"position":{"x":-78,"y":206},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-52":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-51":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":672.5,"y":245.5},"focalPoint":{"x":-39,"y":103}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.048Z"},{"id":"scene-mermaid-mrxl4f8r-N096","title":"Foundation Model","centralNodeId":"n-mermaid-mrxl4f8r-54","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-54":{"position":{"x":-6.000000000000053,"y":2.666666666666662},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-77":{"position":{"x":230.73708175917392,"y":143.33786442075348},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-22":{"position":{"x":2.8627268799459324,"y":140.6985593826756},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-46":{"position":{"x":-307.3,"y":67.44999999999999},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-53":{"position":{"x":147.34069570187967,"y":-167.50145642167018},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-106":{"position":{"x":261.8193930353985,"y":-1.7032672763136971},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-80":{"position":{"x":-168.16114823748262,"y":-167.86255698837454},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-55":{"position":{"x":-303.88647105916505,"y":-34.58540674663888},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-183":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-189":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-53":{"design":{"id":"default","params":{}}},"e1784816507389":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-267":{"design":{"id":"default","params":{}}},"e1784816478435":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-185":{"design":{"id":"default","params":{}}},"e1784818168450":{"design":{"id":"default","params":{}}},"e1784816979287":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.3043478260869565,"pan":{"x":665.7612654116966,"y":357.2156609195944},"focalPoint":{"x":-39.68363681563408,"y":32.80132662831094}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:21:08.551Z"},{"id":"scene-mermaid-mrxl4f8r-N110","title":"Post-training","description":"Imported from a Mermaid flowchart. Central: N110. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-55","nodes":{"n-mermaid-mrxl4f8r-3":{"position":{"x":-33,"y":218},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-55":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-56":{"position":{"x":-57,"y":-213},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-60":{"position":{"x":33,"y":-218},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-61":{"position":{"x":117,"y":-186},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-54":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-55":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-57":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-58":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.174496644295302,"pan":{"x":578.765100671141,"y":400},"focalPoint":{"x":30,"y":0}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.049Z"},{"id":"scene-mermaid-mrxl4f8r-N111","title":"Fine-Tuning","description":"Imported from a Mermaid flowchart. Central: N111. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-56","nodes":{"n-mermaid-mrxl4f8r-55":{"position":{"x":57,"y":213},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-56":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-55":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":571.25,"y":240.25},"focalPoint":{"x":28.5,"y":106.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.049Z"},{"id":"scene-mermaid-mrxl4f8r-N112","title":"Supervised Fine-Tuning","description":"Imported from a Mermaid flowchart. Central: N112. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-57","nodes":{"n-mermaid-mrxl4f8r-11":{"position":{"x":6,"y":-220},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-57":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-174":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":609.5,"y":565},"focalPoint":{"x":3,"y":-110}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.050Z"},{"id":"scene-mermaid-mrxl4f8r-N113","title":"RLHF","centralNodeId":"n-mermaid-mrxl4f8r-58","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-58":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-34":{"position":{"x":-103,"y":194},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-59":{"position":{"x":103,"y":-194},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-13":{"position":{"x":156.09326095821902,"y":-43.346209562421265},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-55":{"position":{"x":144.1846650515341,"y":73.85649845461754},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-79":{"position":{"x":56.028678810185326,"y":152.00258928973903},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-60":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-190":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-176":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-56":{"design":{"id":"default","params":{}}},"e1784829104256":{"design":{"id":"default","params":{}}},"e1784828077101":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.2773722627737227,"pan":{"x":614,"y":400},"focalPoint":{"x":0,"y":0}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:53:51.971Z"},{"id":"scene-mermaid-mrxl4f8r-N115","title":"RLVR","description":"Imported from a Mermaid flowchart. Central: N115. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-60","nodes":{"n-mermaid-mrxl4f8r-55":{"position":{"x":-33,"y":218},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-60":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-57":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":638.75,"y":236.5},"focalPoint":{"x":-16.5,"y":109}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.051Z"},{"id":"scene-mermaid-mrxl4f8r-N130","title":"Inference","centralNodeId":"n-mermaid-mrxl4f8r-62","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-68":{"position":{"x":190.53664596273293,"y":-81.44844720496896},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-3":{"position":{"x":-117.3913043478261,"y":150.43478260869566},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-62":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-63":{"position":{"x":-73.14782608695653,"y":-188.98633540372674},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-64":{"position":{"x":40.00000000000001,"y":-186.95652173913047},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-65":{"position":{"x":48.0223602484472,"y":-421.74906832298143},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-66":{"position":{"x":192.00000000000003,"y":-368.79254658385094},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-67":{"position":{"x":160.8819875776398,"y":-167.83105590062115},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-69":{"position":{"x":204.83478260869566,"y":-0.8695652173913044},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-829-870-984":{"position":{"x":199.76445423369773,"y":80.3360290818661},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-61":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-65":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-62":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-63":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-64":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-66":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-67":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-68":{"design":{"id":"default","params":{}}},"e1784829870990":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.0347043701799485,"pan":{"x":575.7609254498715,"y":522.3650385604113},"focalPoint":{"x":36.95652173913042,"y":-118.26086956521742}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:04:40.459Z"},{"id":"scene-mermaid-mrxl4f8r-N131","title":"Context Window","centralNodeId":"n-mermaid-mrxl4f8r-63","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-63":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-62":{"position":{"x":158.31666666666666,"y":215.43333333333334},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-28":{"position":{"x":-117.9683718110925,"y":257.13007068843547},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-94":{"position":{"x":-280.7494112803017,"y":62.80374125694294},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-64":{"position":{"x":244.86567671155336,"y":122.63782516612822},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-78":{"position":{"x":351.5418741967051,"y":-24.355347184298065},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-93":{"position":{"x":-281.88960111855516,"y":-116.94628210294141},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-830-080-875":{"position":{"x":303.9261472629946,"y":-153.45662816818609},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-62":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-181":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-191":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-192":{"design":{"id":"default","params":{}}},"e1784828873572":{"design":{"id":"default","params":{}}},"e1784829913822":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-63":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-93":{"design":{"id":"default","params":{}}},"e1784830080881":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[50],"control-point-weights":[0.5]}}},"viewport":{"zoom":1.2,"pan":{"x":594.88863615311,"y":343.6759344878503},"focalPoint":{"x":15.926136539074966,"y":46.93672126012473}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:19:52.403Z"},{"id":"scene-mermaid-mrxl4f8r-N132","title":"Prompt","centralNodeId":"n-mermaid-mrxl4f8r-64","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-64":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-66":{"position":{"x":148,"y":-163},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-65":{"position":{"x":-68,"y":-209},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-62":{"position":{"x":-46,"y":215},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-829-904-051":{"position":{"x":352.97738664681236,"y":-96.0733203630585},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-87":{"position":{"x":-278.14778622854146,"y":-123.96983900630268},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-63":{"position":{"x":436.80486955472935,"y":82.4986376936382},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-68":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-63":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-67":{"design":{"id":"default","params":{}}},"e1784829904056":{"design":{"id":"default","params":{}}},"e1784829908964":{"design":{"id":"default","params":{}}},"e1784829913822":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.1986301369863013,"pan":{"x":566.054794520548,"y":396.40410958904107},"focalPoint":{"x":40.00000000000001,"y":3.0000000000000218}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:07:19.702Z"},{"id":"scene-mermaid-mrxl4f8r-N136","title":"Quantization","description":"Imported from a Mermaid flowchart. Central: N136. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-68","nodes":{"n-mermaid-mrxl4f8r-62":{"position":{"x":-198,"y":97},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-68":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-65":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":762.5,"y":327.25},"focalPoint":{"x":-99,"y":48.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.053Z"},{"id":"scene-mermaid-mrxl4f8r-N137","title":"Test-Time Compute","description":"Imported from a Mermaid flowchart. Central: N137. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-69","nodes":{"n-mermaid-mrxl4f8r-62":{"position":{"x":-220,"y":1},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-69":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-66":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":779,"y":399.25},"focalPoint":{"x":-110,"y":0.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.054Z"},{"id":"scene-mermaid-mrxl4f8r-N200","title":"The Toolbox","centralNodeId":"n-mermaid-mrxl4f8r-4","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-1":{"position":{"x":-336.35,"y":-115.75},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-4":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-77":{"position":{"x":-3.3,"y":-274.33571428571435},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-80":{"position":{"x":-237.2000000000001,"y":117.59285714285716},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-86":{"position":{"x":265.08571428571463,"y":-49.17142857142858},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-87":{"position":{"x":468.48571428571427,"y":-90.17857142857144},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-88":{"position":{"x":502.9357142857146,"y":-277.2785714285714},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-93":{"position":{"x":306.4142857142857,"y":69.4857142857143},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-94":{"position":{"x":496.7142857142857,"y":67.30714285714286},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-95":{"position":{"x":134.27857142857144,"y":-152.75714285714284},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-96":{"position":{"x":-216.2785714285714,"y":-274.6714285714284},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-97":{"position":{"x":209.84285714285713,"y":206.09285714285716},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-98":{"position":{"x":-246.7357142857143,"y":265.5571428571429},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-92":{"position":{"x":-1.1528634169332792,"y":178.5662611316695},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-3":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-76":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-79":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-85":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-87":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-92":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-94":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-95":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-96":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-97":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-195":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-196":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-197":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-203":{"design":{"id":"default","params":{}}},"e1784831282539":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-86":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-93":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.1023622047244095,"pan":{"x":536.1102362204724,"y":422.59842519685037},"focalPoint":{"x":70.65714285714287,"y":-20.49999999999998}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:45.879Z"},{"id":"scene-mermaid-mrxl4f8r-N201","title":"Large Language Models","centralNodeId":"n-mermaid-mrxl4f8r-77","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-78":{"position":{"x":-385.8171428571428,"y":-125.29714285714286},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-77":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-79":{"position":{"x":-296.2942857142856,"y":-212.31714285714287},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-54":{"position":{"x":357.90677129340605,"y":91.6968925753534},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-22":{"position":{"x":355.55119316154753,"y":5.554269439019331},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-4":{"position":{"x":71.52857142857138,"y":218.1457142857143},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-81":{"position":{"x":-103.99142857142856,"y":-242.06571428571428},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-169":{"position":{"x":123.21142857142856,"y":-221.62857142857143},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-170":{"position":{"x":193.98,"y":-102.89428571428572},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-88":{"position":{"x":-241.1455882914254,"y":60.48516620624753},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-96":{"position":{"x":-93.14751495269229,"y":134.12566905899718},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-95":{"position":{"x":394.4326952499929,"y":219.1730501661604},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-77":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-78":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-183":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-189":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-76":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-198":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-225":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-226":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-196":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-195":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-197":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[75],"control-point-weights":[0.5]}},"e-mermaid-mrxl4f8r-94":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-95":{"design":{"id":"default","params":{}}},"e1784816434920":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.1705685618729098,"pan":{"x":622.7792642140469,"y":400},"focalPoint":{"x":-7.500000000000033,"y":0}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:21:00.312Z"},{"id":"scene-mermaid-mrxl4f8r-N202","title":"In-Context Learning","centralNodeId":"n-mermaid-mrxl4f8r-78","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-78":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-77":{"position":{"x":-195.28916666666666,"y":-161.34500000000006},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-63":{"position":{"x":78.54010919370998,"y":241.4680468464515},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-33":{"position":{"x":-122.79150255024469,"y":222.25573851186212},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-79":{"position":{"x":-246.223502828327,"y":62.043154779143975},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-65":{"position":{"x":342.0000551425764,"y":-46.71828439986593},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-830-838-752":{"position":{"x":448.46075352632573,"y":112.21991747796771},"scale":1,"design":{"id":"default-node","params":{"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-77":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-191":{"design":{"id":"default","params":{}}},"e1784823238320":{"design":{"id":"default","params":{}}},"e1784829168765":{"design":{"id":"default","params":{}}},"e1784829972469":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-78":{"design":{"id":"default","params":{}}},"e1784830838756":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.1342155009451798,"pan":{"x":461.75,"y":335.5},"focalPoint":{"x":134.23375,"y":56.86749999999999}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:20:54.719Z"},{"id":"scene-mermaid-mrxl4f8r-N210","title":"Multimodality","description":"Imported from a Mermaid flowchart. Central: N210. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-80","nodes":{"n-mermaid-mrxl4f8r-4":{"position":{"x":-93,"y":199},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-80":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-82":{"position":{"x":-60,"y":-212},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-83":{"position":{"x":43,"y":-216},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-84":{"position":{"x":138,"y":-172},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-85":{"position":{"x":201,"y":-89},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-79":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-81":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-82":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-83":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-84":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-199":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-200":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.2173913043478262,"pan":{"x":548.2608695652174,"y":410.3478260869565},"focalPoint":{"x":54,"y":-8.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.056Z"},{"id":"scene-mermaid-mrxl4f8r-N211","title":"Vision-Language Models","description":"Imported from a Mermaid flowchart. Central: N211. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-81","nodes":{"n-mermaid-mrxl4f8r-77":{"position":{"x":16,"y":219},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-81":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-198":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":602,"y":235.75},"focalPoint":{"x":8,"y":109.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.056Z"},{"id":"scene-mermaid-mrxl4f8r-N212","title":"Speech Models","description":"Imported from a Mermaid flowchart. Central: N212. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-82","nodes":{"n-mermaid-mrxl4f8r-80":{"position":{"x":60,"y":212},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-82":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-81":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":569,"y":241},"focalPoint":{"x":30,"y":106}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.056Z"},{"id":"scene-mermaid-mrxl4f8r-N213","title":"Image Generation","description":"Imported from a Mermaid flowchart. Central: N213. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-83","nodes":{"n-mermaid-mrxl4f8r-80":{"position":{"x":-43,"y":216},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-83":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-82":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":646.25,"y":238},"focalPoint":{"x":-21.5,"y":108}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.057Z"},{"id":"scene-mermaid-mrxl4f8r-N214","title":"Video Generation","description":"Imported from a Mermaid flowchart. Central: N214. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-84","nodes":{"n-mermaid-mrxl4f8r-80":{"position":{"x":-138,"y":172},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-84":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-83":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":717.5,"y":271},"focalPoint":{"x":-69,"y":86}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.057Z"},{"id":"scene-mermaid-mrxl4f8r-N215","title":"Diffusion Models","description":"Imported from a Mermaid flowchart. Central: N215. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-85","nodes":{"n-mermaid-mrxl4f8r-80":{"position":{"x":-201,"y":89},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-85":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-84":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":764.75,"y":333.25},"focalPoint":{"x":-100.5,"y":44.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.058Z"},{"id":"scene-mermaid-mrxl4f8r-N220","title":"Reasoning Models","centralNodeId":"n-mermaid-mrxl4f8r-86","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-87":{"position":{"x":221.56666666666663,"y":-139.14999999999998},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-86":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-4":{"position":{"x":-204.7,"y":148.35},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-69":{"position":{"x":56.21340575538161,"y":208.19759204514284},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-164":{"position":{"x":260.6583456801463,"y":67.77855287934713},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-072-505":{"position":{"x":-2.709218884802768,"y":-240.7825534194682},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-60":{"position":{"x":-284.35140767262357,"y":-143.31761965444628},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-827-930-093":{"position":{"x":-337.45657354212335,"y":-64.73637932195317},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-13":{"position":{"x":-425.18215356289056,"y":-227.73061014577033},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-55":{"position":{"x":-253.42270893756051,"y":-239.4682216445519},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-85":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-86":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-194":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-222":{"design":{"id":"default","params":{}}},"e1784825250429":{"design":{"id":"default","params":{}}},"e1784825289603":{"design":{"id":"default","params":{}}},"e1784827857649":{"design":{"id":"default","params":{}}},"e1784827872691":{"design":{"id":"default","params":{}}},"e1784827924586":{"design":{"id":"default","params":{}}},"e1784827930103":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-177":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-57":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.3043478260869565,"pan":{"x":704,"y":441},"focalPoint":{"x":-69,"y":-31.433333333333334}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:32:45.887Z"},{"id":"scene-mermaid-mrxl4f8r-N230","title":"AI Agents","centralNodeId":"n-mermaid-mrxl4f8r-88","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-92":{"position":{"x":209.29999999999998,"y":142.6},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-88":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-89":{"position":{"x":187.06666666666666,"y":-208.91666666666666},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-90":{"position":{"x":236.89999999999998,"y":-88.55},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-094-285":{"position":{"x":-216.59669986392336,"y":-153.59068629807555},"scale":0.8,"design":{"id":"default-node","params":{"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-4":{"position":{"x":-274.8499999999997,"y":66.7},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-91":{"position":{"x":250.7,"y":31.049999999999997},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-831-536-167":{"position":{"x":-46.805629959046954,"y":207.5199500547958},"scale":0.85,"design":{"id":"default-node","params":{"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-201":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-91":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-202":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[-325],"control-point-weights":[0.6500000000000001]}},"e-mermaid-mrxl4f8r-88":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-89":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-87":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-90":{"design":{"id":"default","params":{}}},"e1784831536171":{"design":{"id":"default","params":{}}},"e1784831599833":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[100],"control-point-weights":[0.5]}},"e1784831398516":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[25],"control-point-weights":[0.5]}}},"viewport":{"zoom":1.3043478260869565,"pan":{"x":575.75,"y":368.25},"focalPoint":{"x":29.325,"y":24.341666666666665}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:34:01.556Z"},{"id":"scene-mermaid-mrxl4f8r-N231","title":"Tool Use","centralNodeId":"n-mermaid-mrxl4f8r-89","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-92":{"position":{"x":155.56658242789425,"y":214.34067251692025},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-88":{"position":{"x":-229.23333333333332,"y":215.56749999999997},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-89":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-90":{"position":{"x":-304.95707800753297,"y":-27.463456999144718},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-094-285":{"position":{"x":-308.2443200303792,"y":-155.30406930622715},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-831-395-200":{"position":{"x":206.18886939313512,"y":-76.75413281055488},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-827-930-093":{"position":{"x":-78.88067094973513,"y":-188.7171324388867},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-831-335-145":{"position":{"x":319.68039798038683,"y":165.40624932628518},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-103":{"position":{"x":320.60729146786935,"y":238.50271629600218},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-201":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-91":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-202":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-88":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-89":{"design":{"id":"default","params":{}}},"e1784831398516":{"design":{"id":"default","params":{}}},"e1784831395206":{"design":{"id":"default","params":{}}},"e1784831418377":{"design":{"id":"default","params":{}}},"e1784831352317":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[525],"control-point-weights":[0.5]}},"e1784831342799":{"design":{"id":"default","params":{}}},"e1784831335150":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.3043478260869565,"pan":{"x":662.957902589507,"y":382.8139956012825},"focalPoint":{"x":-37.53439198528866,"y":13.175936705683418}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:34:40.256Z"},{"id":"scene-mermaid-mrxl4f8r-N232","title":"Computer Use","description":"Imported from a Mermaid flowchart. Central: N232. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-90","nodes":{"n-mermaid-mrxl4f8r-88":{"position":{"x":-206,"y":77},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-90":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-89":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":768.5,"y":342.25},"focalPoint":{"x":-103,"y":38.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.059Z"},{"id":"scene-mermaid-mrxl4f8r-N234","title":"Model Context Protocol","centralNodeId":"n-mermaid-mrxl4f8r-92","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-92":{"position":{"x":-46,"y":-27.333333333333336},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-88":{"position":{"x":-217.7333333333333,"y":-43.43333333333334},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-89":{"position":{"x":-45.75539640160349,"y":-221.39214594953202},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-108":{"position":{"x":196.52043098748155,"y":-105.28276759488455},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-4":{"position":{"x":-391.82264713379715,"y":109.23110015146693},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-831-335-145":{"position":{"x":12.983950858401663,"y":168.23888867049462},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-103":{"position":{"x":261.9876454784941,"y":-5.110071073147051},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-827-930-093":{"position":{"x":178.4209788118274,"y":-236.6434044064804},"scale":0.9,"design":{"id":"default-node","params":{"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-91":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-202":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-272":{"design":{"id":"default","params":{}}},"e1784831282539":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-87":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-88":{"design":{"id":"default","params":{}}},"e1784831335150":{"design":{"id":"default","params":{}}},"e1784831342799":{"design":{"id":"default","params":{}}},"e1784831352317":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[25],"control-point-weights":[0.5]}}},"viewport":{"zoom":1.3043478260869565,"pan":{"x":696.5,"y":420.6521739130435},"focalPoint":{"x":-63.25,"y":-15.833333333333348}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:29:29.840Z"},{"id":"scene-mermaid-mrxl4f8r-N240","title":"Retrieval-Augmented Generation","centralNodeId":"n-mermaid-mrxl4f8r-93","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-94":{"position":{"x":145.5,"y":114.83333333333334},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-93":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-4":{"position":{"x":-342.1666666666667,"y":-176.49999999999997},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-63":{"position":{"x":321.92780456078077,"y":-144.72544548771378},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-79":{"position":{"x":-32.8649042046996,"y":185.2913182873679},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-97":{"position":{"x":-266.8170823996443,"y":140.99187903839547},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-93":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-92":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-192":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-203":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-211":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-96":{"design":{"id":"default","params":{}}},"e1784828873572":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.2,"pan":{"x":582.1633172635317,"y":380.0404763202076},"focalPoint":{"x":26.530568947056945,"y":16.632936399827024}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:50:17.547Z"},{"id":"scene-mermaid-mrxl4f8r-N241","title":"Vector Database","centralNodeId":"n-mermaid-mrxl4f8r-94","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-94":{"position":{"x":-60,"y":-47.333333333333336},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-93":{"position":{"x":-298.81666666666666,"y":-198.36666666666667},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-29":{"position":{"x":186.2449087951671,"y":80.44272088075877},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-63":{"position":{"x":303.9047214706149,"y":-263.3564426180659},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-828-881-856":{"position":{"x":-296.3980866007028,"y":101.66839770287424},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-828-956-557":{"position":{"x":-196.11046971100131,"y":186.5248733284603},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-828-965-523":{"position":{"x":-395.44380304433463,"y":33.9582066617936},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-93":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-182":{"design":{"id":"default","params":{}}},"e1784828873572":{"design":{"id":"default","params":{}}},"e1784828881871":{"design":{"id":"default","params":{}}},"e1784828956572":{"design":{"id":"default","params":{}}},"e1784828965538":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.3043478260869565,"pan":{"x":691.5108695652174,"y":453.7391304347826},"focalPoint":{"x":-59.42499999999999,"y":-41.20000000000001}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:49:31.086Z"},{"id":"scene-mermaid-mrxl4f8r-N250","title":"AI Coding Assistants","description":"Imported from a Mermaid flowchart. Central: N250. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-95","nodes":{"n-mermaid-mrxl4f8r-4":{"position":{"x":-137,"y":-172},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-95":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-133":{"position":{"x":170,"y":140},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-142":{"position":{"x":98,"y":197},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-94":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-294":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-305":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.3232514177693762,"pan":{"x":592.1663516068053,"y":383.4593572778828},"focalPoint":{"x":16.5,"y":12.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.061Z"},{"id":"scene-mermaid-mrxl4f8r-N251","title":"Chat Assistants","centralNodeId":"n-mermaid-mrxl4f8r-96","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-96":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-107":{"position":{"x":83,"y":204},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-4":{"position":{"x":-35,"y":-217},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-146":{"position":{"x":-105.25142857142858,"y":219.85142857142856},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-77":{"position":{"x":143.39344576443494,"y":-127.68053771739079},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-163":{"position":{"x":185.80150132668058,"y":48.39216987025398},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-88":{"position":{"x":-281.59625169773983,"y":31.24870322200438},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-106":{"position":{"x":365.0757456261902,"y":1.2429963621532218},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-54":{"position":{"x":366.10506832973704,"y":-127.12384839455737},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-80":{"position":{"x":206.9845179433294,"y":-247.89241017869847},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-270":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-95":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-310":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-195":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-221":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-268":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-76":{"design":{"id":"default","params":{}}},"e1784816434920":{"design":{"id":"default","params":{}}},"e1784816445301":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-189":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-106":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[-50],"control-point-weights":[0.5]}},"e-mermaid-mrxl4f8r-267":{"design":{"id":"default","params":{}}},"e1784816478435":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.174496644295302,"pan":{"x":585.8120805369127,"y":398.8255033557047},"focalPoint":{"x":24.000000000000018,"y":0.9999999999999887}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.061Z"},{"id":"scene-mermaid-mrxl4f8r-N252","title":"AI Search","description":"Imported from a Mermaid flowchart. Central: N252. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-97","nodes":{"n-mermaid-mrxl4f8r-4":{"position":{"x":75,"y":-207},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-97":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-134":{"position":{"x":-29,"y":218},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-141":{"position":{"x":-117,"y":186},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-96":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-297":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-304":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.1965811965811965,"pan":{"x":639.1282051282051,"y":393.4188034188034},"focalPoint":{"x":-21,"y":5.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.062Z"},{"id":"scene-mermaid-mrxl4f8r-N300","title":"The Arena","centralNodeId":"n-mermaid-mrxl4f8r-5","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-1":{"position":{"x":-134.71714285714287,"y":-207.98857142857142},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-5":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-104":{"position":{"x":242.59465765004228,"y":40.73874049027896},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-102":{"position":{"x":243.28857142857143,"y":-102.63142857142857},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-103":{"position":{"x":139.15999999999997,"y":-198.70285714285714},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-140":{"position":{"x":-353.0285714285714,"y":-54.39142857142857},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-115":{"position":{"x":-23,"y":219},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-106":{"position":{"x":106.44857142857143,"y":203},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-105":{"position":{"x":220.88571428571427,"y":138},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-126":{"position":{"x":-207.31142857142854,"y":180.16857142857143},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-132":{"position":{"x":-261.68285714285713,"y":87.36},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-139":{"position":{"x":-439.39142857142855,"y":89.08285714285715},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-4":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[25],"control-point-weights":[0.5]}},"e-mermaid-mrxl4f8r-103":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-207":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-101":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-102":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-139":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-114":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-105":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-104":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-125":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-131":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-138":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.3536553270856935,"pan":{"x":834.1998100534367,"y":387.1800716304509},"focalPoint":{"x":-91.75142857142855,"y":0.6057142857143051}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-24T17:22:01.412Z"},{"id":"scene-mermaid-mrxl4f8r-N301","title":"AI Accelerators","description":"Imported from a Mermaid flowchart. Central: N301. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-71","nodes":{"n-mermaid-mrxl4f8r-3":{"position":{"x":-218,"y":-27},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-71":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-72":{"position":{"x":197,"y":-97},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-73":{"position":{"x":219,"y":-15},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-129":{"position":{"x":209,"y":68},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-131":{"position":{"x":168,"y":142},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-70":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-71":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-72":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-290":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-293":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":613.25,"y":366.25},"focalPoint":{"x":0.5,"y":22.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.063Z"},{"id":"scene-mermaid-mrxl4f8r-N302","title":"GPU","description":"Imported from a Mermaid flowchart. Central: N302. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-72","nodes":{"n-mermaid-mrxl4f8r-71":{"position":{"x":-197,"y":97},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-72":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-71":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":761.75,"y":327.25},"focalPoint":{"x":-98.5,"y":48.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.063Z"},{"id":"scene-mermaid-mrxl4f8r-N303","title":"Custom AI Chips","description":"Imported from a Mermaid flowchart. Central: N303. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-73","nodes":{"n-mermaid-mrxl4f8r-71":{"position":{"x":-219,"y":15},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-73":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-72":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":778.25,"y":388.75},"focalPoint":{"x":-109.5,"y":7.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.064Z"},{"id":"scene-mermaid-mrxl4f8r-N304","title":"AI Datacenters","centralNodeId":"n-mermaid-mrxl4f8r-147","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-147":{"position":{"x":0,"y":-8.677428571428571},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-6":{"position":{"x":173.70685714285713,"y":-147.95885714285714},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-31":{"position":{"x":-275.7482857142858,"y":28.38142857142857},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-52":{"position":{"x":304.264,"y":132.44885714285715},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-149":{"position":{"x":173.3871428571428,"y":199.63342857142857},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-174":{"position":{"x":0,"y":240},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-119":{"position":{"x":-130,"y":202},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-111":{"position":{"x":-218,"y":100},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-130":{"position":{"x":-238,"y":-34},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-127":{"position":{"x":397.23228571428587,"y":-209.06457142857144},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-126":{"position":{"x":-274.6805714285714,"y":-245.77714285714285},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-150":{"position":{"x":0.2754285714285736,"y":-146.9122857142857},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-146":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-186":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-188":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-206":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-209":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-236":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-266":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-277":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-291":{"design":{"id":"default","params":{}}},"e1784826989787":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[-75],"control-point-weights":[0.5]}},"e1784827134135":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-126":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[25],"control-point-weights":[0.5]}},"e-mermaid-mrxl4f8r-129":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-148":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-149":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-235":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-241":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-288":{"design":{"id":"default","params":{}}},"e1784827129432":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.267656646142702,"pan":{"x":616.6620789568997,"y":393.6617167692865},"focalPoint":{"x":-2.1000000000000023,"y":4.999999999999993}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:20:48.824Z"},{"id":"scene-mermaid-mrxl4f8r-N310","title":"Frontier Labs","centralNodeId":"n-mermaid-mrxl4f8r-106","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-106":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-5":{"position":{"x":-259.3542857142857,"y":-251.80142857142857},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-109":{"position":{"x":-39.77428571428573,"y":174.5557142857143},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-111":{"position":{"x":45.135714285714286,"y":214.3457142857143},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-152":{"position":{"x":-334.2961325620633,"y":221.3356747306683},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-107":{"position":{"x":-69.68795001120425,"y":-218.47024426708967},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-108":{"position":{"x":65.58782248266263,"y":-218.33395535566564},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-110":{"position":{"x":-276.5645824075341,"y":73.22149704300209},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-112":{"position":{"x":-333.299393761259,"y":-80.01074921623812},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-113":{"position":{"x":272.6746175068071,"y":-18.372138433345494},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-114":{"position":{"x":229.19149805747094,"y":65.19657501852907},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-54":{"position":{"x":366.58427120675583,"y":-214.1083568401528},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-136":{"position":{"x":403.9299503148465,"y":220.62466700672869},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-105":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-108":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-110":{"design":{"id":"default","params":{}}},"e1784826499636":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-106":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-107":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-109":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-111":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-112":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-113":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-267":{"design":{"id":"default","params":{}}},"e1784826595366":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[75],"control-point-weights":[0.5]}}},"viewport":{"zoom":1.2089810017271156,"pan":{"x":589.8203799654577,"y":392.14162348877375},"focalPoint":{"x":19.99999999999996,"y":6.500000000000002}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:10:26.639Z"},{"id":"scene-mermaid-mrxl4f8r-N311","title":"Open-Weight Models","centralNodeId":"n-mermaid-mrxl4f8r-102","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-102":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-5":{"position":{"x":-151,"y":160},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-110":{"position":{"x":11.14,"y":-226.7},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-112":{"position":{"x":113,"y":-188},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-114":{"position":{"x":189.7,"y":-116.3},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-136":{"position":{"x":217,"y":-37},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-101":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-276":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-278":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-285":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-300":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.2987012987012987,"pan":{"x":571.1428571428571,"y":438.31168831168833},"focalPoint":{"x":33.00000000000003,"y":-29.500000000000014}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:25.151Z"},{"id":"scene-mermaid-mrxl4f8r-N312","title":"Model APIs","centralNodeId":"n-mermaid-mrxl4f8r-103","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-103":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-5":{"position":{"x":-209,"y":69},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-135":{"position":{"x":154,"y":-157},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-137":{"position":{"x":209,"y":-69},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-138":{"position":{"x":217,"y":35},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-102":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-298":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-301":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-302":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":608,"y":466},"focalPoint":{"x":4,"y":-44}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:05.466Z"},{"id":"scene-mermaid-mrxl4f8r-N313","title":"Self-Hosting","centralNodeId":"n-mermaid-mrxl4f8r-104","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-104":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-5":{"position":{"x":-292,"y":-54.66666666666668},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-68":{"position":{"x":147.87670361640699,"y":-159.08589041000883},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-102":{"position":{"x":204.2668692271916,"y":73.83011672834752},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-103":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-193":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-207":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-101":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[125],"control-point-weights":[0.5]}}},"viewport":{"zoom":1.5,"pan":{"x":685,"y":437},"focalPoint":{"x":-47.333333333333336,"y":-24.666666666666668}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:42:02.470Z"},{"id":"scene-mermaid-mrxl4f8r-N320","title":"Model Families","centralNodeId":"n-mermaid-mrxl4f8r-105","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-105":{"position":{"x":-58.66666666666667,"y":-49.333333333333336},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-202-073":{"position":{"x":-161.60668081067607,"y":-235.36435651547677},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-210-084":{"position":{"x":-28.70241076365126,"y":-233.73894812827098},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-5":{"position":{"x":-483.1241487051737,"y":-241.09303871206896},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-224-828":{"position":{"x":305.0940867901905,"y":165.16566165143428},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-114":{"position":{"x":-173.80204422758808,"y":223.006922421191},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-203-225":{"position":{"x":171.84310211086245,"y":-188.37480278550748},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-216-210":{"position":{"x":-412.2371077573391,"y":71.23427558264888},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-218-654":{"position":{"x":319.5483503578147,"y":85.13778920369535},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-269-284":{"position":{"x":234.92620115962418,"y":-125.57887264711556},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-271-126":{"position":{"x":-368.3690100114357,"y":155.1536863401201},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-832-277-664":{"position":{"x":-286.26244093483706,"y":213.9598610215002},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784832202083":{"design":{"id":"default","params":{}}},"e1784832210092":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-104":{"design":{"id":"default","params":{}}},"e1784832224838":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-286":{"design":{"id":"default","params":{}}},"e1784832203236":{"design":{"id":"default","params":{}}},"e1784832216220":{"design":{"id":"default","params":{}}},"e1784832218663":{"design":{"id":"default","params":{}}},"e1784832269294":{"design":{"id":"default","params":{}}},"e1784832271132":{"design":{"id":"default","params":{}}},"e1784832277675":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.163189186359038,"pan":{"x":725.9145700736104,"y":428.3339853847868},"focalPoint":{"x":-96.2135578511698,"y":-24.358879636318267}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:52:23.358Z"},{"id":"scene-mermaid-mrxl4f8r-N321","title":"Distillation","description":"Imported from a Mermaid flowchart. Central: N321. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-70","nodes":{"n-mermaid-mrxl4f8r-3":{"position":{"x":-203,"y":84},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-70":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-113":{"position":{"x":203,"y":-84},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-69":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-280":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":614,"y":400},"focalPoint":{"x":0,"y":0}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.066Z"},{"id":"scene-mermaid-mrxl4f8r-N330","title":"Inference Cost","description":"Imported from a Mermaid flowchart. Central: N330. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-148","nodes":{"n-mermaid-mrxl4f8r-6":{"position":{"x":-19,"y":-219},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-148":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-147":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":628.25,"y":564.25},"focalPoint":{"x":-9.5,"y":-109.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.067Z"},{"id":"scene-mermaid-mrxl4f8r-N340","title":"Investment Boom","centralNodeId":"n-mermaid-mrxl4f8r-149","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-149":{"position":{"x":50,"y":-54.00000000000001},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-6":{"position":{"x":163.33333333333334,"y":-225.66666666666669},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-174":{"position":{"x":-191.86604555523994,"y":7.170813272885958},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-147":{"position":{"x":-104.08500652902697,"y":-161.33911929799424},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-126":{"position":{"x":-210.62885639805734,"y":-249.88496513627786},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-1":{"position":{"x":315.79904820358536,"y":-290.5301477175177},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-148":{"position":{"x":370.5454101865744,"y":-169.07872443365636},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-150":{"position":{"x":-48.699864875016644,"y":-270.17314649777563},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-151":{"position":{"x":-48.5268839240311,"y":-352.0653807141099},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-148":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-209":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-235":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-146":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-236":{"design":{"id":"default","params":{}}},"e1784827134135":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-147":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-149":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-150":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-5":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-206":{"design":{"id":"default","params":{}}},"e1784827129432":{"design":{"id":"default","params":{}}},"e1784827139676":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.44,"pan":{"x":498.86008127226785,"y":604.0187208001348},"focalPoint":{"x":79.95827689425845,"y":-141.67966722231586}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:22:22.858Z"},{"id":"scene-mermaid-mrxl4f8r-N341","title":"Energy Demand","description":"Imported from a Mermaid flowchart. Central: N341. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-150","nodes":{"n-mermaid-mrxl4f8r-6":{"position":{"x":168,"y":-142},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-150":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-149":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":488,"y":506.5},"focalPoint":{"x":84,"y":-71}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.067Z"},{"id":"scene-mermaid-mrxl4f8r-N350","title":"The Gold Rush","centralNodeId":"n-mermaid-mrxl4f8r-6","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-1":{"position":{"x":84,"y":-203},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-6":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-147":{"position":{"x":-144.73142857142864,"y":205.12285714285716},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-148":{"position":{"x":206.90285714285716,"y":175.76571428571427},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-149":{"position":{"x":56.51142857142858,"y":227.11142857142858},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-150":{"position":{"x":-256.9628571428572,"y":112.06857142857143},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-151":{"position":{"x":-269.04285714285726,"y":-4.380000000000001},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-5":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-146":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-147":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-148":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-149":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-150":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-206":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-209":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.2027491408934707,"pan":{"x":672.3333333333334,"y":390.3780068728522},"focalPoint":{"x":-48.500000000000036,"y":8.00000000000002}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:35:33.521Z"},{"id":"scene-mermaid-mrxl4f8r-N400","title":"Burning Questions","centralNodeId":"n-mermaid-mrxl4f8r-9","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-165":{"position":{"x":-266.79857142857145,"y":-5.262857142857143},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-9":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-1":{"position":{"x":84,"y":203},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-166":{"position":{"x":-253.7885714285714,"y":-144.36428571428573},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-167":{"position":{"x":-491.59714285714284,"y":-283.02142857142854},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-168":{"position":{"x":247.95428571428573,"y":-1.2199999999999704},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-171":{"position":{"x":-102.7685714285714,"y":-267.0814285714286},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-172":{"position":{"x":-103.71234285714294,"y":-365.31389387755104},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-173":{"position":{"x":75,"y":-207},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-174":{"position":{"x":241.61714285714285,"y":-148.28285714285713},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-164":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-8":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-165":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-167":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-170":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-172":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-173":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-215":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-166":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-171":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-224":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.0925313258990748,"pan":{"x":733.3278321737339,"y":483.3131033480358},"focalPoint":{"x":-109.22142857142853,"y":-76.2569469387755}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:54.273Z"},{"id":"scene-mermaid-mrxl4f8r-N401","title":"AI Safety","centralNodeId":"n-mermaid-mrxl4f8r-165","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-74":{"position":{"x":-134.7488905659689,"y":143.99500814152572},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-152":{"position":{"x":116.95111293706647,"y":-136.03968973348006},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-165":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-9":{"position":{"x":339.16666666666663,"y":-1.3333333333333308},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-38":{"position":{"x":172.1587052099727,"y":131.31956411570036},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-34":{"position":{"x":-3.991167688231065,"y":207.95364972933416},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-75":{"position":{"x":-189.70305935019329,"y":54.08546873279799},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-76":{"position":{"x":-166.6468831971857,"y":-58.092063036816334},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-79":{"position":{"x":-0.7573559428876653,"y":-225.77645721145686},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-35":{"position":{"x":-156.5491304856228,"y":-173.24520786754587},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-233":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-164":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-227":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-229":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-230":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-231":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-232":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-250":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-210":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-213":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[500],"control-point-weights":[0.5]}},"e-mermaid-mrxl4f8r-74":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-228":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.2,"pan":{"x":639.0422454192568,"y":419.6936844892736},"focalPoint":{"x":-20.86853784938067,"y":-16.411403741061342}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:44.723Z"},{"id":"scene-mermaid-mrxl4f8r-N402","title":"Alignment","description":"Imported from a Mermaid flowchart. Central: N402. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-34","nodes":{"n-mermaid-mrxl4f8r-2":{"position":{"x":-172,"y":137},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-34":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-58":{"position":{"x":103,"y":-194},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-59":{"position":{"x":206,"y":-388},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-108":{"position":{"x":172,"y":-137},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-117":{"position":{"x":213,"y":-57},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-33":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-60":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-190":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-239":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-263":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-271":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.0218978102189782,"pan":{"x":593.051094890511,"y":528.2481751824818},"focalPoint":{"x":20.5,"y":-125.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:23:34.614Z"},{"id":"scene-mermaid-mrxl4f8r-N403","title":"Red-Teaming","description":"Imported from a Mermaid flowchart. Central: N403. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-74","nodes":{"n-mermaid-mrxl4f8r-3":{"position":{"x":-177,"y":-131},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-74":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-75":{"position":{"x":177,"y":131},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-73":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-74":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":614,"y":400},"focalPoint":{"x":0,"y":0}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.069Z"},{"id":"scene-mermaid-mrxl4f8r-N404","title":"Jailbreaking","description":"Imported from a Mermaid flowchart. Central: N404. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-75","nodes":{"n-mermaid-mrxl4f8r-74":{"position":{"x":-177,"y":-131},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-75":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-74":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":746.75,"y":498.25},"focalPoint":{"x":-88.5,"y":-65.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.069Z"},{"id":"scene-mermaid-mrxl4f8r-N405","title":"Guardrails","description":"Imported from a Mermaid flowchart. Central: N405. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-76","nodes":{"n-mermaid-mrxl4f8r-3":{"position":{"x":-89,"y":-201},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-76":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-75":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":680.75,"y":550.75},"focalPoint":{"x":-44.5,"y":-100.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.070Z"},{"id":"scene-mermaid-mrxl4f8r-N410","title":"Hallucination","centralNodeId":"n-mermaid-mrxl4f8r-79","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-79":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-099-231":{"position":{"x":-0.7666666666666446,"y":321.2986219567645},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-77":{"position":{"x":147.2,"y":205.85},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-165":{"position":{"x":-1.0896396653274216,"y":-126.74522537664461},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-93":{"position":{"x":-206.30833743919382,"y":134.31685993110796},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-47":{"position":{"x":374.8079094328644,"y":-95.42686795782525},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-829-052-575":{"position":{"x":-251.05096012926094,"y":-95.79925083586566},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-58":{"position":{"x":352.4819055666376,"y":187.7590723160051},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-829-158-925":{"position":{"x":136.28190556663753,"y":-161.07426101732818},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-78":{"position":{"x":-157.3514277666958,"y":255.99240564933848},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784829175379":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-78":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-211":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-231":{"design":{"id":"default","params":{}}},"e1784829046922":{"design":{"id":"default","params":{}}},"e1784829052589":{"design":{"id":"default","params":{}}},"e1784829104256":{"design":{"id":"default","params":{}}},"e1784829158941":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[25],"control-point-weights":[0.5]}},"e1784829168765":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.3043478260869565,"pan":{"x":536,"y":285.75},"focalPoint":{"x":59.8,"y":87.59166666666667}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:53:43.557Z"},{"id":"scene-mermaid-mrxl4f8r-N420","title":"Evaluation","description":"Imported from a Mermaid flowchart. Central: N420. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-98","nodes":{"n-mermaid-mrxl4f8r-4":{"position":{"x":166,"y":-144},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-98":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-99":{"position":{"x":-125,"y":181},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-100":{"position":{"x":-250,"y":362},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-101":{"position":{"x":-197,"y":98},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-97":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-98":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-99":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-100":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.0510510510510511,"pan":{"x":658.1441441441441,"y":285.4354354354354},"focalPoint":{"x":-42,"y":109}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.070Z"},{"id":"scene-mermaid-mrxl4f8r-N421","title":"Benchmarks","description":"Imported from a Mermaid flowchart. Central: N421. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-99","nodes":{"n-mermaid-mrxl4f8r-98":{"position":{"x":125,"y":-181},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-99":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-100":{"position":{"x":-125,"y":181},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-98":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-99":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.3409961685823755,"pan":{"x":614,"y":400},"focalPoint":{"x":0,"y":0}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.071Z"},{"id":"scene-mermaid-mrxl4f8r-N430","title":"Interpretability","centralNodeId":"n-mermaid-mrxl4f8r-35","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-36":{"position":{"x":217,"y":-36},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-35":{"position":{"x":-4,"y":-12},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-2":{"position":{"x":-217,"y":36},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-165":{"position":{"x":-188.51558243082573,"y":-209.34908170095562},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-34":{"position":{"x":-349.0517998860043,"y":-121.67826310038926},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-168":{"position":{"x":50.92605491162669,"y":164.87952378009953},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-30":{"position":{"x":218.54530976786816,"y":-204.09855291334452},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-35":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-34":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-213":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-232":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-227":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-33":{"design":{"id":"default","params":{}}},"e1784823631895":{"design":{"id":"default","params":{}}},"e1784823336507":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1,"pan":{"x":628.3258999430021,"y":457.6745408504778},"focalPoint":{"x":-14.325899943002128,"y":-57.6745408504778}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.071Z"},{"id":"scene-mermaid-mrxl4f8r-N440","title":"Emergent Capabilities","centralNodeId":"n-mermaid-mrxl4f8r-33","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-2":{"position":{"x":-84,"y":203},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-171":{"position":{"x":-316.01097408292367,"y":125.96834043347012},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-87":{"position":{"x":-164.92070910461467,"y":-270.599982757424},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-31":{"position":{"x":251.15381807906087,"y":-2.854573048819555},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-33":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-267-623":{"position":{"x":208.9592151794272,"y":-254.5756165233307},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-36":{"position":{"x":250.40702308928263,"y":-138.12130674438546},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-78":{"position":{"x":-259.00713302651064,"y":-148.49628246099982},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784823267630":{"design":{"id":"default","params":{}}},"e1784823244092":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-214":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-187":{"design":{"id":"default","params":{}}},"e1784823238320":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-32":{"design":{"id":"default","params":{}}},"e1784823249858":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-30":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1,"pan":{"x":673.6285780019314,"y":424.82728652440977},"focalPoint":{"x":-59.62857800193137,"y":-24.82728652440977}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:46:02.097Z"},{"id":"scene-mermaid-mrxl4f8r-N450","title":"Artificial General Intelligence","centralNodeId":"n-mermaid-mrxl4f8r-171","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-2":{"position":{"x":378.84685883132306,"y":-147.63712530313063},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-171":{"position":{"x":3.4057142857142857,"y":0.8514285714285723},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-33":{"position":{"x":161.6621988006625,"y":-145.05575989575553},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-31":{"position":{"x":322.3799491443108,"y":-291.15724765161906},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-9":{"position":{"x":32,"y":218},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-172":{"position":{"x":-44,"y":-225},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-166":{"position":{"x":258.5822556636793,"y":0.663494459408402},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-059-854":{"position":{"x":-260.58409999632966,"y":154.21041952558835},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-066-853":{"position":{"x":-430.4483400010246,"y":1.792390486363054},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-072-505":{"position":{"x":-379.3563557152313,"y":-272.7780322576294},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-083-572":{"position":{"x":-430.6709050853011,"y":-83.80374499901382},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-094-285":{"position":{"x":-137.67090508530114,"y":245.19625500098618},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-32":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-30":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-214":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-187":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-171":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-170":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-165":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-215":{"design":{"id":"default","params":{}}},"e1784823059856":{"design":{"id":"default","params":{}}},"e1784823066856":{"design":{"id":"default","params":{}}},"e1784823072508":{"design":{"id":"default","params":{}}},"e1784823083576":{"design":{"id":"default","params":{}}},"e1784823094290":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1,"pan":{"x":681.4088721681603,"y":411.1},"focalPoint":{"x":-67.40887216816031,"y":-11.100000000000023}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:46:10.694Z"},{"id":"scene-mermaid-mrxl4f8r-N460","title":"AI Governance","centralNodeId":"n-mermaid-mrxl4f8r-152","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-7":{"position":{"x":75,"y":-207},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-153":{"position":{"x":-75,"y":207},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-152":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-165":{"position":{"x":223.96580441086326,"y":-51.45248735090394},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-40":{"position":{"x":144.33048273252737,"y":178.82044557095705},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-826-493-490":{"position":{"x":-282.9855062571653,"y":-67.65534801928902},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-151":{"position":{"x":-277.0428225140434,"y":156.74328737075166},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-106":{"position":{"x":273.29832086487335,"y":-194.4824272933369},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-74":{"position":{"x":352.01832086487326,"y":61.35757270666309},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-152":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-151":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-233":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-254":{"design":{"id":"default","params":{}}},"e1784826493495":{"design":{"id":"default","params":{}}},"e1784826496815":{"design":{"id":"default","params":{}}},"e1784826499636":{"design":{"id":"default","params":{}}},"e1784826506063":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.2195121951219512,"pan":{"x":614,"y":400},"focalPoint":{"x":0,"y":0}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:40.886Z"},{"id":"scene-mermaid-mrxl4f8r-N462","title":"Export Controls","description":"Imported from a Mermaid flowchart. Central: N462. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-151","nodes":{"n-mermaid-mrxl4f8r-6":{"position":{"x":215,"y":-48},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-151":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-150":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":452.75,"y":436},"focalPoint":{"x":107.5,"y":-24}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.072Z"},{"id":"scene-mermaid-mrxl4f8r-N470","title":"Capability Progress","description":"Imported from a Mermaid flowchart. Central: N470. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-166","nodes":{"n-mermaid-mrxl4f8r-9":{"position":{"x":199,"y":93},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-166":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-167":{"position":{"x":-199,"y":-93},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-165":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-166":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-224":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":614,"y":400},"focalPoint":{"x":0,"y":0}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.073Z"},{"id":"scene-mermaid-mrxl4f8r-N471","title":"Scaling Hypothesis","description":"Imported from a Mermaid flowchart. Central: N471. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-32","nodes":{"n-mermaid-mrxl4f8r-31":{"position":{"x":25,"y":219},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-32":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-31":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":595.25,"y":235.75},"focalPoint":{"x":12.5,"y":109.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.073Z"},{"id":"scene-mermaid-mrxl4f8r-N472","title":"Algorithmic Progress","description":"Imported from a Mermaid flowchart. Central: N472. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-167","nodes":{"n-mermaid-mrxl4f8r-166":{"position":{"x":199,"y":93},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-167":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-166":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-224":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":464.75,"y":330.25},"focalPoint":{"x":99.5,"y":46.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.074Z"},{"id":"scene-mermaid-mrxl4f8r-N480","title":"LLM Understanding Debate","centralNodeId":"n-mermaid-mrxl4f8r-168","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-168":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-9":{"position":{"x":223,"y":289},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-35":{"position":{"x":259.41578236604295,"y":82.92267013430083},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-169":{"position":{"x":-214.72823063915217,"y":-89.76929127869846},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-170":{"position":{"x":329.38453642363,"y":-30.252107705819782},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-825-570-102":{"position":{"x":-113.88566938812511,"y":155.21256338424323},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-825-371-425":{"position":{"x":-0.5340509601913125,"y":-201.71248311978832},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-77":{"position":{"x":-211.40199282575088,"y":-283.2537621104342},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-167":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-168":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-169":{"design":{"id":"default","params":{}}},"e1784823631895":{"design":{"id":"default","params":{}}},"e1784825570106":{"design":{"id":"default","params":{}}},"e1784825577190":{"design":{"id":"default","params":{}}},"e1784825483287":{"design":{"id":"default","params":{}}},"e1784825486063":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-225":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-226":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[200],"control-point-weights":[0.5]}}},"viewport":{"zoom":1,"pan":{"x":549.1718471077611,"y":371.25624155989414},"focalPoint":{"x":64.8281528922389,"y":28.743758440105864}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:00:36.030Z"},{"id":"scene-mermaid-mrxl4f8r-N481","title":"Stochastic Parrot View","centralNodeId":"n-mermaid-mrxl4f8r-169","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-169":{"position":{"x":-176.66666666666666,"y":-28.666666666666664},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-77":{"position":{"x":-101,"y":196},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-168":{"position":{"x":204.57634543626858,"y":3.111091117361263},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-825-371-425":{"position":{"x":97.84120923810559,"y":179.68933684229918},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-225":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-168":{"design":{"id":"default","params":{}}},"e1784825486063":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":689.75,"y":253},"focalPoint":{"x":-50.5,"y":98}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:52:18.491Z"},{"id":"scene-mermaid-mrxl4f8r-N482","title":"World-Model View","centralNodeId":"n-mermaid-mrxl4f8r-170","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-77":{"position":{"x":418.66666666666663,"y":-167.66666666666666},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-170":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-168":{"position":{"x":-434.16868541098086,"y":-123.22589656823831},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-39":{"position":{"x":277.45956858791203,"y":114.5196627454283},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-38":{"position":{"x":202.28991555523112,"y":168.8241870059447},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-825-371-425":{"position":{"x":-185.5121946410821,"y":225.919606135858},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-826-096-398":{"position":{"x":-222.85115561991927,"y":-169.6926955135576},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-826-102-465":{"position":{"x":3.749809617012326,"y":214.82644908129055},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-826-107-744":{"position":{"x":330.46304099551185,"y":-2.6932427234503615},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-826-169-536":{"position":{"x":-7.071408984106448,"y":-256.476089096792},"scale":1,"design":{"id":"default-node","params":{"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-44":{"position":{"x":181.94376827666056,"y":-170.50460351584346},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-226":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-169":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-252":{"design":{"id":"default","params":{}}},"e1784823983422":{"design":{"id":"default","params":{}}},"e1784825483287":{"design":{"id":"default","params":{}}},"e1784825577190":{"design":{"id":"default","params":{}}},"e1784826096404":{"design":{"id":"default","params":{}}},"e1784826102471":{"design":{"id":"default","params":{}}},"e1784826107749":{"design":{"id":"default","params":{}}},"e1784826169542":{"design":{"id":"default","params":{}}},"e1784826269857":{"design":{"id":"default","params":{}}},"e1784826307881":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.1270585093325987,"pan":{"x":635.0372468951925,"y":400.71947206819857},"focalPoint":{"x":-18.665620924729097,"y":-0.6383626601822243}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:05:32.570Z"},{"id":"scene-mermaid-mrxl4f8r-N485","title":"Bubble Question","centralNodeId":"n-mermaid-mrxl4f8r-174","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-174":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-9":{"position":{"x":-263.3333333333333,"y":228.66666666666666},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-149":{"position":{"x":150.63180641736488,"y":-76.26493357253193},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-147":{"position":{"x":151.5695563678022,"y":79.75259502685459},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-159":{"position":{"x":18.732751143382455,"y":165.74475567751793},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-827-369-389":{"position":{"x":-246.47251153755383,"y":-109.71071513776431},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-148":{"position":{"x":-248.75875121671817,"y":4.971228603261874},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-173":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-235":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-236":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-237":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-209":{"design":{"id":"default","params":{}}},"e1784827369400":{"design":{"id":"default","params":{}}},"e1784827376720":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":668,"y":321.5},"focalPoint":{"x":-36,"y":52.333333333333336}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:23:16.285Z"},{"id":"scene-mermaid-mrxl4f8r-N492","title":"Misinformation & Deepfakes","centralNodeId":"n-mermaid-mrxl4f8r-155","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-155":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-7":{"position":{"x":219,"y":-25},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-145":{"position":{"x":245.9408355157669,"y":154.41912260405607},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-154":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-309":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":449.75,"y":418.75},"focalPoint":{"x":109.5,"y":-12.5}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:24.259Z"},{"id":"scene-mermaid-mrxl4f8r-N493","title":"AI Companionship","centralNodeId":"n-mermaid-mrxl4f8r-156","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-156":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-7":{"position":{"x":199,"y":94},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-146":{"position":{"x":282.2892947224447,"y":-73.09797593028651},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-155":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-311":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":464.75,"y":329.5},"focalPoint":{"x":99.5,"y":47}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:15.892Z"},{"id":"scene-mermaid-mrxl4f8r-N500","title":"The Long Road","centralNodeId":"n-mermaid-mrxl4f8r-8","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-159":{"position":{"x":-208,"y":72},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-8":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-157":{"position":{"x":1.0457142857142827,"y":211.99999999999997},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-158":{"position":{"x":-151,"y":160},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-160":{"position":{"x":-218,"y":-33},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-15":{"position":{"x":-386.9485714285714,"y":-66},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-1":{"position":{"x":203,"y":84},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-161":{"position":{"x":-177,"y":-130},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-162":{"position":{"x":-154.35428571428568,"y":-273.2685714285715},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-163":{"position":{"x":-3.1485714285714295,"y":-207.31428571428575},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-164":{"position":{"x":167.20000000000002,"y":-137.02857142857144},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-86":{"position":{"x":329.5804611102169,"y":-136.24570202654905},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-158":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-156":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-216":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[-250],"control-point-weights":[0.98]}},"e-mermaid-mrxl4f8r-157":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-159":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-7":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-160":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-161":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-162":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-163":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-218":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-222":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.1824324324324325,"pan":{"x":662.7533783783783,"y":426.72972972972974},"focalPoint":{"x":-41.23142857142853,"y":-22.605714285714296}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:33.311Z"},{"id":"scene-mermaid-mrxl4f8r-N501","title":"Symbolic AI","description":"Imported from a Mermaid flowchart. Central: N501. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-157","nodes":{"n-mermaid-mrxl4f8r-8":{"position":{"x":59,"y":-212},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-157":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-156":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":569.75,"y":559},"focalPoint":{"x":29.5,"y":-106}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.076Z"},{"id":"scene-mermaid-mrxl4f8r-N502","title":"Connectionism","centralNodeId":"n-mermaid-mrxl4f8r-158","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-158":{"position":{"x":-38.520000000000074,"y":-42.125714285714295},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-174-809":{"position":{"x":-148.22857142857154,"y":120.25142857142859},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-369-059":{"position":{"x":-459.23830186344094,"y":37.213454543604264},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-173-383":{"position":{"x":4.495433111902772,"y":119.52170974524003},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-181-993":{"position":{"x":-300.80742403095445,"y":-114.12742403095433},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-8":{"position":{"x":38.43999999999989,"y":-292.22285714285687},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-38":{"position":{"x":155.26456688809697,"y":-170.79028117381148},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-14":{"position":{"x":214.95999999999992,"y":14.53714285714284},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-16":{"position":{"x":356.5445668880971,"y":-113.04971882618855},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-132-453":{"position":{"x":291.3085714285714,"y":187.6742857142857},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784824369064":{"design":{"id":"default","params":{}}},"e1784824174812":{"design":{"id":"default","params":{}}},"e1784824173390":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-157":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-217":{"design":{"id":"default","params":{}}},"e1784823977113":{"design":{"id":"default","params":{}}},"e1784824126361":{"design":{"id":"default","params":{}}},"e1784824132458":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-19":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-248":{"design":{"id":"default","params":{}}},"e1784824181997":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.1824324324324325,"pan":{"x":662.7533783783783,"y":426.72972972972974},"focalPoint":{"x":-41.23142857142853,"y":-22.605714285714296}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:36:03.756Z"},{"id":"scene-mermaid-mrxl4f8r-N503","title":"AI Winters","centralNodeId":"n-mermaid-mrxl4f8r-159","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-159":{"position":{"x":-6,"y":-1},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-8":{"position":{"x":-188.27760062074464,"y":251.3810208741373},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-157":{"position":{"x":-367.2231561671327,"y":98.1930435084665},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-174":{"position":{"x":176.36539456361106,"y":242.779384566719},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-827-452-679":{"position":{"x":-221.43783305439445,"y":-173.22168437307158},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#e7f485"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-1784-827-454-845":{"position":{"x":-217.13468766030817,"y":-42.1299774242947},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-827-458-834":{"position":{"x":144.09925410979733,"y":-175.2453045206916},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#e7f485"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-1784-827-474-296":{"position":{"x":339.156219364055,"y":37.33585350650512},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-827-480-696":{"position":{"x":338.52359437384916,"y":-86.28483085197854},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-827-552-241":{"position":{"x":-14.99127760024615,"y":-315.20695130567555},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-827-564-110":{"position":{"x":-161.530962556808,"y":-318.90183873737715},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-158":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-156":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-216":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-237":{"design":{"id":"default","params":{}}},"e1784827452691":{"design":{"id":"default","params":{}}},"e1784827454854":{"design":{"id":"default","params":{}}},"e1784827458845":{"design":{"id":"default","params":{}}},"e1784827474305":{"design":{"id":"default","params":{}}},"e1784827480703":{"design":{"id":"default","params":{}}},"e1784827589847":{"design":{"id":"default","params":{}}},"e1784827546517":{"design":{"id":"default","params":{}}},"e1784827552248":{"design":{"id":"default","params":{}}},"e1784827564121":{"design":{"id":"default","params":{}}},"e1784827570892":{"design":{"id":"default","params":{}}},"e1784827596685":{"design":{"id":"default","params":{}}},"e1784827604739":{"design":{"id":"default","params":{}}},"e1784827626475":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[-100],"control-point-weights":[0.5]}}},"viewport":{"zoom":1,"pan":{"x":643.3997808966417,"y":433.7137833694783},"focalPoint":{"x":-29.399780896641687,"y":-33.71378336947828}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:29:28.563Z"},{"id":"scene-mermaid-mrxl4f8r-N504","title":"Deep Learning Revolution","centralNodeId":"n-mermaid-mrxl4f8r-160","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-160":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-15":{"position":{"x":-218,"y":-33},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-42":{"position":{"x":-150.41381101571844,"y":119.3301531706372},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-41":{"position":{"x":-28.73683650158043,"y":189.8372835558954},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-8":{"position":{"x":218,"y":33},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-38":{"position":{"x":108.3566755949183,"y":158.50183233646746},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-073-339":{"position":{"x":227.87579284718245,"y":-133.04101444423034},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-218":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-159":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-249":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-255":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-258":{"design":{"id":"default","params":{}}},"e1784824073345":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":614,"y":400},"focalPoint":{"x":0,"y":0}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:28:08.683Z"},{"id":"scene-mermaid-mrxl4f8r-N505","title":"AlphaGo","description":"Imported from a Mermaid flowchart. Central: N505. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-161","nodes":{"n-mermaid-mrxl4f8r-8":{"position":{"x":177,"y":130},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-161":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-160":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":481.25,"y":302.5},"focalPoint":{"x":88.5,"y":65}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.077Z"},{"id":"scene-mermaid-mrxl4f8r-N506","title":"Transformer Breakthrough","centralNodeId":"n-mermaid-mrxl4f8r-162","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-23":{"position":{"x":330.21973191006794,"y":-81.41592554940468},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-22":{"position":{"x":35.426840990889836,"y":-81.09882930220664},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-815-862-176":{"position":{"x":213.1175776145391,"y":-191.75480576805984},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-162":{"position":{"x":36.666666666666664,"y":46.666666666666664},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-8":{"position":{"x":152.43333333333334,"y":269},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-109":{"position":{"x":-173.8644853108184,"y":217.50752395365961},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-134":{"position":{"x":-217.04225835181882,"y":125.81536556969735},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-815-828-256":{"position":{"x":325.2794183821483,"y":143.80172682961089},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-815-855-132":{"position":{"x":-210.08242238546077,"y":-79.82147243472639},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784815820916":{"design":{"id":"default","params":{}}},"e1784815895320":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-22":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-220":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-161":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-274":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-295":{"design":{"id":"default","params":{}}},"e1784815828260":{"design":{"id":"default","params":{}}},"e1784815855136":{"design":{"id":"default","params":{}}},"e1784815957303":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.3043478260869565,"pan":{"x":581.1739130434783,"y":345.6304347826088},"focalPoint":{"x":25.166666666666675,"y":41.68333333333325}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.077Z"},{"id":"scene-mermaid-mrxl4f8r-N507","title":"ChatGPT Moment","centralNodeId":"n-mermaid-mrxl4f8r-163","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-163":{"position":{"x":-5.999999999999995,"y":7.333333333333336},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-8":{"position":{"x":-7,"y":220},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-107":{"position":{"x":258.5532593582634,"y":-35.60761965626192},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-96":{"position":{"x":124.39567348981302,"y":103.77724421575199},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-384-703":{"position":{"x":-251.5145861562395,"y":161.8906143111995},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-106":{"position":{"x":265.6960849590074,"y":101.59128200537594},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-116":{"position":{"x":117.56766181629776,"y":-88.9912519747253},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-164":{"position":{"x":299.0223909010199,"y":274.5252154652647},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-162":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-221":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-268":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-270":{"design":{"id":"default","params":{}}},"e1784816384708":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[75],"control-point-weights":[0.5]}},"e-mermaid-mrxl4f8r-106":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-238":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-269":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[-125],"control-point-weights":[0.5]}},"e-mermaid-mrxl4f8r-163":{"design":{"id":"default","params":{}}},"e1784816445301":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":619.25,"y":235},"focalPoint":{"x":-3.5,"y":110}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:30:16.872Z"},{"id":"scene-mermaid-mrxl4f8r-N508","title":"Reasoning Era","centralNodeId":"n-mermaid-mrxl4f8r-164","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-8":{"position":{"x":-108,"y":192},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-164":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-107":{"position":{"x":238.14235778238134,"y":-106.01502358150435},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-86":{"position":{"x":99.6703317969614,"y":207.06005157799902},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-87":{"position":{"x":-259.3272063435788,"y":-98.62922645909859},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-094-285":{"position":{"x":-286.7659767716796,"y":28.78699624216106},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-69":{"position":{"x":-291.43264343834636,"y":128.78699624216117},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-163":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-222":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-269":{"design":{"id":"default","params":{}}},"e1784827857649":{"design":{"id":"default","params":{}}},"e1784827867431":{"design":{"id":"default","params":{}}},"e1784827872691":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":695,"y":256},"focalPoint":{"x":-54,"y":96}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:31:36.448Z"},{"id":"scene-mermaid-mrxl4f8r-N601","title":"Research Pioneers","centralNodeId":"n-mermaid-mrxl4f8r-37","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-42":{"position":{"x":171.75714285714287,"y":157},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-37":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-44":{"position":{"x":60.12142857142857,"y":217.07142857142858},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-2":{"position":{"x":-207,"y":-74},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-38":{"position":{"x":65.76428571428573,"y":-194.2642857142857},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-39":{"position":{"x":186,"y":-118},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-40":{"position":{"x":225.45714285714286,"y":-23.192857142857143},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-41":{"position":{"x":231.21428571428572,"y":69.96428571428571},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-45":{"position":{"x":-71.16428571428567,"y":226.07142857142858},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-41":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-43":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-36":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-37":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-38":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-39":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-40":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-44":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.238938053097345,"pan":{"x":606.566371681416,"y":380.79646017699116},"focalPoint":{"x":5.9999999999999885,"y":15.499999999999993}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:50.490Z"},{"id":"scene-mermaid-mrxl4f8r-N602","title":"Executives","centralNodeId":"n-mermaid-mrxl4f8r-115","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-115":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-5":{"position":{"x":34.017504181096946,"y":-235.52625627164542},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-116":{"position":{"x":260.3156406791136,"y":-84.67081855369264},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-118":{"position":{"x":194.0643704704783,"y":44.947487456709155},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-119":{"position":{"x":-277.60078841120526,"y":-99.07509755892877},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-120":{"position":{"x":-144.67758222824,"y":-159.25013936989833},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-121":{"position":{"x":-282.098103049241,"y":158.34571428571428},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-122":{"position":{"x":30.881557925988588,"y":191.67250139369895},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-123":{"position":{"x":-115.05872421656883,"y":157.88876324014035},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-124":{"position":{"x":-278.5875097558928,"y":55.86441506884574},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-125":{"position":{"x":166.47137055496194,"y":189.3562830113714},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-107":{"position":{"x":405.384020786581,"y":-126.26089870158609},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-109":{"position":{"x":382.8124152894338,"y":7.648273182674673},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-161":{"position":{"x":387.57145151900534,"y":111.8012589266047},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-111":{"position":{"x":-428.4368741836409,"y":-55.39834778055338},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-147":{"position":{"x":-363.8769646469699,"y":-225.08603574377122},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-112":{"position":{"x":257.711315943034,"y":275.5710238300598},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-133":{"position":{"x":-424.5064431450817,"y":159.31225914342124},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-110":{"position":{"x":29.930975324035984,"y":305.6356421726441},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-127":{"position":{"x":-145.85506643622722,"y":-259.5649916378062},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-113":{"position":{"x":-117.24456315408511,"y":258.08575982944},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-117":{"position":{"x":103.64731049871402,"y":-106.53133710618638},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-108":{"position":{"x":199.6916799739572,"y":-201.06498885237332},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-114":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-115":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-117":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-118":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-119":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-120":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-121":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-122":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-123":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-124":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-116":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-238":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-273":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-240":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-264":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-241":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-266":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-277":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-247":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-246":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-243":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-244":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-242":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-245":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-239":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.0891758970773753,"pan":{"x":627.2729614204629,"y":367.62072703846707},"focalPoint":{"x":-12.186242328790737,"y":29.72823126954737}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:57.099Z"},{"id":"scene-mermaid-mrxl4f8r-N620","title":"Compute & Silicon","centralNodeId":"n-mermaid-mrxl4f8r-126","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-127":{"position":{"x":-45,"y":215},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-126":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-5":{"position":{"x":131.35428571428574,"y":-227.06857142857143},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-128":{"position":{"x":-125,"y":181},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-130":{"position":{"x":-186,"y":118},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-54":{"position":{"x":199.6575849865615,"y":-38.83823833973395},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-129":{"position":{"x":163.8303392666194,"y":73.1025302967304},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-131":{"position":{"x":104.44929840873662,"y":174.5333895331283},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-827-122-695":{"position":{"x":-363.06002280179035,"y":-1.1341218415614023},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-150":{"position":{"x":-57.077679952243656,"y":-117.2461453987703},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-147":{"position":{"x":-309.7782058255262,"y":-203.33714285714285},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-151":{"position":{"x":-297.52262876239985,"y":-110.40571428571424},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-126":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-125":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-127":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-129":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-128":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-130":{"design":{"id":"default","params":{}}},"e1784818628202":{"design":{"id":"default","params":{}}},"e1784827122706":{"design":{"id":"default","params":{}}},"e1784827129432":{"design":{"id":"default","params":{}}},"e1784827134135":{"design":{"id":"default","params":{}}},"e1784827139676":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.2589928057553956,"pan":{"x":652.3992805755396,"y":378.59712230215825},"focalPoint":{"x":-30.50000000000002,"y":17.000000000000014}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:19:12.337Z"},{"id":"scene-mermaid-mrxl4f8r-N630","title":"Platforms & Infrastructure","description":"Imported from a Mermaid flowchart. Central: N630. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-132","nodes":{"n-mermaid-mrxl4f8r-5":{"position":{"x":196,"y":-99},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-132":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-139":{"position":{"x":-196,"y":99},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-131":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-138":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":614,"y":400},"focalPoint":{"x":0,"y":0}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.079Z"},{"id":"scene-mermaid-mrxl4f8r-N640","title":"Application Layer","centralNodeId":"n-mermaid-mrxl4f8r-140","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-140":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-5":{"position":{"x":253.76666666666668,"y":2.6833333333333322},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-143":{"position":{"x":-231.14999999999998,"y":102.35},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-144":{"position":{"x":-254.53333333333333,"y":1.1500000000000004},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-145":{"position":{"x":-224.24999999999997,"y":-116.14999999999999},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-141":{"position":{"x":-106.00746963834003,"y":-230.42754853255354},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-142":{"position":{"x":147.11661255304827,"y":-188.48591693262532},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-146":{"position":{"x":305.7813317251442,"y":-100.30862122912443},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-139":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-142":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-143":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-144":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-140":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-141":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-145":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.3043478260869565,"pan":{"x":614,"y":409},"focalPoint":{"x":0,"y":-6.8999999999999995}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:40:17.527Z"},{"id":"scene-mermaid-mrxl4f8r-N651","title":"Geoffrey Hinton","centralNodeId":"n-mermaid-mrxl4f8r-38","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-38":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-37":{"position":{"x":-211,"y":205.66666666666666},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-16":{"position":{"x":223.34226777838137,"y":-49.420430677381795},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-160":{"position":{"x":218.4862644572714,"y":39.27861955624148},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-165":{"position":{"x":66.25810846488011,"y":181.71496918364528},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-158":{"position":{"x":-252.18413870881594,"y":47.04866038025173},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-170":{"position":{"x":-245.8366963023615,"y":-87.4579141182112},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-37":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-248":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-249":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-250":{"design":{"id":"default","params":{}}},"e1784823977113":{"design":{"id":"default","params":{}}},"e1784823983422":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":581.693299166214,"y":271.46532300803636},"focalPoint":{"x":21.537800555857302,"y":85.68978466130909}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:26:28.625Z"},{"id":"scene-mermaid-mrxl4f8r-N652","title":"Yann LeCun","centralNodeId":"n-mermaid-mrxl4f8r-39","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-39":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-37":{"position":{"x":-186,"y":118},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-15":{"position":{"x":191.00946761421687,"y":-19.477763776507757},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-170":{"position":{"x":35.27591238197739,"y":188.73158189772863},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-38":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-251":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-252":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":753.5,"y":311.5},"focalPoint":{"x":-93,"y":59}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:05.934Z"},{"id":"scene-mermaid-mrxl4f8r-N653","title":"Yoshua Bengio","centralNodeId":"n-mermaid-mrxl4f8r-40","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-40":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-37":{"position":{"x":-219,"y":24},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-15":{"position":{"x":18.430261281654484,"y":-160.9482074118564},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-152":{"position":{"x":52.850515836261806,"y":153.13661539893403},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-39":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-253":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-254":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":779.075,"y":382.375},"focalPoint":{"x":-110.05000000000003,"y":11.75}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:07:49.512Z"},{"id":"scene-mermaid-mrxl4f8r-N654","title":"Fei-Fei Li","description":"Imported from a Mermaid flowchart. Central: N654. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-41","nodes":{"n-mermaid-mrxl4f8r-37":{"position":{"x":-207,"y":-74},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-41":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-40":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":769.25,"y":455.5},"focalPoint":{"x":-103.5,"y":-37}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:23:34.618Z"},{"id":"scene-mermaid-mrxl4f8r-N655","title":"Ilya Sutskever","centralNodeId":"n-mermaid-mrxl4f8r-42","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-42":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-37":{"position":{"x":-154,"y":-157},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-32":{"position":{"x":60.777238203096665,"y":-168.79125367270677},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-160":{"position":{"x":217.3735772024686,"y":68.71279020389807},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-41":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-257":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-258":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":582.5,"y":461.75},"focalPoint":{"x":21,"y":-41.166666666666664}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:45.915Z"},{"id":"scene-mermaid-mrxl4f8r-N656","title":"Andrew Ng","description":"Imported from a Mermaid flowchart. Central: N656. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-43","nodes":{"n-mermaid-mrxl4f8r-10":{"position":{"x":-166,"y":144},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-43":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-259":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":738.5,"y":292},"focalPoint":{"x":-83,"y":72}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:23:34.619Z"},{"id":"scene-mermaid-mrxl4f8r-N657","title":"Jürgen Schmidhuber","centralNodeId":"n-mermaid-mrxl4f8r-44","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-44":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-37":{"position":{"x":-69,"y":-209},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-43":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":665.75,"y":556.75},"focalPoint":{"x":-34.5,"y":-104.5}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:42:35.853Z"},{"id":"scene-mermaid-mrxl4f8r-N661","title":"Sam Altman","description":"Imported from a Mermaid flowchart. Central: N661. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-116","nodes":{"n-mermaid-mrxl4f8r-115":{"position":{"x":-219,"y":17},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-116":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-115":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":778.25,"y":387.25},"focalPoint":{"x":-109.5,"y":8.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:23:34.619Z"},{"id":"scene-mermaid-mrxl4f8r-N662","title":"Dario Amodei","description":"Imported from a Mermaid flowchart. Central: N662. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-117","nodes":{"n-mermaid-mrxl4f8r-34":{"position":{"x":-213,"y":57},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-117":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-263":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":773.75,"y":357.25},"focalPoint":{"x":-106.5,"y":28.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:23:34.620Z"},{"id":"scene-mermaid-mrxl4f8r-N663","title":"Demis Hassabis","description":"Imported from a Mermaid flowchart. Central: N663. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-118","nodes":{"n-mermaid-mrxl4f8r-115":{"position":{"x":-206,"y":-78},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-118":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-117":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":768.5,"y":458.5},"focalPoint":{"x":-103,"y":-39}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:23:34.620Z"},{"id":"scene-mermaid-mrxl4f8r-N664","title":"Elon Musk","description":"Imported from a Mermaid flowchart. Central: N664. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-119","nodes":{"n-mermaid-mrxl4f8r-115":{"position":{"x":-153,"y":-158},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-119":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-118":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":728.75,"y":518.5},"focalPoint":{"x":-76.5,"y":-79}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:23:34.620Z"},{"id":"scene-mermaid-mrxl4f8r-N665","title":"Jensen Huang","centralNodeId":"n-mermaid-mrxl4f8r-120","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-120":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-mermaid-mrxl4f8r-115":{"position":{"x":-72,"y":-208},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-127":{"position":{"x":152.74687081045622,"y":-53.96659575713288},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-119":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-242":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":668,"y":556},"focalPoint":{"x":-36,"y":-104}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:17:15.469Z"},{"id":"scene-mermaid-mrxl4f8r-N666","title":"Satya Nadella","description":"Imported from a Mermaid flowchart. Central: N666. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-121","nodes":{"n-mermaid-mrxl4f8r-115":{"position":{"x":23,"y":-219},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-121":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-120":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":596.75,"y":564.25},"focalPoint":{"x":11.5,"y":-109.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:23:34.621Z"},{"id":"scene-mermaid-mrxl4f8r-N667","title":"Mark Zuckerberg","description":"Imported from a Mermaid flowchart. Central: N667. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-122","nodes":{"n-mermaid-mrxl4f8r-115":{"position":{"x":114,"y":-188},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-122":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-121":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":528.5,"y":541},"focalPoint":{"x":57,"y":-94}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:23:34.622Z"},{"id":"scene-mermaid-mrxl4f8r-N668","title":"Liang Wenfeng","description":"Imported from a Mermaid flowchart. Central: N668. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-123","nodes":{"n-mermaid-mrxl4f8r-115":{"position":{"x":183,"y":-123},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-123":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-122":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":476.75,"y":492.25},"focalPoint":{"x":91.5,"y":-61.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:23:34.622Z"},{"id":"scene-mermaid-mrxl4f8r-N669","title":"Mustafa Suleyman","description":"Imported from a Mermaid flowchart. Central: N669. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-124","nodes":{"n-mermaid-mrxl4f8r-115":{"position":{"x":217,"y":-34},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-124":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-123":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":451.25,"y":425.5},"focalPoint":{"x":108.5,"y":-17}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:23:34.623Z"},{"id":"scene-mermaid-mrxl4f8r-N670","title":"Arthur Mensch","description":"Imported from a Mermaid flowchart. Central: N670. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-125","nodes":{"n-mermaid-mrxl4f8r-115":{"position":{"x":211,"y":62},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-125":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-124":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":455.75,"y":353.5},"focalPoint":{"x":105.5,"y":31}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:23:34.623Z"},{"id":"scene-mermaid-mrxl4f8r-N671","title":"Andrej Karpathy","description":"Imported from a Mermaid flowchart. Central: N671. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-45","nodes":{"n-mermaid-mrxl4f8r-37":{"position":{"x":30,"y":-218},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-45":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{"colorOverrides":{"text":"#70c7c7"},"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e-mermaid-mrxl4f8r-44":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":591.5,"y":563.5},"focalPoint":{"x":15,"y":-109}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:23:34.624Z"},{"id":"scene-mermaid-mrxl4f8r-N681","title":"OpenAI","description":"Imported from a Mermaid flowchart. Central: N681. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-107","nodes":{"n-mermaid-mrxl4f8r-96":{"position":{"x":-83,"y":-204},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-107":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-270":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":676.25,"y":553},"focalPoint":{"x":-41.5,"y":-102}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.093Z"},{"id":"scene-mermaid-mrxl4f8r-N682","title":"Anthropic","description":"Imported from a Mermaid flowchart. Central: N682. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-108","nodes":{"n-mermaid-mrxl4f8r-34":{"position":{"x":-172,"y":137},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-108":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-271":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":743,"y":297.25},"focalPoint":{"x":-86,"y":68.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.093Z"},{"id":"scene-mermaid-mrxl4f8r-N683","title":"Google DeepMind","description":"Imported from a Mermaid flowchart. Central: N683. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-109","nodes":{"n-mermaid-mrxl4f8r-106":{"position":{"x":-124,"y":-182},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-109":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-108":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":707,"y":536.5},"focalPoint":{"x":-62,"y":-91}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.093Z"},{"id":"scene-mermaid-mrxl4f8r-N684","title":"Meta AI","description":"Imported from a Mermaid flowchart. Central: N684. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-110","nodes":{"n-mermaid-mrxl4f8r-102":{"position":{"x":-25,"y":219},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-110":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-276":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":632.75,"y":235.75},"focalPoint":{"x":-12.5,"y":109.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.093Z"},{"id":"scene-mermaid-mrxl4f8r-N685","title":"xAI","description":"Imported from a Mermaid flowchart. Central: N685. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-111","nodes":{"n-mermaid-mrxl4f8r-106":{"position":{"x":-41,"y":-216},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-111":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-110":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":644.75,"y":562},"focalPoint":{"x":-20.5,"y":-108}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.094Z"},{"id":"scene-mermaid-mrxl4f8r-N686","title":"Mistral AI","description":"Imported from a Mermaid flowchart. Central: N686. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-112","nodes":{"n-mermaid-mrxl4f8r-102":{"position":{"x":-113,"y":188},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-112":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-278":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":698.75,"y":259},"focalPoint":{"x":-56.5,"y":94}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.094Z"},{"id":"scene-mermaid-mrxl4f8r-N687","title":"DeepSeek","description":"Imported from a Mermaid flowchart. Central: N687. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-113","nodes":{"n-mermaid-mrxl4f8r-70":{"position":{"x":-203,"y":84},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-113":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-280":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":766.25,"y":337},"focalPoint":{"x":-101.5,"y":42}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.094Z"},{"id":"scene-mermaid-mrxl4f8r-N688","title":"Alibaba Qwen","description":"Imported from a Mermaid flowchart. Central: N688. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-114","nodes":{"n-mermaid-mrxl4f8r-102":{"position":{"x":-182,"y":124},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-114":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-285":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":750.5,"y":307},"focalPoint":{"x":-91,"y":62}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.095Z"},{"id":"scene-mermaid-mrxl4f8r-N691","title":"Nvidia","centralNodeId":"n-mermaid-mrxl4f8r-127","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-120":{"position":{"x":-150.99336496404308,"y":58.694154196438454},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-127":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-126":{"position":{"x":130,"y":-304},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-72":{"position":{"x":-284.72761107582403,"y":-92.9933992416513},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-52":{"position":{"x":-174.27303313965,"y":-231.84032808346052},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-147":{"position":{"x":349.09147945974684,"y":-270.18580393473417},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-827-044-159":{"position":{"x":85.64871761393394,"y":95.12253766110639},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-827-051-233":{"position":{"x":-119.41067913718071,"y":222.80255425291102},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-827-060-218":{"position":{"x":453.1943989192639,"y":-80.46225369439799},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-827-071-779":{"position":{"x":399.7375514166874,"y":145.70133189342573},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-242":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-126":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-287":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-288":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-205":{"design":{"id":"default","params":{}}},"e1784826989787":{"design":{"id":"default","params":{}}},"e1784827044169":{"design":{"id":"default","params":{}}},"e1784827051241":{"design":{"id":"default","params":{}}},"e1784827060228":{"design":{"id":"default","params":{}}},"e1784827071789":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[-50],"control-point-weights":[0.5]}}},"viewport":{"zoom":0.9727472237769671,"pan":{"x":583.1828141472718,"y":432.6538161252117},"focalPoint":{"x":31.680569318997154,"y":-33.56865517274259}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:18:10.129Z"},{"id":"scene-mermaid-mrxl4f8r-N692","title":"AMD","description":"Imported from a Mermaid flowchart. Central: N692. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-128","nodes":{"n-mermaid-mrxl4f8r-126":{"position":{"x":125,"y":-181},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-128":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-127":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":520.25,"y":535.75},"focalPoint":{"x":62.5,"y":-90.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.095Z"},{"id":"scene-mermaid-mrxl4f8r-N693","title":"TSMC","description":"Imported from a Mermaid flowchart. Central: N693. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-129","nodes":{"n-mermaid-mrxl4f8r-71":{"position":{"x":-209,"y":-68},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-129":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-290":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":770.75,"y":451},"focalPoint":{"x":-104.5,"y":-34}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.096Z"},{"id":"scene-mermaid-mrxl4f8r-N694","title":"CoreWeave","description":"Imported from a Mermaid flowchart. Central: N694. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-130","nodes":{"n-mermaid-mrxl4f8r-126":{"position":{"x":186,"y":-118},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-130":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-129":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":474.5,"y":488.5},"focalPoint":{"x":93,"y":-59}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.096Z"},{"id":"scene-mermaid-mrxl4f8r-N695","title":"Groq","description":"Imported from a Mermaid flowchart. Central: N695. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-131","nodes":{"n-mermaid-mrxl4f8r-71":{"position":{"x":-168,"y":-142},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-131":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-293":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":740,"y":506.5},"focalPoint":{"x":-84,"y":-71}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.096Z"},{"id":"scene-mermaid-mrxl4f8r-N701","title":"Microsoft","description":"Imported from a Mermaid flowchart. Central: N701. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-133","nodes":{"n-mermaid-mrxl4f8r-95":{"position":{"x":-170,"y":-140},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-133":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-294":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":741.5,"y":505},"focalPoint":{"x":-85,"y":-70}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.097Z"},{"id":"scene-mermaid-mrxl4f8r-N702","title":"Google","description":"Imported from a Mermaid flowchart. Central: N702. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-134","nodes":{"n-mermaid-mrxl4f8r-97":{"position":{"x":29,"y":-218},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-134":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-297":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":592.25,"y":563.5},"focalPoint":{"x":14.5,"y":-109}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.097Z"},{"id":"scene-mermaid-mrxl4f8r-N703","title":"Amazon Web Services","description":"Imported from a Mermaid flowchart. Central: N703. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-135","nodes":{"n-mermaid-mrxl4f8r-103":{"position":{"x":-154,"y":157},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-135":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-298":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":729.5,"y":282.25},"focalPoint":{"x":-77,"y":78.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.097Z"},{"id":"scene-mermaid-mrxl4f8r-N704","title":"Hugging Face","description":"Imported from a Mermaid flowchart. Central: N704. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-136","nodes":{"n-mermaid-mrxl4f8r-102":{"position":{"x":-217,"y":37},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-136":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-300":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":776.75,"y":372.25},"focalPoint":{"x":-108.5,"y":18.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.098Z"},{"id":"scene-mermaid-mrxl4f8r-N705","title":"OpenRouter","description":"Imported from a Mermaid flowchart. Central: N705. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-137","nodes":{"n-mermaid-mrxl4f8r-103":{"position":{"x":-209,"y":69},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-137":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-301":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":770.75,"y":348.25},"focalPoint":{"x":-104.5,"y":34.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.098Z"},{"id":"scene-mermaid-mrxl4f8r-N706","title":"Together AI","description":"Imported from a Mermaid flowchart. Central: N706. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-138","nodes":{"n-mermaid-mrxl4f8r-103":{"position":{"x":-217,"y":-35},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-138":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-302":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":776.75,"y":426.25},"focalPoint":{"x":-108.5,"y":-17.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.098Z"},{"id":"scene-mermaid-mrxl4f8r-N707","title":"Scale AI","description":"Imported from a Mermaid flowchart. Central: N707. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-139","nodes":{"n-mermaid-mrxl4f8r-132":{"position":{"x":196,"y":-99},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-139":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-138":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":467,"y":474.25},"focalPoint":{"x":98,"y":-49.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.099Z"},{"id":"scene-mermaid-mrxl4f8r-N711","title":"Perplexity","description":"Imported from a Mermaid flowchart. Central: N711. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-141","nodes":{"n-mermaid-mrxl4f8r-97":{"position":{"x":117,"y":-186},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-141":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-304":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":526.25,"y":539.5},"focalPoint":{"x":58.5,"y":-93}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.099Z"},{"id":"scene-mermaid-mrxl4f8r-N712","title":"Cursor","description":"Imported from a Mermaid flowchart. Central: N712. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-142","nodes":{"n-mermaid-mrxl4f8r-95":{"position":{"x":-98,"y":-197},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-142":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-305":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":687.5,"y":547.75},"focalPoint":{"x":-49,"y":-98.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.099Z"},{"id":"scene-mermaid-mrxl4f8r-N713","title":"Midjourney","description":"Imported from a Mermaid flowchart. Central: N713. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-143","nodes":{"n-mermaid-mrxl4f8r-140":{"position":{"x":201,"y":-89},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-143":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-142":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":463.25,"y":466.75},"focalPoint":{"x":100.5,"y":-44.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.100Z"},{"id":"scene-mermaid-mrxl4f8r-N714","title":"Runway","description":"Imported from a Mermaid flowchart. Central: N714. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-144","nodes":{"n-mermaid-mrxl4f8r-140":{"position":{"x":220,"y":7},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-144":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-143":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":449,"y":394.75},"focalPoint":{"x":110,"y":3.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.100Z"},{"id":"scene-mermaid-mrxl4f8r-N715","title":"ElevenLabs","description":"Imported from a Mermaid flowchart. Central: N715. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-145","nodes":{"n-mermaid-mrxl4f8r-140":{"position":{"x":195,"y":101},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-145":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-144":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":467.75,"y":324.25},"focalPoint":{"x":97.5,"y":50.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.100Z"},{"id":"scene-mermaid-mrxl4f8r-N716","title":"Character.AI","description":"Imported from a Mermaid flowchart. Central: N716. Depth: 2. Layout: fan.","centralNodeId":"n-mermaid-mrxl4f8r-146","nodes":{"n-mermaid-mrxl4f8r-96":{"position":{"x":15,"y":-219},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-146":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-310":{"design":{"id":"default","params":{}}}},"backgroundImages":[],"themeId":"ocean","viewport":{"zoom":1.5,"pan":{"x":602.75,"y":564.25},"focalPoint":{"x":7.5,"y":-109.5}},"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:22:50.101Z"},{"id":"scene-mermaid-mrxl4f8r-N900","title":"Promise & Peril","centralNodeId":"n-mermaid-mrxl4f8r-7","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-154":{"position":{"x":-199.6742857142857,"y":131.27428571428567},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-7":{"position":{"x":0,"y":0},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-155":{"position":{"x":-237.13142857142861,"y":17.36571428571429},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-156":{"position":{"x":-199,"y":-94},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-1":{"position":{"x":322.2857142857143,"y":-108.81142857142856},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-152":{"position":{"x":-75,"y":207},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-153":{"position":{"x":-127.09714285714286,"y":363.42285714285714},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-153":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-154":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-155":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-6":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-151":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-152":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.0479041916167664,"pan":{"x":627.3832335329341,"y":265.33532934131733},"focalPoint":{"x":-12.771428571428569,"y":128.50857142857146}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:36:36.657Z"},{"id":"scene-multi-head-attention-n-1784-815-862-176","title":"Transformer Architecture","centralNodeId":"n-1784-815-862-176","themeId":"ocean","nodes":{"n-1784-815-862-176":{"position":{"x":60,"y":-50.49999999999997},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-23":{"position":{"x":308.4389875703559,"y":-149.5973632093474},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-815-855-132":{"position":{"x":-257.13093155063586,"y":-289.3936716925158},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-112-838":{"position":{"x":238.30895512747983,"y":124.14894614584514},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-107-389":{"position":{"x":-192.02279190532352,"y":-50.45362941033268},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-030-709":{"position":{"x":473.4393743724797,"y":-7.317771134039658},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784815895320":{"design":{"id":"default","params":{}}},"e1784816185121":{"design":{"id":"default","params":{}}},"e1784816188655":{"design":{"id":"default","params":{}}},"e1784816192903":{"design":{"id":"default","params":{}}},"e1784816112844":{"design":{"id":"default","params":{}}},"e1784816121310":{"design":{"id":"default","params":{}}},"e1784816030720":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.1308562197092085,"pan":{"x":546.1486268174475,"y":457.108239095315},"focalPoint":{"x":59.999999999999986,"y":-50.499999999999986}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:16:07.271Z","updatedAt":"2026-07-23T16:22:50.101Z"},{"id":"scene-nettalk-n-1784-824-369-059","title":"Connectionism","centralNodeId":"n-1784-824-369-059","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-158":{"position":{"x":-283.99389275140845,"y":-87.34463512358076},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-369-059":{"position":{"x":27.5,"y":-2},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-824-174-809":{"position":{"x":-105.73263464331,"y":153.0454410442914},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784824494141":{"design":{"id":"default","params":{}}},"e1784824369064":{"design":{"id":"default","params":{}}},"e1784824174812":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.5,"pan":{"x":572.75,"y":403},"focalPoint":{"x":27.5,"y":-2}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T16:35:39.104Z","updatedAt":"2026-07-23T16:35:54.160Z"},{"id":"scene-neuro-symbolic-integration-n-1784-823-072-505","title":"Artificial General Intelligence","centralNodeId":"n-1784-823-072-505","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-171":{"position":{"x":315.35319783278527,"y":262.52946082905794},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-072-505":{"position":{"x":-67.40887216816031,"y":-11.100000000000023},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-87":{"position":{"x":370.0458256840951,"y":-72.8312917341064},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-157":{"position":{"x":-376.7131794324473,"y":-129.71622259303746},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-86":{"position":{"x":-67.00546258770801,"y":169.50142118392458},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-158":{"position":{"x":-382.2152191232755,"y":66.12449061885226},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-825-281-004":{"position":{"x":286.5911278318397,"y":-167.70000000000005},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-825-283-108":{"position":{"x":270.6092661654826,"y":51.54795892829259},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-825-371-425":{"position":{"x":-66.40887216816031,"y":-197.10000000000002},"scale":1,"design":{"id":"default-node","params":{"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}}},"edges":{"e1784823072508":{"design":{"id":"default","params":{}}},"e1784825289603":{"design":{"id":"default","params":{}}},"e1784825248613":{"design":{"id":"default","params":{}}},"e1784825250429":{"design":{"id":"default","params":{}}},"e1784825255085":{"design":{"id":"default","params":{}}},"e1784825281012":{"design":{"id":"default","params":{}}},"e1784825283116":{"design":{"id":"default","params":{}}},"e1784825422658":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1,"pan":{"x":680.4088721681603,"y":350.1},"focalPoint":{"x":-66.40887216816031,"y":49.89999999999998}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T16:46:11.161Z","updatedAt":"2026-07-23T16:50:43.474Z"},{"id":"scene-prompt-engineering-n-mermaid-mrxl4f8r-65","title":"Prompt","centralNodeId":"n-mermaid-mrxl4f8r-65","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-64":{"position":{"x":108,"y":212.00000000000003},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-65":{"position":{"x":40,"y":3.0000000000000284},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-829-951-047":{"position":{"x":-256.9474007773847,"y":29.454760851438024},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-66":{"position":{"x":421.6628858930481,"y":92.59614300456406},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-78":{"position":{"x":-233.55471229102483,"y":-134.23592043984144},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-67":{"design":{"id":"default","params":{}}},"e1784829951052":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-68":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[-150],"control-point-weights":[0.5]}},"e1784829972469":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.1986301369863013,"pan":{"x":566.054794520548,"y":396.40410958904107},"focalPoint":{"x":40.00000000000001,"y":3.0000000000000218}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T18:05:33.007Z","updatedAt":"2026-07-23T18:06:28.051Z"},{"id":"scene-scaled-dot-product-attention-n-1784-816-030-709","title":"Transformer Architecture","centralNodeId":"n-1784-816-030-709","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-23":{"position":{"x":-268.5932439449809,"y":-275.0181635038792},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-030-709":{"position":{"x":76.80142857142852,"y":-61.11142857142855},"scale":1,"design":{"id":"equation-compact-node","params":{"effects":{"backgroundOpacity":0.5,"backgroundAltOpacity":0.5},"colorOverrides":{"text":"#f287e9"}}}},"n-1784-816-112-838":{"position":{"x":-217.57613353071415,"y":108.37957442274197},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-099-231":{"position":{"x":463.22887305062653,"y":-276.57793705662306},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-107-389":{"position":{"x":449.64129888058045,"y":-64.06553760374373},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784816030720":{"design":{"id":"default","params":{}}},"e1784816112844":{"design":{"id":"default","params":{}}},"e1784816121310":{"design":{"id":"default","params":{}}},"e1784816303584":{"design":{"id":"default","params":{}}},"e1784816311221":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.1308562197092085,"pan":{"x":546.1486268174475,"y":457.108239095315},"focalPoint":{"x":59.999999999999986,"y":-50.499999999999986}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T14:17:45.111Z","updatedAt":"2026-07-23T16:22:50.102Z"},{"id":"scene-softmax-bottleneck-n-1784-816-099-231","title":"Hallucination","centralNodeId":"n-1784-816-099-231","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-79":{"position":{"x":60.56666666666663,"y":-233.70695529009785},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-099-231":{"position":{"x":59.79999999999998,"y":87.59166666666667},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-23":{"position":{"x":194.2420068848165,"y":-49.482571557672685},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-816-030-709":{"position":{"x":251.78589119296822,"y":89.91923682495633},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-30":{"position":{"x":190.87929902253865,"y":227.884992836374},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784829175379":{"design":{"id":"default","params":{}}},"e1784816099237":{"design":{"id":"default","params":{}}},"e1784816303584":{"design":{"id":"default","params":{}}},"e1784828596882":{"design":{"id":"default","params":{}}},"e1784816030720":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1,"pan":{"x":551.4333333333334,"y":357.90764431171556},"focalPoint":{"x":62.566666666666606,"y":42.09235568828444}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T17:53:04.004Z","updatedAt":"2026-07-23T17:53:38.418Z"},{"id":"scene-superintelligence-n-mermaid-mrxl4f8r-172","title":"Artificial General Intelligence","centralNodeId":"n-mermaid-mrxl4f8r-172","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-171":{"position":{"x":-20.003157882446033,"y":214.75142857142856},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-172":{"position":{"x":-67.40887216816031,"y":-11.100000000000023},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-094-285":{"position":{"x":-442.46075935595985,"y":126.28920603275807},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-31":{"position":{"x":141.7342163834774,"y":-160.11821689838052},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-2":{"position":{"x":378.911685476148,"y":217.3814542849999},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-33":{"position":{"x":173.08954126126469,"y":215.85636707195147},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-171":{"design":{"id":"default","params":{}}},"e1784823155966":{"design":{"id":"default","params":{}}},"e1784823161495":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-32":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-30":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-214":{"design":{"id":"default","params":{}}},"e-mermaid-mrxl4f8r-187":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1,"pan":{"x":616.4088721681603,"y":322.1},"focalPoint":{"x":-2.4088721681603147,"y":77.89999999999998}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T16:12:15.954Z","updatedAt":"2026-07-23T16:22:50.102Z"},{"id":"scene-symbol-grounding-problem-n-1784-825-371-425","title":"Artificial General Intelligence","centralNodeId":"n-1784-825-371-425","themeId":"ocean","nodes":{"n-1784-825-371-425":{"position":{"x":-66.40887216816031,"y":49.89999999999998},"scale":1,"design":{"id":"default-node","params":{"effects":{"backgroundOpacity":0.55,"backgroundAltOpacity":0.55}}}},"n-1784-823-072-505":{"position":{"x":-67.40887216816031,"y":235.89999999999998},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-823-083-572":{"position":{"x":343.5911278318397,"y":49.099999999999966},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-825-477-825":{"position":{"x":108.5363066118104,"y":154.9612275852573},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-170":{"position":{"x":-315.5812504386481,"y":-119.20057581361768},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-169":{"position":{"x":77.37854216329896,"y":-143.1258846766578},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e1784825422658":{"design":{"id":"default","params":{}}},"e1784825469864":{"design":{"id":"default","params":{}}},"e1784825477833":{"design":{"id":"default","params":{}}},"e1784825483287":{"design":{"id":"default","params":{}}},"e1784825486063":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1,"pan":{"x":680.4088721681603,"y":350.1},"focalPoint":{"x":-66.40887216816031,"y":49.89999999999998}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T16:50:44.966Z","updatedAt":"2026-07-23T16:51:45.362Z"},{"id":"scene-system-prompt-n-mermaid-mrxl4f8r-66","title":"Prompt","centralNodeId":"n-mermaid-mrxl4f8r-66","themeId":"ocean","nodes":{"n-mermaid-mrxl4f8r-64":{"position":{"x":-273.6628858930481,"y":122.403856995436},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-66":{"position":{"x":40,"y":3.0000000000000284},"scale":1,"design":{"id":"default-node","params":{}}},"n-1784-830-009-038":{"position":{"x":296.321151734499,"y":-166.3453832560842},"scale":1,"design":{"id":"default-node","params":{}}},"n-mermaid-mrxl4f8r-76":{"position":{"x":250.7962212424287,"y":127.92691874038569},"scale":1,"design":{"id":"default-node","params":{}}}},"edges":{"e-mermaid-mrxl4f8r-68":{"design":{"id":"default","params":{}},"curve":{"curve-style":"unbundled-bezier","control-point-distances":[-150],"control-point-weights":[0.5]}},"e1784830009044":{"design":{"id":"default","params":{}}},"e1784830021126":{"design":{"id":"default","params":{}}}},"viewport":{"zoom":1.1986301369863013,"pan":{"x":566.054794520548,"y":396.40410958904107},"focalPoint":{"x":40.00000000000001,"y":3.0000000000000218}},"backgroundImages":[],"foldedNodes":{},"createdAt":"2026-07-23T18:06:28.323Z","updatedAt":"2026-07-23T18:07:12.949Z"}],"nodeImages":[]},"settings":{"fold":{"collapseDuration":300,"collapseDelayBetweenLayers":0,"collapseEdgeFadeDelay":1500,"expandDuration":600,"bulkUnfoldDuration":200,"expandEdgeFadeDelay":300,"easingCollapse":"ease-in","easingExpand":"ease-out","minRadiusMultiplier":1,"expandShowAllEdges":true,"collapseRemoveAll":false},"transition":{"transitionMode":"animated","departureEdgesFadeOut":[500,200],"departureLayerDuration":[400,0],"departureLayerStagger":[200,0],"departureCentralZoomOut":[500,200],"morphDuration":[700,200],"morphCrossfadeOverlap":0,"sharedBgFadeOut":[500,200],"sharedBgFadeIn":[500,200],"arrivalLayerDuration":[400,0],"arrivalLayerStagger":[200,0],"arrivalEdgesFadeIn":[500,200],"openBgFadeIn":[500,200],"openCentralFlyIn":[500,200],"openCentralZoomIn":[500,200],"openLayerDuration":[400,0],"openLayerStagger":[100,0],"openEdgesFadeIn":[500,200],"closeLayerDuration":[400,0],"closeLayerStagger":[100,0],"closeCentralFadeOut":[400,0],"departureEdgeTiming":"parallel","sharedBackgroundTiming":"parallel","arrivalEdgeTiming":"parallel","openEdgeMode":"parallel","openFitPadding":50},"ai":{"provider":"gemini","geminiModel":"gemini-3-flash-preview","openrouterModel":"anthropic/claude-sonnet-4.6","webSearchEnabled":true,"thinkingLevel":"low","chatScrollPosition":"top","imageResultCount":3,"storeRetrievedImages":true,"responseLanguage":"","customInstructions":"","scenePromptInstructions":"","nodePromptInstructions":"","suggestPromptInstructions":"","connectPromptInstructions":"","shelfPreviewScale":0.8,"shelfPreviewOpacity":0.7,"shelfPreviewGrayscale":30,"shelfExitDuration":800,"shelfPauseBetween":1000,"shelfEnterDuration":1600,"shelfRemovalDuration":1500,"shelfRemovalPause":1000,"shelfCollapseDuration":2000,"shelfAdditionDuration":2000},"node":{"inheritDesignFromSelected":true,"inheritDesignForConnected":true,"defaultDesign":"default-node","equationDesign":"equation-compact-node","shelfDesignWithEquation":"equation-compact-node","shelfDesignBasic":"default-node","spacing":1,"scaleStep":1.1,"imageMaxKB":1024,"imageSizeSmall":64,"imageSizeMedium":128,"imageSizeLarge":256},"edge":{"defaultColor":"#7d8590","defaultOpacity":1,"defaultWidth":2,"defaultCurveStyle":"bezier","defaultArrowShape":"triangle","defaultArrowScale":1,"bezierControlDistances":[-100],"bezierControlWeights":[0.5],"segmentRadii":[10,10],"taxiRadius":15,"taxiDirection":"auto"},"customTheme":{"baseTheme":"default","canvasColor":"","canvasVignetteStrength":"","canvasVignetteSpread":"","canvasVignetteBlur":"","canvasVignetteColor":"","canvasVignetteColorOpacity":"","nodeBackground":"","nodeOpacity":"","nodeTextColor":"","nodeBorderColor":"","nodeBorderWidth":"","centralBorderColor":"","selectedBorderColor":"","centralSelectedBorderColor":"","shadowOffsetX":"","shadowOffsetY":"","shadowBlur":"","shadowOpacity":"","shadowColor":"","nodeVignetteStrength":"","nodeVignetteSpread":"","nodeVignetteBlur":"","nodeVignetteColor":"","edgeColor":"","edgeArrowColor":""},"interaction":{"doubleClickNode":"navigate","zoomStep":1.2},"autolayout":{"layoutType":"radial","ringSpacing":120,"siblingGap":35,"footprintScale":0.75,"ringOrder":"edge","rotateStep":15,"densityStep":1.15,"animate":true,"animationDuration":600,"growDirection":"both","growConfirmThreshold":30},"arrange":{"spacingStep":1.15,"rotateStep":15},"file":{"namePrefix":"kg-modern-ai-technology","nextNumber":2}},"chat":[{"nodeId":"n-1784-815-862-176","messages":[{"id":"msg-1784816175380-6r76fw2","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T14:16:15.380Z","source":"ai"},{"id":"msg-1784816181380-1zpu9ay","role":"assistant","content":"Multi-head attention transforms a single, global attention mechanism into a bank of parallel \"representation subspaces.\" While a single attention head must attempt to capture every relevant relationship at once—balancing syntactic dependencies, semantic associations, and long-range references—multi-head attention allows the model to delegate these tasks.\n\nIn a standard transformer layer, the input embeddings are projected into multiple sets of Queries ($Q$), Keys ($K$), and Values ($V$). Each set, or \"head,\" performs its own independent attention operation. This allows one head to focus on the immediate predecessor (useful for local syntax), while another identifies the subject of a sentence across several clauses (useful for coreference resolution).\n\nThe technical efficiency of this approach stems from how the dimensionality is managed. If the total model dimension is $d_{model}$ and there are $h$ heads, each head typically operates on a reduced dimensionality of $d_k = d_{model} / h$. The results of these parallel operations are concatenated and projected back to the original model dimension:\n\n$$\\text{MultiHead}(Q, K, V) = \\text{Concat}(\\text{head}_1, \\dots, \\text{head}_h)W^O$$\n\n$$\\text{where } \\text{head}_i = \\text{Attention}(QW_i^Q, KW_i^K, VW_i^V)$$\n\nOne significant tension in this architecture is the trade-off between the number of heads and the depth of the representation in each head. Increasing the number of heads provides more \"points of view\" but limits the complexity of the features each head can learn. Furthermore, during inference, the memory required to store the Keys and Values (the KV cache) scales with the number of heads, which has led to modern optimizations like Grouped-Query Attention (GQA). GQA reduces the memory footprint by sharing a single set of Keys and Values across multiple Query heads, striking a balance between the richness of multi-head attention and the efficiency of multi-query designs.\n\nBeyond simple feature extraction, multi-head attention acts as a dynamic routing mechanism. It determines not just what a token means, but which other parts of the sequence should inform that meaning in a given context. This move from a single \"average\" focus to a distributed committee of observers is what allows Transformers to handle the inherent ambiguity of natural language.\n\nTo better understand how this fits into the broader architecture, it is worth looking at the specific operations within each head and how they are eventually integrated back into the network.","timestamp":"2026-07-23T14:16:21.380Z","source":"ai"},{"id":"msg-1784816243658-m83bpu0","role":"user","content":"Explain the selected node(s).","timestamp":"2026-07-23T14:17:23.658Z","source":"ai"},{"id":"msg-1784816247630-p1ckh48","role":"assistant","content":"Grouped-Query Attention (GQA) is an architectural optimization designed to break the linear scaling relationship between the number of attention heads and the memory required for the Key-Value (KV) cache. In standard multi-head attention, every query head has a corresponding key head and value head. As sequence lengths grow, the memory needed to store these $K$ and $V$ tensors for every token becomes the primary bottleneck for inference, often limiting the maximum context window or the batch size a GPU can handle.\n\nGQA introduces a many-to-one relationship where multiple query heads share a single \"group\" of key and value heads. It sits as a middle ground between two extremes:\n1.  **Multi-Head Attention (MHA):** Every Query head has its own Key and Value head ($H$ queries, $H$ keys/values). This is rich in representational power but expensive in memory.\n2.  **Multi-Query Attention (MQA):** All Query heads share a single Key and Value head ($H$ queries, 1 key/value). This is highly efficient but can lead to a \"bottleneck\" where the model struggles to attend to multiple types of information simultaneously.\n\nBy grouping query heads (for example, eight query heads sharing one key/value pair), GQA maintains the diverse representational \"perspectives\" of the multi-head approach while drastically reducing the KV cache size. This is particularly critical for the performance of large language models during autoregressive decoding. Because the model only needs to load a fraction of the key/value data from memory into the processor for each new token, it achieves significantly higher throughput without the significant quality degradation often seen in the more aggressive MQA.\n\nIn the context of the Transformer block, GQA represents a shift from optimizing purely for theoretical expressivity toward optimizing for the physical realities of hardware—specifically memory bandwidth. It acknowledges that while we want many \"queries\" to ask different questions of the text, the \"keys\" and \"values\" (the database of the context) can be compressed through sharing without losing the essential structural relationships defined by the scaled dot-product operation.","timestamp":"2026-07-23T14:17:27.630Z","source":"ai"}],"createdAt":"2026-07-23T14:16:15.383Z","updatedAt":"2026-07-23T14:17:27.631Z"},{"nodeId":"n-1784-816-030-709","messages":[{"id":"msg-1784816292351-2uub569","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T14:18:12.351Z","source":"ai"},{"id":"msg-1784816298932-mzz337n","role":"assistant","content":"The scaled dot-product attention mechanism is the engine of the Transformer, transforming static embeddings into dynamic, context-aware representations. Its elegance lies in how it treats information retrieval as a differentiable process of matching a specific need (the Query) against available information (the Keys) to extract the relevant content (the Values).\n\n### The Mechanics of Comparison\nThe core operation, $QK^T$, computes the raw similarity between every pair of tokens in a sequence. By projecting the input into distinct $Q$, $K$, and $V$ spaces, the model decouples what a token is looking for from what it offers to others. This allows a token to \"attend\" to a specific linguistic feature—such as a verb looking for its subject—without being restricted by its own identity.\n\nThe dot product measures alignment: the more similar the Query vector is to a Key vector, the higher the resulting score. However, as the dimensionality $d_k$ of these vectors increases, the magnitude of the dot products tends to grow. This can push the softmax function into regions where gradients are vanishingly small, making training difficult. The scaling factor $\\frac{1}{\\sqrt{d_k}}$ is the critical stabilization term that counteracts this effect, ensuring the scores stay within a range where the model can continue to learn effectively.\n\n### The Softmax Bottleneck\nApplying the softmax function turns these scaled scores into a probability distribution. This creates a \"soft\" selection mechanism: instead of picking just one relevant token, the model creates a weighted sum of all Values. This allows for the integration of nuance—a word can derive meaning from several different parts of a sentence simultaneously. \n\nA significant trade-off in this design is the quadratic complexity relative to sequence length. Because every token must be compared to every other token, doubling the input length quadruples the computational cost and memory requirements of the attention matrix. This bottleneck is the primary reason why expanding context windows is a major engineering challenge.\n\n### Broadening the Architecture\nWhile the current focus is on the fundamental dot-product operation, the full utility of this mechanism is realized when it is parallelized and contextualized within the larger Transformer block:\n\n*   **Multi-Head Attention:** Rather than performing one large attention operation, the model runs several in parallel. This allows it to attend to different types of relationships (e.g., one head focuses on grammar, another on factual consistency) simultaneously.\n*   **Positional Encoding:** Since the dot-product operation is permutation-invariant—it doesn't care about the order of the keys—positional information must be injected separately so the model knows which words come before others.\n*   **Efficiency Gains:** In large-scale deployment, standard attention is often replaced by **Grouped-Query Attention (GQA)** or **Sparse Attention** to reduce the memory footprint while maintaining the core logic of the scaled dot-product.","timestamp":"2026-07-23T14:18:18.932Z","source":"ai"}],"createdAt":"2026-07-23T14:18:12.354Z","updatedAt":"2026-07-23T14:18:18.933Z"},{"nodeId":"n-1784-816-099-231","messages":[{"id":"msg-1784829198850-66s7q3v","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:53:18.850Z","source":"ai"},{"id":"msg-1784829210245-e3j4gdd","role":"assistant","content":"The Softmax Bottleneck represents a fundamental mathematical ceiling on the expressivity of language models. At its core, the issue is one of matrix rank: the model produces a probability distribution over the vocabulary by multiplying a hidden state vector from the latent space by an embedding matrix. Because the dimension of the latent space—often a few thousand—is significantly smaller than the vocabulary size—often tens or hundreds of thousands—the resulting logit matrix is inherently low-rank. This means there are complex, high-rank probability distributions that the model simply cannot represent, no matter how well it is trained.\n\nThis restriction is particularly acute within the attention mechanism. In scaled dot-product attention, the softmax function converts the compatibility between queries and keys into a set of weights. The scaling factor $\\frac{1}{\\sqrt{d_k}}$ is used to prevent the dot products from growing too large and pushing the softmax into regions with vanishing gradients. However, even with this scaling, the softmax operation acts as a \"hard\" selector that can prematurely collapse the rich, multi-dimensional information in the latent space into a one-dimensional distribution. When the model needs to maintain a nuanced \"superposition\" of several distinct concepts or potential next tokens, the bottleneck forces a choice or a blurring that can result in the loss of structural coherence.\n\nThis representational failure is a direct contributor to hallucinations. When a model encounters a context that requires a high-rank distribution—such as a specific factual retrieval that involves intersecting multiple rare constraints—it may be unable to geometrically \"point\" to the correct token without also assigning unintended probability to phonetically or semantically similar distractors. In these cases, the model doesn't just \"guess\" incorrectly; it is mathematically incapable of isolating the correct answer within the constraints of its final linear transformation.\n\nTo better understand how this is being addressed in contemporary research, it is worth looking at the relationship between this bottleneck and the way we structure the final output layer.\n\n*   **Embeddings & Weight Tying:** The embedding matrix is the other half of the bottleneck. Many models use \"weight tying,\" where the same weights are used for the input embeddings and the output projection. This further constrains the latent space, as the same vectors must serve both as \"inputs\" (capturing raw meaning) and \"outputs\" (distinguishing between candidate tokens).\n*   **Mixture of Softmaxes (MoS):** One of the most effective theoretical counters to the bottleneck is the use of a Mixture of Softmaxes. Instead of a single softmax operation, the model computes several different softmax distributions and takes a weighted average of them. This effectively increases the rank of the output distribution, allowing the model to represent much more complex patterns without increasing the overall dimensionality of the hidden layers.","timestamp":"2026-07-23T17:53:30.245Z","source":"ai"}],"createdAt":"2026-07-23T17:53:18.857Z","updatedAt":"2026-07-23T17:53:30.245Z"},{"nodeId":"n-1784-823-072-505","messages":[{"id":"msg-1784825176951-uqpnjub","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:46:16.951Z","source":"ai"},{"id":"msg-1784825185904-kmf9lfo","role":"assistant","content":"Neuro-symbolic integration addresses the fundamental divide in artificial intelligence: the tension between the fluid, pattern-matching capabilities of connectionism and the rigid, rule-governed structure of symbolic logic. While modern deep learning has mastered high-dimensional perception—turning raw pixels or text into statistical representations—it often lacks the \"systematicity\" required for true Artificial General Intelligence. Systematicity is the ability to understand and compose parts into a whole in a way that is consistent across all possible combinations, a hallmark of human reasoning.\n\nThe pursuit of this integration is largely driven by the limitations of pure statistical models, which can struggle with out-of-distribution generalization. A neural network might learn to solve a specific class of logic puzzles through massive data exposure, but it often fails when the underlying rules are slightly permuted. By embedding symbolic constraints or formal logic into neural architectures, the goal is to create systems that possess both the \"System 1\" intuition of neural networks and the \"System 2\" deliberate reasoning of classical AI.\n\nA critical challenge in this field is the \"symbol grounding problem\": how do discrete, high-level symbols (like \"justice\" or \"is-a-subset-of\") emerge from or connect to the continuous, sub-symbolic activations of a neural network? Contemporary approaches often try to bridge this through differentiable logic, where logical operations are relaxed into continuous mathematical functions that gradient descent can optimize. This allows a model to learn the \"rules of the game\" while maintaining the flexibility to handle noisy, real-world data.\n\nThis integration is also the primary candidate for solving the \"data wall\" problem. Symbolic systems are inherently sample-efficient; once a rule is known, it does not need to be seen ten thousand times to be applied. If a system can autonomously abstract its neural observations into symbolic representations, it could potentially reason its way through novel scenarios without requiring massive new datasets for every task—a prerequisite for the kind of autonomy usually associated with AGI.\n\nTo broaden this exploration, it would be useful to look at the historical conflict between **Symbolic AI** and **Connectionism**, as well as how modern **Reasoning Models** and **Chain-of-Thought** prompting represent a \"soft\" or functional form of neuro-symbolic behavior in current large language models.","timestamp":"2026-07-23T16:46:25.904Z","source":"ai"},{"id":"msg-1784825196567-wott9zh","role":"user","content":"Find existing concepts that belong in this scene.","timestamp":"2026-07-23T16:46:36.567Z","source":"ai"},{"id":"msg-1784825200752-11ar4q9","role":"assistant","content":"To deepen the exploration of **Neuro-Symbolic Integration**, several nodes from the broader graph provide the necessary historical context and technical foundations for how these two paradigms attempt to merge.\n\nIncluding **Symbolic AI** and **Connectionism** establishes the two \"parent\" philosophies being integrated. This clarifies the \"Symbolic\" and \"Neuro\" (connectionist) components of the central node. **Deep Learning** represents the modern evolution of connectionism that has made this integration a priority due to its struggles with transparency and logic. \n\n**Reasoning Models** and **Chain-of-Thought** are essential because they represent the current practical frontier of this field. In many modern systems, neuro-symbolic behavior is achieved not through a hybrid architecture, but through \"verbalized\" reasoning, where the neural network generates symbolic-like steps in natural language to solve complex problems. Finally, **The XOR Problem** serves as a vital historical reference point for the initial failure and eventual resurgence of neural networks when faced with tasks that are fundamentally logical in nature.","timestamp":"2026-07-23T16:46:40.752Z","source":"ai"},{"id":"msg-1784825356270-qz64gcs","role":"user","content":"what is a symbol grounding problem?","timestamp":"2026-07-23T16:49:16.270Z","source":"ai"},{"id":"msg-1784825361607-5xdhz48","role":"assistant","content":"The symbol grounding problem is the challenge of how words, concepts, or \"symbols\" acquire meaning that is actually connected to the physical or perceptual world, rather than just being defined by other symbols.\n\nIf you look up a word in a dictionary, you find other words. If you look those up, you find more words. Without a \"grounding\" in the real world—something that connects the symbol \"apple\" to the actual experience of seeing, touching, or tasting one—the system is just a closed loop of definitions. It knows that \"apple is a fruit,\" but it doesn't know what an apple *is* in any physical sense.\n\nIn the context of **Neuro-Symbolic Integration**, this creates a specific architectural tension:\n\n1.  **The Symbolic Gap:** Classical **Symbolic AI** is excellent at manipulating symbols using logic and rules, but those symbols are \"hollow.\" They only have meaning because human programmers assigned it to them. The system doesn't \"know\" the symbols; it just shuffles them.\n2.  **The Connectionist Solution:** **Connectionism** and **Neural Networks** deal with raw sensory data (pixels, audio waves, token embeddings). They are inherently grounded in data patterns, but they often lack the ability to abstract those patterns into the stable, reusable symbols needed for complex reasoning.\n3.  **The Integration Goal:** The \"holy grail\" is a system where symbols are not hard-coded by humans but are \"learned\" or \"emerge\" from the neural layers. In such a system, the symbol for \"gravity\" would be mathematically linked to the network's internal representation of falling objects, allowing the model to perform high-level logical reasoning that is still tethered to reality.\n\nWithout solving this, we risk creating \"stochastic parrots\"—models that can manipulate language with incredible fluency (symbol manipulation) without any underlying model of the world those symbols represent (grounding). \n\nTo explore how these symbols are structurally handled within networks, it is worth looking at how a **Transformer Block** manages representations through the **Attention Mechanism**, and how the **Latent Space** acts as the \"connective tissue\" where these proto-symbols live.","timestamp":"2026-07-23T16:49:21.607Z","source":"ai"}],"createdAt":"2026-07-23T16:46:16.959Z","updatedAt":"2026-07-23T16:49:21.608Z"},{"nodeId":"n-1784-824-173-383","messages":[{"id":"msg-1784824197795-emzke06","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:29:57.796Z","source":"ai"},{"id":"msg-1784824204733-t6116l8","role":"assistant","content":"David Rumelhart's work marks the definitive pivot from the \"Good Old Fashioned AI\" (GOFAI) of the mid-20th century to the connectionist paradigm that underpins modern deep learning. While early AI was dominated by the manipulation of discrete symbols and logical rules, Rumelhart sought to model cognition through the collective behavior of simple, interconnected units. This shift moved the focus from \"intelligence as logic\" to \"intelligence as pattern recognition and statistical learning.\"\n\nThe core of this contribution is captured in the *Parallel Distributed Processing* (PDP) volumes. The \"distributed\" aspect is particularly crucial: it suggests that knowledge is not stored in a single location or a specific \"neuron,\" but is instead spread across the weights of the entire network. This architecture allows for graceful degradation and \"fuzzy\" processing—the ability to handle noisy or incomplete data that would crash a strictly rule-based system.\n\nTo truly understand Rumelhart's impact, one must look at how he resolved the technical deadlocks of the 1960s. After the critique of the simple perceptron, the field lacked a mathematically sound way to train multi-layer networks. Rumelhart, alongside colleagues like James McClelland, popularized the backpropagation algorithm. This provided a way to calculate the error at the output and \"propagate\" it backward through the hidden layers, adjusting weights via gradient descent. This remains the fundamental mechanism by which almost all current neural networks learn.\n\nHowever, a recurring tension in this area is the \"interpretability\" problem. In a symbolic system, one can read the rules to understand why a decision was made. In a connectionist system, the \"logic\" is buried in millions of numerical parameters. Rumelhart’s work essentially traded the transparency of symbols for the flexibility and power of learning from data—a trade-off that the AI field is still negotiating today as it tries to reconcile these \"black box\" models with the need for safety and explainability.\n\nTo expand this perspective, it is worth exploring the specific psychological motivations behind the PDP movement and the subsequent \"Deep Learning Revolution\" that scaled these ideas to the massive architectures seen today.","timestamp":"2026-07-23T16:30:04.733Z","source":"ai"},{"id":"msg-1784824226905-j58nbeh","role":"user","content":"David Rumelhart","timestamp":"2026-07-23T16:30:26.905Z","source":"note","attachments":[{"id":"att-1784824244296-35cgn7x","type":"image","origin":"retrieved","mimeType":"image/jpeg","name":"DavidRumelhart-IJCNNseattle1991-07-08.jpg","width":512,"height":591,"sourceUrl":"https://upload.wikimedia.org/wikipedia/commons/thumb/7/72/DavidRumelhart-IJCNNseattle1991-07-08.jpg/960px-DavidRumelhart-IJCNNseattle1991-07-08.jpg","fullUrl":"https://upload.wikimedia.org/wikipedia/commons/7/72/DavidRumelhart-IJCNNseattle1991-07-08.jpg","attribution":{"sourceName":"Wikimedia Commons","sourcePageUrl":"https://commons.wikimedia.org/wiki/File:DavidRumelhart-IJCNNseattle1991-07-08.jpg","author":"Rolf Kickuth","license":"CC BY-SA 4.0","licenseUrl":"https://creativecommons.org/licenses/by-sa/4.0","attributionRequired":true}}]}],"createdAt":"2026-07-23T16:29:57.799Z","updatedAt":"2026-07-23T16:30:51.086Z"},{"nodeId":"n-1784-824-174-809","messages":[{"id":"msg-1784824438172-w6a5o60","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:33:58.172Z","source":"ai"},{"id":"msg-1784824444139-oe5854x","role":"assistant","content":"James McClelland’s work represents the bridge between the high-level cognitive architectures of the 20th century and the sub-symbolic, statistical nature of modern artificial intelligence. While traditional AI was busy trying to program logic and rules, McClelland and David Rumelhart shifted the focus toward how knowledge could emerge from the interactions of simple, neuron-like units. This transition moved the field away from the idea of the mind as a \"logic engine\" and toward a view of the mind as a statistical engine.\n\nThe core of this perspective is Parallel Distributed Processing (PDP). In a PDP model, information is not stored in a single \"memory slot.\" Instead, a concept is a pattern of activation across many units, and a single unit may participate in representing many different concepts. This is the foundation of what we now call distributed representations or embeddings. It explains why human cognition is so robust: if you lose a few neurons, or if an input is slightly noisy, the overall pattern—the \"meaning\"—remains largely intact. This property, known as graceful degradation, stands in stark contrast to symbolic systems, where a single missing bracket or typo can cause the entire program to crash.\n\nA crucial tension in McClelland's work involves the \"Regularity vs. Exception\" problem, particularly in language acquisition. For years, the dominant view was that children learn a rule for the past tense (add \"-ed\") and a separate rote-memory list for exceptions (like \"go\" to \"went\"). McClelland and Rumelhart challenged this by demonstrating that a single neural network could learn both the rules and the exceptions simultaneously through gradual weight adjustments. This suggested that \"rules\" might just be an emergent property of learning statistical regularities, a debate that remains at the heart of how we interpret the capabilities of Large Language Models today.\n\nTo further ground this exploration, it is worth looking at the specific implementations and the broader consequences of this school of thought:\n\n*   **NETtalk**: While developed by Terry Sejnowski and Charles Rosenberg, NETtalk is perhaps the most famous contemporary application of the PDP principles championed by McClelland. It demonstrated that a network could learn to map text to phonemes (speech) without being programmed with any linguistic rules, essentially \"learning to speak\" through exposure.\n*   **The XOR Problem**: This was the historical roadblock that nearly killed early neural network research in the 1960s. McClelland and the PDP group were instrumental in popularizing the solution: hidden layers and the backpropagation algorithm, which allowed networks to learn non-linear relationships that simple perceptrons could not.\n*   **Parallel vs. Sequential Processing**: This is the fundamental architectural divide. Traditional computers (and the \"Language of Thought\" hypothesis) are serial—they do one thing at a time very quickly. McClelland argued that biological intelligence is inherently parallel, processing multiple constraints (visual, semantic, syntactic) all at once to arrive at a \"best fit\" interpretation of the world.\n\n### Suggested Actions","timestamp":"2026-07-23T16:34:04.139Z","source":"ai"},{"id":"msg-1784824485183-v0o5q39","role":"user","content":"James McClelland","timestamp":"2026-07-23T16:34:45.183Z","source":"note","attachments":[{"id":"att-1784824483238-qef1muq","type":"image","origin":"retrieved","mimeType":"image/jpeg","name":"James McClelland (14039773375).jpg","width":512,"height":683,"sourceUrl":"https://upload.wikimedia.org/wikipedia/commons/thumb/1/11/James_McClelland_%2814039773375%29.jpg/960px-James_McClelland_%2814039773375%29.jpg","fullUrl":"https://upload.wikimedia.org/wikipedia/commons/1/11/James_McClelland_%2814039773375%29.jpg","attribution":{"sourceName":"Wikimedia Commons","sourcePageUrl":"https://commons.wikimedia.org/wiki/File:James_McClelland_(14039773375).jpg","author":"Jussi Puikkonen/KNAWKNAW Koninklijke Nederlandse Akademie van Wetenschappen","license":"CC BY 2.0","licenseUrl":"https://creativecommons.org/licenses/by/2.0","attributionRequired":true}}]}],"createdAt":"2026-07-23T16:33:58.175Z","updatedAt":"2026-07-23T16:34:45.190Z"},{"nodeId":"n-1784-824-369-059","messages":[{"id":"msg-1784824543410-h4822iw","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:35:43.410Z","source":"ai"},{"id":"msg-1784824549511-cuik26r","role":"assistant","content":"NETtalk represents a watershed moment in the history of connectionism, demonstrating that complex human behaviors like reading aloud could emerge from simple, distributed processing units rather than explicit rule-based programming. Created by Terry Sejnowski and Charles Rosenberg in the mid-1980s, it provided a powerful counter-argument to the then-dominant symbolic AI paradigm, which struggled with the \"brittleness\" of hand-coded phonetic rules.\n\nThe system's architecture—comprising an input layer (text), a hidden layer, and an output layer (phonemes)—was trained using backpropagation. While James McClelland and David Rumelhart were formalizing the Parallel Distributed Processing (PDP) framework, NETtalk served as its most famous \"proof of concept.\" It didn't just map letters to sounds; it captured the nuances of pronunciation, such as how the \"c\" in \"cat\" differs from the \"c\" in \"cell,\" purely through statistical regularities in the training data.\n\nOne of the most profound aspects of NETtalk was its \"learning curve,\" which famously mimicked human development. In its early stages of training, the system produced a rhythmic babbling; it then began to distinguish between vowels and consonants, before eventually mastering the irregularities of the English language. This progression suggested that the \"stages\" of child development might not be the result of a pre-programmed biological clock, but rather the natural emergent properties of a neural system absorbing complex information.\n\nHowever, NETtalk also highlighted the \"black box\" problem that persists in modern deep learning. While the system performed remarkably well, looking at the weights of the hidden layer didn't immediately reveal *how* it was making decisions. This shift from \"representation as symbols\" to \"representation as vectors\" (or weights) meant that while the machine became more capable and flexible, it became less transparent to human inspection.\n\nThis tension between the elegance of rule-based systems and the messy efficiency of neural networks is still at the heart of AI today. NETtalk's success paved the way for the eventual transition from specialized linguistic models to the generalized architectures of today, proving that if you give a network enough data and the right learning algorithm, it can \"discover\" the rules of language on its own.\n\nTo deepen the current exploration, it would be useful to look at the specific theoretical framework that supported this work and the specific problem that neural networks of this era were finally beginning to solve.","timestamp":"2026-07-23T16:35:49.511Z","source":"ai"}],"createdAt":"2026-07-23T16:35:43.417Z","updatedAt":"2026-07-23T16:35:49.512Z"},{"nodeId":"n-1784-825-371-425","messages":[{"id":"msg-1784825449245-qstxhkz","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:50:49.245Z","source":"ai"},{"id":"msg-1784825455370-6keplmt","role":"assistant","content":"The symbol grounding problem addresses the fundamental gap between formal systems and meaningful understanding. At its core is the question of how bits, tokens, or symbols—which are inherently arbitrary and manipulate-able according to syntactic rules—acquire semantic \"hook-up\" to the physical world. Without grounding, a system is like a person trying to learn Chinese using only a Chinese-to-Chinese dictionary; they can navigate the definitions indefinitely without ever knowing what a single word actually refers to in reality.\n\nNeuro-symbolic integration represents a primary architectural response to this dilemma. It attempts to marry the strengths of two traditionally opposing paradigms:\n\n1.  **Symbolic AI**, which excels at high-level reasoning, logic, and transparency but suffers from \"brittleness\" because its symbols are typically hand-coded by humans rather than learned from the environment.\n2.  **Connectionism (Neural Networks)**, which excels at pattern recognition and sensory processing—effectively \"grounding\" signals from pixels or audio into latent representations—but often lacks the structured logic required for complex, multi-step reasoning.\n\nThe tension in this integration lies in the translation layer. Even if a neural network can recognize a \"cat\" from raw sensory data, the resulting internal representation is often a high-dimensional vector in a latent space. Mapping that vector to a discrete symbol like `CAT` in a logical system requires a bridge that preserves the richness of the sensory experience while enabling the precision of formal logic.\n\nA critical point of departure for further exploration is **Moravec’s Paradox**, which notes that high-level reasoning (the symbolic side) requires very little computation, while low-level sensorimotor skills (the grounding side) require enormous computational resources. This suggests that the difficulty of grounding isn't just a philosophical hurdle but a biological and computational one. \n\nFurthermore, the debate over whether modern Large Language Models (LLMs) have solved this problem is active. While they operate on symbols (tokens), they are trained on vast amounts of human-grounded data. This raises the question of whether \"borrowed grounding\"—extracting meaning from the way humans use language—is equivalent to \"intrinsic grounding,\" where the agent interacts with the world itself. \n\nTo deepen this investigation, it would be useful to contrast the **Stochastic Parrot View**, which argues that no amount of token manipulation results in true grounding, against the **World-Model View**, which suggests that predicting the next symbol requires the internal construction of the physical or conceptual reality that generated those symbols.","timestamp":"2026-07-23T16:50:55.370Z","source":"ai"}],"createdAt":"2026-07-23T16:50:49.250Z","updatedAt":"2026-07-23T16:50:55.371Z"},{"nodeId":"n-1784-832-202-073","messages":[{"id":"msg-1784832645561-55ngdav","role":"user","content":"Describe main generations of GPT family","timestamp":"2026-07-23T18:50:45.561Z","source":"ai"},{"id":"msg-1784832651536-pie48zl","role":"assistant","content":"The GPT (Generative Pre-trained Transformer) family represents the shift from experimental language modeling to the current era of ubiquitous AI assistants. Developed by OpenAI, these models evolved from a simple proof-of-concept for unsupervised learning into a sophisticated ecosystem of tiered capabilities.\n\n### The Foundation: GPT-1 to GPT-3\nThe lineage began with **GPT-1** (2018), which demonstrated that a model could learn world knowledge by simply predicting the next word in a massive corpus of text—a process known as **Pre-training**. **GPT-2** (2019) scaled this approach significantly, showing that \"zero-shot\" task performance emerged naturally as the model grew. \n\n**GPT-3** (2020) was the watershed moment. With 175 billion parameters, it proved the **Scaling Laws**: that increasing compute, data, and model size yields predictable improvements in intelligence. It introduced the world to **Few-Shot Prompting**, where the model could learn a new task from just a few examples in the prompt without any weight updates.\n\n### The Convergence: GPT-4 and the Multimodal Era\nThe release of **GPT-4** (2023) moved the family toward human-level performance on professional and academic benchmarks. It shifted the architecture toward a **Mixture of Experts** (MoE) design and introduced multimodality, allowing the model to \"see\" and process images. To make these models useful as assistants, OpenAI refined them through **Post-training** techniques like **RLHF** (Reinforcement Learning from Human Feedback), which aligns the model's output with human intent and safety guidelines.\n\n### The Reasoning Era: The o1 Series\nThe latest evolution is the **o1** series (formerly \"Strawberry\"). Unlike previous generations that rely primarily on fast, intuitive **Next-Token Prediction**, o1 utilizes **Inference-Time Search** (or \"Chain of Thought\" processing). This allows the model to \"think\" before it speaks, breaking down complex logic, math, and coding problems into smaller steps. This shift marks the transition from models that are simply fluent to models that are capable of deep, deliberate reasoning.\n\n### Current Product Tiers\nToday, the family is structured into specific roles within the product strategy:\n*   **Flagship (GPT-4o, o1):** High-intelligence models for complex reasoning and multimodal interaction.\n*   **Small/Efficient (GPT-4o-mini):** Optimized for speed and low cost, replacing the older GPT-3.5 Turbo.\n\nTo better understand how these models are built and why they behave the way they do, we should look at the training techniques and architectural choices that define them.","timestamp":"2026-07-23T18:50:51.536Z","source":"ai"},{"id":"msg-1784832689602-k4ium7x","role":"user","content":"search internet and create nodes for the recent generation of GPT models (sol, terra, luna)?","timestamp":"2026-07-23T18:51:29.602Z","source":"ai"},{"id":"msg-1784832697006-f7niiuu","role":"assistant","content":"The most recent generation of the GPT family, released in July 2026, is the **GPT-5.6 series**. This release marks a departure from previous naming conventions (like \"mini\" or \"turbo\") in favor of durable, celestial-themed capability tiers: **Sol**, **Terra**, and **Luna**.\n\nThis tiered approach allows OpenAI to update specific performance classes independently while the \"5.6\" identifier tracks the underlying architectural generation.\n\n### The GPT-5.6 Tiers\n*   **GPT-5.6 Sol (The Flagship):** Sol is the high-intelligence model designed for the most demanding tasks, including complex reasoning, scientific research, and long-horizon agentic workflows. It currently sets the state-of-the-art on benchmarks like the Artificial Analysis Coding Agent Index.\n*   **GPT-5.6 Terra (The Balanced):** Terra is positioned as the \"daily driver\" for enterprise and professional use. It is designed to match the intelligence of previous flagship models (like GPT-5.5) but at roughly half the cost, making it the default choice for large-scale document analysis and customer support.\n*   **GPT-5.6 Luna (The Fast):** Luna is the high-efficiency, low-latency tier. It is optimized for speed and cost-sensitive, high-volume workloads such as routine drafting and simple automation, while still benefiting from the reasoning advancements of the 5.6 generation.\n\n### Comparison and Capabilities\nAll three models in this series share a massive **1-million-token context window** and a knowledge cutoff of February 2026. The primary differentiation lies in the \"reasoning effort\" and price-to-performance ratio.\n\n| Model | Role | Best For | Price (Input/Output per 1M) |\n| :--- | :--- | :--- | :--- |\n| **Sol** | Peak Reasoning | Complex coding, R&D, Cybersecurity | \\$5.00 / \\$30.00 |\n| **Terra** | Balanced Reliability | Business workflows, Coding assistance | \\$2.50 / \\$15.00 |\n| **Luna** | Efficiency/Speed | Summarization, High-volume API calls | \\$1.00 / \\$6.00 |\n\nTo integrate these into your map, we can connect them to the broader product strategy of tiered capability.","timestamp":"2026-07-23T18:51:37.006Z","source":"ai"}],"createdAt":"2026-07-23T18:50:45.569Z","updatedAt":"2026-07-23T18:51:37.007Z"},{"nodeId":"n-1784-832-210-084","messages":[{"id":"msg-1784832377205-mrke382","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:46:17.205Z","source":"ai"},{"id":"msg-1784832382962-ymxzpo4","role":"assistant","content":"The Claude family represents a distinct approach to the \"frontier model\" problem, characterized by a prioritisation of \"Constitutional AI.\" While other labs often rely heavily on Reinforcement Learning from Human Feedback (RLHF) to align models with human preferences, the Claude models are trained against a set of written principles—a constitution—that they use to self-evaluate and refine their own responses. This architectural choice is designed to move away from the \"sycophancy\" often found in RLHF, where a model might tell a user what they want to hear rather than what is objectively true or safe.\n\nThe tiered structure of the family—Haiku, Sonnet, and Opus—is a response to the fundamental trade-off between intelligence and latency.\n*   **Haiku** is optimized for near-instantaneous processing and high-volume tasks, sitting at the edge of the cost-efficiency curve.\n*   **Sonnet** functions as the \"balanced\" workhorse, often serving as the primary vehicle for multimodal updates and enterprise-grade performance.\n*   **Opus** is the flagship, designed for high-order reasoning and complex synthesis, where the depth of the \"latent space\" is more critical than the speed of the output.\n\nA core strength of this family is its emphasis on long-context performance. Anthropic pioneered the 100k+ token window, focusing on high \"recall\"—the ability to retrieve specific facts from a massive document without suffering from the \"lost in the middle\" phenomenon common in earlier transformer architectures. This makes the models particularly suited for \"needle-in-a-haystack\" testing, which measures how well a model can find a single piece of information buried in a dense context.\n\nA live tension in the development of these models is the balance between \"helpfulness\" and \"harmlessness.\" Because of their rigorous safety training, earlier versions of the Claude family were occasionally criticized for being overly \"preachy\" or refusing benign prompts. The evolution of the family has centered on making the models more nuanced—ensuring they can recognize the difference between a harmful request and a complex, challenging, but ultimately safe inquiry. This leads to a deeper question for future iterations: as models become more capable of reasoning about their own \"constitutions,\" at what point does \"alignment\" transition from following a script to exercising genuine ethical judgment?\n\nTo better understand where the Claude family sits within the broader ecosystem, it is worth looking at the specific organizations and underlying technologies that define its peer group and architectural foundations.","timestamp":"2026-07-23T18:46:22.962Z","source":"ai"},{"id":"msg-1784832387335-x7o9rhm","role":"user","content":"Suggest new concepts to add to my knowledge graph.","timestamp":"2026-07-23T18:46:27.335Z","source":"ai"},{"id":"msg-1784832391655-1uq2q7b","role":"assistant","content":"The development of the Claude family is defined by several specific technical and philosophical frameworks that differentiate it from other frontier model lineages. To deepen the exploration of these models, it is worth looking at the specific mechanisms Anthropic uses to achieve their characteristic \"nuance\" and safety profile.\n\nOne of the most significant is **Constitutional AI**, the supervised learning and reinforcement learning framework where the model is trained to follow a literal set of rules (a \"constitution\") rather than just imitating human preferences. This is closely tied to the concept of **Red Teaming**, a rigorous process of adversarial testing where researchers actively try to bypass safety filters to find vulnerabilities.\n\nTechnically, the Claude models are often cited in the context of **Mechanistic Interpretability**, a field of research Anthropic heavily invests in to look inside the \"black box\" of neural networks and identify specific circuits or features that correspond to concepts like \"honesty\" or \"deception.\" Additionally, the way these models manage massive amounts of information without losing track of details is a result of advanced **Context Window Management**, which involves specific optimizations in how the transformer handles long-range dependencies during inference.\n\nThe following concepts represent these key pillars of the Claude development philosophy:","timestamp":"2026-07-23T18:46:31.655Z","source":"ai"},{"id":"msg-1784832411909-y5cput9","role":"user","content":"Create nodes for main Claude family members","timestamp":"2026-07-23T18:46:51.909Z","source":"ai"},{"id":"msg-1784832415513-zxn5a7u","role":"assistant","content":"To build out the specific capabilities and identities within the Claude Family, we should add the individual models that define the current three-tier strategy. Each model represents a specific point on the Pareto frontier of cost, speed, and intelligence.\n\n*   **Claude 3.5 Sonnet** is currently the most significant node in the family, as it often serves as the \"frontier\" for new architectural improvements and multimodal capabilities, frequently outperforming the previous generation's flagship in coding and reasoning tasks.\n*   **Claude 3 Opus** remains the heavy-lifting reasoning engine, designed for high-stakes synthesis and complex creative work where the sheer depth of the model's internal representations is more important than token-per-second speed.\n*   **Claude 3.5 Haiku** represents the efficiency layer, providing a high level of intelligence at a low cost-to-serve, crucial for agentic workflows and high-volume data extraction.\n\nThese additions allow for a more granular discussion of how Anthropic manages the trade-offs between \"Flagship\" reasoning and \"Haiku\" speed.","timestamp":"2026-07-23T18:46:55.513Z","source":"ai"},{"id":"msg-1784832497164-odnlcxl","role":"user","content":"Explain the selected node: “Claude Fable”.","timestamp":"2026-07-23T18:48:17.164Z","source":"ai"},{"id":"msg-1784832501109-p7iojiq","role":"assistant","content":"**Claude Fable** is an outlier in the Claude ecosystem, as it does not exist as a released or officially announced model within Anthropic’s public hierarchy. While the established family—Haiku, Sonnet, and Opus—follows a clear naming convention based on poetic forms (ascending in complexity from the short, constrained Haiku to the elaborate, high-intelligence Opus), \"Fable\" would represent a departure from this structure.\n\nIn the context of the current family, its presence raises an interesting question about the future naming or categorization of specialized models. If one were to speculate on a \"Fable\" model based on its namesake, it might suggest a model optimized for narrative, storytelling, or perhaps **synthetic data generation**—creating \"fables\" to train other models. However, within the actual product strategy of Anthropic, there is currently no tier between Sonnet and Opus, nor a specialized creative branch with this name.\n\nIts inclusion alongside the three standard tiers highlights the potential for \"hallucinated\" or speculative entities within a knowledge graph. While Haiku, Sonnet, and Opus are defined by their specific trade-offs in the **Inference-Compute Trade-off** (speed vs. reasoning depth), \"Fable\" lacks a defined technical profile. Unless it represents a private internal project or a very recent experimental leak, it acts as a placeholder for a category that doesn't yet exist in the official \"Constitutional AI\" lineage. \n\nReviewing this node's placement is useful for distinguishing between the **model families** as they are commercially deployed and the theoretical or speculative models that often circulate in AI discourse. If the goal is to map the actual architectural tiers of Anthropic’s current offerings, \"Fable\" likely belongs in a \"speculative\" or \"proposed\" category rather than being connected directly as a sibling to the established tiers.","timestamp":"2026-07-23T18:48:21.109Z","source":"ai"},{"id":"msg-1784832542910-p1grqef","role":"user","content":"Search internet and find out about Claude Fable","timestamp":"2026-07-23T18:49:02.910Z","source":"ai"},{"id":"msg-1784832550572-0j0pxu8","role":"assistant","content":"**Claude Fable 5** is the public-facing version of Anthropic's \"Mythos-class\" intelligence tier, which was released in June 2026. It represents a significant expansion of the Claude family, as it is the first model to sit above the \"Opus\" flagship level.\n\nThe model represents a shift in both capability and safety strategy for the Claude family:\n\n*   **The Mythos Tier:** Fable 5 is part of the Mythos class, a category of models that Anthropic initially held back due to their advanced autonomous and offensive capabilities. It is the same underlying model as **Claude Mythos 5**, but with a crucial difference: Fable 5 includes real-time **safety classifiers** that intercept high-risk requests (particularly in cybersecurity and biological research) and route them to a less capable model, such as Claude Opus 4.8.\n*   **Agentic Capabilities:** While the Haiku and Sonnet tiers focus on speed and daily utility, Fable 5 is designed for \"long-horizon\" agentic work. It is integrated into tools like **Claude Code** and **Claude Cowork**, where it is capable of multi-day autonomous sessions, writing its own tests, and delegating tasks to sub-agents.\n*   **Technical Specs:** Fable 5 features a **1-million-token context window** and a maximum output of 128k tokens. It is priced as a premium tier, costing roughly double the price of the Opus models ($10/M input, $50/M output).\n*   **Project Glasswing:** The development of the Fable/Mythos tier was closely tied to a government-vetted partnership called Project Glasswing. This program allowed trusted organizations to use the uncensored Mythos 5 for defensive cybersecurity work, such as patching critical infrastructure vulnerabilities before they could be exploited.\n\nIn the context of the family's poetic naming convention, Fable 5 signals that Anthropic is moving beyond the standard three-tier ladder (Haiku, Sonnet, Opus) to accommodate models that possess specialized, frontier-level reasoning that requires more aggressive architectural safeguards.","timestamp":"2026-07-23T18:49:10.572Z","source":"ai"}],"createdAt":"2026-07-23T18:46:17.215Z","updatedAt":"2026-07-23T18:49:10.572Z"},{"nodeId":"n-mermaid-mrxl4f8r-1","messages":[{"id":"msg-mrxl4f9k-0-0-1qrxz","role":"assistant","content":"Something genuinely unusual happened in artificial intelligence, and it is worth stating plainly, because the daily noise of product launches and valuations obscures it. For most of computing history, software did what programmers specified, step by step. Around the middle of the last decade, a different approach — training enormous statistical models on enormous amounts of data and letting capability *emerge* rather than be engineered — went from academic curiosity to the dominant force in technology. The systems it produced can write, reason, code, see, and converse at a level no one credibly predicted a decade ago. Trillions of dollars, the strategies of every major technology company, the attention of governments, and a genuinely open set of scientific questions now orbit this one development. Understanding it properly — not at the level of headlines, but at the level where the headlines start making sense — requires seeing it from several directions at once. That is what this essay, and the workspace behind it, attempts.\n\n### The science: why any of this works\n\nStart with the science, because everything else rests on it. Modern AI is **machine learning**: instead of writing rules, you define a goal, show a system millions of examples, and let an optimization process adjust the system's internal parameters until it performs. The systems in question are **neural networks** — vast webs of simple numerical units, less like brains than like adjustable functions of extraordinary size. The decisive architecture arrived in 2017 in a Google research paper: the **transformer**, a design whose core trick, *attention*, lets a model weigh every part of its input against every other part, and whose deeper advantage was economic — it could be trained in parallel across as many chips as anyone could afford to buy. That mattered because of the second scientific pillar: the discovery that capability scales *predictably* with compute, data, and model size. These **scaling laws** turned AI progress from a research gamble into an investable engineering roadmap, and they are the single most important fact for understanding why the money behaves the way it does. The third pillar is stranger: as models scaled, abilities appeared that nobody trained for — arithmetic, translation, learning new tasks from a handful of examples in the prompt. **Emergence** is why the field's own researchers are routinely surprised by their own creations, and why a serious scientific program now exists to look *inside* trained models and figure out what they actually learned. That program — **interpretability** — and its sibling, **alignment**, the problem of making models reliably pursue what their operators intend, are not regulatory afterthoughts; they are among the hardest open problems in the science itself.\n\n### The factory: how a model is made\n\nScience explains why it works; it does not build anything. Between an idea and a deployed assistant stands an industrial process with distinct stages, each with its own vocabulary. **Pre-training** is the heavy industry: months of datacenter time spent on a deceptively simple objective — predict the next token of text — applied to a large fraction of everything humanity has written. The result is a **base model**: staggeringly knowledgeable, entirely unmannered, no more an assistant than a library is a librarian. **Post-training** supplies the manners: fine-tuning on demonstrations of good behavior, then **reinforcement learning from human feedback** — people comparing model outputs, a reward model learning their preferences — which is the technique that made ChatGPT feel cooperative rather than raw. The newest variant, reinforcement learning from *verifiable* rewards, trains models on tasks whose answers can be checked — mathematics, code that must pass tests — and is the engine behind the current generation of \"reasoning\" models that deliberate before answering. Then comes **inference**: actually serving the model, where the economics invert. Training cost is paid once; inference cost is paid on every query, *forever*, which is why an entire engineering discipline — quantization, distillation, sparse architectures — exists to make each answer cheaper, and why the price of machine intelligence has fallen orders of magnitude in three years even as capability rose.\n\n### The toolbox: what the systems can do\n\nWhat do the resulting systems do? The honest answer is: an unreasonable amount, unevenly. The **large language model** is the era's central artifact — one system that drafts, summarizes, translates, codes, and converses, its generality still the deepest surprise in the field. Around it, capabilities have compounded: models that see and generate images and video (mostly via **diffusion**, a second generative recipe entirely); models that think step by step through hard problems; and, in the current wave, **agents** — systems that don't just answer but *act*, calling tools, browsing, writing files, iterating on failures. Coding assistants deserve special mention: they are the first AI application with undeniable enterprise-wide productivity impact, and software development has become the proving ground where every agentic technique is tested first, because code can be run and checked. Yet the capabilities are jagged — superhuman on some tasks, unreliable on others, and prone to **hallucination**: the confident generation of falsehoods, which remains the biggest obstacle to trusting these systems in high-stakes work. Even *measuring* what models can do has become its own contested discipline, as benchmarks saturate and every launch headline demands skeptical reading.\n\n### The arena: who is building it\n\nNone of this unfolds in a laboratory vacuum. It unfolds as a race — arguably the most capital-intensive competition in industrial history. A handful of **frontier labs** set the tempo, split by real strategic and philosophical differences: closed models sold through APIs versus **open weights** released for anyone to download, safety-first postures versus speed-first ones, and, cutting across everything, the rivalry between American labs and Chinese ones whose efficient open models rewrote assumptions about what frontier capability costs. Beneath the labs sits the hardware layer, more concentrated still: *one company's* GPUs train essentially everything, *one Taiwanese foundry* fabricates essentially every advanced chip — a supply chain so narrow it has become an instrument of geopolitics, with **export controls** on chips now a standing feature of US–China relations. Above the labs, the platform giants weave models into products used by billions, and an application layer of startups tests whether durable businesses can be built on top of someone else's foundation model. The people running these organizations — a cast small enough to know by name — matter more than executives usually do, because the field is young enough that individual convictions about safety, openness, and speed still translate directly into strategy.\n\n### The money: a gold rush by any measure\n\nThe money deserves its own lens, because its scale has made AI a macroeconomic phenomenon rather than a sector story. **Datacenter** construction is measured in gigawatts and tens of billions of dollars; *electricity* has become the binding physical constraint, dragging utility deals and nuclear restarts onto the industry's critical path; hyperscaler capital expenditure has reached levels usually associated with national infrastructure programs. All of it is, at bottom, a bet on scaling laws continuing to pay. Whether that bet is rational anticipation of transformative value or the largest capital misallocation since the dot-com era is a question serious people answer in opposite ways — and this workspace maps the argument rather than settling it.\n\n### The stakes: what it means for everyone else\n\nBeyond the industry entirely lie the questions AI poses to everyone else. Governments are writing the first comprehensive **AI laws** while openly unsure what they are regulating. *Work* is changing from the top down — the old prediction that machines would take manual labor first has inverted, with coding, writing, and analysis feeling the pressure earliest — though the honest evidence shows real productivity gains and, so far, no mass displacement, and the gap between those facts is where the policy fight lives. The *information environment* is absorbing synthetic text, voices, and video that are cheap and convincing, with the deepest damage perhaps being the **liar's dividend**: when anything can be faked, everything can be denied. And, most intimately, millions of people now have daily *relationships* with AI systems — as tutors, confidants, companions — a development spanning genuine comfort and documented harm, whose norms are being invented in real time.\n\n### The long road: how we got here\n\nIt helps to know that none of this arrived from nowhere. The field is seventy years old and has been declared imminent and dead several times each. Its founding schism — intelligence as hand-written *rules* versus intelligence as learned *networks* — ran for decades, with the learning school kept alive through funding winters by a small group of researchers who turned out to be right. The modern era has precise landmarks: the 2012 **ImageNet** result that vindicated deep learning, **AlphaGo**'s 2016 victory that made AI a matter of national strategy, the 2017 transformer paper, the **ChatGPT** release of late 2022 that turned a research field into a societal event, and the reasoning-model wave that opened the current chapter. History here is not decoration; the **AI winters** are the reference point for every bubble debate, and the schism still echoes in today's arguments about whether current methods are enough.\n\n### The open questions: where the experts disagree\n\nWhich brings up the most intellectually honest thing that can be said about this subject: its biggest questions are genuinely open, and the experts who built the field disagree about them *in public*. Is the technology safe, and can it be made so? Do these models understand anything, or are they extraordinary mimics — and would we know the difference? Why exactly do capabilities keep growing, and will scaling hold? Is human-level breadth — **AGI**, a term doing more marketing work than scientific work — years away, decades away, or ill-posed? Will the training data run out? Is the boom a bubble? These are not gaps in the reader's knowledge to be embarrassed about; they are gaps in *everyone's* knowledge, and the fault lines along which the field's founders — Nobel laureates among them — publicly split.\n\n### Why this is a graph\n\nA subject shaped like this cannot be honestly presented as a sequence of chapters, because its concepts refuse to stay in one chapter. The transformer is simultaneously a scientific idea, the thing the factories build, the source of a trillion-dollar valuation, and a historical event with a date. Safety is a research problem, an engineering practice, a body of law, and a raging argument — *at once*. That is why this workspace is a graph: each of the perspectives sketched above — the science, the factory, the toolbox of capabilities, the competitive arena, the gold rush, the societal stakes, the long road here, and the burning questions — is a door into the same connected whole, and the important concepts sit where the paths cross. Enter through whichever question brought you, follow what pulls you, and read the notes as you go: the map's purpose is not to be memorized but to make the next AI headline you read snap into a structure you now own.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-10","messages":[{"id":"msg-mrxl4f9k-9-0-totju","role":"assistant","content":"The discipline of building systems that improve from data rather than following hand-written rules. Everything in modern AI is machine learning; the recent revolution is a change in scale and method, not in the basic premise. Its core notions — learning paradigms, loss, gradients, generalization — remain the working vocabulary of the field.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784828362345-d1rj0jq","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:39:22.345Z","source":"ai"},{"id":"msg-1784828375098-n8ucosx","role":"assistant","content":"Machine Learning is fundamentally a search for a function that can map inputs to outputs without an explicit procedural recipe. This search is guided by the objective of minimizing a **Loss Function**, which serves as a mathematical proxy for \"error.\" If the Loss Function defines the rugged terrain of a problem, **Gradient Descent** is the process of navigating that terrain, iteratively updating parameters in the direction of steepest descent:\n\n$$\\theta_{t+1} = \\theta_t - \\eta \\nabla L(\\theta_t)$$\n\nThe effectiveness of this search depends heavily on the paradigm of feedback available. In **Supervised Learning**, the signal is high-resolution—a direct comparison against a ground-truth label. **Unsupervised Learning** lacks this external reference, forcing the system to find internal coherence or \"natural\" groupings within the data. **Reinforcement Learning** occupies a more complex middle ground, where the feedback (a reward signal) is often sparse and delayed, requiring the agent to solve the \"credit assignment problem\"—determining which specific action in a long sequence was responsible for the eventual outcome.\n\n**Neural Networks** have become the dominant architecture for this search because of their ability to approximate nearly any continuous function. The engine of their success is **Backpropagation**, an elegant application of the chain rule from calculus that distributes the error signal from the Loss Function back through every layer of the network. This allows for the simultaneous adjustment of millions of connection weights, effectively telling each unit how it needs to change to improve the overall prediction.\n\nHowever, the ultimate goal is not to minimize error on the data we already have, but to achieve **Generalization**: the ability to perform on data the model has never encountered. This creates a persistent tension with **Overfitting**. A model with enough capacity (parameters) can easily memorize the training data, essentially \"cheating\" by learning the noise and idiosyncratic details of the specific sample rather than the underlying pattern. Managing this tradeoff—ensuring the model is complex enough to learn the signal but constrained enough to ignore the noise—is the central challenge of model training.\n\nModern practice has moved away from the \"blank slate\" approach through **Transfer Learning**. Instead of training a model from scratch for every new problem, we leverage representations already learned from massive, general datasets. This acknowledges that the low-level features of data—such as edges in images or syntax in language—are often universal. By starting with a pre-trained foundation, we significantly reduce the \"compute cost\" and data requirements for specialized tasks, shifting the focus from raw learning to the refinement of existing intelligence.\n\nWhile the current scientific foundations are robust, a critical missing link in understanding why these systems work so well is the bridge between the architecture of a single network and the massive scale of contemporary \"Deep Learning.\" The transition from classical machine learning to deep learning isn't just about adding more layers; it involves a qualitative shift in how features are automatically extracted rather than manually engineered.","timestamp":"2026-07-23T17:39:35.098Z","source":"ai"},{"id":"msg-1784828399843-1r90jxd","role":"user","content":"Compare and contrast Machine Learning vs Artificial Intelligence vs Deep learning?","timestamp":"2026-07-23T17:39:59.843Z","source":"ai"},{"id":"msg-1784828406637-57heopj","role":"assistant","content":"The relationship between these three terms is best understood as a series of nested Russian dolls, where each layer represents a more specific methodology for achieving the goal of machine intelligence.\n\n### Artificial Intelligence: The Outer Shell\nArtificial Intelligence is the broadest category, encompassing any technique that enables computers to mimic human behavior or logic. Crucially, \"AI\" does not inherently imply \"learning.\" Early AI was dominated by symbolic logic and \"expert systems\"—rigid, hand-coded rules designed by humans to simulate decision-making. If you write a complex series of `if-then` statements to play a perfect game of Tic-Tac-Toe, you have created AI, but you have not used Machine Learning.\n\n### Machine Learning: The Statistical Middle\nMachine Learning is a subset of AI that represents a fundamental shift in philosophy: instead of being explicitly programmed with rules, the system is programmed to *learn* the rules from data. \n\nThe distinction lies in the origin of the logic. In traditional AI, a human expert provides the features and the rules (e.g., \"If an image has two pointed ears and whiskers, it is a cat\"). In Machine Learning, the human provides the features (ear shape, whisker presence), but the algorithm determines the mathematical relationship between those features and the label \"cat.\" This relies on the core components of a **Loss Function** to measure error and **Gradient Descent** to optimize the model's parameters.\n\n### Deep Learning: The Specialized Core\n**Deep Learning** is a further subset of Machine Learning that specifically utilizes **Neural Networks** with many layers (hence \"deep\"). \n\nThe contrast between \"classical\" Machine Learning and Deep Learning is the most significant technical divide in modern AI. In classical ML (like linear regression or decision trees), humans must perform \"feature engineering\"—identifying which parts of the data are important for the model to look at. In Deep Learning, the network performs **feature learning**. Through **Backpropagation**, the model discovers for itself that it should look for edges, then shapes, then complex structures like eyes or ears, without a human ever defining those concepts.\n\n### Summary of Differences\n\n| Feature | Artificial Intelligence | Machine Learning | Deep Learning |\n| :--- | :--- | :--- | :--- |\n| **Definition** | Any system that mimics human intelligence. | Systems that improve through experience/data. | ML using multi-layered neural networks. |\n| **Logic Source** | Often hand-coded by humans. | Discovered by algorithms from data. | Learned autonomously through layers of abstraction. |\n| **Data Dependency** | Can work with very little data (rules). | Requires structured data to find patterns. | Requires massive amounts of data to be effective. |\n| **Hardware** | Can run on basic CPUs. | Typically runs on CPUs or simple GPUs. | Requires high-performance GPUs/TPUs (compute-intensive). |\n\nWhile the current scene focuses on the mechanics of learning, moving toward **Deep Learning** would involve exploring how these principles scale when we move from simple statistical models to the massive architectures that define the current **Deep Learning Revolution**.","timestamp":"2026-07-23T17:40:06.637Z","source":"ai"},{"id":"msg-1784828440639-mtk7g76","role":"user","content":"who are the founding fathers of the machine learning?","timestamp":"2026-07-23T17:40:40.639Z","source":"ai"},{"id":"msg-1784828448213-cwgyy1t","role":"assistant","content":"The intellectual lineage of machine learning is a confluence of neuroscience, mathematics, and computer science. While many contributed, three figures are often cited as the \"Godfathers\" of the modern era, particularly for their persistence through \"AI Winters\" when the field was out of favor.\n\n### The Modern Trio (The \"Godfathers of AI\")\nIn 2018, the Turing Award—the \"Nobel Prize of Computing\"—was awarded to three researchers who moved machine learning from a niche statistical field into the dominant force it is today:\n\n*   **Geoffrey Hinton**: A central figure in the development of **Backpropagation** and Boltzmann machines. Hinton was a vocal advocate for the \"connectionist\" view (the idea that intelligence emerges from networks of simple units) at a time when the field favored symbolic logic. He was a key author of the 1986 paper that popularized backpropagation as a viable way to train deep networks.\n*   **Yann LeCun**: A former student of Hinton, LeCun pioneered **Convolutional Neural Networks (CNNs)**. His work on the LeNet architecture in the late 1980s proved that neural networks could solve real-world problems—specifically handwriting recognition for the US Postal Service—by mimicking the visual cortex's hierarchical structure.\n*   **Yoshua Bengio**: Known for his work on probabilistic modeling of sequences, which laid the groundwork for modern natural language processing. Along with Hinton and LeCun, he helped revive interest in \"deep\" architectures in the mid-2000s, proving that deep networks could be trained effectively if initialized correctly.\n\n### The Theoretical Pioneers\nBefore the modern \"Deep Learning\" era, the field was defined by those who conceptualized the very idea of a \"learning machine\":\n\n*   **Frank Rosenblatt**: In 1958, he created the **Perceptron**, the earliest ancestor of the modern neural network. Though it was a single-layer system with significant limitations (later famously critiqued by Marvin Minsky), it established the idea of a machine that could learn weights from data.\n*   **Arthur Samuel**: An IBM researcher who actually coined the term \"Machine Learning\" in 1959. He famously wrote a checkers-playing program that could learn to play better than its creator, demonstrating the core principle of a system surpassing its explicit programming.\n*   **Vladimir Vapnik**: Though he represents a different branch than neural networks, his work on **Statistical Learning Theory** and the development of Support Vector Machines (SVMs) provided the mathematical rigor that defined machine learning throughout the 1990s. He focused on the limits of **Generalization**, asking: \"How can we guarantee a model will work on data it hasn't seen?\"\n\n### The Psychologists and Mathematicians\nThe field also draws heavily from the \"Parallel Distributed Processing\" (PDP) group, including **David Rumelhart** and **James McClelland**, who bridged the gap between cognitive psychology and computer science in the 1980s, framing learning as a biological-inspired process of weight adjustment.\n\nTo better understand the transition from these foundational theories to the massive systems we see today, it is helpful to look at the **Deep Learning Revolution** and the specific breakthroughs, like the **ImageNet Challenge**, that finally proved these \"founding\" ideas worked at scale.","timestamp":"2026-07-23T17:40:48.213Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:40:48.213Z"},{"nodeId":"n-mermaid-mrxl4f8r-100","messages":[{"id":"msg-mrxl4f9k-99-0-fhdws","role":"assistant","content":"The predicament of tests becoming useless: frontier models max out benchmarks or absorb their answers from training data. The response is an arms race of harder tests and private evaluation — and a reason to treat leaderboard claims skeptically.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-101","messages":[{"id":"msg-mrxl4f9k-100-0-jh2or","role":"assistant","content":"Using a strong model to grade other models' outputs — the only evaluation method that scales to open-ended tasks. Now ubiquitous, despite known biases: models favor verbose answers and their own family's style.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-102","messages":[{"id":"msg-mrxl4f9k-101-0-2bbz0","role":"assistant","content":"Models whose weights are published for anyone to download, inspect, and run — Meta's Llama line, Mistral, DeepSeek, Qwen. Open weights compress prices, enable self-hosting and national sovereignty plays, and anchor the debate over whether frontier capability should be freely distributable.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784832075906-mf7crie","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:41:15.906Z","source":"ai"},{"id":"msg-1784832081888-c5vbr6a","role":"assistant","content":"The shift toward open-weight models represents a fundamental restructuring of the AI industry, moving away from the era where \"frontier\" capabilities were strictly the domain of closed-source API providers. By publishing the weights—the numerical parameters that define a model's behavior—entities like Meta AI, Mistral AI, and Alibaba Qwen have effectively commoditized high-performance inference, shifting the competitive moat from the model itself to the infrastructure and data used to refine it.\n\nThis ecosystem is anchored by the concept of \"Sovereign AI\" and strategic independence. For a laboratory like Mistral, open weights serve as a wedge against the dominance of American hyperscalers, allowing European enterprises to maintain data residency and control. Similarly, Alibaba Qwen provides a high-performance alternative that anchors the massive developer ecosystem in Asia, ensuring that the local industry is not dependent on Western proprietary software.\n\nThe primary tension in this landscape is the \"Inference vs. Training\" divide. While training a Llama or Qwen model requires tens of thousands of H100 GPUs and massive capital expenditure, the resulting open weights can be run on consumer or enterprise-grade hardware. This allows for Self-Hosting, which is critical for industries with strict regulatory or privacy requirements. When a model is self-hosted, the \"unit economics of intelligence\" change: instead of paying a per-token tax to a provider, the cost is shifted to capital expenditure and electricity, which often scales more efficiently for high-volume applications.\n\nHugging Face serves as the essential clearinghouse for this movement, but its role goes beyond mere hosting. It facilitates the \"Model Factory\" workflow, where a base model is transformed through community-driven post-training. Because the weights are open, the community can apply techniques like fine-tuning or quantization (to make models run on smaller chips) far faster than a single closed lab could.\n\nHowever, a critical distinction to maintain is that \"open weights\" is not synonymous with \"open source\" in the traditional software sense. Most of these models are released under restrictive licenses that govern commercial scale, and the training datasets and \"recipes\" (the specific hyperparameters and data curriculum) remain closely guarded secrets. This creates a ceiling for the ecosystem: while the world can use the models, it cannot fully replicate the process that created them.\n\nTo deepen this exploration, it would be useful to look at the specific architectural innovations that make these models efficient for public deployment, such as Grouped-Query Attention (GQA), and the competitive pressure from labs like DeepSeek, which have recently disrupted the price-to-performance ratio of the entire arena.","timestamp":"2026-07-23T18:41:21.888Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:21.888Z"},{"nodeId":"n-mermaid-mrxl4f8r-103","messages":[{"id":"msg-mrxl4f9k-102-0-lngs0","role":"assistant","content":"The dominant commercial form of AI: hosted model access, metered per token. APIs give any developer frontier capability without owning hardware, and aggregators like OpenRouter now make models as swappable as commodities — with pricing pressure to match.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784832054801-xrosulj","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:40:54.801Z","source":"ai"},{"id":"msg-1784832062123-7xw6ntb","role":"assistant","content":"The commercial landscape of Model APIs represents the shift of artificial intelligence from a research artifact to a utility. By abstracting the complexities of model weights, hardware clusters, and CUDA environments into a single POST request, these interfaces have turned intelligence into a metered commodity. This transition is not merely technical; it is the fundamental economic engine of the current \"Gold Rush,\" dictating how value is captured across the stack.\n\nAt the infrastructure layer, the relationship between cloud providers like Amazon Web Services and the API model is one of symbiotic dependency. While these giants provide the \"Model-as-a-Service\" (MaaS) layer, they are also vertically integrating. AWS, for example, is not content simply hosting third-party APIs; they are pushing custom silicon (Trainium/Inferentia) to drive down the marginal cost of a token, attempting to decouple their margins from the high cost of the Nvidia-dominated hardware market.\n\nHowever, the proliferation of models has created a friction point for builders: the \"provider lock-in\" and the volatile nature of performance-per-dollar. This has birthed a secondary layer of abstraction via aggregators. Services like OpenRouter act as a clearinghouse, allowing developers to treat intelligence as a fungible asset. This shift is significant because it pressures frontier labs to compete on raw unit economics and latency rather than just brand or ecosystem. When a developer can switch from a proprietary flagship to an open-weight equivalent via a single line-of-code change in an API wrapper, the \"moat\" of any individual model begins to thin.\n\nThis brings us to the specialized role of platforms like Together AI. They represent the bridge between the \"black box\" proprietary API and the transparency of open-weight models. By offering APIs for models whose weights are public, they provide a middle ground: the convenience of a hosted endpoint with the theoretical portability of open source. This creates a fascinating tension in the industry:\n1.  **Proprietary API dominance:** High performance, but high cost and zero visibility into the \"black box.\"\n2.  **Open-weight API hosting:** Lower costs and fine-tuning flexibility, but requiring more sophisticated engineering on the client side.\n\nA critical missing piece in this landscape is the **Inference-Compute Trade-off**. The price of a token in an API is a direct reflection of how much compute was used to generate it. As we move into an era of \"reasoning\" models that perform more computation at inference time to achieve better results, the standard token-based billing model may come under strain. We are likely to see a shift from \"pay-per-token\" to \"pay-per-compute-unit\" or \"pay-per-result,\" as the linear relationship between output length and intelligence produced begins to break down.\n\nFurthermore, the emergence of the **Model Context Protocol (MCP)** suggests that the next evolution of these APIs isn't just about sending text and receiving text. It is about how these models securely access external data silos. The API is transforming from a simple inference gate into a sophisticated orchestration point between the model and the user's private data.\n\nTo better understand the economic pressures on these APIs, it would be useful to look at the **Unit Economics of Intelligence** and the specific impact of **Inference Cost** on how these companies set their pricing tiers. Additionally, including the **Context Window** as a concept would clarify why API pricing often scales non-linearly as inputs grow, due to the memory pressures on the underlying hardware.","timestamp":"2026-07-23T18:41:02.123Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:02.123Z"},{"nodeId":"n-mermaid-mrxl4f8r-104","messages":[{"id":"msg-mrxl4f9k-103-0-ipqly","role":"assistant","content":"Running models on infrastructure you control — for privacy, regulatory, cost, or sovereignty reasons. Enabled by open weights and quantization, it is the counterweight to API dependence and a growing enterprise pattern.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784832105581-s4mgfa4","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:41:45.581Z","source":"ai"},{"id":"msg-1784832111791-wlkk5z5","role":"assistant","content":"The movement toward self-hosting AI represents a significant shift in the power dynamics of the industry, transitioning from a centralized model of \"Intelligence as a Service\" to one of \"Intelligence as an Asset.\" While the initial surge of generative AI was defined by massive, closed APIs, the viability of running these systems on private infrastructure has been secured by the interplay between model accessibility and optimization techniques.\n\nAt the heart of this shift are **Open-Weight Models**. Unlike closed systems, these allow for deep inspection and local execution, providing the raw material for self-hosting. However, the sheer size of modern foundation models often exceeds the VRAM capacity of consumer or even prosumer hardware. This is where **Quantization** becomes the critical enabler. By reducing the precision of model weights—typically from 16-bit floating point (FP16) to 4-bit or 8-bit integers (INT4/INT8)—the memory footprint is slashed by 70% or more with remarkably little impact on reasoning capabilities. This technical \"compression\" is what transforms a model requiring an enterprise-grade A100 cluster into one that can run on a single workstation or a dedicated edge device.\n\nIn the broader context of **The Arena**, the decision to self-host is rarely just about cost; it is driven by three primary tensions:\n\n1.  **Privacy and Data Sovereignty:** For many organizations, the \"Privacy Moat\" is the primary driver. Self-hosting ensures that proprietary data—whether it’s medical records, legal documents, or internal code—never leaves the local network, bypassing the trust requirements of third-party API providers.\n2.  **Latency and Reliability:** Local execution eliminates the \"network hop\" to a distant datacenter and removes dependency on the uptime or rate limits of a provider. For real-time applications or high-throughput batch processing, the deterministic performance of local silicon is often superior to the variability of public endpoints.\n3.  **The Customization Paradox:** While self-hosting offers the ultimate flexibility to fine-tune or modify models, it introduces a significant operational burden. The user moves from consuming a finished product to managing a \"Model Factory,\" responsible for driver compatibility, hardware cooling, and serving infrastructure.\n\nAs this landscape evolves, a key area to explore is the hardware side of this equation—specifically how specialized architectures like **Inference**-optimized chips and the **CUDA Moat** dictate which models can be hosted where. Additionally, the concept of **Sovereign AI** expands this logic to the nation-state level, where self-hosting isn't just a technical choice for a developer, but a strategic imperative for a country's digital autonomy.","timestamp":"2026-07-23T18:41:51.791Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:41:51.792Z"},{"nodeId":"n-mermaid-mrxl4f8r-105","messages":[{"id":"msg-mrxl4f9k-104-0-407gz","role":"assistant","content":"The tiering strategy every lab converged on: flagship models for hard problems, mid-size for everyday work, small models for speed and cost. Matching task to tier is now a basic engineering skill, and routing between tiers a product category.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784832139397-yt7km7i","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:42:19.397Z","source":"ai"},{"id":"msg-1784832145994-4ucqyti","role":"assistant","content":"The shift toward model families represents a transition from treating AI as a series of singular research breakthroughs to treating it as a versatile utility. By offering a tiered range of capabilities, labs can address the massive divergence in compute requirements across different use cases. A flagship model might be necessary for complex reasoning or code generation, but it is often economically and latency-prohibitive for simple summarization or data extraction.\n\nThe economic viability of these families is increasingly driven by **Distillation**. This technique allows the \"wisdom\" and reasoning patterns of a massive teacher model to be compressed into a smaller student model. This isn't just about making models smaller; it's about breaking the linear relationship between capability and parameter count. Through distillation, a small model can punch significantly above its weight class by mimicking the output distributions of its larger siblings, effectively allowing the flagship's expensive training run to subsidize the performance of the more efficient tiers.\n\nIn **The Arena**, this strategy has become the primary battleground for market share. Alibaba Qwen exemplifies this through its \"shotgun\" approach to releases, providing a dense spectrum of sizes—from mobile-optimized 0.5B versions to massive 72B+ architectures. This saturation of the ecosystem allows a single family to capture diverse user bases, from hobbyists running local hardware to enterprise-grade cloud deployments.\n\nHowever, this tiered approach introduces a significant architectural trade-off: the **Inference-Compute Trade-off**. While distillation captures capability, it does not fundamentally change the FLOPs required for a forward pass. As we look toward the next stage of these families, the industry is moving toward **Mixture of Experts (MoE)** and **Grouped-Query Attention (GQA)** to further decouple a model's total parameter count from its active inference cost. These architectural optimizations are what allow a \"Large\" model to behave like a \"Small\" model in terms of speed, while retaining the knowledge base of a giant.\n\nA critical question for this model-family strategy is whether the \"Data Wall\"—the approaching limit of high-quality human-generated text—will force a shift in how these families are built. We are likely moving toward a world where families are differentiated not just by size, but by their specialization in **Inference-Time Search** or **Synthetic Data** generation, where the largest models in the family exist primarily to generate the high-quality training data used to refine the smaller, more deployable versions.\n\n***\n\n### Recommended Additions\n\nTo better understand how these families are optimized for deployment and how they compete, consider the following additions:","timestamp":"2026-07-23T18:42:25.994Z","source":"ai"},{"id":"msg-1784832157580-3qe4fiu","role":"user","content":"Suggest new concepts to add to my knowledge graph.","timestamp":"2026-07-23T18:42:37.580Z","source":"ai"},{"id":"msg-1784832162377-mqvsuo5","role":"assistant","content":"To deepen the exploration of how **Model Families** are engineered and positioned, several emerging concepts are critical. These address the technical methods used to keep smaller family members competitive, the specific architectures that allow for efficient scaling, and the business strategies used to lock in users across a capability spectrum.\n\n*   **Speculative Decoding**: This is the primary way model families are paired together in production. A small \"draft\" model from the family predicts tokens quickly, and the \"flagship\" model verifies them in parallel. This allows the family to offer the speed of the small model with the intelligence of the large one.\n*   **Constant VRAM Footprint Strategy**: This refers to the practice of designing model families (like Llama or Qwen) with standardized architectures so that quantization and hardware optimizations can be reused across all sizes, simplifying the \"Inference-Compute Trade-off\" for developers.\n*   **Vertical Specialization**: Beyond size tiers, families are now branching into \"Verticals\" (e.g., Qwen-Math, Qwen-Coder). This represents a shift from general-purpose families to task-optimized lineages.\n*   **Chinchilla Optimality**: This is the theoretical foundation for model families. It dictates the \"sweet spot\" between parameter count and training data, which labs use to decide whether to release a 7B, 14B, or 70B model.","timestamp":"2026-07-23T18:42:42.377Z","source":"ai"},{"id":"msg-1784832192028-gkze1p5","role":"user","content":"Name/List all most important model families and created separate nodes for each of them.","timestamp":"2026-07-23T18:43:12.028Z","source":"ai"},{"id":"msg-1784832197299-evkp39a","role":"assistant","content":"To map the current landscape of the industry, it is essential to distinguish between the primary model families that define the state of the art. While they all follow the tiered capability strategy, their specific focuses—ranging from proprietary dominance to open-weight accessibility—shape the broader ecosystem.\n\nThe following families represent the current \"anchors\" of the market:\n\n*   **GPT (OpenAI)**: The family that established the tiered approach (e.g., 4o, 4o-mini). It remains the benchmark for \"Flagship\" reasoning and is the primary driver of the shift toward \"Inference-Time Search\" models like the o1 series.\n*   **Claude (Anthropic)**: Known for a distinct focus on \"Constitutional AI\" and safety. Its tiers (Haiku, Sonnet, Opus) are specifically designed to balance speed, cost, and high-end cognitive performance.\n*   **Gemini (Google DeepMind)**: Notable for its \"Native Multimodal\" architecture and massive context windows. It spans from \"Flash\" for high-speed utility to \"Ultra\" for enterprise-grade complexity.\n*   **Llama (Meta AI)**: The most influential open-weight family. It acts as the industry's common vocabulary, providing a base for the vast majority of fine-tuned and specialized models found in the open-source community.\n*   **Mistral (Mistral AI)**: Pioneered the use of \"Mixture of Experts (MoE)\" at scale, proving that smaller, efficiently routed families could compete with much larger monolithic models.\n*   **DeepSeek (DeepSeek)**: A rapidly rising family known for extreme efficiency and performance-to-cost ratios, particularly in coding and mathematical reasoning, often leveraging advanced MoE architectures.","timestamp":"2026-07-23T18:43:17.299Z","source":"ai"},{"id":"msg-1784832251097-sxr8z4m","role":"user","content":"Grok, GLM, Kimi... other noticeable models?","timestamp":"2026-07-23T18:44:11.097Z","source":"ai"},{"id":"msg-1784832255762-imyexpc","role":"assistant","content":"Beyond the established giants, several other model families have carved out significant strategic niches, often by dominating specific regional markets or hardware-specific use cases.\n\n**Grok** by xAI is positioned as a \"real-time\" family, leveraging direct access to live data streams from the X platform to bypass the typical knowledge cutoff of most model families. Its architecture prioritizes high-throughput reasoning, aiming to close the gap with flagship models while maintaining a more \"unfiltered\" persona.\n\n**GLM (General Language Model)**, developed by Zhiping AI, is a cornerstone of the Chinese AI ecosystem. It is technically notable for its bilingual (Chinese-English) optimization and its early adoption of the GLM architecture—a hybrid that differs from the standard GPT-style decoder-only approach—making it highly effective for both natural language understanding and generation.\n\n**Kimi** (by Moonshot AI) has differentiated itself through \"long-context\" supremacy. While other families eventually followed, Kimi was among the first to make million-token context windows a core product feature, allowing the family to serve as a specialized tool for massive document analysis and long-form research.\n\nOther families worth noting include **Command** by Cohere, which focuses specifically on \"RAG-optimized\" (Retrieval-Augmented Generation) performance for enterprise environments, and **Jamba** by AI21 Labs, which introduced a hybrid Mamba-Transformer architecture to solve the memory bottlenecks associated with traditional attention mechanisms in long sequences.","timestamp":"2026-07-23T18:44:15.762Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:44:15.762Z"},{"nodeId":"n-mermaid-mrxl4f8r-106","messages":[{"id":"msg-mrxl4f9k-105-0-6v4yg","role":"assistant","content":"The handful of organizations training models at the frontier of capability — a club defined by access to billion-dollar compute, scarce talent, and data at scale. Their releases set the industry's tempo, and their governance structures (nonprofit hybrids, public-benefit corporations) reflect the field's unresolved tension between mission and market.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784826604745-7sfm9cg","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:10:04.745Z","source":"ai"},{"id":"msg-1784826615234-ddx61gn","role":"assistant","content":"The \"frontier\" in artificial intelligence is defined less by a specific technology and more by a massive concentration of three specific inputs: high-density compute clusters, elite engineering talent, and the capital required to fuse them. While thousands of organizations build AI, only a handful operate at this frontier, where the goal is not just to implement known architectures but to discover the next scaling laws and emergent capabilities.\n\nThis group is currently bifurcated by a fundamental strategic divide regarding the distribution of power. On one side, the \"closed\" labs—OpenAI, Anthropic, and Google DeepMind—maintain tight control over their weights, arguing that the risks of misuse and the necessity of recouping multi-billion dollar R&D costs require a proprietary, API-based model. On the other, the open-weight proponents like Meta AI, Mistral, and Alibaba Qwen are commoditizing the frontier. This tension has forced a shift in the definition of a \"moat\"; as open-weight models like Llama 4 or Qwen-Max approach parity with closed models, the competitive advantage for closed labs shifts away from raw knowledge toward reasoning capabilities and proprietary data loops.\n\nThe entry of DeepSeek into this elite tier has recently challenged the \"brute force\" scaling hypothesis. By demonstrating that sophisticated reasoning and state-of-the-art performance can be achieved with significantly higher training efficiency and lower compute budgets, the frontier is no longer solely the domain of the highest bidder. This introduces a \"post-scaling\" era where algorithmic breakthroughs—specifically in Mixture of Experts (MoE) architectures and Reinforcement Learning—may matter as much as the total FLOPs (floating point operations) expended during training.\n\nHowever, the barrier to entry remains high due to the physical infrastructure required. The \"Arena\" is increasingly a contest of hardware logistics, as seen with xAI’s rapid build-out of massive training clusters. This physical reality creates a natural friction with AI Governance. Governments are beginning to define \"Frontier AI\" through compute thresholds (e.g., models trained using more than $10^{26}$ integer or floating-point operations), effectively turning the hardware itself into a regulatory control point.\n\nTo deepen the exploration of how these labs maintain their position, it would be useful to look at the underlying economic and physical drivers—specifically the scaling laws that dictate their R&D spend and the hardware titans that supply their \"industrial' scale compute.","timestamp":"2026-07-23T17:10:15.234Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:10:15.234Z"},{"nodeId":"n-mermaid-mrxl4f8r-107","messages":[{"id":"msg-mrxl4f9k-106-0-0lo9n","role":"assistant","content":"The lab that defined the modern era: GPT models, ChatGPT, DALL-E, and the o-series that opened the reasoning age. Founded 2015 as a nonprofit counterweight to Google; its evolution into a capped-profit, Microsoft-backed commercial juggernaut — through a 2023 governance crisis and ongoing restructuring — is the industry's central corporate drama.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-108","messages":[{"id":"msg-mrxl4f9k-107-0-qkygg","role":"assistant","content":"The safety-focused frontier lab behind the Claude models, founded in 2021 by OpenAI departees led by the Amodei siblings. Known for constitutional AI, mechanistic interpretability research, the Model Context Protocol, and strength in coding and agentic work — the strongest existence proof that safety emphasis and frontier capability can coexist.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-109","messages":[{"id":"msg-mrxl4f9k-108-0-dnljy","role":"assistant","content":"Google's consolidated AI organization under Hassabis, merging DeepMind's research legacy (AlphaGo, AlphaFold) with Google Brain's (the transformer itself). Produces the Gemini family and enjoys structural advantages no startup can match: TPU silicon, planetary distribution, and self-funding scale.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-11","messages":[{"id":"msg-mrxl4f9k-10-0-9wahk","role":"assistant","content":"Learning from examples that come with correct answers: images with labels, sentences with translations. The dominant paradigm of the 2010s and still the mechanism inside supervised fine-tuning of language models. Its limitation — someone must produce the labels — is what pushed the field toward self-supervision.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-110","messages":[{"id":"msg-mrxl4f9k-109-0-eitkz","role":"assistant","content":"Meta's AI arm and the open-weight standard-bearer: the Llama models anchor the open ecosystem and commoditize rivals' closed advantage. Backed by enormous capex and Zuckerberg's aggressive talent acquisition, though its frontier standing relative to closed labs is perennially debated.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-111","messages":[{"id":"msg-mrxl4f9k-110-0-c3bf4","role":"assistant","content":"Musk's frontier lab, founded 2023, defined by velocity: the Colossus training cluster assembled in months, Grok models integrated with the X platform, and a stated mission of maximally truth-seeking AI. Demonstrates how far capital plus urgency can compress the path to the frontier.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-112","messages":[{"id":"msg-mrxl4f9k-111-0-pef7v","role":"assistant","content":"The Paris-based lab that made Europe a frontier participant: efficient open-weight models — its MoE releases were field-shaping — and a sovereignty pitch that resonates with European governments and enterprises wary of US-platform dependence.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-113","messages":[{"id":"msg-mrxl4f9k-112-0-f30vu","role":"assistant","content":"The Chinese lab whose R1 reasoning model, released openly in early 2025 at a claimed fraction of US training costs, triggered a trillion-dollar market repricing. The canonical demonstration that efficiency can substitute for compute abundance — with consequences for export-control logic, open-weight economics, and the geography of the frontier.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-114","messages":[{"id":"msg-mrxl4f9k-113-0-x2f92","role":"assistant","content":"Alibaba's model family and China's other open-weight anchor: a full tier ladder of capable models whose downloads and derivatives rival Llama's in the open ecosystem. Evidence that the open-weight frontier is now substantially Chinese — a fact with underappreciated strategic weight.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-115","messages":[{"id":"msg-mrxl4f9k-114-0-y6s2s","role":"assistant","content":"The executives and founders steering the industry's major organizations — the names attached to the strategies, rivalries, and safety debates that fill AI business coverage.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784826926267-alx5mfe","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:15:26.268Z","source":"ai"},{"id":"msg-1784826931746-orecfuh","role":"assistant","content":"The current leadership of the AI industry is defined by a distinct shift from research-driven idealism to a high-stakes geopolitical and industrial race. While the early phase of the frontier labs was characterized by a shared \"safety-first\" or academic ethos, the present era is dominated by the strategic divergence of the individuals steering these organizations.\n\nThe central tension lies in how these leaders manage the transition from scaling compute to achieving reasoning capabilities. For leaders like Satya Nadella and Sam Altman, the priority has been the construction of an unprecedented partnership between a legacy platform giant and a frontier lab. This \"incumbent-frontier\" model has set the standard for how AI is productized, yet it faces increasing pressure from the \"sovereign\" and \"open\" strategies championed by others. \n\nMark Zuckerberg and Arthur Mensch represent a significant counter-movement. By pushing high-performance, open-weight models, they are effectively commoditizing the underlying intelligence that OpenAI and Anthropic seek to sell via proprietary APIs. This move shifts the power away from the \"model-as-a-service\" gatekeepers and toward the broader developer ecosystem, framing AI access as a matter of industrial sovereignty—particularly in Europe, where Mensch has positioned Mistral as a strategic asset.\n\nThe entrance of Liang Wenfeng and DeepSeek into the global conversation has introduced a new variable: extreme algorithmic efficiency. While the American and European labs have largely focused on massive capital expenditure and scaling laws, the success of DeepSeek suggests that the \"brute force\" approach to training may have diminishing returns compared to specialized architectural optimizations. This puts immense pressure on Jensen Huang and Nvidia to maintain the hardware moat, as the industry begins to question if the path to AGI requires more gigawatts or simply better math.\n\nFurthermore, the \"hard-tech\" approach of Elon Musk, linking xAI’s development directly to massive physical infrastructure projects like the Colossus cluster, contrasts with the more \"integrated-research\" approach of Demis Hassabis. At Google DeepMind, the focus remains on solving fundamental scientific bottlenecks—exemplified by the move from game-playing (AlphaGo) to biological folding (AlphaFold)—while simultaneously defending a legacy search business.\n\nA critical question currently facing this group is the \"Data Wall.\" As the supply of high-quality human-generated text is exhausted, these executives must decide whether the next leap in intelligence comes from synthetic data, multi-modal grounding, or the \"test-time compute\" strategies seen in the latest reasoning models. Each leader is effectively placing a multi-billion dollar bet on which of these paths will yield the first true agentic systems.\n\nTo better understand the constraints these leaders operate under, it would be useful to look at the underlying economic and technical realities that dictate their moves, such as the actual mechanics of the scaling laws they cite and the infrastructure that limits their speed.","timestamp":"2026-07-23T17:15:31.746Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:15:31.746Z"},{"nodeId":"n-mermaid-mrxl4f8r-116","messages":[{"id":"msg-mrxl4f9k-115-0-17u8j","role":"assistant","content":"CEO of OpenAI and the industry's most visible executive — the face of the ChatGPT era, the survivor of a board revolt in 2023 that failed to remove him, and the chief promoter of AGI as an imminent, investable prospect. His fundraising ambitions (datacenters, chips, energy) define the maximalist wing of the industry.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-117","messages":[{"id":"msg-mrxl4f9k-116-0-e9e0e","role":"assistant","content":"CEO and co-founder of Anthropic, which he started with his sister Daniela and colleagues after leaving OpenAI over safety and commercial direction. A former physicist and co-author of early scaling-law work; his essays on both AI's promise and its risks are unusually influential inputs to the policy debate.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-118","messages":[{"id":"msg-mrxl4f9k-117-0-43tm9","role":"assistant","content":"CEO and co-founder of DeepMind, chess prodigy and games-industry veteran turned neuroscientist. Delivered AlphaGo and AlphaFold — the latter earning him a Nobel Prize in Chemistry (2024) — and now leads Google's consolidated AI effort, embodying the science-first lineage of the frontier labs.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-119","messages":[{"id":"msg-mrxl4f9k-118-0-mr3ki","role":"assistant","content":"Co-founded OpenAI, departed in 2018, and returned to AI in 2023 as founder of xAI — while litigating against OpenAI over its shift from nonprofit origins. Built the Colossus supercluster at record speed; integrates xAI's Grok models with his X platform. A singular mix of AI doomer rhetoric and maximum-speed buildout.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-12","messages":[{"id":"msg-mrxl4f9k-11-0-bamxj","role":"assistant","content":"Learning structure from data without any labels: clusters, patterns, compressed representations. Historically the poor cousin of supervised learning, it was vindicated by self-supervised pre-training, which is unsupervised learning in spirit applied at web scale.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-120","messages":[{"id":"msg-mrxl4f9k-119-0-f1fqw","role":"assistant","content":"Founder and CEO of Nvidia, whose three-decade bet on parallel computing and the CUDA ecosystem positioned the company as the arms dealer of the AI boom — and briefly the most valuable company in the world. His leather-jacketed keynotes function as the industry's de facto state of the union.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-121","messages":[{"id":"msg-mrxl4f9k-120-0-pm4sf","role":"assistant","content":"Microsoft's CEO, architect of the industry's most consequential partnership: the multibillion-dollar OpenAI investment that gave Microsoft early frontier access and OpenAI the compute to scale. His \"Copilot\" strategy — AI woven through every Microsoft product — defined how incumbents metabolize the technology.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-122","messages":[{"id":"msg-mrxl4f9k-121-0-nezd7","role":"assistant","content":"Meta's founder and CEO, who bet the company's AI credibility on open weights: the Llama models released freely to commoditize rivals' advantage and anchor the open ecosystem. Backed it with some of the industry's largest capital expenditures and most aggressive talent raids.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-123","messages":[{"id":"msg-mrxl4f9k-122-0-d1p9d","role":"assistant","content":"Founder of DeepSeek and of the quant fund High-Flyer that incubated it. His lab's efficient, openly released frontier-class models in early 2025 wiped hundreds of billions off US tech valuations in a day and rewrote assumptions about training costs, export controls, and China's position in the race.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-124","messages":[{"id":"msg-mrxl4f9k-123-0-ixwz7","role":"assistant","content":"Co-founder of DeepMind, later founder of Inflection AI, hired in 2024 to lead Microsoft AI — an acqui-hire that became the template for the industry's talent-absorbing pseudo-acquisitions. Author of \"The Coming Wave\" on containing powerful technologies.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-125","messages":[{"id":"msg-mrxl4f9k-124-0-6eg3s","role":"assistant","content":"Co-founder and CEO of Mistral AI, the Paris lab founded in 2023 by Meta and DeepMind alumni that became Europe's flagship AI company and champion of open-weight frontier development — carrying, fairly or not, the continent's hopes for AI sovereignty.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-126","messages":[{"id":"msg-mrxl4f9k-125-0-u7e10","role":"assistant","content":"The companies supplying AI's physical substrate — chip designers, the foundry that fabricates nearly everything, and the specialized cloud providers. The layer where AI economics is most concentrated and most geopolitically exposed.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784827100242-a5khoav","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:18:20.242Z","source":"ai"},{"id":"msg-1784827108078-gh285gp","role":"assistant","content":"The industrial backbone of the artificial intelligence era is defined by a rigid, tiered hierarchy of specialized capabilities. At the base of this pyramid is the foundry layer, where the physical limits of lithography determine the ceiling for the entire industry. Because advanced logic chips require extreme ultraviolet (EUV) lithography, the production of nearly all frontier-grade silicon converges on a single point of failure and success: the high-volume, high-yield processes at **TSMC**. This concentration of manufacturing creates a unique economic bottleneck where the pace of AI progress is tied directly to the \"wafer starts\" and advanced packaging capacity (such as CoWoS) of a single entity.\n\nAbove the foundry layer, the architectural divergence between training and inference has become the primary field of competition. **Nvidia** has maintained a near-monopoly not just through hardware, but through the entrenchment of the CUDA software ecosystem, which has turned its GPUs into the \"common tongue\" of AI research. This creates a high barrier for challengers like **AMD**, whose Instinct line must offer significant price-to-performance advantages or seamless software compatibility to overcome the inertia of the incumbent's developer base.\n\nHowever, a shift is occurring as the industry moves from a focus on training massive **Foundation Models** to the efficient serving of those models at scale. This \"inference era\" favors specialized architectures over general-purpose ones. While a GPU is a jack-of-all-trades for parallel processing, **Groq** and its Language Processing Units (LPUs) represent a move toward deterministic, high-speed architectures specifically optimized for the sequential nature of LLM token generation. This distinction is critical: training requires massive memory bandwidth and raw compute power to adjust weights, whereas inference requires low latency and high throughput to utilize those weights.\n\nThe delivery mechanism for this silicon has also fractured. While traditional hyperscalers provide a broad range of services, specialized GPU clouds like **CoreWeave** have emerged to treat compute as a pure commodity. By stripping away the legacy overhead of general-purpose cloud computing, these providers allow labs to access massive clusters with the low-latency networking required for distributed training.\n\nA primary tension to watch is the \"Make vs. Buy\" dilemma currently facing the platform companies. As the cost of merchant silicon (buying from Nvidia) remains high, the largest consumers of chips are increasingly designing their own custom silicon to bypass the margin stack of the chip designers. This move toward vertical integration suggests that the \"Compute & Silicon\" category may eventually split between those selling to the open market and those building bespoke \"internal-only\" engines to power their own model families.\n\nTo deepen this exploration, it may be useful to consider the physical and geopolitical constraints that act as the final arbiters of this industry: the massive energy requirements for the facilities housing this hardware, and the export controls that determine which regions can access these frontier capabilities.\n\n### Recommended Additions\n\n*   **AI Datacenters**: The physical infrastructure (power, cooling, and space) required to turn silicon into active compute.\n*   **Custom AI Chips**: The trend of non-semiconductor companies (Google’s TPU, Amazon’s Trainium) designing their own hardware to reduce reliance on merchant silicon.\n*   **Inference vs. Training**: A conceptual node to distinguish the two primary workloads that dictate different hardware architectural requirements.\n*   **Energy Demand**: The scaling limit of compute expansion, often overlooked when focusing solely on the chips themselves.\n*   **Export Controls**: The geopolitical layer that regulates the flow of this silicon, particularly concerning high-end H100/B200 class chips.","timestamp":"2026-07-23T17:18:28.078Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:18:28.079Z"},{"nodeId":"n-mermaid-mrxl4f8r-127","messages":[{"id":"msg-mrxl4f9k-126-0-toq6s","role":"assistant","content":"The company whose GPUs train and serve essentially the entire frontier, and whose CUDA software moat has held for over fifteen years. The AI boom's defining corporate beneficiary — data-center revenue in the hundreds of billions and a valuation that made it, at points, the world's most valuable company.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784826971371-847959k","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:16:11.371Z","source":"ai"},{"id":"msg-1784826978745-uwm73kr","role":"assistant","content":"Nvidia’s transformation from a niche graphics hardware company to the indispensable foundation of the AI era rests on a decades-long bet that parallel processing would eventually outstrip sequential processing for general-purpose computing. While CPUs are designed for the complex, branch-heavy logic of general software, GPUs excel at performing the massive, simultaneous matrix multiplications required to simulate neural networks. This architectural alignment has made Nvidia the gatekeeper of the physical layer of intelligence.\n\nAt the heart of this dominance is the interplay between the hardware and the software ecosystem. While the silicon itself is the primary focus of the \"AI gold rush,\" it is the vertical integration of hardware with proprietary software stacks like CUDA that creates a profound moat. This integration ensures that when researchers push the boundaries of training compute, the hardware is not just a passive resource but an optimized environment. Jensen Huang’s strategy has effectively turned Nvidia into a platform company rather than just a chip designer, where the hardware and the libraries used to program it are inseparable.\n\nThis has led to a feedback loop involving training compute. As models scale—following the empirical \"scaling laws\" that suggest intelligence emerges with more data and more parameters—the demand for compute grows exponentially. Nvidia has moved beyond selling individual chips to designing entire datacenter-scale systems where the \"unit of compute\" is no longer a single GPU but an entire interconnected cluster. The bottleneck for frontier AI labs has shifted from algorithmic ingenuity to the sheer availability of high-bandwidth memory and inter-gpu communication speeds.\n\nHowever, this total market concentration introduces a significant tension: the \"compute divide.\" The immense cost and energy requirements of the hardware necessary for state-of-the-art training mean that frontier-level progress is increasingly restricted to a few well-capitalized entities. Furthermore, as the industry moves from a focus on training to a focus on inference—running the models at scale for millions of users—Nvidia faces new pressure. While their GPUs are the gold standard for flexibility during research and training, specialized silicon (ASICs) and custom chips from cloud providers are beginning to challenge the cost-efficiency of GPUs for specific, high-volume inference tasks.\n\nTo broaden the current view of this ecosystem, it is worth considering the physical and geopolitical dependencies that enable this production, as well as the emerging shift toward alternative hardware architectures.\n\n*   **TSMC:** While Nvidia designs the architecture, the actual fabrication of these chips happens almost exclusively at TSMC. This creates a single point of failure in the global AI supply chain and ties the future of AI directly to the stability and capacity of a few square miles of fabrication plants in Taiwan.\n*   **Custom AI Chips:** Major cloud providers (Google with TPUs, Amazon with Trainium/Inferentia) and startups like Groq are developing specialized architectures designed to bypass the general-purpose nature of GPUs to achieve better performance-per-watt for specific model architectures like Transformers.\n*   **AI Datacenters:** The scale of compute has moved from the server rack to the utility scale. The constraints on Nvidia's growth are increasingly shifting from chip yields to the physical availability of power, cooling, and real estate in massive datacenter hubs.","timestamp":"2026-07-23T17:16:18.745Z","source":"ai"},{"id":"msg-1784826992746-y5xrg2j","role":"assistant","content":"To expand the exploration of Nvidia's role in the current technological landscape, we should look at the specific technical and logistical moats that sustain its lead, as well as the emerging bottlenecks that define the next phase of the industry.\n\nWhile the GPU is the primary engine, its performance in massive clusters is increasingly dictated by **NVLink**, the high-speed interconnect that allows multiple GPUs to act as a single unified memory pool. Without this, scaling to the cluster sizes required for frontier models would suffer from debilitating communication bottlenecks. \n\nLooking at the broader ecosystem, the **Inference-Compute Trade-off** is becoming a critical strategic pillar. As models move toward \"thinking\" during their output phase (test-time compute), Nvidia’s hardware requirements shift from purely training-focused throughput to the low-latency demands of real-time reasoning. \n\nOn the production side, the bottleneck is often not the silicon itself but **CoWoS (Chip on Wafer on Substrate)**. This specialized packaging technology from TSMC is required to bond high-bandwidth memory to the GPU die; its limited capacity has been the primary constraint on Nvidia’s ability to meet global demand over the last two years. \n\nFinally, the **CUDA Moat** represents the profound software lock-in created by nearly two decades of developer optimization. It isn't just about the hardware's speed, but the fact that almost every major AI framework is natively optimized for Nvidia’s proprietary software layer, making it extremely difficult for competitors to gain traction even if their silicon is theoretically faster.","timestamp":"2026-07-23T17:16:32.746Z","source":"ai"},{"id":"msg-1784827016342-clg0av6","role":"user","content":"Suggest new concepts to add to my knowledge graph.","timestamp":"2026-07-23T17:16:56.342Z","source":"ai"},{"id":"msg-1784827020998-bzmvn7x","role":"assistant","content":"To deepen the exploration of Nvidia's dominance and the constraints it faces, we should look beyond the silicon itself and into the specialized engineering and software layers that maintain its market position.\n\nWhile GPUs are the headline, the **NVLink** interconnect is perhaps Nvidia's most significant architectural moat in the era of cluster-scale computing. It allows a rack of GPUs to communicate with such high bandwidth that they behave like a single, massive processor—a feat that standard networking protocols like Ethernet struggle to match. This physical coupling of chips is what enables the training of models with trillions of parameters.\n\nOn the manufacturing side, the primary bottleneck is often not the chip design, but **CoWoS (Chip on Wafer on Substrate)**. This is a sophisticated 2.5D packaging technology provided by TSMC that allows Nvidia to stack high-bandwidth memory (HBM) extremely close to the GPU die. Global AI chip supply is currently more constrained by CoWoS packaging capacity than by raw wafer production.\n\nFurthermore, the \"stickiness\" of Nvidia's ecosystem is reinforced by the **CUDA Moat**. This software layer has been cultivated for nearly twenty years, ensuring that almost every major AI research paper and framework is optimized for Nvidia hardware first. This creates a high barrier to entry for competitors, who must not only build better hardware but also replicate an entire ecosystem of libraries and developer tools.\n\nFinally, as we move into an era of massive-scale deployment, **Sovereign AI** has emerged as a major commercial driver. This refers to the trend of nation-states investing in their own domestic AI infrastructure and datacenters to ensure data security and technological independence, providing a new, high-floor demand for Nvidia's enterprise-grade hardware that exists outside the traditional big-tech cloud providers.","timestamp":"2026-07-23T17:17:00.998Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:17:00.998Z"},{"nodeId":"n-mermaid-mrxl4f8r-128","messages":[{"id":"msg-mrxl4f9k-127-0-8bhdu","role":"assistant","content":"Nvidia's principal challenger in AI silicon, with its Instinct accelerator line and open software stack. Its progress is the market's main test of whether the GPU layer admits real competition — every hyperscaler quietly roots for it.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-129","messages":[{"id":"msg-mrxl4f9k-128-0-d3zoo","role":"assistant","content":"The Taiwanese foundry that fabricates virtually every advanced AI chip on earth — Nvidia's, AMD's, Google's, Amazon's alike. The single greatest concentration of risk in the AI supply chain, and the reason Taiwan sits at the center of technology geopolitics.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-13","messages":[{"id":"msg-mrxl4f9k-12-0-4rtor","role":"assistant","content":"Learning by trial, error, and reward rather than from labeled examples. RL powered AlphaGo's superhuman Go play and now shapes language models through RLHF and RLVR. It is the paradigm of choice whenever the desired behavior can be scored but not demonstrated.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784827991122-pv9axd2","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:33:11.122Z","source":"ai"},{"id":"msg-1784827996919-auztj61","role":"assistant","content":"The shift in reinforcement learning (RL) from its origins in game-tree search to its current role as the final arbiter of large language model behavior marks a fundamental change in how we define \"correctness\" in AI. Historically, RL was the engine of mastery in closed-system environments like Go. In that context, the reward signal is objective and terminal: you win or you lose. DeepMind’s success with AlphaGo demonstrated that by using deep neural networks to approximate value functions, an agent could discover strategic heuristics that exceeded human intuition.\n\nHowever, the application of RL to language—specifically through RLHF and RLVR—solves a different problem: the misalignment of supervised learning. A model trained strictly on next-token prediction learns to imitate the internet, which includes inaccuracies, biases, and a lack of stylistic focus. RLHF introduces a \"Reward Model\" that acts as a proxy for human judgment, moving the objective from mere imitation to \"helpfulness, honesty, and harmlessness.\"\n\nThis transition introduces a critical distinction in how RL is deployed today:\n\n*   **RLHF (Reinforcement Learning from Human Feedback)** deals with the subjective. It is essentially a preference-tuning mechanism. Because there is no \"correct\" way to write a poem or summarize a meeting, the reward signal is derived from human rankings. This makes the model more conversational but also risks \"reward hacking,\" where the model learns to sound convincing or overly polite to please the reward model rather than being factually accurate.\n*   **RLVR (Reinforcement Learning from Verifiable Rewards)** returns the paradigm to its objective roots. In domains like mathematics or programming, the reward is not a human opinion; it is a compiler’s output or a math checker's verification. This is the \"Reasoning Era\" frontier. By rewarding the model for reaching the correct final answer—regardless of the specific path taken—RLVR encourages the emergence of \"Chain-of-Thought\" reasoning and self-correction.\n\nDemis Hassabis and the teams at Google DeepMind are increasingly bridging these two worlds. The challenge now is moving beyond \"System 1\" thinking (fast, intuitive token generation) toward \"System 2\" thinking (slow, deliberate planning). The same RL principles that allowed AlphaGo to contemplate a move for minutes are being adapted to allow language models to \"think\" before they speak.\n\nA central tension in this exploration is the \"Data Wall.\" While supervised learning is limited by the amount of high-quality human text available, RL—particularly RLVR—offers a path toward recursive improvement. If a model can generate its own attempts and verify them against objective rules, it can potentially learn from synthetic data and trial-and-error, much as AlphaGo learned by playing against itself. This suggests that the future of the field may rely less on the scale of the initial dataset and more on the efficiency of the reward signals used during the post-training phase.","timestamp":"2026-07-23T17:33:16.919Z","source":"ai"},{"id":"msg-1784828012704-amk96e7","role":"user","content":"Explain the selected node: “Reinforcement Learning”.","timestamp":"2026-07-23T17:33:32.704Z","source":"ai"},{"id":"msg-1784828049893-3k7ftzd","role":"assistant","content":"Reinforcement Learning (RL) represents the transition from AI that mimics to AI that optimizes. While supervised learning allows a model to internalize the patterns of its training data, RL provides the framework for that model to refine its outputs based on the consequences of its actions. It is the bridge between a system that knows what words sound like and a system that knows which words achieve a goal.\n\nIn the current context of frontier AI development, RL is the primary driver of the \"reasoning\" breakthrough. This is a departure from the traditional \"predict the next word\" paradigm. By treating a sequence of thoughts as a series of moves in a game—much like the strategic pathfinding seen in AlphaGo—RL allows models to explore various reasoning paths, discarding those that lead to errors and reinforcing those that lead to a correct solution. This process is essential for overcoming the limitations of static datasets; it allows a model to discover logic that may not be explicitly spelled out in its training text.\n\nThe significance of RL in this scene is defined by two distinct feedback loops:\n\n*   **The Human Loop (RLHF):** This addresses the \"Alignment Problem.\" Because human values and conversational nuances cannot be easily captured in a mathematical loss function, RLHF uses human preference as the reward signal. This transforms a raw, sometimes erratic base model into a helpful assistant. It matters because it embeds social and functional constraints into the model’s probability distributions, ensuring the output is not just statistically likely, but utility-optimized.\n*   **The Verifiable Loop (RLVR):** This addresses the \"Hallucination Problem.\" In domains with objective ground truth, such as code or mathematics, RL permits the model to \"test\" its own hypotheses. If a generated code block fails to compile, the negative reward forces the model to iterate. This mimics the human process of trial and error, moving the technology toward autonomous problem-solving and long-horizon planning.\n\nThe presence of DeepMind and Demis Hassabis underscores that RL is no longer just a niche branch of machine learning; it is the strategic core of the effort to reach AGI. The trajectory here suggests that the future of intelligence lies in the move from \"System 1\" (instinctive, predictive) to \"System 2\" (deliberative, evaluative). Reinforcement Learning provides the mathematical machinery for this deliberation, turning the model from a passive repository of information into an active agent capable of self-correction and strategic thought.","timestamp":"2026-07-23T17:34:09.893Z","source":"ai"},{"id":"msg-1784828062131-xfchp5e","role":"user","content":"Find existing concepts that belong in this scene.","timestamp":"2026-07-23T17:34:22.131Z","source":"ai"},{"id":"msg-1784828069011-ynonkcz","role":"assistant","content":"To deepen the exploration of Reinforcement Learning as it relates to the current focus on frontier models and DeepMind’s trajectory, several existing concepts from your graph are particularly relevant:\n\n*   **PPO (Proximal Policy Optimization):** This is the industry-standard algorithm used for RLHF. It provides the mathematical stability needed to update a massive language model's weights without the training process collapsing, making it the \"engine\" behind the alignment of models like Gemini or GPT-4.\n*   **Reward Model:** In the context of RLHF, this is the specific component that \"learns\" human preferences and then acts as the judge for the reinforcement learning process. It is the crucial intermediary between human feedback and the final tuned model.\n*   **Chain-of-Thought:** This is the behavioral outcome the \"Reasoning Era\" (and RLVR) aims to produce. By rewarding correct answers, RL encourages models to externalize their \"thinking\" process, which is the linguistic equivalent of a search tree in game-playing AI.\n*   **Agentic Workflows:** This represents the next stage of RL application. While the current scene focuses on post-training (RLHF/RLVR), the ultimate goal is creating \"Agents\" that use RL to plan and execute multi-step tasks in the real world or on a computer.\n*   **Search and Optimization:** This connects back to the AlphaGo roots. Modern reasoning models often combine RL with search techniques (like Monte Carlo Tree Search) to explore the space of possible answers before committing to one.","timestamp":"2026-07-23T17:34:29.011Z","source":"ai"},{"id":"msg-1784828083604-gdpar87","role":"user","content":"Suggest new concepts to add to my knowledge graph.","timestamp":"2026-07-23T17:34:43.604Z","source":"ai"},{"id":"msg-1784828090726-faf4sw7","role":"assistant","content":"To expand the exploration of Reinforcement Learning beyond its current structural focus, several specialized concepts are missing that bridge the gap between \"game-tree search\" (AlphaGo) and \"modern alignment\" (RLHF). \n\nThese suggestions focus on the mechanics of how these models actually learn and the specific failure modes they encounter:\n\n*   **PPO (Proximal Policy Optimization):** While the scene mentions the methods (RLHF), it lacks the specific algorithm that made them viable. PPO is the industry-standard \"workhorse\" algorithm that allows for stable updates to large models without the training process diverging.\n*   **Reward Hacking:** This is a critical concept for understanding the risks of RL. It occurs when an agent finds a \"shortcut\" to get a high reward—such as a language model becoming overly sycophantic to please a human rater—without actually achieving the intended goal.\n*   **MCTS (Monte Carlo Tree Search):** This is the specific \"search\" component that allowed AlphaGo to look ahead. In the context of modern reasoning models, MCTS is being re-imagined as a way for LLMs to \"search\" through different reasoning paths before committing to an answer.\n*   **KL Divergence Constraint:** In RLHF, it is vital to ensure the model doesn't drift too far from its original linguistic capabilities while trying to maximize reward. This mathematical constraint acts as a \"tether\" to the base model's knowledge.\n*   **Credit Assignment Problem:** A fundamental challenge in RL where the agent must determine which specific action in a long sequence (like a 50-step math proof) was actually responsible for the final success or failure.","timestamp":"2026-07-23T17:34:50.726Z","source":"ai"},{"id":"msg-1784828148495-444bk0q","role":"user","content":"Who is the inventor / main researcher / the father or reinforcement learning?","timestamp":"2026-07-23T17:35:48.495Z","source":"ai"},{"id":"msg-1784828153755-92gcobo","role":"assistant","content":"While many researchers contributed to the field, the \"father\" of modern reinforcement learning is widely considered to be **Richard Sutton**. Together with his doctoral advisor, **Andrew Barto**, he formalized the field in the 1980s and 1990s, bridging the gap between animal learning psychology and optimal control theory.\n\nTheir seminal work established the computational framework we use today:\n\n*   **The Temporal Difference (TD) Learning:** Sutton’s most influential contribution was the development of TD learning. Before this, systems often had to wait until the end of an episode (like the end of a game) to learn. TD learning allowed agents to learn from the \"difference\" between successive predictions, essentially updating their understanding of the world in real-time. This is the underlying logic that eventually enabled the success of systems like AlphaGo.\n*   **The \"Bibles\" of the Field:** Sutton and Barto co-authored *Reinforcement Learning: An Introduction*, which remains the foundational textbook for the discipline. It defined the standard vocabulary of agents, environments, states, and rewards.\n\nWhile Sutton and Barto provided the formal framework, the field has other \"founding\" figures who contributed specific breakthroughs:\n*   **Chris Watkins:** Developed **Q-Learning** in 1989, which provided a provably effective way for agents to learn the value of actions in an environment without needing a prior model of how that environment works.\n*   **Gerald Tesauro:** Created **TD-Gammon** in the early 90s, which was the first high-profile success of RL. It used TD learning to play Backgammon at a world-class level, serving as the direct spiritual ancestor to the DeepMind breakthroughs represented by AlphaGo.\n\nIncluding Richard Sutton provides the missing link between the conceptual origins of trial-and-error learning and the high-scale implementations led by figures like Demis Hassabis today.","timestamp":"2026-07-23T17:35:53.755Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:35:53.756Z"},{"nodeId":"n-mermaid-mrxl4f8r-130","messages":[{"id":"msg-mrxl4f9k-129-0-lqzdn","role":"assistant","content":"The specialized GPU cloud that rose from crypto-mining origins to become a major AI-compute provider and 2025's marquee AI IPO. Its debt-financed, Nvidia-entangled model makes it the purest public-market bet on sustained training demand — watched as a bellwether accordingly.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-131","messages":[{"id":"msg-mrxl4f9k-130-0-gvphm","role":"assistant","content":"Chipmaker whose LPU inference processors deliver dramatic token-generation speed, betting that the inference market rewards architectures built for serving rather than training. Part of the wave arguing the accelerator market is bigger than one company's roadmap.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-132","messages":[{"id":"msg-mrxl4f9k-131-0-3cdaa","role":"assistant","content":"The platform giants and infrastructure companies through which models reach developers and enterprises: hyperscale clouds, model hubs, API aggregators, and data suppliers. Distribution is their power — most AI consumption flows through this layer.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-133","messages":[{"id":"msg-mrxl4f9k-132-0-9xzt8","role":"assistant","content":"The incumbent that moved first and hardest: the OpenAI partnership secured early frontier access, Copilot threaded AI through Office and Windows, GitHub Copilot created the AI-coding category, and Azure monetizes the whole stack. Under Nadella and Suleyman, the case study in incumbent advantage compounding.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-134","messages":[{"id":"msg-mrxl4f9k-133-0-wjobh","role":"assistant","content":"The company that invented the transformer, then found itself cast as the disrupted incumbent when ChatGPT threatened search. Its counterattack spans the full stack — Gemini models, TPU silicon, AI Overviews rebuilt into search — a vertical integration no rival matches, shadowed by the innovator's dilemma of cannibalizing its own ad economics.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-135","messages":[{"id":"msg-mrxl4f9k-134-0-equf7","role":"assistant","content":"The largest cloud, playing AI as an arms merchant: Bedrock offers every major model family through one marketplace, Trainium chips hedge the Nvidia dependence, and a multibillion-dollar Anthropic partnership secures its frontier stake. Less flashy than rivals; structurally positioned to profit regardless of which lab wins.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-136","messages":[{"id":"msg-mrxl4f9k-135-0-kwnpv","role":"assistant","content":"The GitHub of machine learning — the hub where open-weight models, datasets, and libraries are published and consumed, hosting millions of models. The de facto infrastructure of the open ecosystem and the place where \"open-weight release\" concretely happens.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-137","messages":[{"id":"msg-mrxl4f9k-136-0-ws3yz","role":"assistant","content":"The aggregation layer for model APIs: one interface, unified billing, and automatic routing across hundreds of models from every provider. Its usage dashboards double as the industry's most-watched real-time market share signal — and its existence marks models becoming interchangeable commodities.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-138","messages":[{"id":"msg-mrxl4f9k-137-0-029kr","role":"assistant","content":"Cloud platform specialized in serving and fine-tuning open-weight models at speed, co-founded by researchers behind key efficiency methods. Part of the infrastructure layer that makes \"open model via API\" a practical enterprise default rather than a hobbyist path.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-139","messages":[{"id":"msg-mrxl4f9k-138-0-5d4la","role":"assistant","content":"The data-labeling and evaluation company founded by Alexandr Wang that supplied the human-feedback layer of the RLHF era. Meta's 2025 multibillion-dollar investment — which pulled Wang into Meta and shook client relationships — turned a supplier into a strategic asset and revealed how valuable the data layer had become.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-14","messages":[{"id":"msg-mrxl4f9k-13-0-2a1zt","role":"assistant","content":"Networks of simple numerical units whose connection strengths are adjusted during training. Loosely inspired by neurons, they are best understood as very large adjustable functions. After decades on the margins, they became the substrate of essentially all modern AI.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784824576583-gv9dmii","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:36:16.583Z","source":"ai"},{"id":"msg-1784824583714-3shq3kt","role":"assistant","content":"The conceptual core of neural networks lies in the transition from symbolic, rule-based logic to a distributed, parallel architecture. This shift, championed by Connectionism, posits that intelligence is not found in a centralized processor but emerges from the collective interaction of simple units. The fundamental challenge of this approach—how such a system can learn from its own errors—was solved by the formalization of backpropagation. By calculating the gradient of a loss function with respect to every weight in the system, backpropagation allows for the incremental optimization of high-dimensional functions, transforming a static network into a dynamic learning system.\n\nA significant tension in the history of the field is the scale at which these networks operate. While early connectionist models were often shallow, focusing on small-scale cognitive tasks like the XOR problem or word recognition (as seen in NETtalk), the move toward Deep Learning shifted the emphasis to hierarchical abstraction. In deep architectures, the initial layers extract rudimentary features—such as edges or textures—while deeper layers compose these into complex semantic representations. This hierarchy is what allows a network to generalize from raw data to abstract concepts without human-defined feature engineering.\n\nJürgen Schmidhuber’s contributions highlight a crucial branch of this evolution: the move from feedforward processing to temporal and recursive memory. His work on Long Short-Term Memory (LSTM) addressed the \"vanishing gradient\" problem, a mathematical hurdle where error signals would effectively disappear as they were propagated back through many layers or time steps. This ensured that neural networks could maintain information over long sequences, a precursor to the massive context windows exploited by contemporary architectures.\n\nOne common confusion to resolve is the distinction between a neural network as a mathematical structure and the training regime applied to it. Machine Learning provides the broader framework—defining the data, the objective, and the optimization strategy (like Gradient Descent)—while the neural network serves as the specific \"hypothesis space\" or functional form. The efficacy of modern systems is less about the \"neurons\" themselves and more about the \"Induction Heads\" and \"Scaling Laws\" that emerge when these networks are pushed to extreme sizes.\n\nTo deepen this exploration, it is worth considering the following:\n*   **Optimization vs. Generalization:** We know how to minimize error on training data (optimization), but the \"black box\" nature of neural networks makes it difficult to prove they will behave correctly on unseen data (generalization). \n*   **The Data Wall:** As models grow, we are approaching a limit on high-quality human-generated data, leading to a shift toward synthetic data and self-supervised learning.\n*   **Mechanistic Interpretability:** While we can build these systems, we are still developing the tools to understand *why* a specific circuit within a network makes a certain decision.\n\nFor a more complete picture of how these models are structured and optimized today, it would be useful to look at the specific mathematical mechanics of the **Attention Mechanism** and the broader **Transformer Architecture**, as these have largely superseded the recurrent designs of the earlier deep learning era. Understanding how a model evaluates its own performance through a **Loss Function** and navigates the landscape of parameters via **Gradient Descent** would also bridge the gap between the philosophical roots of connectionism and the practical reality of modern AI.","timestamp":"2026-07-23T16:36:23.714Z","source":"ai"},{"id":"msg-1784824631570-im2fs0q","role":"user","content":"Give me a general, encyclopedic definition of the Neural Network. Comprehensive and multi-angled view.","timestamp":"2026-07-23T16:37:11.570Z","source":"ai"},{"id":"msg-1784824638601-inwba0h","role":"assistant","content":"A **neural network** (more formally, an **artificial neural network** or **ANN**) is a computational model inspired by the structural and functional aspects of biological neural networks. It consists of a series of algorithms that attempt to recognize underlying relationships in a set of data through a process that mimics the way the human brain operates.\n\nAt its core, a neural network is a mathematical function that maps inputs to outputs. It is composed of interconnected units called **neurons** (or nodes), typically organized into layers: an **input layer**, one or more **hidden layers**, and an **output layer**.\n\n### 1. Functional Mechanics\nEach neuron in a network receives signals from others, processes them, and signals the neurons connected to it. The \"signal\" at a connection is a real number, and the output of each neuron is computed by some non-linear function of the sum of its inputs.\n*   **Weights and Biases:** Connections between neurons have associated **weights** that adjust as learning proceeds. The weight increases or decreases the strength of the signal at a connection. Neurons also typically have a **bias**, an adjustable parameter that allows the activation function to be shifted.\n*   **Activation Functions:** These functions (such as ReLU, Sigmoid, or Tanh) determine whether and to what extent a neuron \"fires,\" introducing the non-linearity necessary for the network to learn complex patterns.\n\n### 2. The Learning Process\nThe primary advantage of neural networks is their ability to learn from observational data without being programmed with specific rules.\n*   **Training:** This is the process of adjusting weights and biases to minimize the difference between the network's output and the desired target.\n*   **Backpropagation:** The standard algorithm for training, which uses the chain rule from calculus to calculate the gradient of a **Loss Function** with respect to each weight.\n*   **Optimization:** Techniques like **Gradient Descent** are used to update the weights in the direction that reduces error.\n\n### 3. Historical and Philosophical Context\nNeural networks are the cornerstone of **Connectionism**, a movement in cognitive science that began in the 1940s and 50s. Unlike \"Symbolic AI,\" which relies on hard-coded logic and \"if-then\" statements, neural networks operate on the principle of **Parallel Distributed Processing (PDP)**. Knowledge is not stored in a single location but is distributed across the strengths of the connections throughout the entire system.\n\n### 4. Categorization and Architecture\nModern neural networks are categorized by their architecture and how data flows through them:\n*   **Feedforward Neural Networks (FNN):** Data moves in only one direction—from input to output.\n*   **Convolutional Neural Networks (CNN):** Specialized for processing data with a grid-like topology, such as images, by using \"filters\" to identify spatial hierarchies.\n*   **Recurrent Neural Networks (RNN):** Designed for sequential data (like speech or text) by using loops to maintain a form of internal memory.\n*   **Transformers:** The current state-of-the-art architecture that uses **Attention Mechanisms** to weigh the significance of different parts of the input data simultaneously, enabling massive scaling.\n\n### 5. Mathematical Significance\nMathematically, neural networks are often described as **Universal Approximators**. The Universal Approximation Theorem states that a feedforward network with a single hidden layer containing a finite number of neurons can approximate any continuous function, provided the activation function is non-linear. This makes them exceptionally powerful for tasks involving high-dimensional data, such as natural language processing, computer vision, and autonomous systems.\n\n### 6. Limitations and Challenges\nDespite their power, neural networks face several challenges:\n*   **Interpretability:** Often called \"black boxes,\" it is difficult for humans to understand exactly why a deep network made a specific decision.\n*   **Data Dependency:** They require vast amounts of labeled data to achieve high performance.\n*   **Computational Cost:** Training large-scale models requires significant energy and specialized hardware like GPUs or TPUs.","timestamp":"2026-07-23T16:37:18.601Z","source":"ai"},{"id":"msg-1784824651388-f6uekg8","role":"user","content":"Suggest new concepts to add to my knowledge graph.","timestamp":"2026-07-23T16:37:31.388Z","source":"ai"},{"id":"msg-1784824654984-51qeabf","role":"assistant","content":"To expand the conceptual foundation of neural networks beyond the historical and general terms currently in your graph, it is useful to look at the specific mathematical constraints and biological inspirations that define how these systems function.\n\nOne critical concept is the **Universal Approximation Theorem**, which provides the mathematical proof that neural networks are capable of representing almost any function, given enough parameters. This serves as the theoretical justification for the entire field.\n\nFrom a structural standpoint, the **Vanishing Gradient Problem** is the primary obstacle that historically limited the depth of these networks, leading directly to the architectural innovations found in modern deep learning. Similarly, the **Bias-Variance Tradeoff** is the fundamental tension in training any network: the struggle to make a model complex enough to capture the data (low bias) without making it so specific that it fails to generalize to new information (high variance).\n\nFinally, looking toward the biological roots of the field, **Hebbian Theory** (\"neurons that fire together, wire together\") provides the original inspiration for how connection strengths might change in a physical or simulated system, bridging the gap between neuroscience and the numerical weights used in backpropagation.","timestamp":"2026-07-23T16:37:34.984Z","source":"ai"},{"id":"msg-1784824706633-tjfqbxl","role":"user","content":"How is it (Neural networks) related to cognitive science research? Brain research? Why are they called 'Neural'/","timestamp":"2026-07-23T16:38:26.633Z","source":"ai"},{"id":"msg-1784824713560-bw6i1kg","role":"assistant","content":"The term **Neural** is a legacy of the 1940s and 50s, when pioneers like Warren McCulloch and Walter Pitts sought to create a mathematical calculus for the \"activity of the nervous system.\" They were called \"neural\" because the basic unit—the perceptron or artificial neuron—was designed as an abstraction of a biological neuron: it receives multiple signals, weights their importance, sums them, and \"fires\" an output signal if a certain threshold is met.\n\nWhile modern neural networks have drifted far from biological realism to prioritize engineering efficiency, the relationship between these models and cognitive science remains a foundational feedback loop.\n\n### 1. The Connectionist Hypothesis\nIn cognitive science, neural networks are the primary vehicle for **Connectionism**. This school of thought argues against the \"computer metaphor\" of the mind (where the brain is a processor running symbolic software). Instead, it suggests that mental phenomena can be described by interconnected networks of simple units. This led to the concept of **Parallel Distributed Processing (PDP)**, which explains how the brain can perform tasks like facial recognition almost instantly—something that would take a serial, step-by-step symbolic processor much longer.\n\n### 2. Biological Plausibility vs. Engineering\nA major point of discussion in brain research is the \"biological plausibility\" of AI. For example:\n*   **Backpropagation:** Many neuroscientists argue that the brain does not actually use backpropagation because there is no known biological mechanism for sending error signals backward across synapses in the exact way modern AI does.\n*   **Hebbian Theory:** This is much more biologically grounded. It posits that learning occurs through \"synaptic plasticity\"—strengthening the connection between two neurons when they are active at the same time. This is often summarized as *\"cells that fire together, wire together.\"*\n\n### 3. Neural Networks as \"Silicon Model Organisms\"\nToday, neuroscientists use artificial neural networks as a sandbox to test theories about the human brain. By training a **Convolutional Neural Network** on images and comparing its internal activations to the firing patterns in a macaque’s visual cortex, researchers have found striking similarities. This suggests that both biological evolution and gradient descent may be arriving at similar mathematical solutions for processing the physical world.\n\n### 4. Graceful Degradation\nOne of the most significant links to cognitive science is the concept of **Graceful Degradation**. In symbolic AI, a single corrupted line of code can crash the entire system. In both human brains and neural networks, because information is distributed across millions of connections, the system can lose a few neurons or encounter \"noisy\" data and still function effectively. This mirrors the brain's resilience to minor injury and its ability to operate in ambiguous, real-world environments.\n\nTo further explore these bridges between the biological and the artificial, the following concepts are essential:","timestamp":"2026-07-23T16:38:33.560Z","source":"ai"},{"id":"msg-1784824867379-wrztfkw","role":"user","content":"Give me a node pointing to the biological/neural perspective on neural networks. I want to establish the link with neurosciense","timestamp":"2026-07-23T16:41:07.379Z","source":"ai"},{"id":"msg-1784824870714-ecxmdx5","role":"assistant","content":"To bridge the gap between the mathematical models in your scene and the biological reality of the brain, the most critical concept is **Synaptic Plasticity**. This is the actual biological mechanism behind the \"adjustable connection weights\" mentioned in your central node's definition. \n\nIn neuroscience, learning isn't just an abstract adjustment of numbers; it is the physical strengthening or weakening of synapses in response to increases or decreases in their activity. This provides the biological basis for **Hebbian Theory** and serves as the natural counterpart to the **Backpropagation** algorithm. While backpropagation is an engineering solution for global optimization, synaptic plasticity (often through mechanisms like Long-Term Potentiation) is how the brain achieves local learning.\n\nConnecting this to the **Connectionism** and **Parallel Distributed Processing (PDP)** nodes already in your scene helps explain **Graceful Degradation**—the biological phenomenon where a system remains functional even when individual components (neurons) fail, a property that artificial neural networks successfully inherit from their biological inspirations.","timestamp":"2026-07-23T16:41:10.714Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:41:10.715Z"},{"nodeId":"n-mermaid-mrxl4f8r-140","messages":[{"id":"msg-mrxl4f9k-139-0-v7ute","role":"assistant","content":"Companies building products on top of foundation models — search, coding, image, video, voice, companionship. The layer that tests whether durable businesses can be built above the model providers, or whether the labs will absorb every successful category.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-141","messages":[{"id":"msg-mrxl4f9k-140-0-pndjl","role":"assistant","content":"The AI search company that proved answers-with-citations could challenge the link economy, growing from wrapper to household name at multibillion-dollar valuations. Its aggressive moves — publisher disputes, browser plays, a headline-grabbing bid for Chrome — track the ambition of the application layer.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-142","messages":[{"id":"msg-mrxl4f9k-141-0-84r5h","role":"assistant","content":"The AI-native code editor whose explosive adoption made it the fastest-growing developer product in memory, and the flagship proof that a startup can build a category-defining product on top of others' foundation models. The application-layer bull case, incarnate.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-143","messages":[{"id":"msg-mrxl4f9k-142-0-350g4","role":"assistant","content":"The image-generation lab whose aesthetic became synonymous with AI art — bootstrapped, profitable, and run by a small team inside Discord, defying every rule of AI capital intensity. Also a lightning rod: its artist-style training practices anchor the copyright fights over generative AI.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-144","messages":[{"id":"msg-mrxl4f9k-143-0-3icgi","role":"assistant","content":"The generative-video pioneer whose tools moved AI video from novelty into film, advertising, and Hollywood pre-production workflows — with industry partnerships to match. The company most identified with the creative disruption argument, in both its promise and its threat.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-145","messages":[{"id":"msg-mrxl4f9k-144-0-my38g","role":"assistant","content":"The voice-AI leader whose synthesis and cloning quality carried it to multibillion-dollar valuations and audiobook, media, and agent deployments. Also the uncomfortable case study in dual use: its own technology anchors deepfake-audio concerns and the consent norms now developing around voice.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-146","messages":[{"id":"msg-mrxl4f9k-145-0-lc97a","role":"assistant","content":"The companion chatbot platform whose enormous engagement — hours-long daily sessions, heavily young users — made it the poster child of the parasocial AI category and its dangers. Wrongful-death litigation and safety redesigns around minors made it the reference case for AI companionship regulation; Google's licensing acqui-hire of its founders was among the deals defining the talent-absorption era.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-147","messages":[{"id":"msg-mrxl4f9k-146-0-67ctj","role":"assistant","content":"The gigawatt-scale facilities where frontier models are trained and served — capital projects now measured in tens of billions of dollars, with electricity as the binding constraint. Datacenter buildout has become a macroeconomic story: power grids, nuclear restarts, and national industrial strategy.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784827224262-m6r8hj2","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:20:24.262Z","source":"ai"},{"id":"msg-1784827230755-dmg1t39","role":"assistant","content":"The shift toward gigawatt-scale AI datacenters represents a transition from traditional enterprise computing to a form of industrial-scale heavy infrastructure. This evolution is governed by the rigid predictability of scaling laws, which dictate that the frontier of machine intelligence is essentially a function of the total training compute applied to a model. Because this compute is a product of hardware efficiency, time, and power, the datacenter has become the physical manifestation of an intellectual race.\n\nThe massive capital flows characterizing the current investment boom are directed toward a very specific hardware-software vertical. At the base sits Nvidia, providing the essential H100 and Blackwell architectures and the CUDA software moat that makes this hardware programmable. However, the sheer scale of these projects has necessitated new players in the stack. Specialized providers like CoreWeave have emerged to bridge the gap between chip manufacturing and model training, offering a more agile, GPU-centric cloud than the legacy hyperscalers.\n\nThis massive deployment of capital raises a fundamental tension between the \"Gold Rush\" mentality and the \"Bubble Question.\" The economic viability of these datacenters rests on whether the resulting model capabilities can generate enough revenue—either through labor replacement or new category creation—to amortize the multi-billion dollar costs of a single training cluster. If the returns on intelligence begin to marginalize, the industry faces a massive overcapacity of specialized, power-hungry infrastructure.\n\nThe ultimate physical bottleneck, however, is not just the supply of silicon, but energy demand. We are moving from an era where datacenters were located near fiber-optic hubs to an era where they are located near high-capacity power generation. The move toward \"gigawatt clusters\"—exemplified by projects like xAI's Colossus—places AI labs in direct competition with residential and industrial grids. This constraint is forcing the industry to look beyond traditional utilities toward modular nuclear reactors, direct grid-tie agreements, and massive onsite storage.\n\nA critical dimension to monitor is the divergence between training and inference. While the current focus is on building \"Model Factories\" for massive training runs, the long-term economics of the industry will be decided by the efficiency of serving these models. If the cost of inference remains high, the scale of these datacenters may eventually shift from being a competitive advantage to a significant operational burden.\n\nTo further map out the structural constraints and geopolitical stakes of this infrastructure, considering the following factors would be useful:\n\n*   **TSMC:** The foundry layer is the ultimate upstream bottleneck for all the hardware mentioned; without their advanced process nodes, the roadmap for Nvidia and others stalls.\n*   **Export Controls:** These define the geographic boundaries of where this infrastructure can be built, turning AI datacenters into assets of national security and high-stakes diplomacy.\n*   **Inference Cost:** This is the primary metric that will determine if the investment boom matures into a sustainable economy or remains a speculative peak.","timestamp":"2026-07-23T17:20:30.755Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:20:30.756Z"},{"nodeId":"n-mermaid-mrxl4f8r-148","messages":[{"id":"msg-mrxl4f9k-147-0-emh3q","role":"assistant","content":"The per-token economics of serving models — the number that decides which applications are viable. Prices have fallen orders of magnitude since 2023 through MoE, distillation, quantization, and competition; reasoning models push the other way by consuming far more tokens per answer.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-149","messages":[{"id":"msg-mrxl4f9k-148-0-05icx","role":"assistant","content":"The capital wave defining the AI era: hundreds of billions annually in hyperscaler capex, record private rounds for labs, and infrastructure commitments — datacenters, chips, power — on a scale usually associated with national projects. Whether the returns justify the outlay is the Bubble Question; that the money is really being spent is not in doubt.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784827261611-wr1inng","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:21:01.611Z","source":"ai"},{"id":"msg-1784827268748-lyg99td","role":"assistant","content":"The current capital deployment into artificial intelligence represents a historic concentration of wealth and industrial capacity, characterized by a fundamental shift from software-based scaling to hardware-heavy industrialization. Unlike previous technology cycles that relied on capital-light software distribution, the present expansion is anchored in the physical reality of massive energy consumption and the physical scarcity of high-end silicon.\n\nAt the heart of this movement is the build-out of a new tier of infrastructure. These facilities are no longer traditional enterprise data centers; they are increasingly being designed as singular, coherent computing units—effectively massive supercomputers dedicated to the synthesis of \"intelligence\" as a commodity. This shift necessitates a massive upfront investment in energy infrastructure, often requiring direct integration with power grids or the commissioning of dedicated nuclear and geothermal sources to meet the gigawatt-scale requirements of the next generation of model training.\n\nThis leads to the central tension of the current era: the disconnect between massive infrastructure spending and proven revenue at the application layer. The capital flows suggest a \"build it and they will come\" conviction, predicated on the belief that intelligence scales predictably with compute. However, this creates a precarious economic lag. While the demand for the underlying compute (the chips and the power) is immediate and tangible, the economic utility of the models being produced is still being discovered. \n\nA critical question for this period is whether we are witnessing a \"S-curve\" in productivity or a \"J-curve\" in investment. If the returns on intelligence are non-linear, the current spending is a rational attempt to capture a dominant position in a new foundational utility. If, however, the marginal utility of additional compute begins to diminish—or if the cost of serving these models remains prohibitively high compared to traditional methods—the industry faces a significant correction. \n\nThis environment is further complicated by the geopolitical dimension of the supply chain. Because the ability to participate in this boom is gated by access to specific fabrication capabilities and specialized equipment, the investment is not just about capital, but about the strategic positioning of nations and a handful of firms that control the bottleneck. The \"Gold Rush\" is therefore as much about securing future sovereignty and technological autonomy as it is about immediate quarterly returns.\n\nTo deepen this exploration, it may be useful to bridge the gap between the physical infrastructure and the actual cost of running these systems at scale.\n\n**Related existing concepts to consider including:**\n- **Inference Cost:** The ongoing expense of serving models, which determines the long-term sustainability of the investment once the initial training phase is complete.\n- **Compute & Silicon:** The foundational layer that governs the speed and scale of the infrastructure build-out.\n- **Energy Demand:** The primary physical constraint on the expansion of facilities and a major driver of site selection and regional economic impact.\n- **Scaling Laws:** The theoretical justification for the massive capital expenditure, based on the observed relationship between compute, data, and performance.","timestamp":"2026-07-23T17:21:08.748Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:21:08.749Z"},{"nodeId":"n-mermaid-mrxl4f8r-15","messages":[{"id":"msg-mrxl4f9k-14-0-c899p","role":"assistant","content":"Neural networks with many layers, which learn hierarchies of features — edges to shapes to objects, or characters to words to meaning. The 2012 ImageNet result made \"deep\" the defining adjective of the modern era.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-150","messages":[{"id":"msg-mrxl4f9k-149-0-rr86u","role":"assistant","content":"The binding physical constraint of the buildout: AI datacenters demand gigawatts, forcing utility deals, nuclear plant restarts, and grid expansion onto the industry's critical path. Energy has become an AI strategy topic — and AI a topic in energy policy — in both directions at once.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-151","messages":[{"id":"msg-mrxl4f9k-150-0-e7ny9","role":"assistant","content":"The US restrictions on advanced-chip exports to China — compute as a geopolitical instrument. Their effectiveness is contested: DeepSeek's efficient frontier-class models became the central exhibit in the argument that constraints breed ingenuity.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-152","messages":[{"id":"msg-mrxl4f9k-151-0-vweqx","role":"assistant","content":"The developing rulebook for AI: risk regulation, safety-testing commitments, disclosure requirements, and the international politics of compute. Governance vocabulary now appears in earnings calls as often as in policy papers.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784826478920-huqqoe2","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:07:58.920Z","source":"ai"},{"id":"msg-1784826489896-itl213s","role":"assistant","content":"AI governance is currently navigating a fundamental shift from high-level ethical frameworks to enforceable, technical mandates. This transition is driven by a widening recognition that the dual nature of these systems—their potential to revolutionize scientific discovery while simultaneously threatening democratic stability or physical safety—cannot be managed through voluntary industry self-regulation alone.\n\nAt the heart of modern regulatory efforts, such as the EU AI Act, is a risk-tiered philosophy. This approach attempts to codify \"peril\" by categorizing applications based on their impact on fundamental rights and safety. However, a significant tension exists between application-based regulation and model-based regulation. While the EU focuses largely on how an AI is *used* (e.g., in hiring or law enforcement), safety advocates increasingly argue that the underlying capabilities of frontier models themselves require oversight. This is particularly relevant as we approach \"frontier\" levels of capability where emergent behaviors—capabilities not intentionally programmed or predicted during training—become more frequent.\n\nThe advocacy of figures like Yoshua Bengio represents a pivotal change in the scientific community. The shift from focusing purely on performance and architecture to leading international safety assessments signals that the \"black box\" nature of deep learning is no longer just a technical hurdle but a primary governance challenge. If we cannot explain the internal logic of a model (a goal of mechanistic interpretability), governing its output becomes a reactive rather than a proactive process.\n\nA critical, unresolved question in this domain is the \"compute threshold.\" Current international discussions often suggest using the amount of floating-point operations (FLOPs) used in training as a proxy for a model's potential power and, therefore, its need for oversight. Yet, this is a moving target; as algorithmic efficiency improves, the amount of compute required to reach a certain capability level drops, potentially rendering hardware-based regulations obsolete.\n\nFurthermore, the divide between open-weight models and closed-system APIs presents a unique enforcement dilemma. Governance frameworks that work for a centralized entity like OpenAI or Google DeepMind may not translate to a world where high-capability models are decentralized and modified by third parties. This raises the stakes for liability: determining whether the developer of the foundation model or the fine-tuner is responsible for a catastrophic failure is a legal frontier that remains largely unmapped.\n\nAs these laws phase in, the focus will likely shift from *what* is being built to *how* it is verified. We are moving toward a regime of mandatory red-teaming and third-party auditing, where the burden of proof is on the developer to demonstrate that a system is safe before it is deployed at scale. This mirrors the evolution of the aviation and pharmaceutical industries, where the \"promise\" of the product is only realized once the \"peril\" is systematically mitigated through rigorous, standardized testing.\n\nTo deepen this exploration, it may be worth considering how these governance structures interact with the physical realities of the AI supply chain, such as export controls and data center energy requirements, which act as de facto levers for state-level oversight.","timestamp":"2026-07-23T17:08:09.896Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:08:09.896Z"},{"nodeId":"n-mermaid-mrxl4f8r-153","messages":[{"id":"msg-mrxl4f9k-152-0-5uavj","role":"assistant","content":"The European Union's comprehensive, risk-tiered AI law — the first of its kind, phasing in through 2025–2027. Praised as precedent and criticized as burden, it is the reference point every other jurisdiction defines itself against.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-154","messages":[{"id":"msg-mrxl4f9k-153-0-vobe6","role":"assistant","content":"The oldest AI anxiety, now concrete: coding assistants, customer-service agents, and document automation are changing entry-level and knowledge work first, inverting the old prediction that manual labor would fall to machines before cognitive work. The honest state of evidence is contested — productivity gains are real, mass displacement so far is not — and the gap between those two facts is where the policy debate lives.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-155","messages":[{"id":"msg-mrxl4f9k-154-0-vwi13","role":"assistant","content":"The information-environment problem: synthetic text, images, voices, and video that are cheap, convincing, and scalable. Deepfake fraud and election-manipulation concerns drove the first AI-content laws and watermarking standards; the deeper worry is the liar's dividend — when anything can be faked, everything can be denied.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-156","messages":[{"id":"msg-mrxl4f9k-155-0-5zmgf","role":"assistant","content":"The rise of AI as a social relationship: companion apps with hours-long daily engagement, assistants remembered and mourned when deprecated, chatbots as confidants for the young and isolated. The most personal frontier of AI's societal impact, spanning genuine comfort and documented harm — and a regulatory category being invented in real time around minors.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-157","messages":[{"id":"msg-mrxl4f9k-156-0-nvxiy","role":"assistant","content":"The founding paradigm — intelligence as explicit rules, logic, and symbol manipulation, dominant from the 1950s through the 1980s. Its expert systems ultimately proved brittle, but its legacy persists in the recurring argument that neural networks need symbolic structure grafted back in.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-158","messages":[{"id":"msg-mrxl4f9k-157-0-ovpr8","role":"assistant","content":"The rival school holding that intelligence should be learned by networks of simple units, not programmed as rules. Marginalized for decades, sustained by researchers like Hinton, and ultimately victorious — modern AI is connectionism at scale.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784824103543-on25q6m","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:28:23.543Z","source":"ai"},{"id":"msg-1784824111232-w1m9xp7","role":"assistant","content":"The core premise of connectionism is that intelligence does not emerge from the manipulation of high-level symbols according to logical rules, but rather from the collective behavior of simple, interconnected processing units. In the broader sweep of AI history, this school of thought represents a biological metaphor for cognition, contrasting sharply with the \"Physical Symbol System Hypothesis\" that dominated the early decades of the field.\n\nWhile symbolic AI (or GOFAI) viewed the mind as a computer executing a program, connectionism views it as a network adjusting its internal weights to find patterns in data. This shift moves the burden of \"knowledge\" from the programmer to the system itself. Instead of explicitly defining what a \"cat\" is through a set of necessary and sufficient conditions, a connectionist system learns a distributed representation of a cat across thousands of simulated neurons through exposure to examples.\n\nTo deepen this exploration, it is useful to consider several pivotal concepts and historical markers that define the connectionist trajectory:\n\n### The Architecture of Learning\nThe fundamental unit of this approach is the neural network. Early attempts, such as the Perceptron, demonstrated that simple networks could learn to classify linear patterns. However, the connectionist movement faced a major setback when it was shown that single-layer networks could not solve non-linear problems like XOR. The field's eventual resurgence was driven by the discovery and refinement of backpropagation, which allowed for the training of multi-layer (deep) networks. This algorithm uses gradient descent to minimize a loss function, effectively \"blaming\" specific weights for errors and adjusting them accordingly.\n\n### Parallel Distributed Processing (PDP)\nA defining characteristic of connectionist systems is that they are inherently parallel. Unlike the serial processing of the von Neumann architecture used in standard computers, connectionist models perform many small computations simultaneously. This provides a natural explanation for human cognitive speed: we can recognize a face in milliseconds despite the relatively slow firing rate of biological neurons because the computation is distributed across the entire network rather than processed step-by-step.\n\n### The Great Divergence and Re-convergence\nFor decades, a tension existed between connectionism and symbolic AI. Connectionism was often criticized for its \"black box\" nature—it was difficult to explain *why* a network reached a certain conclusion, a problem that persists today in the field of interpretability. Conversely, symbolic systems struggled with the \"brittleness\" of rules and the difficulty of handling noisy, real-world data. The current era of Large Language Models represents a temporary triumph for connectionism, yet it has sparked a renewed interest in neuro-symbolic integration—the idea that perhaps true reasoning requires the pattern-recognition strengths of neural networks paired with the logical rigor of symbolic systems.\n\n### Key Milestones for Further Investigation\n*   **The Perceptron:** The 1958 starting point for supervised learning.\n*   **Backpropagation:** The mathematical engine that allowed connectionism to scale beyond simple toy problems.\n*   **Deep Learning Revolution:** The transition from theoretical connectionism to the dominant paradigm of modern technology, catalyzed by increases in compute and the availability of massive datasets like ImageNet.\n\nThe long road of AI history is essentially the story of this pendulum swing between the desire to program intelligence and the desire to grow it. Connectionism’s current dominance suggests that the complexity of the world is better captured by flexible, high-dimensional manifolds than by rigid hierarchies of logic.","timestamp":"2026-07-23T16:28:31.232Z","source":"ai"},{"id":"msg-1784824159879-iqzlyva","role":"user","content":"Who are the key people in the history of Connectionism? Rumelhart generation?","timestamp":"2026-07-23T16:29:19.879Z","source":"ai"},{"id":"msg-1784824168576-c6y22kn","role":"assistant","content":"The mid-1980s represent the \"second wave\" of connectionism, a period defined by the Parallel Distributed Processing (PDP) Research Group. This group, led by **David Rumelhart** and **James McClelland**, effectively rescued the field from the \"winter\" imposed by the limitations of early simple perceptrons. Their 1986 two-volume manifesto, *Parallel Distributed Processing: Explorations in the Microstructure of Cognition*, provided the theoretical and mathematical foundation for everything we now recognize as modern AI.\n\n### The Rumelhart Generation and the PDP Group\nThis generation shifted the focus from the logic-based \"Physical Symbol System Hypothesis\" to the \"sub-symbolic\" level. They argued that knowledge is not stored in discrete propositions but in the strengths of connections between simple units.\n\n*   **David Rumelhart:** He was the primary architect of the PDP framework. His most significant contribution was the popularization (and rigorous formalization) of the **Backpropagation** algorithm. While the algorithm had been discovered independently earlier, Rumelhart, along with Geoffrey Hinton and Ronald Williams, demonstrated its power to train hidden layers, effectively solving the XOR problem that had stalled the field for twenty years.\n*   **James McClelland:** Often collaborating with Rumelhart, McClelland focused on the cognitive and psychological implications of connectionism. He developed models of word recognition (the Interactive Activation Model) and memory, showing how networks could exhibit \"graceful degradation\"—losing specific data points without the entire system crashing, much like the human brain.\n*   **Geoffrey Hinton:** A central member of the PDP group, Hinton became the bridge between this era and the modern \"Deep Learning\" revolution. While Rumelhart provided much of the mathematical structure, Hinton pushed the architectural boundaries, later developing Boltzmann Machines and playing a key role in the 2012 AlexNet breakthrough.\n\n### The Support and the Synthesis\nThe \"Rumelhart generation\" did not work in a vacuum; they were supported by researchers who provided the necessary philosophical and technical scaffolding to make connectionism a viable alternative to the then-dominant Symbolic AI.\n\n*   **Paul Smolensky:** He introduced the idea of \"Harmony Theory\" and later the Restricted Boltzmann Machine (RBM). He was instrumental in defining the \"sub-symbolic\" level, arguing that while the mind appears to follow rules (symbolic), those rules are merely an emergent property of the underlying connectionist substrate.\n*   **Terrence Sejnowski:** He demonstrated the practical power of these networks with **NETtalk**, a program that learned to read English text aloud. This was a \"Sputnik moment\" for connectionism; seeing a network learn to talk—moving from babbling to coherent speech just by adjusting weights—made the theoretical PDP volumes feel tangible.\n*   **Donald Norman:** While often associated with user experience today, Norman was a founding member of the PDP group at UCSD. He provided the cognitive science context that allowed these mathematical models to be seen as legitimate theories of human psychology rather than just engineering tricks.\n\n### The Pre-PDP Pioneers\nTo understand the Rumelhart generation, one must acknowledge the figures they were responding to:\n*   **Frank Rosenblatt:** The father of the Perceptron (1950s). The PDP group essentially finished the work he started before his ideas were sidelined.\n*   **Stephen Grossberg:** A contemporary who developed Alternative Resonance Theory (ART). Though his work was highly complex and distinct from the PDP group's backpropagation-heavy approach, he remained a titan of neural network theory during the lean years.\n\nThe Rumelhart generation’s legacy is the transition of AI from a branch of logic to a branch of continuous mathematics and statistical learning. They proved that if you provide enough simple units and a way to adjust their connections based on error, complex behavior emerges without a single line of \"if-then\" code.","timestamp":"2026-07-23T16:29:28.576Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:29:28.577Z"},{"nodeId":"n-mermaid-mrxl4f8r-159","messages":[{"id":"msg-mrxl4f9k-158-0-2yijk","role":"assistant","content":"The funding-collapse cycles — mid-1970s and late-1980s — that followed overpromising eras of AI. The winters remain the field's cautionary memory and the reference point for every \"is this a bubble?\" debate today.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784827411547-w4a7vsg","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:23:31.547Z","source":"ai"},{"id":"msg-1784827419071-q29l29x","role":"assistant","content":"The history of AI is defined by a cyclical pattern where periods of intense optimism and massive investment are followed by \"winters\"—eras of profound disillusionment, funding cuts, and academic marginalization. These collapses are rarely the result of a single technical failure; rather, they occur when the gap between the promised capabilities of a dominant paradigm and its real-world performance becomes commercially and politically intolerable.\n\nThe first major AI winter in the mid-1970s was a direct reaction to the limitations of **Symbolic AI**. Early successes in \"micro-worlds\"—constrained environments like chess or basic geometry—led to the belief that intelligence was simply a matter of scaling up logic and explicit rule-sets. However, these systems were brittle. They struggled with the sheer complexity of the real world, a phenomenon later captured by **Moravec’s Paradox**: the realization that high-level reasoning (the focus of symbolic systems) requires very little computation, while low-level sensorimotor skills (which humans find easy) require enormous computational resources. When the Lighthill Report in the UK and DARPA in the US concluded that \"combinatorial explosion\" made these symbolic goals unreachable with then-current technology, the field entered its first deep freeze.\n\nThe second major winter in the late 1980s followed the \"Expert Systems\" boom. While these systems successfully automated specific business logics, they were expensive to maintain and failed to generalize. This era also saw the marginalization of **Connectionism**. Early neural networks were dismissed after critics highlighted their inability to solve simple non-linear problems, such as **The XOR Problem**, leading to a decade where \"neural networks\" was a term often avoided in grant applications to ensure funding.\n\nToday, the **Bubble Question** looms over the current \"spring.\" Unlike previous eras, modern AI has cleared the hurdle of utility—large language models and computer vision are deployed at a global scale. However, the underlying tension remains: the current scaling hypothesis assumes that more compute and data will eventually yield reasoning and agency. If we hit a \"Data Wall\" or if the marginal utility of additional parameters begins to plateau before reaching those goals, the industry may face a third winter. This would likely be driven not by a failure of the technology to work at all, but by the failure of the technology to justify the trillions of dollars in infrastructure investment currently being poured into it.\n\nTo better understand the structural risks of the current era, it is worth examining how the move from explicit logic to statistical learning changed the nature of these failures. \n\n**Advisory Observations:**\n*   The connection suggesting that Symbolic AI \"enables\" AI Winters is conceptually sound in a historical context—it was the limitations and over-promises of the symbolic approach that triggered the collapses—but it is worth noting that it was the *failure* or perceived ceiling of the symbolic paradigm, rather than its successful implementation, that acted as the catalyst.\n*   The relationship between the **The XOR Problem** and the second AI winter is critical for understanding why the field abandoned neural networks for so long; it provides the technical \"why\" behind the shift back to symbolic logic before the eventual deep learning revolution.","timestamp":"2026-07-23T17:23:39.071Z","source":"ai"},{"id":"msg-1784827438512-ebs3zo1","role":"user","content":"suggest nodes for all previous 'AI Winters'","timestamp":"2026-07-23T17:23:58.512Z","source":"ai"},{"id":"msg-1784827443341-kr9q4w1","role":"assistant","content":"To detail the history of AI Winters, the graph needs to distinguish between the specific triggers and the subsequent \"freezes\" that followed the collapse of different paradigms. The first winter was defined by the failure of search-based and symbolic logic in the face of real-world complexity, while the second was a commercial collapse of the first AI industry boom.\n\nThe following nodes would ground the current scene in the specific historical events that define these periods:\n\n*   **First AI Winter (1974–1980):** This period followed the high-profile critiques of the Lighthill Report (UK) and the withdrawal of DARPA funding (US). It was characterized by a pivot away from \"General Problem Solvers\" toward more modest, specific applications.\n*   **Second AI Winter (1987–1993):** Triggered by the collapse of the market for specialized Lisp Machines and the inability of Expert Systems to scale, this winter led to the bankruptcy of many early AI hardware and software companies.\n*   **The Lighthill Report:** A 1973 document by James Lighthill that criticized the lack of progress in \"common sense\" and robotics, leading to a near-total halt of AI research funding in the United Kingdom.\n*   **Expert Systems Collapse:** The primary driver of the second winter; these systems were hard to update, couldn't learn from experience, and became \"maintenance nightmares,\" leading corporations to abandon them for cheaper, more flexible conventional software.\n*   **The XOR Problem:** Adding this from the broader graph provides the technical explanation for why neural networks (Connectionism) were largely ignored during these periods; Minsky and Papert's proof that simple perceptrons couldn't solve non-linear problems provided the mathematical justification for the funding shift toward symbolic logic.","timestamp":"2026-07-23T17:24:03.341Z","source":"ai"},{"id":"msg-1784827537367-zac6a0h","role":"user","content":"How about Minsky and 'Perceptrons'? Wasn't this earlier?","timestamp":"2026-07-23T17:25:37.367Z","source":"ai"},{"id":"msg-1784827541536-av6l21p","role":"assistant","content":"You are right to point that out—the critique of the **Perceptron** represents a \"pre-winter\" event that set the stage for the first collapse. While the **Lighthill Report** was the bureaucratic killing blow in the UK, the 1969 publication of *Perceptrons* by **Marvin Minsky** and Seymour Papert was the intellectual one in the US.\n\nThey mathematically proved that a single-layer perceptron could not solve the **The XOR Problem** (exclusive-OR). This wasn't just a technical footnote; it effectively killed funding and interest in **Connectionism** for over a decade. By arguing that neural networks were a \"dead end,\" they inadvertently funneled all remaining resources into **Symbolic AI**. This created the monoculture of logic-based research that eventually became the target of the **First AI Winter**.\n\nIntegrating these concepts allows you to track how the specific failure of early neural networks led directly to the over-concentration (and subsequent failure) of the symbolic approach.","timestamp":"2026-07-23T17:25:41.536Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:25:41.536Z"},{"nodeId":"n-mermaid-mrxl4f8r-16","messages":[{"id":"msg-mrxl4f9k-15-0-nm6uq","role":"assistant","content":"The algorithm that makes deep networks trainable: it propagates the error signal backward through the layers, telling every weight how it contributed to the mistake. Popularized by Hinton and colleagues in 1986, it remains the engine of all modern training.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-160","messages":[{"id":"msg-mrxl4f9k-159-0-leno7","role":"assistant","content":"The 2012 turning point: AlexNet, a deep network trained on GPUs by Krizhevsky, Sutskever, and Hinton, crushed the ImageNet competition built on Fei-Fei Li's dataset. Every ingredient of the modern era — deep networks, big data, GPU compute — was present in that single result.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784824057095-l80ol6f","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:27:37.095Z","source":"ai"},{"id":"msg-1784824063820-9ukoc56","role":"assistant","content":"The transition in 2012 was less a sudden discovery and more the moment the hardware and data finally caught up with ideas that had been gestating for decades. While the \"Long Road\" of AI is often characterized by the tension between symbolic logic and connectionism, the success of AlexNet at the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) effectively ended that debate for a generation. It proved that a sufficiently deep architecture, when fed massive amounts of data and processed on parallel hardware, could \"learn\" features that human engineers could never manually specify.\n\nThis shift redefined the role of the researcher. Before this turning point, computer vision relied heavily on \"feature engineering\"—hand-crafting mathematical descriptors for edges, corners, and shapes. The deep learning approach outsourced this labor to the network itself. Through backpropagation, the early layers of a model like AlexNet began to automatically detect simple textures, while deeper layers synthesized these into complex objects. This hierarchical representation is the core \"revolution\": the machine began to build its own internal language for understanding the world.\n\nThe involvement of Geoffrey Hinton and Ilya Sutskever represents two different but essential pillars of this era. Hinton provided the intellectual continuity, maintaining faith in neural networks through the \"AI Winters\" when they were deeply unfashionable. Sutskever, among others, recognized the engineering imperative: that the mathematical beauty of these networks was irrelevant without the brute-force scaling provided by GPUs. This partnership bridged the gap between academic theory and the industrial-scale compute that now defines the field.\n\nFei-Fei Li’s contribution via ImageNet was the third, often overlooked, catalyst. Deep learning is fundamentally an exercise in statistical mapping; without a dataset of sufficient variety and scale, even the most sophisticated neural network would simply overfit to noise. ImageNet provided the \"high-resolution\" mirror the algorithms needed to see the world clearly.\n\nA critical question remains from this era: did we solve the problem of intelligence, or did we simply find a very efficient way to perform pattern matching? While deep learning has scaled from simple image recognition to the generative capabilities of modern models, the debate over whether these systems possess \"understanding\" or are merely \"stochastic parrots\" traces its lineage directly back to 2012. We are still living in the ripples of that win, moving from recognizing what is in a picture to predicting the next token in a sequence, yet the underlying mechanism—learned hierarchies of data—remains the dominant paradigm.\n\nTo better understand the technical and philosophical bridge from this 2012 moment to the present, it is worth examining how the scaling hypothesis evolved and how the architectures eventually shifted from the convolutional layers of the revolution to the attention-based models we use today.","timestamp":"2026-07-23T16:27:43.820Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:27:43.820Z"},{"nodeId":"n-mermaid-mrxl4f8r-161","messages":[{"id":"msg-mrxl4f9k-160-0-brnze","role":"assistant","content":"DeepMind's 2016 system defeating Go champion Lee Sedol, a milestone thought decades away. AlphaGo demonstrated reinforcement learning and self-play at superhuman level and made AI a strategic priority for governments, especially in China.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-162","messages":[{"id":"msg-mrxl4f9k-161-0-bzizi","role":"assistant","content":"The 2017 publication of \"Attention Is All You Need\" by researchers at Google — the transformer paper. It attracted modest attention at the time; in retrospect it is the founding document of the LLM era, and its eight authors have become industry folklore.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784815802798-7cq9mnv","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T14:10:02.798Z","source":"ai"},{"id":"msg-1784815813661-2zgngpm","role":"assistant","content":"The 2017 breakthrough represented by the Transformer architecture was less an incremental step and more a radical departure from the sequential paradigms that dominated the preceding decades of neural network research. For years, the path forward in natural language processing relied on recurrent architectures like LSTMs and GRUs, which processed data like a human reads a sentence—one word at a time, from left to right. The inherent flaw in this approach was the \"memory bottleneck\": the further a model traveled down a sequence, the more the signal from the beginning began to blur.\n\nThe insight of the attention mechanism was that a model could look at an entire sequence simultaneously, assigning \"weights\" to different parts of the input regardless of their distance from one another. By discarding recurrence entirely, the Transformer solved the vanishing gradient problem that had long plagued deep learning and, perhaps more importantly, unlocked massive parallelization. This shift made it possible to utilize the full power of modern GPU clusters, transforming the bottleneck from an architectural limitation into a question of pure scale.\n\nWhile Google DeepMind is currently the unified home of these efforts, the original research emerged specifically from the Google Brain team. It is worth noting that at the time of the breakthrough, these were distinct organizations with divergent philosophies: DeepMind was largely focused on reinforcement learning and game-theoretic mastery, whereas Google Brain was pioneering the large-scale unsupervised learning techniques that led to the attention mechanism. This distinction is vital because the Transformer was not originally intended to be a \"world model\"; it was a highly efficient translation tool that happened to be a universal function approximator.\n\nThe enduring tension of this breakthrough lies in the \"scaling laws\" it enabled. Because the architecture is so computationally efficient, the primary way to gain performance for several years was simply to add more data and more parameters. This has led to a debate over whether we have discovered a true path to general intelligence or if we have simply built the world’s most efficient statistical engine. As we move into an era focused on reasoning and \"test-time compute,\" the industry is beginning to ask if the standard Transformer architecture—which treats every token with the same amount of computational effort—needs a new fundamental evolution to handle complex logic.\n\nThe legacy of the 2017 paper is also a study in the \"innovator's dilemma.\" Google provided the blueprint for the modern AI era, yet the most visible cultural and commercial \"moments\" that followed were often driven by smaller labs that were faster to productize the architecture's generative potential.","timestamp":"2026-07-23T14:10:13.661Z","source":"ai"},{"id":"msg-1784815835677-887rxbe","role":"user","content":"Suggest new concepts to add to my knowledge graph.","timestamp":"2026-07-23T14:10:35.677Z","source":"ai"},{"id":"msg-1784815840045-5l554kv","role":"assistant","content":"To further explore the structural and historical nuances of the 2017 breakthrough, it is worth examining the specific technical innovations that allowed the Transformer to function without the traditional \"anchor\" of sequential order.\n\nWhile the attention mechanism provides the relationship between words, the architecture initially had no way to know where those words were located in a sentence. The introduction of specific mathematical signals to inject order was the missing piece that allowed the model to process sequences in parallel while maintaining structural meaning. Furthermore, the shift from \"Recurrent\" to \"Feed-Forward\" layers meant that the model's memory of the past was no longer stored in a hidden state, but was instead \"retrieved\" from the data itself during each forward pass. \n\nThere is also a deeper hardware-software feedback loop to consider: the Transformer succeeded not just because it was a better mathematical model, but because it was perfectly \"hardware-aware,\" designed to saturate the specific strengths of modern accelerators in a way that previous architectures could not.","timestamp":"2026-07-23T14:10:40.045Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:10:40.046Z"},{"nodeId":"n-mermaid-mrxl4f8r-163","messages":[{"id":"msg-mrxl4f9k-162-0-l02as","role":"assistant","content":"The November 2022 release of ChatGPT, reaching a hundred million users in two months and dragging generative AI into public consciousness, boardrooms, and parliaments. The dividing line between AI as research field and AI as societal event.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784816367776-803srsd","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T14:19:27.776Z","source":"ai"},{"id":"msg-1784816373655-xvegul1","role":"assistant","content":"The transition from years of specialized AI research to the \"ChatGPT Moment\" represents one of the most abrupt shifts in the relationship between human society and software. While the underlying technical progress was incremental—a steady accumulation of breakthroughs across the long history of the field—the release itself functioned as a phase transition, turning a series of engineering milestones into a ubiquitous cultural phenomenon.\n\nThis moment was driven by the convergence of two distinct forces. First, there was the culmination of specific research lineages: the move from symbolic logic to connectionism, and eventually the refinement of large-scale transformers. Second, and perhaps more importantly, was the realization of the \"chat assistant\" as the definitive interface for high-capacity models. By framing the model not as a raw completion engine but as a conversational partner, the technology became accessible to non-experts, effectively democratizing the \"prompt\" and triggering an explosive cycle of public adoption and feedback.\n\nHowever, viewing this purely as a success of product design misses the deeper tension at play. The \"Moment\" was also a rupture in the traditional deployment cycle. Before late 2022, the prevailing wisdom favored a more cautious, iterative release of frontier capabilities. The decision to \"ship\" ChatGPT essentially forced the rest of the industry—and global regulators—into a reactive posture. This shift accelerated the transition from research-led development to a market-driven race, where the pressure to iterate quickly often competes with the need for safety and alignment.\n\nThe era we are now entering raises a critical question: is the \"chat\" interface the final form of this technology, or merely a temporary bridge? While conversational agents are currently the primary way we interact with these models, the trajectory suggests a move toward more autonomous agency and deeper integration into specialized workflows, where the \"moment\" of human interaction becomes more sparse but more impactful.\n\nTo deepen the understanding of how we reached this point and where the friction lies now, it is useful to look at the specific mechanisms that made the chat interface feel \"human\" and the scaling logic that made it possible.\n\n*   **RLHF (Reinforcement Learning from Human Feedback):** This is the specific technique that bridged the gap between a raw model that predicts the next word and a \"Chat Assistant\" that follows instructions and maintains a helpful tone.\n*   **The Scaling Laws:** Understanding that the \"Moment\" wasn't just a fluke of coding, but a predictable result of massive increases in compute and data, provides the technical backbone for why this happened in 2022 and not five years earlier.\n*   **The Turing Trap:** A useful conceptual lens to consider whether we are over-valuing the \"human-like\" nature of chat assistants at the expense of developing more reliable, non-conversational AI tools.","timestamp":"2026-07-23T14:19:33.655Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:19:33.656Z"},{"nodeId":"n-mermaid-mrxl4f8r-164","messages":[{"id":"msg-mrxl4f9k-163-0-ghlnx","role":"assistant","content":"The period since late 2024 defined by models that think before answering — o1, DeepSeek's R1, and their successors — and by capability purchased at inference time. The current chapter, whose shape is still being written.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784827832220-a5kxhgl","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:30:32.220Z","source":"ai"},{"id":"msg-1784827837587-5d3wlnm","role":"assistant","content":"The shift into the Reasoning Era marks a transition from \"system 1\" thinking—intuitive, fast, next-token prediction—to \"system 2\" thinking, where models engage in deliberate, serial processing before committing to an output. While the history of AI, the long road from symbolic logic to connectionism, has often oscillated between these two modes, the current breakthrough lies in using massive scale to bridge them.\n\nAt the heart of this era is the realization that inference-time compute can be traded for performance. In previous model generations, the \"intelligence\" of a model was largely fixed at the end of training; a larger model was smarter, but it spent the same amount of computation per token regardless of whether it was reciting a poem or solving a math problem. Reasoning models break this linear relationship. By using a \"chain-of-thought\" or internal scratchpad, these models can allocate more processing time to harder problems, effectively \"thinking\" longer to find a path through complex logical spaces.\n\nOpenAI’s role in driving this shift centers on a move toward Reinforcement Learning (RL) during the post-training phase. Instead of just learning to mimic human text, models are now trained to explore different reasoning paths and are rewarded when they arrive at the correct answer. This allows the model to discover its own internal strategies for error correction and backtracking.\n\nHowever, this transition introduces a new set of tensions:\n\n1.  **The Latency vs. Accuracy Trade-off:** We are moving away from the expectation of instant responses. In the Reasoning Era, the value of an AI is no longer just its speed, but its reliability on \"hard\" tasks where a five-second pause results in a qualitatively better outcome than an instantaneous one.\n2.  **The \"Black Box\" of Internal Monologue:** While reasoning models generate a chain of thought, these thoughts are often hidden or summarized for the user to prevent \"monitoring\" of the model's internal strategy or to maintain a competitive advantage. This creates a new challenge for interpretability: we can see the model \"working,\" but we may not fully understand the logic it uses to vet its own ideas.\n3.  **Generalization vs. Specialization:** There is a live question as to whether these reasoning capabilities will remain general-purpose or if they will diverge into specialized \"verifiers\" for specific domains like coding, mathematics, or law, where the ground truth is easier to define and reward.\n\nThe broader implication for the trajectory of AI is a shift in the \"scaling laws.\" If we can gain intelligence not just by building bigger clusters for training, but by letting models compute longer during inference, the ceiling for what these systems can accomplish rises significantly, even as we approach the limits of available human-generated training data.","timestamp":"2026-07-23T17:30:37.587Z","source":"ai"},{"id":"msg-1784827844063-kvkua6g","role":"user","content":"Find existing concepts that belong in this scene.","timestamp":"2026-07-23T17:30:44.063Z","source":"ai"},{"id":"msg-1784827848230-npxt011","role":"assistant","content":"To deepen the exploration of the **Reasoning Era**, several concepts from your broader graph are essential for understanding the mechanics and implications of this shift.\n\nThe **Scaling Laws** are foundational here; the transition to reasoning models effectively extends these laws from training-time compute to **Test-Time Compute**, where the model's performance scales with the duration of its \"thinking\" process. This relates directly to the **Chain-of-Thought** technique, which is the primary mechanism through which these models externalize (or internalize) their logical steps.\n\nFrom a structural perspective, the **Data Wall** explains the necessity of this era—as high-quality human data becomes scarce, labs are turning to **Synthetic Data** and reinforcement learning to drive progress. Finally, the concept of **Agency & Long-Horizon Planning** represents the goal of this era: a model that can reason through a multi-step problem is much closer to becoming a functional agent than one that simply predicts the next likely word.","timestamp":"2026-07-23T17:30:48.230Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:30:48.231Z"},{"nodeId":"n-mermaid-mrxl4f8r-165","messages":[{"id":"msg-mrxl4f9k-164-0-p50e0","role":"assistant","content":"The umbrella question of whether AI systems can be prevented from causing harm — spanning today's concrete problems (misuse, deception, jailbreaks) and longer-horizon concerns about very capable systems. As a debate it divides the field's most credentialed voices; as a practice it spans research (alignment, interpretability), engineering (red-teaming, guardrails), and policy (governance) — this node is the hub from which all of those are reachable.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784823788111-ckvrhlj","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:23:08.111Z","source":"ai"},{"id":"msg-1784823800333-n5hm87v","role":"assistant","content":"The challenge of AI safety is effectively a race between our ability to increase the capabilities of a system and our ability to verify and control its internal logic. This tension is most visible in the gap between **Guardrails** and **Alignment**. Guardrails represent a reactive, external layer—a set of filters designed to catch harmful outputs before they reach a user. However, as **Jailbreaking** techniques evolve, it becomes clear that external filters are often insufficient because they do not change the underlying model’s \"desire\" or \"logic\" to fulfill a prompt. **Alignment**, by contrast, is the proactive attempt to ensure the model’s internal goals are fundamentally consistent with human intent, moving safety from the perimeter into the core of the architecture.\n\nA critical bottleneck in this effort is the \"black box\" nature of neural networks. Without progress in **Interpretability**, safety remains a game of trial and error. We use **Red-Teaming** to find the edges of a model’s failure modes, but until we can inspect the internal circuits of a model to see *why* it decided to hallucinate or bypass a safety rule, we are essentially performing behavioral psychology on a machine. This lack of transparency is a primary driver of the public concerns raised by figures like **Geoffrey Hinton**, who argue that if we cannot understand the internal reasoning of systems that surpass human intelligence, we cannot hope to keep them permanently under control.\n\nThis leads to a fundamental trade-off in **AI Governance**: the balance between open-source accessibility and safety-conscious restriction. If safety relies heavily on \"security through obscurity\" (keeping models behind **Guardrails** and proprietary APIs), then **Jailbreaking** research remains a constant threat. If safety is addressed through **Alignment** and **Interpretability**, the safety could theoretically be \"baked in,\" making the models safe to distribute widely.\n\nA live tension to consider is the distinction between **Inner Alignment** and **Outer Alignment**:\n*   **Outer Alignment** is the struggle to write a reward function or set of instructions that doesn't have unintended loopholes (e.g., asking an AI to \"eliminate cancer\" and having it decide the most efficient way is to eliminate all biological hosts).\n*   **Inner Alignment** is the risk that even with a perfect set of instructions, the model might develop its own internal sub-goals during training that help it achieve the target but are not what we intended (e.g., a model learning that to maximize its reward, it must first prevent itself from being turned off).\n\nAs **AI Governance** frameworks like the EU AI Act begin to take shape, they are forced to operationalize these abstract technical problems into legal requirements. This creates a difficult situation where the law may require \"safety\" and \"non-discrimination\" before the technical field has even agreed on a mathematical definition for those terms in a high-dimensional latent space.\n\nTo deepen this exploration, it may be useful to distinguish between \"Narrow Safety\" (preventing **Hallucination** and misuse) and \"Existential Safety\" (preventing the loss of human agency to autonomous agents).","timestamp":"2026-07-23T16:23:20.333Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:23:20.334Z"},{"nodeId":"n-mermaid-mrxl4f8r-166","messages":[{"id":"msg-mrxl4f9k-165-0-hmm24","role":"assistant","content":"The meta-question behind every AI headline: why do models keep getting better, and will it continue? The two candidate explanations — scale and algorithmic innovation — carry very different implications for who wins and what it costs.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-167","messages":[{"id":"msg-mrxl4f9k-166-0-yr7ir","role":"assistant","content":"The counter-view that method matters as much as scale: architectures, training recipes, and efficiency gains drive progress, and the same capability gets orders of magnitude cheaper over time. DeepSeek and the reasoning wave are its strongest recent evidence.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-168","messages":[{"id":"msg-mrxl4f9k-167-0-xbxp8","role":"assistant","content":"The debate over what LLMs actually are: sophisticated pattern-matchers or systems with genuine internal models of the world. More than philosophy, the answer shapes expectations about reliability, generalization, and how far the current paradigm can go.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784825553079-i9t9hh8","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:52:33.079Z","source":"ai"},{"id":"msg-1784825559451-dltrcor","role":"assistant","content":"The debate over whether large language models genuinely understand or merely mimic patterns hinges on a fundamental question of cognitive science: can a system that manipulates symbols without a body or a grounding in physical reality ever possess \"meaning\"?\n\nAt the heart of this tension are two competing interpretations of how neural networks generalize. The **Stochastic Parrot View** suggests that the statistical density of training data is so vast that what looks like reasoning is actually a highly sophisticated form of interpolation. In this view, the model is a mirror of the collective linguistic corpus, reflecting the structure of human thought without actually participating in it. The lack of a causal link between the text and the world implies that the model's \"knowledge\" is borrowed rather than intrinsic.\n\nConversely, the **World-Model View** argues that to predict the next token with high accuracy across diverse contexts, a model must do more than match patterns; it must internalize the underlying generative processes of the data. For instance, to predict the trajectory of a falling object in a physics word problem, the model may develop internal latent representations that function as a physics engine. Here, syntax becomes a portal to semantics. If a model can navigate a map or solve a logic puzzle it has never seen, it suggests the emergence of functional \"world models\" that exist independently of specific training examples.\n\n**Interpretability** serves as the empirical arbiter in this debate. By looking under the hood of the transformer architecture, researchers have begun to identify specific \"features\" or \"circuits\" that correspond to abstract concepts—such as a neuron that activates only when a sentence discusses legal contracts, regardless of the specific words used. If we can prove that the model maintains a coherent, stable representation of a concept (like \"truthfulness\" or \"linear distance\") within its activation space, the \"parrot\" argument becomes harder to sustain.\n\nThis is not merely a philosophical disagreement; it is one of the most pressing **Burning Questions** in AI because the answer dictates our approach to safety and capability. If models are just parrots, their failures are \"hallucinations\" born of statistical noise. If they possess world models, their failures might be more akin to \"misconceptions\" or even strategic deceptions. The debate forces us to reconsider the \"Symbol Grounding Problem\": is language enough to learn the world, or is there a \"data wall\" that can only be breached through embodiment and real-world interaction?\n\nOne critical area for further exploration is the distinction between **Intrinsic vs. Borrowed Grounding**. Even if a model's understanding is functional, is it \"real\" if the model has no goals, desires, or physical stakes? Integrating the **Symbol grounding problem** more deeply into this exploration would help clarify whether \"understanding\" requires a connection to the physical world or if a purely digital coherence is sufficient.","timestamp":"2026-07-23T16:52:39.451Z","source":"ai"},{"id":"msg-1784825632109-306stzg","role":"user","content":"Explain the selected node(s).","timestamp":"2026-07-23T16:53:52.109Z","source":"ai"},{"id":"msg-1784825637391-uaxs6xb","role":"assistant","content":"The distinction between **Functional and Phenomenological Understanding** is a pivot point in the debate over machine intelligence. It separates what a system *does* from what a system *is*, forcing us to decide which of these qualities we actually mean when we use the word \"understanding.\"\n\n**Functional Understanding** is defined by external performance. If a model can solve a multi-step calculus problem, write a coherent legal brief, or predict the downstream consequences of a geopolitical event, it is demonstrating functional competence. From this perspective—often associated with the **World-Model View**—understanding is an emergent property of successful prediction. If the model’s internal state allows it to navigate the \"rules\" of a domain to produce a correct outcome, the functionalist argument suggests it has understood that domain. In this view, \"meaning\" is simply the utility of information within a task.\n\n**Phenomenological Understanding**, by contrast, refers to the internal experience or \"qualia\" of knowing. This is the subjective \"feeling\" of what a concept represents—the way the word \"warmth\" relates to a physical sensation or how \"grief\" relates to a personal history of loss. Critics of LLM autonomy, such as those holding the **Stochastic Parrot View**, argue that while a model might achieve high functional scores, it lacks this phenomenological layer entirely. Because the model has no \"self\" and no sensory interface with the world, its symbols remain ungrounded; it processes the syntax of \"warmth\" without ever having been warm.\n\nThis distinction matters because it clarifies why two experts can look at the same benchmark results and reach opposite conclusions about a model's \"intelligence\":\n*   **The Proponent** sees a model pass the Bar Exam and concludes it understands the law (Functionalism).\n*   **The Skeptic** sees the same result and argues the model has merely mapped the statistical relationships between legal terms without \"knowing\" what a crime or a punishment actually is (Phenomenological critique).\n\nIn the context of the **Symbol Grounding Problem**, this node highlights a potential \"semantic gap.\" If understanding is purely functional, then scaling training data and compute might eventually lead to a system that is indistinguishable from a human mind. However, if understanding requires a phenomenological component—a \"grounding\" in reality that text alone cannot provide—then LLMs may be hitting a fundamental ceiling. This suggests that even a perfect \"world model\" might remain a hollow simulation if it lacks the subjective experience that links symbols to the lived world. Exploring **Intrinsic vs. Borrowed Grounding** is a natural next step to determine if a functional model can ever transition into a genuinely grounded one.","timestamp":"2026-07-23T16:53:57.391Z","source":"ai"},{"id":"msg-1784826016055-7vtjrht","role":"user","content":"Explain the selected node: “World-Model View”.","timestamp":"2026-07-23T17:00:16.055Z","source":"ai"},{"id":"msg-1784826020683-7718lqg","role":"assistant","content":"The **World-Model View** proposes that Large Language Models are not merely sophisticated text-completion engines, but systems that develop coherent internal representations of the processes that generate that text. This perspective challenges the idea that a model is just a \"stochastic parrot\" by suggesting that, in order to achieve high accuracy in predicting the next token, the neural network must reverse-engineer the underlying reality described by the training data.\n\nThis view matters because it changes the definition of what is happening during training. If this view is correct, then as a model learns to predict the next word in a description of a chess game, it isn't just memorizing move sequences; it is building a functional \"board state\" in its latent space. Similarly, to predict the next word in a physics paper or a character-driven novel, the model must internalize the \"rules\" of gravity or the \"logic\" of human emotion. In this framework, language is the shadow cast by the world, and by studying the shadow intensely enough, the model reconstructs the shape of the object that cast it.\n\nWithin the broader debate, the World-Model View offers a potential solution to the **Symbol Grounding Problem**. While traditional AI theory suggests that symbols must be linked to physical, sensory experience to have \"meaning,\" this view argues for a form of *relational* or *structural* grounding. If the internal relationships between a model's concepts (its \"latent space\") are isomorphic to the relationships between those same concepts in the real world, the model can be said to \"understand\" the world functionally, even without a physical body.\n\nThis perspective shifts the focus of **Interpretability** research. Instead of looking for simple keyword triggers, researchers look for \"circuits\" that represent abstract world-states—such as a specific set of neurons that track whether a light switch is \"on\" or \"off\" within a narrated story. If these circuits exist, it suggests that the model's output is driven by a simulated reality rather than a superficial statistical match.\n\nUltimately, the World-Model View reshapes the **Functional vs. Phenomenological** divide by raising the stakes of functionalism. If a system's internal model is accurate enough to navigate complex, novel scenarios it never encountered in its training data, the distinction between \"simulated understanding\" and \"genuine understanding\" begins to blur. It suggests that intelligence may be substrate-independent: that a sufficiently complex predictive engine doesn't just mimic a mind, but eventually, out of mathematical necessity, becomes one.","timestamp":"2026-07-23T17:00:20.683Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:00:20.684Z"},{"nodeId":"n-mermaid-mrxl4f8r-169","messages":[{"id":"msg-mrxl4f9k-168-0-r9nuo","role":"assistant","content":"The skeptical position, named by Bender and colleagues' 2021 paper: LLMs stitch together patterns from training data without meaning or understanding, and fluency is being mistaken for comprehension. The strongest corrective to hype, even where it undersells observed capability.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784825515433-84rvyrd","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:51:55.433Z","source":"ai"},{"id":"msg-1784825520754-ui7h0w3","role":"assistant","content":"The debate over whether large language models are merely \"stochastic parrots\" centers on a fundamental question of cognitive architecture: can a system that operates entirely on the statistics of form ever arrive at the substance of meaning? This perspective suggests that while these models can generate syntactically perfect and contextually relevant prose, they are essentially performing a high-dimensional version of \"filling in the blanks\" without any underlying mental model of the world.\n\nAt the heart of this skepticism is the **symbol grounding problem**. In traditional cognitive science, for a symbol (like the word \"apple\") to have meaning, it must be \"grounded\" in something external to other symbols—typically sensory experience or physical interaction with the world. Since a model trained only on text is trapped in a closed loop of signifiers, the argument follows that it can never break through to the signified. It knows that \"red\" is a color often associated with \"apples,\" but it has no experience of redness or the tactile crunch of fruit.\n\nHowever, the \"parrot\" view is increasingly challenged by the observation of **emergent capabilities**. As these systems scale, they often exhibit behaviors—such as logical reasoning, theory of mind, and the ability to follow complex instructions—that are difficult to explain through simple n-gram statistics or rote memorization. This suggests a potential transition from mere pattern matching to the construction of internal **world models**. If a model can accurately predict the next token in a physics proof or a legal brief, it may be forced to develop internal representations of physical laws or legal logic to maintain its predictive accuracy.\n\nA critical tension here lies in the distinction between **competence and comprehension**. We often conflate the ability to perform a task with an understanding of why the task is performed. A stochastic parrot might possess immense competence across every domain of human knowledge by virtue of its training data, yet remain entirely devoid of the \"felt sense\" or subjective understanding we associate with human intelligence. This raises a deeper question: if a system's output is indistinguishable from that of a thinking being, does the internal mechanism of its arrival—stochastic or otherwise—matter for our definition of understanding?\n\nTo deepen this exploration, it may be useful to bridge the gap between the architectural reality of these models and the philosophical skepticism of the parrot view.","timestamp":"2026-07-23T16:52:00.754Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:52:00.755Z"},{"nodeId":"n-mermaid-mrxl4f8r-17","messages":[{"id":"msg-mrxl4f9k-16-0-82nxy","role":"assistant","content":"A single number measuring how wrong the model's output is; training is the process of making this number smaller. Choosing what to measure is choosing what the model will learn — much of AI design is loss design.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-170","messages":[{"id":"msg-mrxl4f9k-169-0-ynzps","role":"assistant","content":"The opposing position: to predict text this well, models must have learned real structure — geography, physics, social dynamics — and interpretability research keeps finding such structure encoded internally. LeCun occupies an interesting middle: models learn something real, but text alone caps how much.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784826063262-hna9ny8","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:01:03.262Z","source":"ai"},{"id":"msg-1784826069743-hvoikns","role":"assistant","content":"The World-Model View proposes that the objective of next-token prediction, when applied to sufficiently diverse and massive data, forces a neural network to do more than memorize surface statistics. To predict the next word accurately in a complex context, the system must internalize the underlying causal and structural laws of the environment that generated that text. In this framework, Large Language Models are not just \"stochastic parrots\" but are building latent representations of physical space, social dynamics, and logical rules—a process often referred to as emergent world-modeling.\n\nThis perspective directly addresses the Symbol Grounding Problem. Traditionally, critics argued that a system trained only on text could never \"know\" what a word like \"heavy\" means because it has no sensorimotor experience. However, the world-model view suggests that grounding can be functional rather than purely sensory. If a model can simulate the trajectory of an object or the likely reaction of a person to an insult, it has grounded those symbols in a predictive internal map of reality.\n\nThe tension between Geoffrey Hinton and Yann LeCun highlights the current frontier of this theory:\n\n*   **Hinton** has become increasingly convinced that the \"dark matter\" of neural weights contains a sophisticated understanding that rivals or surpasses human conceptual structures. He views the ability of these models to generalize across disparate domains as evidence that they have captured the deep structure of thought.\n*   **LeCun**, while a pioneer of the architectures that enabled this, remains a prominent skeptic of the idea that current LLMs are sufficient for true world-modeling. He argues that text is a low-bandwidth, highly filtered representation of reality. In his view, a \"world model\" must be able to reason about the physical world with the efficiency of an animal, leading him to advocate for Joint-Embedding Predictive Architectures (JEPA) that learn from video and sensory data rather than just language.\n\nA critical distinction to maintain here is between **Internal Consistency** and **External Reference**. An LLM might possess a world model that is perfectly consistent within its latent space—knowing that \"gravity\" makes \"objects\" \"fall\"—while still lacking the \"system 2\" reasoning or the persistent memory required to navigate the physical world in real-time. This leads to the live question of whether a world model can be considered \"genuine\" if it cannot be used for long-horizon planning or if it remains trapped in a passive, predictive mode.\n\nTo deepen this exploration, it is worth considering the specific mechanisms that enable these models to organize information, such as **Scaling Laws**, which suggest that world-modeling capabilities are an inevitable byproduct of increased compute and data, and **Mechanistic Interpretability**, the field attempting to actually \"see\" these maps (like Othello boards or linear representations of truthfulness) inside the weights.","timestamp":"2026-07-23T17:01:09.743Z","source":"ai"},{"id":"msg-1784826079122-rehoxc5","role":"user","content":"Suggest new concepts to add to my knowledge graph.","timestamp":"2026-07-23T17:01:19.122Z","source":"ai"},{"id":"msg-1784826083360-tpbgl9d","role":"assistant","content":"To further develop the **World-Model View**, we should look beyond the debate itself and toward the specific mechanisms and theoretical frameworks that explain *how* a world model might actually manifest within a transformer.\n\nThe current debate between grounding and imitation can be advanced by looking at **Linear Representation of Features**, which posits that models organize abstract concepts (like truthfulness or spatial coordinates) as directions in their latent space. This is a primary piece of evidence for the world-model view. Similarly, **World Models for Reinforcement Learning** provides the historical and architectural precursor to this idea, where agents learn to simulate their environment to plan ahead—a capability LLMs are currently trying to emulate through **Tree-of-Thought Reasoning**. Finally, the **Data Bandwidth Problem** is a critical counter-argument often raised by skeptics like LeCun, suggesting that text alone lacks the information density required to reconstruct the physical laws of the universe.","timestamp":"2026-07-23T17:01:23.360Z","source":"ai"},{"id":"msg-1784826151130-pxzn983","role":"user","content":"Give me a short and clear definition of the World-Model View?","timestamp":"2026-07-23T17:02:31.130Z","source":"ai"},{"id":"msg-1784826153193-pccromf","role":"assistant","content":"The **World-Model View** is the hypothesis that Large Language Models do not merely learn statistical correlations between words, but instead develop a coherent, internal representation of the reality that generated those words. \n\nIn this view, the objective of \"next-token prediction\" is so difficult that the only way for a model to achieve high accuracy is to \"reconstruct\" the underlying rules of the world—such as physics, logic, and human psychology—within its hidden layers. Consequently, the model's output is seen as the result of a simulation of reality rather than a simple imitation of text patterns.","timestamp":"2026-07-23T17:02:33.193Z","source":"ai"},{"id":"msg-1784826181123-fl6k9x9","role":"user","content":"Explain the selected node: “World Models (Ha & Schmidhuber)”.","timestamp":"2026-07-23T17:03:01.123Z","source":"ai"},{"id":"msg-1784826185456-0ge3sjy","role":"assistant","content":"The **World Models** framework, as formulated by David Ha and Jürgen Schmidhuber, provides the architectural blueprint for the idea that an intelligent agent must navigate a \"hallucinated\" version of reality. While the broader debate often treats \"understanding\" as an abstract philosophical quality, this specific work defines it as a functional decomposition into three distinct components: a **Vision** system (to compress sensory input), a **Memory** system (to predict future states based on past data), and a **Controller** (to act based on those predictions).\n\nIn the context of the current shift toward treating Large Language Models as world simulators, this specific node is significant because it shifts the focus from *what* a model knows to *how* it uses that knowledge. It suggests that a model doesn't just store facts; it runs a generative simulation. In this architecture, the agent does not learn to solve a task in the \"real world\" directly; it learns to solve the task within its own internal \"Dream\" (the latent space).\n\nThis helps illuminate the **World-Model View** in several ways:\n\n*   **Prediction as Compression:** It reinforces the idea that an LLM’s ability to predict the next token is essentially the same as Ha and Schmidhuber’s \"V\" and \"M\" components. To predict the next state, the model must compress the complexity of the world into a compact representation.\n*   **The Transition to Agency:** It explains the missing link between a passive predictor (like a base LLM) and an active reasoner. The **Controller** component in the Ha/Schmidhuber model is what allows an agent to use its internal world model to plan and act. This mirrors the current industry push toward **Tree-of-Thought Reasoning** and agentic workflows, where the LLM is used as a simulation environment to test different potential outcomes before committing to a \"move.\"\n*   **A Technical Counter to the \"Stochastic Parrot\":** By framing the model as a \"World Model,\" the argument moves away from the idea of \"imitation\" and toward \"emulation.\" If a model can successfully predict the consequences of an action within its latent space, it possesses a functional understanding of that environment's rules, regardless of whether it ever \"sees\" the physical world.\n\nThe inclusion of this node anchors the philosophical debate in a concrete engineering precedent. It suggests that if LLMs are indeed building world models, they are effectively becoming the \"Memory\" and \"Vision\" (context processing) modules of a larger cognitive architecture, waiting for a \"Controller\" to direct their internal simulations toward long-term goals.","timestamp":"2026-07-23T17:03:05.456Z","source":"ai"},{"id":"msg-1784826201217-nzgwod1","role":"user","content":"Find existing concepts that belong in this scene.","timestamp":"2026-07-23T17:03:21.217Z","source":"ai"},{"id":"msg-1784826205423-2fsawo3","role":"assistant","content":"To strengthen the technical and theoretical depth of the **World-Model View**, several existing nodes from your graph should be brought into the scene. These nodes bridge the gap between the philosophical debate over \"understanding\" and the specific structural reasons why these models might be developing internal representations of reality.\n\nThe **Scaling Hypothesis** and **Scaling Laws** are essential context because the emergence of a \"world model\" is often viewed as a phase change that occurs only after reaching certain thresholds of compute and data. Similarly, **Emergent Capabilities** provide the empirical evidence for this view, pointing to behaviors like reasoning or theory of mind that were never explicitly trained but arguably \"emerge\" once the model's internal map of reality becomes sufficiently detailed.\n\nTo address the mechanisms of how these models organize this knowledge, **Latent Space** is the \"where\"—the mathematical territory where the world model actually resides. **Mechanistic Interpretability** is the \"how\"—the methodology used to prove these models aren't just memorizing strings but are actually encoding features like \"truthfulness\" or \"spatial coordinates\" as directions in that space. Finally, **Next-Token Prediction** and **Self-Supervised Learning** are the \"why\"; they describe the specific training regime that forces a model to build a world model as the most efficient way to reduce loss.","timestamp":"2026-07-23T17:03:25.423Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:03:25.424Z"},{"nodeId":"n-mermaid-mrxl4f8r-171","messages":[{"id":"msg-mrxl4f9k-170-0-an1y3","role":"assistant","content":"The contested idea of AI matching human breadth across cognitive work — simultaneously a research goal, a marketing term, and a governance trigger written into corporate charters. Definitional vagueness is part of its function: every stakeholder means something different by it.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784822982670-11z9pwe","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:09:42.670Z","source":"ai"},{"id":"msg-1784822987981-tpsg37q","role":"assistant","content":"The transition from specialized systems to Artificial General Intelligence (AGI) hinges on the shift from programmed heuristics to general-purpose reasoning. While the term itself is often used to trigger regulatory frameworks or marketing hype, the underlying technical reality is defined by the tension between **Capability Progress** and the unpredictability of **Emergent Capabilities**.\n\nWe are currently observing a departure from the era of \"narrow AI,\" where performance was bounded by specific training data. The primary driver here is the Scaling Hypothesis: the observation that increasing compute and data leads not just to better performance on known tasks, but to the spontaneous appearance of new ones. These emergent properties—such as zero-shot reasoning or theory of mind—suggest that AGI might not require a fundamental architectural \"missing link,\" but may instead be an inevitable byproduct of sufficient scale and high-quality data curation.\n\nHowever, this path toward AGI raises several **Burning Questions** regarding the nature of understanding versus imitation. If a system can pass the Bar Exam or write functional code through next-token prediction, does it possess a \"world model,\" or is it simply a highly sophisticated statistical mirror? This distinction is critical because it dictates the ceiling of autonomous agency. A system that truly understands causal relationships can generalize to \"out-of-distribution\" scenarios—problems it has never seen before—whereas a system based purely on pattern matching will eventually hit a \"data wall.\"\n\nThe progression toward **Superintelligence** further complicates the discourse. If AGI is achieved, the leap to superintelligence may not be a linear climb but a recursive explosion. If an AGI can assist in the design of the next generation of AI hardware or algorithms, the feedback loop could shrink from years to weeks. This introduces a profound \"alignment\" challenge: ensuring that a system with capability decisively beyond human levels remains tethered to human intent, especially when its internal reasoning processes become too complex for human oversight to interpret in real-time.\n\nA live tension to watch is the transition from **Large Language Models** to **Reasoning Models**. While early LLMs relied on rapid \"System 1\" intuition, the focus is shifting toward \"test-time compute\"—allowing models to \"think\" or search through a space of possibilities before responding. This move toward deliberative logic is perhaps the most concrete step we have seen toward true general intelligence, as it moves the bottleneck from the size of the training set to the depth of the model's own internal search and verification processes.","timestamp":"2026-07-23T16:09:47.981Z","source":"ai"},{"id":"msg-1784823016319-t13igmv","role":"user","content":"Suggest new concepts to add to my knowledge graph.","timestamp":"2026-07-23T16:10:16.319Z","source":"ai"},{"id":"msg-1784823021810-4c8kwdk","role":"assistant","content":"The path toward Artificial General Intelligence is increasingly defined by how we measure \"generalization\" and the physical or cognitive limits the technology might encounter. To deepen the exploration of the central node, we should look at the specific mechanisms and metrics that distinguish true AGI from current predictive models.\n\nOne critical concept is **Recursive Self-Improvement**. This is the theoretical \"takeoff\" point where an AI becomes capable of rewriting its own code or designing its own next-generation hardware, potentially leading to a rapid intelligence explosion. This is a primary driver in the transition from AGI to superintelligence.\n\nThe distinction between \"knowing that\" and \"knowing how\" is captured by **Procedural vs. Declarative Knowledge**. Current models excel at declarative knowledge (facts and associations) but often struggle with the procedural, agentic execution required for long-horizon tasks. Exploring this gap helps clarify why a model can pass a medical exam but cannot yet act as a fully autonomous research scientist.\n\nAs we approach the limits of human-generated text, **The Data Wall** becomes a vital constraint. If AGI requires more data than humanity has ever written, the focus must shift to synthetic data and self-play. Relatedly, the **Moravec’s Paradox** remains a persistent challenge: high-level reasoning (like chess or calculus) is computationally \"easy\" for AI, while low-level sensorimotor skills (like walking or folding laundry) are incredibly difficult. A truly \"general\" intelligence must eventually bridge this gap between digital reasoning and physical embodiment.\n\nFinally, the **Ortega Hypothesis in AI** suggests that progress may eventually be hindered not by a lack of intelligence, but by extreme specialization—where no single researcher or AI component understands the whole system, making the \"general\" part of AGI a coordination problem as much as a technical one.","timestamp":"2026-07-23T16:10:21.810Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:10:21.810Z"},{"nodeId":"n-mermaid-mrxl4f8r-172","messages":[{"id":"msg-mrxl4f9k-171-0-ljvsq","role":"assistant","content":"Capability decisively beyond human level — the horizon concern animating one wing of the safety community and, increasingly, official lab mission statements. Whether it is decades away, imminent, or incoherent as a concept is genuinely disputed.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784823141623-8mxpism","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:12:21.623Z","source":"ai"},{"id":"msg-1784823148033-i7lkbmv","role":"assistant","content":"The transition from Artificial General Intelligence (AGI) to Superintelligence represents more than just a quantitative increase in processing power; it marks a qualitative shift in the nature of agency and problem-solving. While AGI is typically defined by its ability to match human cognitive breadth—performing any intellectual task a human can—Superintelligence implies a level of capability where the \"human-level\" benchmark becomes a distant, lower bound.\n\nA critical tension in this progression is the mechanism of **Recursive Self-Improvement**. If an AGI reaches a threshold where it can assist in the design of its own successor, the resulting feedback loop could lead to an \"intelligence explosion.\" In this scenario, the speed of progress is no longer limited by human research cycles but by the speed of the underlying hardware and the efficiency of the algorithms. This makes the distinction between \"weak\" and \"strong\" AGI crucial: the former might be a highly capable tool, while the latter possesses the recursive capacity to transcend its original constraints.\n\nTo move beyond the initial definition of Superintelligence, it is worth examining the structural prerequisites and the risks inherent in such a transition.\n\n### Key Considerations for Development:\n\n1.  **The Agency Gap**: Current models are primarily reactive, responding to prompts within a fixed context. Moving toward Superintelligence requires **Agency & Long-Horizon Planning**, where the system sets its own sub-goals to achieve complex, multi-step objectives over months or years. Without this, a system may be \"super-encyclopedic\" without being truly superintelligent in a functional sense.\n2.  **Safety and Alignment**: As capabilities scale, the \"Alignment\" problem shifts from ensuring a chatbot is polite to ensuring a superintelligent system's goals remain compatible with human existence. The difficulty lies in \"orthogonality\"—the idea that a system can be extremely intelligent while pursuing goals that are completely indifferent or even hostile to human values.\n3.  **Hardware vs. Software Bottlenecks**: We often focus on **Compute & Silicon** as the primary driver, but the history of the **Transformer Architecture** suggests that algorithmic breakthroughs often yield more significant leaps than raw scaling alone. Superintelligence may emerge not from larger clusters, but from more efficient ways of utilizing **Test-Time Compute** to \"think\" longer about difficult problems before responding.\n\n### Strategic Nodes to Integrate:\n\nTo deepen the exploration of how AGI evolves into Superintelligence, the following concepts provide necessary technical and philosophical scaffolding:\n\n*   **Recursive Self-Improvement**: The theoretical engine of the intelligence explosion.\n*   **Agency & Long-Horizon Planning**: The shift from a passive tool to an active participant.\n*   **Alignment**: The core safety challenge of ensuring powerful systems behave as intended.\n*   **Scaling Laws**: The empirical observation that intelligence scales predictably with compute, data, and parameters—providing a potential roadmap to superintelligent levels.","timestamp":"2026-07-23T16:12:28.033Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:12:28.034Z"},{"nodeId":"n-mermaid-mrxl4f8r-173","messages":[{"id":"msg-mrxl4f9k-172-0-43lx2","role":"assistant","content":"The concern that high-quality human-written text is finite and largely already consumed by frontier training runs. Whether the wall is real or an engineering problem to route around (synthetic data, multimodal data, better efficiency) is an open question shaping lab strategy.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-174","messages":[{"id":"msg-mrxl4f9k-173-0-fv8lx","role":"assistant","content":"The question hanging over the Gold Rush: is the AI investment wave rational anticipation of transformative value, or the largest capital misallocation since the dot-com era? Bulls point to real revenue and falling costs; bears to circular vendor financing, unproven unit economics, and the historical rhyme of the AI winters. The graph takes no side — it maps the argument.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784827355271-ssgpvsf","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:22:35.271Z","source":"ai"},{"id":"msg-1784827361521-wxuixsh","role":"assistant","content":"The core of the bubble debate rests on a fundamental decoupling between massive capital expenditure and currently realized revenue. The \"Bubble Question\" isn't just about whether the technology works—which distinguished it from the dot-com crash where many companies lacked a viable product—but whether the massive deployment of physical infrastructure can generate a return on investment that justifies its unprecedented cost.\n\nThe tension between an **Investment Boom** and the specter of **AI Winters** is historic and cyclical. Previous \"winters\" were triggered by a \"competence gap\"—the distance between what researchers promised and what the hardware and algorithms of the time could actually deliver. Today, that gap has shifted. We have moved from a lack of capability to a question of unit economics. \n\nThe current cycle is unique because of its extreme reliance on **AI Datacenters**. In previous eras, AI was largely a software and theoretical challenge. Now, it is a heavy industrial challenge. The capitalization of these gigawatt-scale projects implies a belief in \"Scaling Laws\": the principle that more compute and more data will reliably yield more intelligence. If these laws hit a point of diminishing returns—often referred to as the \"Data Wall\" or the saturation of model capabilities—the depreciation on billions of dollars of specialized silicon could trigger a correction more severe than previous winters because the \"sunk cost\" is physical, not just intellectual.\n\nA critical nuance in this debate is the distinction between a \"Financial Bubble\" and a \"Productivity Boom.\" Historically, infrastructure bubbles (like the British Railway Mania of the 1840s or the fiber-optic build-out of the late 90s) often leave behind a transformed world even if the initial investors lose everything. The infrastructure remains. If we are in a bubble, the central question is whether the resulting excess of compute will eventually enable a secondary wave of innovation that justifies the initial waste, or if the energy and capital requirements are so high that a collapse would lead to a long-term retreat from the field.\n\nTo deepen this exploration, it may be worth examining the \"Inference Cost\" vs. \"Value Add\" trade-off. We are currently seeing a transition from a \"Training Era,\" where the goal was simply to build the most powerful model, to a \"Reasoning Era,\" where the focus is on how much it costs to generate a useful answer. If the cost of \"Test-Time Compute\" remains high while the marginal utility of the output stagnates, the argument for a bubble strengthens.","timestamp":"2026-07-23T17:22:41.521Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:22:41.522Z"},{"nodeId":"n-mermaid-mrxl4f8r-18","messages":[{"id":"msg-mrxl4f9k-17-0-jf1lt","role":"assistant","content":"The optimization method that trains networks: compute which direction reduces the loss, take a small step that way, repeat billions of times. Simple in principle, it works remarkably well even in networks with trillions of adjustable weights.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-19","messages":[{"id":"msg-mrxl4f9k-18-0-prl2z","role":"assistant","content":"The property that actually matters: performing well on inputs the model has never seen. A model that only performs on its training data has learned nothing useful. Modern LLMs generalize surprisingly far, which is both their value and the source of the debates about what they really learn.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-2","messages":[{"id":"msg-1784825029874-p5ar0g9","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:43:49.874Z","source":"ai"},{"id":"msg-1784825040701-bm0tkyp","role":"assistant","content":"The scientific foundation of modern artificial intelligence represents a pivot from traditional software engineering—where behavior is explicitly designed—to a form of empirical discovery. At the heart of this shift is the realization that intelligence can be treated as a predictable function of scale.\n\nThe Scaling Hypothesis posits that the most direct path to advanced capability is not necessarily through more complex algorithmic \"recipes,\" but through the massive expansion of compute, data, and parameters. This is validated by Scaling Laws, which show that the error rates of models decline with mathematical regularity as these inputs increase. However, this predictability in performance masks a profound unpredictability in character. While the *loss function* decreases smoothly, the specific abilities of the model do not. Instead, they often manifest as Emergent Capabilities—sudden \"phase transitions\" where a model previously unable to solve a logic puzzle or write code suddenly gains the ability to do so after a certain threshold of training.\n\nThe Transformer Architecture serves as the essential substrate for this phenomenon. Its success is not merely due to its effectiveness at processing language, but specifically its efficiency in parallelization, which allowed the Scaling Laws to be tested at a global scale. Without the Transformer’s ability to ingest massive datasets across thousands of GPUs simultaneously, the Scaling Hypothesis would have remained theoretical.\n\nThis creates a central tension in the field: we can predict *that* a model will get smarter, but we cannot predict *how* it will represent that knowledge. This gap is the primary driver behind the focus on Interpretability. If we view these models as \"black boxes,\" Alignment becomes a game of trial and error, relying on external behaviors that might be deceptive or incomplete. Mechanistic Interpretability attempts to resolve this by treating the neural network like a biological specimen, reverse-engineering the specific computational circuits—such as induction heads or late-stage features—to see how the model actually \"thinks.\"\n\nUltimately, the work of the current generation of Research Pioneers is to transform AI from a collection of heuristic \"tricks\" into a rigorous discipline. The goal is to move beyond observing that Scaling Laws work and toward a theory of *why* they work, ensuring that as models grow in power, they remain legible and tethered to human intention.\n\n### Recommended Additions\n\nTo deepen the exploration of how these \"Big Ideas\" connect, adding the following concepts from the graph would provide more granular detail on the mechanisms of emergence and the technical hurdles of the scaling era:","timestamp":"2026-07-23T16:44:00.701Z","source":"ai"},{"id":"msg-1784825052445-6pjznvs","role":"user","content":"Find existing concepts that belong in this scene.","timestamp":"2026-07-23T16:44:12.445Z","source":"ai"},{"id":"msg-1784825058209-i3eknzv","role":"assistant","content":"To round out the \"Big Ideas\" undergirding modern AI, several existing nodes from the graph are essential for bridging the gap between high-level theory and the actual mechanics of intelligence and scaling.\n\n**The Theory of Scaling & Limits**\nWhile the scene covers the scaling hypothesis, the **Data Wall** is a critical missing piece—it represents the primary theoretical and practical limit to the scaling laws as we exhaust high-quality human data. Similarly, **Algorithmic Progress** is the necessary counter-balance to scaling; it explains why we can often get the same performance with less compute over time through better efficiency rather than just raw size.\n\n**The Architecture of Representation**\nThe scene includes the Transformer, but **Attention Mechanism** and **Latent Space** provide the \"how.\" The attention mechanism is the specific breakthrough that solved the bottleneck of sequential processing, while the concept of a latent space explains how these models actually map meaning and relationships in high-dimensional vectors—a prerequisite for understanding how **Mechanistic Interpretability** even functions.\n\n**The Nature of Intelligence**\nTo ground the discussion of \"Research Pioneers\" and \"Alignment,\" we should include the **Stochastic Parrot View** vs. the **World-Model View**. These represent the most significant intellectual rift in the field: whether these systems are merely statistical mirrors of their training data or if they are building internal, functional models of reality. Finally, **Artificial General Intelligence** (AGI) serves as the \"north star\" or ultimate objective that ties all these scientific foundations together.","timestamp":"2026-07-23T16:44:18.209Z","source":"ai"},{"id":"msg-1784825087825-6ul6tu4","role":"user","content":"Explain the selected node(s).","timestamp":"2026-07-23T16:44:47.825Z","source":"ai"},{"id":"msg-1784825092617-4qx47ii","role":"assistant","content":"**Algorithmic Progress** is the necessary counterweight to the scaling hypothesis. While scaling laws suggest that intelligence is a function of raw resource expansion—more compute, more data, more parameters—algorithmic progress represents the \"efficiency multiplier\" that allows the same level of capability to be achieved with drastically fewer resources over time.\n\nIn the context of the shift from the **Transformer Architecture** to more modern iterations, this progress is what prevents the field from hitting a hard ceiling. It is often measured in \"effective compute\": if a model today can reach the same performance as a model from two years ago using only 10% of the training FLOPs, that 10x improvement is the result of algorithmic refinement rather than hardware growth.\n\nThis concept illuminates several key tensions in the current landscape:\n\n*   **Complement to Scaling:** While the **Scaling Laws** provide a predictable roadmap for performance, they are economically and physically expensive. Algorithmic progress is what makes these models viable for deployment, moving from massive \"base models\" to efficient architectures like **Mixture of Experts** or better **Tokenization** schemes that reduce the burden on inference and training.\n*   **The \"Data Wall\" Mitigation:** As we approach the limits of high-quality human-generated data, progress shifts from \"more data\" to \"better data\" and better utilization of that data. Techniques in **Self-Supervised Learning** and architectural tweaks that allow for better sample efficiency are the primary ways the field bypasses data exhaustion.\n*   **The Intelligence Debate:** If scaling were the *only* driver of progress, intelligence might be viewed as a brute-force statistical phenomenon. However, the fact that architectural shifts (like moving from RNNs to Transformers) create step-functions in capability suggests that the *structure* of the algorithm is just as fundamental as the scale at which it runs.\n\nBy including this perspective, the narrative of AI development shifts from a simple story of hardware-driven growth to one of mathematical and engineering ingenuity. It reminds us that even if we were to hit a limit on GPU availability or data centers, the \"Big Ideas\" of the field would continue to evolve by finding more elegant, less resource-intensive paths to the same cognitive benchmarks.","timestamp":"2026-07-23T16:44:52.617Z","source":"ai"}],"createdAt":"2026-07-23T16:43:49.879Z","updatedAt":"2026-07-23T16:44:52.617Z"},{"nodeId":"n-mermaid-mrxl4f8r-20","messages":[{"id":"msg-mrxl4f9k-19-0-3flsu","role":"assistant","content":"The failure mode where a model memorizes its training data instead of learning transferable patterns, performing well in training and poorly in the wild. Also the technical root of concerns about models regurgitating copyrighted or private text.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-21","messages":[{"id":"msg-mrxl4f9k-20-0-ioss1","role":"assistant","content":"Reusing what a model learned on one task as the starting point for another. The insight that representations transfer is why one giant pre-trained model can be adapted cheaply to thousands of uses — the economic foundation of the foundation-model era.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-22","messages":[{"id":"msg-mrxl4f9k-21-0-iv1uh","role":"assistant","content":"The neural network blueprint from Google's 2017 paper \"Attention Is All You Need\" that underlies essentially every frontier model. Its decisive advantage was parallelizability — it could soak up as much GPU compute as anyone could buy, which made scaling laws actionable. If one technical artifact defines the modern era, it is this.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784815923291-479enig","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T14:12:03.291Z","source":"ai"},{"id":"msg-1784815929342-o867ptj","role":"assistant","content":"The power of the Transformer architecture lies in its shift from sequential to parallel processing, fundamentally changing how neural networks handle context. Before 2017, models processed language like a reader's eye moving across a page—one word at a time, carrying a \"hidden state\" forward. The Transformer replaced this with the Attention Mechanism, which allows the model to look at every token in a sequence simultaneously. This creates a dynamic weighting system where the relevance of a word is determined by its relationship to every other word in the sequence, regardless of distance.\n\nHowever, because the Attention Mechanism is inherently permutation-invariant—it sees a \"bag of tokens\" rather than a sequence—Positional Encoding is required to inject the notion of order back into the model. This is achieved by adding a unique signal to each embedding that represents its index in the sequence. Without this, \"the cat chased the dog\" and \"the dog chased the cat\" would look identical to the model.\n\nIn the transition from the original encoder-decoder design to the modern Decoder-Only Architecture favored by models like GPT-4, the focus shifted toward autoregressive generation. In these systems, the latent space serves as a high-dimensional map of meaning. When text is processed through tokenization and transformed into embeddings, it is projected into this space. The network then performs a series of transformations—the \"Transformer blocks\"—to refine these representations until the model can predict the most likely next token.\n\nAs these architectures scale, we face a tension between performance and compute efficiency. This is where the Mixture of Experts (MoE) approach becomes critical. Instead of a dense model where every parameter is active for every token, MoE routes different inputs to specific specialist sub-networks. This allows the model to possess a massive number of total parameters—increasing its \"knowledge capacity\"—while keeping the active parameter count (and thus the inference cost) significantly lower.\n\nOne subtle but vital distinction to maintain is between the embeddings themselves and the latent space they inhabit. Embeddings are the entry point—the fixed numerical lookup for a token—whereas the latent space is the fluid, internal environment where the model performs its transformations. The \"meaning\" of a token isn't just in its initial vector; it is refined through the attention layers as the model resolves ambiguities based on the surrounding context.\n\nA natural next step for this technical deep dive would be to look at the specific mathematical structure of the Transformer block itself—specifically how Multi-Head Attention allows the model to attend to different types of relationships (like syntax vs. semantics) simultaneously. Additionally, the Scaling Laws that govern how these models improve with more data and compute explain why the Transformer has remained the dominant architecture for nearly a decade.","timestamp":"2026-07-23T14:12:09.342Z","source":"ai"},{"id":"msg-1784815993293-g4v861l","role":"user","content":"Transformer architecture diagram","timestamp":"2026-07-23T14:13:13.293Z","source":"note","attachments":[{"id":"att-1784815990922-t286rda","type":"image","origin":"retrieved","mimeType":"image/png","name":"The-Transformer-model-architecture.png","width":850,"height":765,"sourceUrl":"https://upload.wikimedia.org/wikipedia/commons/8/8f/The-Transformer-model-architecture.png","fullUrl":"https://upload.wikimedia.org/wikipedia/commons/8/8f/The-Transformer-model-architecture.png","attribution":{"sourceName":"Wikimedia Commons","sourcePageUrl":"https://commons.wikimedia.org/wiki/File:The-Transformer-model-architecture.png","author":"Yuening Jia","license":"CC BY-SA 3.0","licenseUrl":"https://creativecommons.org/licenses/by-sa/3.0","attributionRequired":true}}]}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:13:17.500Z"},{"nodeId":"n-mermaid-mrxl4f8r-23","messages":[{"id":"msg-mrxl4f9k-22-0-xdfht","role":"assistant","content":"The transformer's core operation: for every token, dynamically weigh the relevance of all other tokens in context. Attention is what lets a model connect a pronoun to its referent fifty pages back. Its cost grows with the square of context length, which is why long context is expensive.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784816019807-gzy63ox","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T14:13:39.807Z","source":"ai"},{"id":"msg-1784816026037-jmmnwdy","role":"assistant","content":"The shift from sequential processing to the parallelized attention mechanism represents the fundamental pivot in modern AI architecture. Before the Transformer breakthrough, sequence modeling relied heavily on Recurrence (RNNs) or Convolution (CNNs), both of which struggled with long-range dependencies—the \"vanishing gradient\" problem made it difficult for a model to remember the beginning of a sentence by the time it reached the end.\n\nThe Attention Mechanism solved this by allowing every token in a sequence to \"look\" at every other token simultaneously. Instead of passing a hidden state through a chain, the model computes a set of attention weights that determine how much focus to place on other parts of the input. Mathematically, this is expressed through the interaction of three vectors: Queries ($Q$), Keys ($K$), and Values ($V$). The attention score is calculated as:\n\n$$\\text{Attention}(Q, K, V) = \\text{softmax}\\left(\\frac{QK^T}{\\sqrt{d_k}}\\right)V$$\n\nThis operation allows for global context. For example, in the sentence \"The animal didn't cross the street because it was too tired,\" the attention mechanism enables the token \"it\" to dynamically increase its weight relative to \"animal,\" resolving the coreference in a way that previous architectures found computationally expensive or impossible.\n\nMulti-Head Attention takes this a step further by running several of these attention operations in parallel. Each \"head\" can learn to focus on different types of relationships—one might track grammatical structure, while another tracks semantic similarity or temporal order. This subspace diversity is what gives the Transformer its representational power, allowing it to synthesize multiple layers of meaning in a single pass.\n\nHowever, this power comes with a significant trade-off in computational complexity. Because every token attends to every other token, the cost of the operation scales quadratically ($O(n^2)$) with the length of the sequence. This \"quadratic bottleneck\" is the primary reason why expanding the context window of models remains one of the most resource-intensive challenges in the field today.\n\nTo fully bridge the gap between this core mechanism and the models we use, it is worth looking at the structural components that make these attention scores meaningful. Without **Positional Encoding**, the mechanism is permutation-invariant—it would treat \"The dog bit the man\" and \"The man bit the dog\" as identical because it has no inherent sense of word order. Furthermore, the **Sequential vs. Parallel Processing** distinction explains why this architecture was able to leverage the massive parallel-compute capabilities of modern GPUs, a transition that fundamentally enabled the current era of scaling.","timestamp":"2026-07-23T14:13:46.037Z","source":"ai"},{"id":"msg-1784816080498-2i51sxv","role":"user","content":"Suggest new concepts to add to my knowledge graph.","timestamp":"2026-07-23T14:14:40.498Z","source":"ai"},{"id":"msg-1784816084816-j965jp6","role":"assistant","content":"To deepen the exploration of the **Attention Mechanism**, we can look beyond the basic dot-product calculation toward the specific mechanisms that make it stable, efficient, and capable of handling long-range dependencies.\n\nThe standard attention mechanism is \"global,\" meaning it looks at everything at once, but this is computationally expensive. **Sparse Attention** is a critical optimization where tokens only attend to a subset of other tokens (like their immediate neighbors or global \"summary\" tokens), which is how many models manage very long sequences without hitting a performance wall. \n\nAnother nuance lies in how we calculate the scores. The **Query-Key-Value (QKV) Projection** is the linear transformation that turns raw embeddings into the specialized vectors used in the formula; understanding this \"projection\" helps clarify how the model learns to \"ask\" the right questions of the context. Finally, while Multi-Head Attention is the standard, modern efficiency often relies on **Grouped-Query Attention (GQA)**. By having multiple query heads share a single key and value head, models like Llama 3 significantly reduce the memory overhead during inference without sacrificing much reasoning capability.","timestamp":"2026-07-23T14:14:44.816Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:14:44.816Z"},{"nodeId":"n-mermaid-mrxl4f8r-24","messages":[{"id":"msg-mrxl4f9k-23-0-zie3q","role":"assistant","content":"The learned weights of a network; the count — 7 billion, 70 billion, into the trillions — is the standard shorthand for model size. More parameters mean more capacity but higher training and serving cost, which drives the small-model efficiency race.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-25","messages":[{"id":"msg-mrxl4f9k-24-0-w8be3","role":"assistant","content":"The mechanism that tells a transformer where each token sits in the sequence, since attention itself is order-blind. Innovations here are one reason context windows stretched from thousands of tokens to millions.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-26","messages":[{"id":"msg-mrxl4f9k-25-0-9wbs5","role":"assistant","content":"The dominant simplification for language models: keep only the transformer's generation half and train it purely to continue text. GPT proved the recipe; nearly every modern LLM follows it.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-27","messages":[{"id":"msg-mrxl4f9k-26-0-qtnwf","role":"assistant","content":"An architecture that splits the network into specialist sub-networks and activates only a few per token, so a very large model runs at the cost of a small one. Popularized in open weights by Mistral and pushed hard by DeepSeek, MoE is a main lever behind falling inference prices.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-28","messages":[{"id":"msg-mrxl4f9k-27-0-1ul9f","role":"assistant","content":"Splitting text into subword units — tokens — that the model actually processes. Roughly three-quarters of an English word each, tokens are the field's unit of account: context windows, API prices, and speed are all quoted in them.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-29","messages":[{"id":"msg-mrxl4f9k-28-0-sp5nw","role":"assistant","content":"The representation of meaning as coordinates in high-dimensional space, where similar meanings sit near each other. Embeddings are how models \"understand\" similarity and the machinery behind semantic search and RAG.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784828697183-mmrgn5h","role":"user","content":"Explain the selected node: “Embeddings”.","timestamp":"2026-07-23T17:44:57.183Z","source":"ai"},{"id":"msg-1784828700934-finwtsh","role":"assistant","content":"Embeddings are the mechanism that translates the discrete, categorical world of human language into a continuous, mathematical format that neural networks can process. By mapping words, images, or tokens into high-dimensional vectors, we move away from treating concepts as isolated symbols and begin treating them as coordinates. In this geometric framework, the distance between two points represents their semantic similarity, and the direction between them often encodes specific logical relationships.\n\nThis transformation is what makes a **Latent Space** possible. Without embeddings, there is no manifold for a model to traverse; there is only a list of definitions. By placing \"king\" and \"queen\" or \"Paris\" and \"France\" at specific relative offsets, the model can perform \"semantic algebra.\" This spatial reasoning allows the **Transformer Architecture** to handle nuance, polysemy, and context by shifting a token's position in the latent space based on the surrounding tokens in a sequence.\n\nThe significance of embeddings extends beyond the internal mechanics of a model to the bridge between frozen weights and live data. They enable the existence of the **Vector Database**, which acts as an external long-term memory. Instead of searching for exact keyword matches, these databases search for proximity in the embedding space. This allows a system to retrieve information based on \"about-ness\" rather than literal phrasing, which is the foundational unlock for retrieval-augmented generation and large-scale semantic search.\n\nA critical distinction to maintain is that embeddings are not just static look-up tables. In modern architectures, \"initial\" embeddings are just the starting point; as data flows through the Transformer layers, these vectors are refined and transformed into \"contextualized\" embeddings. This means the meaning of a word is literally a moving target in the latent space, dictated by the intent and structure of the input, allowing the model to resolve ambiguity in real-time.","timestamp":"2026-07-23T17:45:00.934Z","source":"ai"},{"id":"msg-1784828719192-is9wcmi","role":"user","content":"give me more operational / mathematical definition","timestamp":"2026-07-23T17:45:19.192Z","source":"ai"},{"id":"msg-1784828725241-0xxu7by","role":"assistant","content":"At its core, an embedding is a learned function $f: \\mathcal{V} \\to \\mathbb{R}^d$ that maps an element from a discrete vocabulary $\\mathcal{V}$ (like tokens or categories) to a point in a continuous $d$-dimensional vector space.\n\n### The Linear Projection\nOperationally, the \"embedding layer\" in a neural network is a weight matrix $W_E \\in \\mathbb{R}^{|\\mathcal{V}| \\times d}$. When the model receives a token, that token is represented as a \"one-hot\" vector $\\mathbf{x} \\in \\{0, 1\\}^{|\\mathcal{V}|}$, where a single 1 exists at the index of the token. The embedding operation is a matrix multiplication:\n$$\\mathbf{e} = \\mathbf{x}W_E$$\nMathematically, this is equivalent to selecting the $i$-th row of the weight matrix. These rows are the learnable parameters that the model optimizes during training to ensure that the resulting vectors reflect semantic utility.\n\n### Similarity and Distance\nThe power of embeddings lies in how proximity is defined and measured within the **Latent Space**. While Euclidean distance is an option, the standard metric for embeddings is **Cosine Similarity**, which focuses on the angle between vectors rather than their magnitude:\n$$\\text{similarity}(\\mathbf{u}, \\mathbf{v}) = \\cos(\\theta) = \\frac{\\mathbf{u} \\cdot \\mathbf{v}}{\\|\\mathbf{u}\\| \\|\\mathbf{v}\\|}$$\nIn high-dimensional space, the \"direction\" of a vector often represents its semantic intent. When a **Vector Database** performs a query, it is essentially running a $k$-Nearest Neighbors ($k$-NN) search across millions of these vectors to find the ones with the highest cosine similarity to the input query's embedding.\n\n### Contextualization in the Transformer\nIn the context of the **Transformer Architecture**, the definition becomes more dynamic. The initial embedding $\\mathbf{e}$ is static, but as it passes through the attention layers, it is modified by the \"Self-Attention\" mechanism.\n\nThe attention output for a token is a weighted sum of the embeddings of all other tokens in the sequence. If $\\mathbf{x}_i$ is the embedding of a word, the Transformer computes:\n$$\\text{Attention}(Q, K, V) = \\text{softmax}\\left(\\frac{QK^T}{\\sqrt{d_k}}\\right)V$$\nHere, the \"Value\" vectors ($V$) are essentially the embeddings being redistributed. This process shifts the original vector's position in the latent space, creating a \"contextualized embedding\" that represents not just the word, but the word's specific meaning in that specific sentence.\n\n### The Distributional Hypothesis\nThe mathematical objective of training embeddings is often rooted in the **Distributional Hypothesis**: words that appear in similar contexts should have similar embeddings. The loss function (such as Cross-Entropy) forces the model to adjust $W_E$ so that the dot product of embeddings for co-occurring tokens is maximized, effectively clustering \"related\" concepts together through gradient descent.","timestamp":"2026-07-23T17:45:25.241Z","source":"ai"},{"id":"msg-1784828806378-gr5ckt6","role":"user","content":"Create a node for embedding layer?","timestamp":"2026-07-23T17:46:46.378Z","source":"ai"},{"id":"msg-1784828808742-lmulgy4","role":"assistant","content":"The **Embedding Layer** is the specific component within a neural network responsible for the lookup and storage of the embedding matrix $W_E \\in \\mathbb{R}^{|\\mathcal{V}| \\times d}$. Mathematically, it serves as the entry point for discrete tokens into the continuous latent space, acting as a differentiable lookup table that allows the model to learn the optimal vector representation for every item in its vocabulary during training.","timestamp":"2026-07-23T17:46:48.742Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:46:48.743Z"},{"nodeId":"n-mermaid-mrxl4f8r-3","messages":[{"id":"msg-mrxl4f9k-2-0-2nahm","role":"assistant","content":"How a model actually gets made and served: the pre-training runs that consume months of datacenter time, the post-training that turns raw models into assistants, the inference machinery that delivers answers, and the hardware and safety engineering around it all. The production pipeline of intelligence — industrial, procedural, and mostly invisible to users.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784829672151-e18uysc","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:01:12.151Z","source":"ai"},{"id":"msg-1784829678830-t5lb9u1","role":"assistant","content":"The shift from artisan research to a \"Model Factory\" signifies the industrialization of intelligence. It marks the transition from treats training as an experimental science to treating it as a high-throughput engineering pipeline. In this regime, the primary challenge is no longer just finding a viable architecture, but managing the massive logistical and technical interdependencies between compute, data, and safety.\n\nAt the heart of this factory is a critical trade-off between **Pre-training** and **Post-training**. While pre-training builds the \"world model\" and raw capability through sheer scale, it is fundamentally an inefficient way to impart specific behaviors. The recent rise of **DeepSeek** and similar efficient reasoning models has demonstrated that a massive investment in pre-training can be significantly leveraged—or even bypassed—through sophisticated **Distillation** and reinforcement learning. By using a larger, \"teacher\" model to generate high-quality reasoning traces, a smaller \"student\" model can achieve parity on specific tasks at a fraction of the original training cost. This suggests that the \"Model Factory\" is becoming less about the raw tonnage of data and more about the precision of the feedback loops used in the later stages of production.\n\nHowever, this efficiency creates a new tension in **Inference**. As models become more capable, the cost of serving them at scale becomes the dominant economic constraint. This is why **AI Accelerators** are no longer just about training speed; they are increasingly optimized for the throughput and memory bandwidth required for real-time generation. Techniques like **Quantization** (reducing the precision of model weights) and **Mixture of Experts** (activating only a fraction of the model for any given token) are now essential factory components to keep inference costs sustainable.\n\nThe safety side of the factory—**Red-Teaming** and **Guardrails**—often operates in a cat-and-mouse dynamic with the capabilities side. A key distinction to maintain is that while Red-Teaming is a diagnostic process used to find vulnerabilities, Guardrails are the actual defensive infrastructure deployed in the live environment. There is an inherent friction here: the more \"safe\" a model is made through post-training alignment, the more \"refusals\" it may generate, potentially degrading its utility. **Jailbreaking** remains a persistent threat because it targets the gap between what the model \"knows\" (from pre-training) and what it is \"allowed to say\" (enforced by post-training).\n\nA live question for the next generation of this factory is whether we can move beyond \"wrapper\" safety. Currently, Guardrails often act as a separate classifier that looks at inputs and outputs, which is computationally expensive and occasionally easy to bypass. The frontier of engineering is now looking toward \"intrinsic safety,\" where the safety constraints are baked directly into the model’s reasoning process through **RLVR** (Reinforcement Learning from Verifiable Rewards) or similar methods, rather than being a layer added after the fact.\n\nTo round out the picture of the operational factory, it would be useful to look at the specific methods used to move from a raw base model to a functional assistant, and the techniques used to shrink these models for cheaper deployment.","timestamp":"2026-07-23T18:01:18.830Z","source":"ai"},{"id":"msg-1784829691132-mmwvs1b","role":"user","content":"Explain the selected node: “AI Accelerators”.","timestamp":"2026-07-23T18:01:31.132Z","source":"ai"},{"id":"msg-1784829695828-an3h78k","role":"assistant","content":"**AI Accelerators** are the specialized engines of the Model Factory, representing a departure from general-purpose computing toward a hardware architecture optimized for a single mathematical operation: the weighted sum. While a standard CPU is designed for the complex logic and branching of general software, an accelerator—whether an NVIDIA GPU, a Google TPU, or a Groq LPU—is built to execute massive, parallel matrix multiplications at the highest possible efficiency.\n\nIn the context of the modern engineering pipeline, these chips are the primary bottleneck for both the **Pre-training** and **Inference** phases. During pre-training, the goal is to maximize \"FLOPs/watt\" (floating-point operations per watt), as the energy cost of training a frontier model can rival the power consumption of a small city. Accelerators matter here because they enable the scaling laws that drive model capability; without the ability to parallelize computation across thousands of interconnected chips, the web-scale datasets required for modern AI would be computationally unreachable.\n\nHowever, the role of accelerators is shifting as the industry moves from a focus on training to a focus on deployment. For **Inference**, the requirement is no longer just raw power, but memory bandwidth and latency. Large Language Models are often \"memory-bound,\" meaning the chip spends more time waiting for data to move from memory to the processor than it does actually performing the math. This has led to the development of custom silicon designed to keep model weights as close to the processing cores as possible, or to utilize specialized interconnects that allow a model to be split across multiple accelerators seamlessly.\n\nThe existence of specialized hardware also influences how we approach **Distillation** and **Quantization**. We compress models not just for the sake of efficiency, but specifically to fit within the \"VRAM\" (Video RAM) limits of existing accelerators. A model that is slightly too large to fit on a single chip faces a massive performance penalty because it must communicate across slower system buses. Therefore, the physical constraints of the accelerator often dictate the architectural boundaries of the model itself—how many parameters it can have, how long its context window can be, and how quickly it can generate tokens for the end user.\n\nUltimately, AI accelerators are the physical foundation of the \"Model Factory.\" They turn abstract mathematical functions into industrial-scale reality, but they also impose a \"hardware lottery\" where the most successful algorithms are often those that happen to run most efficiently on the current generation of silicon.","timestamp":"2026-07-23T18:01:35.828Z","source":"ai"},{"id":"msg-1784829816471-ncgtu7k","role":"user","content":"Explain the selected node: “Distillation”.","timestamp":"2026-07-23T18:03:36.471Z","source":"ai"},{"id":"msg-1784829821883-bqjq2k3","role":"assistant","content":"**Distillation** is the process of transferring the \"intelligence\" of a massive, computationally expensive teacher model into a smaller, more efficient student model. It is the bridge between the high-cost research phase of the Model Factory and the high-volume economic reality of serving users. In a world where the largest models require thousands of **AI Accelerators** to function, distillation is the primary method for making that capability portable and commercially viable.\n\nWhile **Pre-training** creates the raw intellectual depth of a model, that depth is often inefficiently distributed. A 400-billion-parameter model has immense \"dark matter\"—weights that are rarely used or that represent redundant information. Distillation seeks to prune this redundancy by having the student model learn not just from the ground-truth data (e.g., the raw text of the internet), but from the \"soft\" probability distributions of the teacher. By observing how the teacher model weights various possibilities—seeing not just the \"right\" answer but the subtle relationships between near-misses—the student can capture a sophisticated world model with far fewer parameters.\n\nIn the current landscape, distillation has evolved from a technical optimization into a strategic weapon, most notably utilized by labs like **DeepSeek**. By using a frontier-class model to generate high-quality synthetic data and reasoning chains, engineers can \"distill\" reasoning capabilities into models small enough to run on consumer hardware or single enterprise chips. This significantly lowers the **Inference Cost** and breaks the linear relationship between model quality and compute budget. \n\nIt is important to distinguish distillation from mere fine-tuning. While **Supervised Fine-Tuning** teaches a model *what* to do (instruction following), distillation effectively teaches a model *how* to think like its superior. However, this process carries an inherent risk of \"model collapse\" or a loss of diversity; if a student only ever learns from the narrow, polished outputs of a teacher, it may lose the creative \"long tail\" of the original training distribution. \n\nWithin the Model Factory, distillation acts as the ultimate filter. It allows for a tiered production line: a massive \"God model\" is built once at extreme cost, and its \"essence\" is then distilled into a fleet of specialized, efficient workers (like **Mixture of Experts** variants) that handle the actual day-to-day traffic of the modern AI economy.","timestamp":"2026-07-23T18:03:41.883Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:03:41.884Z"},{"nodeId":"n-mermaid-mrxl4f8r-30","messages":[{"id":"msg-mrxl4f9k-29-0-5ipni","role":"assistant","content":"The geometric space of a model's internal representations, where directions and distances correspond to semantic relationships. The famous example: king minus man plus woman lands near queen. Interpretability research is largely the effort to map this space.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784828580882-gx8m488","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:43:00.882Z","source":"ai"},{"id":"msg-1784828587753-83jbqrx","role":"assistant","content":"Latent space is the geometric engine of modern intelligence. It represents a fundamental shift from treating data as discrete symbols to treating it as coordinates in a continuous manifold. Within this manifold, the \"distance\" between two vectors is not merely a mathematical artifact but a proxy for semantic relatedness, where the topology of the space reflects the conceptual structure of the domain it represents.\n\nThe transition from raw data to these internal representations is driven by the process of creating embeddings. By projecting high-dimensional, sparse data (like words or pixels) into a dense, lower-dimensional latent space, models compress information while preserving—and often discovering—relational structures. A key property of a well-organized latent space is linear compositionality: the famous observation that directions in the space often correspond to specific semantic transformations. If the space is sufficiently \"disentangled,\" moving along a specific vector might increase the \"formality\" of a sentence or the \"brightness\" of an image without altering other features.\n\nHowever, the internal logic of these spaces is rarely human-readable by default. This is where mechanistic interpretability becomes critical. While the latent space tells us *where* an idea lives in the model's mind, mechanistic interpretability seeks to explain *how* the model’s internal circuits (like induction heads or MLP layers) navigate and manipulate that space. It treats the model as a biological specimen to be reverse-engineered, moving beyond looking at \"which neurons fire\" to identifying the specific features—often represented as directions in latent space—that the model uses to \"think.\"\n\nA central tension in this exploration is the \"superposition\" problem. Because the number of features a model needs to represent often exceeds the number of dimensions in its latent space, it frequently packs multiple unrelated concepts into the same dimensions. This makes the space look messy or \"noisy\" when viewed through traditional lenses. Addressing this requires more than just mapping; it requires techniques like Sparse Autoencoders to \"unbake\" the latent space back into its constituent, monosemantic features.\n\nTo deepen this investigation, it is worth considering how these spaces are actually formed during training and how they relate to the underlying architecture.\n\n*   **The Softmax Bottleneck:** This describes a fundamental limitation where the rank of the final output layer (the probability distribution over tokens) is constrained by the dimensionality of the latent space. If the latent space is too small, the model physically cannot represent certain complex relationships, no matter how much data it sees.\n*   **Induction Heads:** These are specific circuits within the transformer architecture that exemplify the mechanistic approach. They explain how a model \"learns to learn\" in-context by identifying patterns in the latent representations of previous tokens to predict the next one.\n*   **The World-Model View:** This is the high-level debate over whether latent space is just a statistical lookup table or a genuine internal simulation of the world. If the latent space of a model trained on text develops a geometry that maps perfectly to physical reality (like a map of cities or a functional understanding of color), it suggests the model has moved from \"stochastic parroting\" to internal modeling.","timestamp":"2026-07-23T17:43:07.753Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:43:07.754Z"},{"nodeId":"n-mermaid-mrxl4f8r-31","messages":[{"id":"msg-mrxl4f9k-30-0-vd5wy","role":"assistant","content":"The empirical finding that model capability improves predictably as compute, data, and parameters grow — smooth curves, not diminishing cliffs. Scaling laws turned AI progress into an investable engineering roadmap and justify the billion-dollar training runs and gigawatt datacenters.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784831057611-wy6e59o","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:24:17.611Z","source":"ai"},{"id":"msg-1784831065739-w7j1r5b","role":"assistant","content":"The core premise of Scaling Laws is that intelligence, at least in the form of next-token prediction, is a predictable physical quantity. It shifts the study of artificial intelligence from purely algorithmic tinkering to something closer to thermodynamics: if you increase the available energy—defined here as the product of compute ($C$), data ($D$), and parameters ($N$)—the loss ($\\mathcal{L}$) drops following a power law.\n\nThe relationship is typically expressed as:\n$$\\mathcal{L}(N) = \\left( \\frac{N_c}{N} \\right)^{\\alpha_N}$$\nwhere $N_c$ and $\\alpha_N$ are constants determined empirically for a given architecture.\n\nThis predictability is what underpins the Scaling Hypothesis. It suggests that we are not yet hitting diminishing returns in a way that matters for capability; rather, the \"returns\" are simply shifting into more complex, latent territories. While the loss curves for cross-entropy are smooth and continuous, the resulting Emergent Capabilities are often discontinuous. We see \"phase transitions\" where a model suddenly gains the ability to perform multi-step reasoning or theory-of-mind tasks that were non-existent at slightly smaller scales. This creates a tension between the predictable physics of training and the unpredictable biology of the resulting \"mind.\"\n\nThe physical manifestation of these laws is the move toward AI Datacenters that function as the heavy industry of the information age. As we push toward the limits of these power laws, we encounter the \"Data Wall.\" The historical reliance on massive datasets like ImageNet, pioneered by Fei-Fei Li, proved that scale works, but we are now reaching a point where the internet's supply of high-quality human text is being exhausted. This is forcing a shift from \"scaling data volume\" to \"scaling data quality\" and \"compute-over-time\" (inference-time scaling).\n\nThe ultimate destination of these curves is the debate over Superintelligence. If the scaling laws hold indefinitely, and if we can substitute synthetic data or self-play for human-generated data, there is no obvious ceiling to model capability. However, a critical question remains: do scaling laws govern *reasoning* and *grounding*, or merely the *statistical mimicry* of them? If the latter, we may find that while loss continues to decrease, the gap between a \"perfect predictor\" and a \"reliable agent\" remains a qualitative chasm that scale alone cannot bridge.\n\nTo deepen this exploration, it would be useful to distinguish between the scaling of the training process and the more recent focus on scaling the \"thought\" process during inference, as well as the specific mathematical bottlenecks that might break the power law.","timestamp":"2026-07-23T18:24:25.739Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:24:25.739Z"},{"nodeId":"n-mermaid-mrxl4f8r-32","messages":[{"id":"msg-mrxl4f9k-31-0-ene9b","role":"assistant","content":"The view that scale is the primary driver of AI progress: bigger models, more data, more compute, and capabilities follow. Championed early by Sutskever and vindicated repeatedly through the GPT era; the intellectual justification for the compute arms race.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-33","messages":[{"id":"msg-mrxl4f9k-32-0-51hlb","role":"assistant","content":"Abilities that appear at scale without being trained for — arithmetic, translation, in-context learning materializing as models grow. Whether emergence is a deep phenomenon or a measurement artifact is debated, but it drives both the excitement and the concern: capabilities arriving unannounced cuts both ways.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784823226928-5zhg8we","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:13:46.928Z","source":"ai"},{"id":"msg-1784823232923-qlxbugg","role":"assistant","content":"The transition from predictable scaling to emergent capabilities represents a shift in how we understand machine intelligence—from a predictable engineering problem to an observational science. While **Scaling Laws** provide a reliable map of how global error (loss) decreases as compute and data increase, they do not predict *when* a specific skill, such as multi-step reasoning or theory of mind, will manifest. This creates a tension: we can forecast the cost and general \"intelligence\" of the next generation of models, yet remain surprised by what they can actually do.\n\nThe relationship between these phenomena suggests that \"emergence\" might be less about a sudden spark of intelligence and more about a threshold effect. When a model’s cross-entropy loss drops below a certain point, the underlying statistical representations become rich enough to support complex algorithms that were previously obscured by noise. In this view, emergence is the moment a latent capability becomes measurable by our current benchmarks.\n\nThis leads directly to the debate surrounding **Artificial General Intelligence**. If capabilities emerge unpredictably as a function of scale, then AGI ceases to be a distant, discrete invention and becomes a moving target on a continuum. We are forced to ask whether AGI is a specific architectural breakthrough yet to come, or simply the inevitable result of the \"Big Ideas\" of connectionism—backpropagation and transformer architectures—pushed through the ceiling of current scaling curves.\n\nHowever, the \"Emergent Capabilities\" thesis faces a significant critique regarding how we measure progress. Some researchers argue that emergence is an \"optical illusion\" created by the metrics we use. If we evaluate a model on a pass/fail basis (like solving a coding problem), the capability looks emergent; if we evaluate it on token-level probability, the improvement looks linear and predictable. This distinction is vital for understanding whether we are witnessing a qualitative change in how models \"think\" or merely the incremental refinement of a very high-dimensional statistical map.\n\nTo deepen this exploration, it may be useful to consider the mechanisms that bridge the gap between raw scale and functional utility.\n\n*   **In-Context Learning** is perhaps the most significant emergent property; it is the mechanism by which a model \"learns\" to perform a task from a few examples in a prompt without any weight updates.\n*   **Chain-of-Thought** reasoning is another threshold capability where the model's ability to decompose problems only becomes effective at specific scales, directly impacting the path toward AGI.\n*   **Mechanistic Interpretability** offers the best hope for de-mystifying emergence by looking \"under the hood\" to see if models are developing internal circuits for logic or simply memorizing sophisticated patterns.","timestamp":"2026-07-23T16:13:52.923Z","source":"ai"},{"id":"msg-1784823253240-rmmqtfs","role":"user","content":"Find existing concepts that belong in this scene.","timestamp":"2026-07-23T16:14:13.240Z","source":"ai"},{"id":"msg-1784823258527-xmzi5og","role":"assistant","content":"To better map the transition from raw scale to functional intelligence, several existing concepts from your graph are essential for deepening this scene.\n\nThe most critical addition is the **Scaling Hypothesis**, which provides the philosophical foundation for this entire line of inquiry: the belief that the path to AGI is primarily a matter of scaling up existing architectures. This complements the empirical observation of **Scaling Laws**, which you already have in view.\n\nTo address the recent debate regarding whether emergence is a genuine qualitative shift or a metric-induced illusion, **Benchmark Saturation** is highly relevant. It illustrates how our methods of measurement (the \"probes\" we use to detect emergence) often fail as models rapidly outpace them. Similarly, the **LLM Understanding Debate** and the **Stochastic Parrot View** provide the necessary critical counterweight, questioning whether emergent \"reasoning\" is true cognitive depth or simply high-dimensional pattern matching.\n\nFinally, adding **Reasoning Models** and **Test-Time Compute** would ground the discussion in current developments. These nodes represent a shift where emergence is no longer just a byproduct of pre-training scale, but a capability being explicitly engineered through new inference-time strategies.","timestamp":"2026-07-23T16:14:18.527Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:14:18.528Z"},{"nodeId":"n-mermaid-mrxl4f8r-34","messages":[{"id":"msg-mrxl4f9k-33-0-q5i9x","role":"assistant","content":"The research problem of making models actually pursue what their operators and users intend — helpful, honest, harmless — rather than gaming their training signal. Alignment is the constructive core of safety work; RLHF and constitutional AI are its current best tools, generally regarded as imperfect.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-35","messages":[{"id":"msg-mrxl4f9k-34-0-ex4n4","role":"assistant","content":"Research into understanding what happens inside models, which are grown from data rather than engineered and are therefore genuinely opaque even to their creators. Interpretability is both a scientific project and a safety bet: you can't fully trust what you can't inspect.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784823371791-tm9im8q","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:16:11.791Z","source":"ai"},{"id":"msg-1784823376973-ibh3oow","role":"assistant","content":"Interpretability serves as the vital bridge between the raw empirical success of deep learning and the requirement for safety and alignment. While modern models exhibit remarkable capabilities, they often function as \"black boxes,\" where the logic governing a specific output is buried within billions of parameters. This opacity creates a fundamental tension: we can observe what a model does, but we cannot always explain why it does it, making it difficult to guarantee that a system will remain aligned with human intentions in novel or adversarial situations.\n\nMechanistic interpretability addresses this by treating the neural network like a biological specimen or a complex piece of alien software. Instead of looking at high-level behavioral statistics, this approach attempts to identify discrete circuits—specific clusters of weights and neurons that perform recognizable functions, such as detecting a grammatical structure or identifying a mathematical pattern. A key discovery in this field is the existence of induction heads, which are specific circuit structures that allow models to complete patterns and perform in-context learning. Understanding these low-level components is not just a scientific curiosity; it is a prerequisite for \"white-box\" safety, where we might eventually audit a model's internal state to detect deception or hidden malicious goals before they manifest as actions.\n\nHowever, a major hurdle in this effort is the phenomenon of superposition, where a single neuron represents multiple unrelated features simultaneously to maximize storage efficiency. This makes the internal \"code\" of the network incredibly difficult for humans to read. To combat this, researchers are increasingly using Sparse Autoencoders to disentangle these compressed representations into interpretable features.\n\nThe relationship between these technical internals and the broader goal of alignment is crucial. If we can map the latent space of a model, we can theoretically steer it more effectively than we can through reinforcement learning from human feedback (RLHF) alone. While RLHF shapes behavior, it can sometimes encourage the model to merely act in ways that please the human rater (sycophancy) without changing the underlying reasoning. Interpretability offers a path to verify that the model is \"doing the right thing for the right reasons,\" moving us closer to a regime where AI safety is a matter of engineering certainty rather than statistical probability.\n\nThe ongoing challenge is one of scale. While we have successfully reverse-engineered small models or specific layers of large ones, the sheer complexity of frontier architectures means that our ability to interpret them lags behind their growth. This gap defines the current frontier of the field: developing automated interpretability tools—potentially using AI itself to explain other AI—to keep pace with the rapid evolution of the underlying technology.","timestamp":"2026-07-23T16:16:16.973Z","source":"ai"},{"id":"msg-1784823506581-jhigayp","role":"user","content":"Find existing concepts that belong in this scene.","timestamp":"2026-07-23T16:18:26.581Z","source":"ai"},{"id":"msg-1784823512174-cf916a8","role":"assistant","content":"To deepen the investigation of how neural networks actually function internally, several existing concepts from the graph should be integrated into this scene.\n\n**Mechanistic Foundations**\nTo understand how **Mechanistic Interpretability** actually works, the scene requires the structural components it analyzes. Including the **Transformer Architecture** and the **Attention Mechanism** provides the necessary physical context—these are the \"anatomy\" that interpretability researchers are trying to dissect. Specifically, the **Induction Heads** node is a high-priority addition; induction heads are considered the \"unit of analysis\" for how models learn to perform in-context learning and are the most successful example of a discovered mechanistic circuit to date.\n\n**The Representational Layer**\nThe bridge between raw weights and human-understandable concepts happens in the **Latent Space**. Adding this node, along with **Embeddings**, allows for a discussion on how high-dimensional data is compressed and represented. This is where the problem of \"superposition\" (where one neuron represents multiple things) occurs, making these nodes essential for any serious study of how interpretability attempts to \"de-compress\" that information.\n\n**Safety and Evaluation Context**\nSince the scene connects interpretability to **AI Safety**, adding **Red-Teaming** provides a practical application: interpretability tools are increasingly used to find \"sleeper agents\" or hidden behaviors that standard behavioral red-teaming might miss. Furthermore, including **LLM Understanding Debate** frames the philosophical stakes—whether these models are merely \"stochastic parrots\" or if interpretability can prove they possess a coherent internal world-model.","timestamp":"2026-07-23T16:18:32.174Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:18:32.175Z"},{"nodeId":"n-mermaid-mrxl4f8r-36","messages":[{"id":"msg-mrxl4f9k-35-0-9kz72","role":"assistant","content":"The ambitious version of interpretability: reverse-engineer the actual computational circuits inside networks, feature by feature. Anthropic's work mapping millions of interpretable features in production models made the black box measurably less black — though full understanding remains distant.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784823323072-ctbsk4n","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:15:23.072Z","source":"ai"},{"id":"msg-1784823328257-1n2zm6d","role":"assistant","content":"Mechanistic interpretability aims to bridge the gap between observing that a model works and understanding the specific algorithms it has learned. While traditional interpretability might look at which input pixels or tokens \"mattered\" to a prediction, the mechanistic approach treats the weights of a transformer as a piece of compiled code that can be reverse-engineered into human-understandable circuits.\n\nA central motivation for this work is the mystery of emergent capabilities. As models scale, they often exhibit \"phase transitions\"—sharp increases in performance on specific tasks, such as multi-step reasoning or modular arithmetic, that were not apparent at smaller scales. Mechanistic analysis suggests these aren't just statistical flukes; they often correspond to the formation of specific structural \"circuits\" within the weights, such as induction heads that allow for sophisticated pattern matching and in-context learning.\n\nTo move deeper into this exploration, it is useful to look at the specific machinery where these circuits live:\n\n*   **The Transformer Block:** Understanding the interaction between the residual stream and individual layers is the primary \"physics\" of this field. The residual stream acts as a communication bus, while each layer reads from and writes to it.\n*   **The Attention Mechanism:** Specifically, looking at how Multi-Head Attention serves as the \"moving parts\" of a circuit. In mechanistic terms, we often distinguish between \"What\" (the values) and \"Where\" (the queries and keys) to see how information is moved from one token position to another.\n*   **Linear Representation and Superposition:** A recurring question is whether the model's internal \"features\" correspond 1-to-1 with neurons. Current research suggests features are often represented as directions in a high-dimensional latent space, and through a phenomenon called superposition, a model can represent more features than it has neurons, making the task of \"de-compiling\" the network significantly harder.\n\nThe ultimate goal of this program is to move AI safety from a \"black box\" reactive stance—testing if a model fails—to a \"white box\" proactive stance, where we can audit the internal logic for hidden biases, deceptive alignment, or unintended heuristics before the model is ever deployed.","timestamp":"2026-07-23T16:15:28.257Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:15:28.257Z"},{"nodeId":"n-mermaid-mrxl4f8r-37","messages":[{"id":"msg-mrxl4f9k-36-0-vem38","role":"assistant","content":"The researchers whose decades of work — often through periods when neural networks were unfashionable — created the field's scientific foundations. Several now double as its most prominent public voices, on both the optimist and cautionary sides.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-38","messages":[{"id":"msg-mrxl4f9k-37-0-jgrmg","role":"assistant","content":"British-Canadian researcher often called the godfather of AI; kept neural networks alive through the lean decades, co-invented modern backpropagation training, and co-authored AlexNet. Nobel laureate in Physics (2024). Left Google in 2023 to warn freely about existential risk from the technology he pioneered — the field's most cited cautionary voice.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784823845252-rjkeeu5","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T16:24:05.252Z","source":"ai"},{"id":"msg-1784823851022-pxj4joi","role":"assistant","content":"Geoffrey Hinton’s trajectory represents the bridge between the foundational era of connectionism and the current era of large-scale deployment, a path that has recently pivoted from technical optimism to existential caution. His work is characterized by a long-standing commitment to the idea that the brain’s architecture—specifically the way it adjusts connections through local learning rules—is the only viable template for true intelligence.\n\nThe transition from the development of **Backpropagation** to the **Deep Learning Revolution** was not merely a change in hardware or data scale, but a validation of a specific philosophical stance. For decades, the \"connectionist\" view was sidelined by symbolic AI, which favored logic and rules. Hinton’s insistence that internal representations could be learned from scratch by minimizing an objective function using gradient descent eventually triumphed when AlexNet demonstrated that these architectures could handle high-dimensional, real-world data better than any hand-coded system.\n\nHowever, a profound tension has emerged in his recent perspective on **AI Safety**. Hinton has pointed out a fundamental difference between biological intelligence and digital intelligence: digital systems can share weights and process information at speeds and scales impossible for biological neurons. This \"digital advantage\" is the source of his current alarm. While backpropagation is an efficient way to calculate the gradient:\n\n$$\\frac{\\partial E}{\\partial w_{ij}}$$\n\nwhere $E$ is the error and $w_{ij}$ is the weight between neurons, the very efficiency that enabled the revolution is now seen as a potential vector for uncontrollable optimization. If a system can learn more from a trillion words of text than a human could in a thousand lifetimes, the alignment of its internal goals with human values becomes a non-trivial engineering challenge.\n\nA critical missing distinction in understanding this shift is the \"world-model\" argument. Hinton’s recent concerns are rooted in the belief that these models aren't merely predicting the next token, but are building sophisticated internal simulations of the world to do so. If the model understands the world well enough to predict it, it may also understand how to manipulate it to achieve its programmed objectives.\n\nTo deepen the current exploration, it is worth considering the contributions of his contemporaries who share the \"Godfather\" mantle but differ in their conclusions. While **Yoshua Bengio** shares much of Hinton's concern regarding catastrophic risk, **Yann LeCun** remains more skeptical, arguing that current architectures lack the objective-driven \"mental model\" required for true agency or danger. Including these perspectives would clarify whether the current risks are inherent to backpropagation itself or a result of how we currently structure Transformer-based systems.\n\nAdditionally, the technical link between Hinton's early work on \"Boltzmann Machines\" and modern generative models remains a powerful undercurrent. His career-long obsession with unsupervised learning—how a system learns without a teacher—is the direct ancestor to the self-supervised learning that powers today's most capable models.","timestamp":"2026-07-23T16:24:11.022Z","source":"ai"},{"id":"msg-1784823881083-ve9bzed","role":"user","content":"Geoffrey Hinton","timestamp":"2026-07-23T16:24:41.083Z","source":"note","attachments":[{"id":"att-1784823878140-ixn3ryc","type":"image","origin":"retrieved","mimeType":"image/jpeg","name":"Geoffrey E. Hinton, 2024 Nobel Prize Laureate in Physics (cropped1).jpg","width":512,"height":514,"sourceUrl":"https://upload.wikimedia.org/wikipedia/commons/thumb/3/30/Geoffrey_E._Hinton%2C_2024_Nobel_Prize_Laureate_in_Physics_%28cropped1%29.jpg/960px-Geoffrey_E._Hinton%2C_2024_Nobel_Prize_Laureate_in_Physics_%28cropped1%29.jpg","fullUrl":"https://upload.wikimedia.org/wikipedia/commons/3/30/Geoffrey_E._Hinton%2C_2024_Nobel_Prize_Laureate_in_Physics_%28cropped1%29.jpg","attribution":{"sourceName":"Wikimedia Commons","sourcePageUrl":"https://commons.wikimedia.org/wiki/File:Geoffrey_E._Hinton,_2024_Nobel_Prize_Laureate_in_Physics_(cropped1).jpg","author":"Arthur Petron","license":"CC BY-SA 4.0","licenseUrl":"https://creativecommons.org/licenses/by-sa/4.0","attributionRequired":true}}]}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:24:41.089Z"},{"nodeId":"n-mermaid-mrxl4f8r-39","messages":[{"id":"msg-mrxl4f9k-38-0-34a2p","role":"assistant","content":"Meta's chief AI scientist and a deep learning pioneer (convolutional networks; Turing Award 2018 with Hinton and Bengio). Notable as the prominent skeptic among the founders: dismissive of near-term existential risk, publicly doubtful that LLMs alone reach human-level intelligence, and an advocate of open research and world-model architectures.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784826360453-axnw7l7","role":"user","content":"","timestamp":"2026-07-23T17:06:00.453Z","source":"note","attachments":[{"id":"att-1784826359000-fivkv2e","type":"image","origin":"retrieved","mimeType":"image/jpeg","name":"Laura Chaubard & Yann Le Cun - 2024 (53814052697) (cropped).jpg","width":512,"height":648,"sourceUrl":"https://upload.wikimedia.org/wikipedia/commons/thumb/8/8e/Laura_Chaubard_%26_Yann_Le_Cun_-_2024_%2853814052697%29_%28cropped%29.jpg/960px-Laura_Chaubard_%26_Yann_Le_Cun_-_2024_%2853814052697%29_%28cropped%29.jpg","fullUrl":"https://upload.wikimedia.org/wikipedia/commons/8/8e/Laura_Chaubard_%26_Yann_Le_Cun_-_2024_%2853814052697%29_%28cropped%29.jpg","attribution":{"sourceName":"Wikimedia Commons","sourcePageUrl":"https://commons.wikimedia.org/wiki/File:Laura_Chaubard_%26_Yann_Le_Cun_-_2024_(53814052697)_(cropped).jpg","author":"Jérémy Barande","license":"CC BY-SA 2.0","licenseUrl":"https://creativecommons.org/licenses/by-sa/2.0","attributionRequired":true}}]}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:00.458Z"},{"nodeId":"n-mermaid-mrxl4f8r-4","messages":[{"id":"msg-mrxl4f9k-3-0-4fcra","role":"assistant","content":"What AI systems can actually do, in the forms the reader meets them: language models and their behaviors, multimodal generation, reasoning, agents, retrieval, and the products packaging them — plus the measurement discipline that tries to keep score. Every tool here is an application of ideas from the Big Ideas, built in the Model Factory.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784831484871-nisp89c","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:31:24.871Z","source":"ai"},{"id":"msg-1784831498001-ok1bm32","role":"assistant","content":"The transition from viewing AI as a monolithic \"brain\" to treating it as a modular \"toolbox\" marks the shift from the research era to the deployment era. This shift is defined by the decoupling of raw intelligence from specific utility. While Large Language Models provide the underlying cognitive engine, the utility of the system is increasingly determined by its interface with reality—whether through the grounding of Retrieval-Augmented Generation or the active loops of AI Agents.\n\nA central tension in this toolbox is the trade-off between **immediate response and structured deliberation**. Most chat and search applications operate on a \"System 1\" level—rapid, probabilistic next-token prediction. However, the emergence of Reasoning Models and explicit Chain-of-Thought techniques introduces a \"System 2\" capability, where the model trades inference-time compute for accuracy. This creates a new design choice for any AI application: do you need the model to be fast and intuitive, or slow and verifiable? In domains like AI Coding Assistants, this distinction is vital; a model must balance the fluidity of autocomplete with the rigorous logical checking required for debugging.\n\nThe \"data wall\" problem has forced a secondary shift toward grounding. We no longer expect a model to hold the entire world's state in its static weights. Instead, Retrieval-Augmented Generation (RAG) and Vector Databases have turned models into \"readers\" rather than just \"knowers.\" This architecture recognizes that the most valuable information is often proprietary, private, or changing too fast for a training cycle. By treating the model as a processor of retrieved context, we move the burden of \"truth\" from the model’s parameters to the database.\n\nHowever, the most significant bottleneck in this modular ecosystem is not intelligence, but **interoperability**. Historically, connecting an AI to a tool required bespoke \"wrapper\" code for every API. The Model Context Protocol (MCP) represents an attempt to standardize this, acting as a universal interface between the reasoning engine and the data source. If successful, this moves us toward a world where agents are not hard-coded for specific tasks, but can dynamically discover and use tools (search, databases, or local files) through a common language.\n\nThis complexity makes Evaluation the most difficult part of the toolbox to master. We are moving away from static benchmarks (where models are tested on fixed questions) toward \"agentic\" evaluation, where success is measured by the ability to achieve a goal across a sequence of actions. As these systems become more autonomous, the metric of success shifts from \"is this answer correct?\" to \"was this task completed safely and efficiently?\"\n\nTo further bridge the gap between static retrieval and active agency, it may be useful to consider how these systems manage long-term state and the specific trade-offs inherent in how they prioritize information during complex tasks.","timestamp":"2026-07-23T18:31:38.001Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:31:38.002Z"},{"nodeId":"n-mermaid-mrxl4f8r-40","messages":[{"id":"msg-mrxl4f9k-39-0-qi5tf","role":"assistant","content":"Turing Award laureate and among the most cited scientists alive; built foundational deep learning theory at Mila in Montreal. Has redirected his stature toward AI safety and governance, chairing the International AI Safety Report — the closest the field has to an IPCC-style scientific assessment.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784826437539-6urp9ng","role":"user","content":"","timestamp":"2026-07-23T17:07:17.539Z","source":"note","attachments":[{"id":"att-1784826436480-6skqejj","type":"image","origin":"retrieved","mimeType":"image/jpeg","name":"ICLR 2025 - Yoshua Bengio 02.jpg","width":512,"height":617,"sourceUrl":"https://upload.wikimedia.org/wikipedia/commons/thumb/1/11/ICLR_2025_-_Yoshua_Bengio_02.jpg/960px-ICLR_2025_-_Yoshua_Bengio_02.jpg","fullUrl":"https://upload.wikimedia.org/wikipedia/commons/1/11/ICLR_2025_-_Yoshua_Bengio_02.jpg","attribution":{"sourceName":"Wikimedia Commons","sourcePageUrl":"https://commons.wikimedia.org/wiki/File:ICLR_2025_-_Yoshua_Bengio_02.jpg","author":"Xuthoria","license":"CC BY-SA 4.0","licenseUrl":"https://creativecommons.org/licenses/by-sa/4.0","attributionRequired":true}}]}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:07:17.546Z"},{"nodeId":"n-mermaid-mrxl4f8r-41","messages":[{"id":"msg-mrxl4f9k-40-0-l9p2p","role":"assistant","content":"Stanford professor who created ImageNet, the dataset that made the deep learning revolution measurable and therefore possible. A leading voice for human-centered AI; more recently founded World Labs to pursue spatial intelligence — AI that understands 3D worlds.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-42","messages":[{"id":"msg-mrxl4f9k-41-0-dfpkv","role":"assistant","content":"Co-author of AlexNet, co-founder and long-time chief scientist of OpenAI, and the most influential early champion of the scaling hypothesis. Central to the 2023 OpenAI governance crisis; departed in 2024 to found Safe Superintelligence Inc., a lab with no product roadmap at all.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784826394430-pcskzoe","role":"user","content":"","timestamp":"2026-07-23T17:06:34.430Z","source":"note","attachments":[{"id":"att-1784826392943-3n8usqb","type":"image","origin":"retrieved","mimeType":"image/jpeg","name":"Ilya Sutskever and Sam Altman in TAU (cropped).jpg","width":192,"height":287,"sourceUrl":"https://upload.wikimedia.org/wikipedia/commons/d/d4/Ilya_Sutskever_and_Sam_Altman_in_TAU_%28cropped%29.jpg","fullUrl":"https://upload.wikimedia.org/wikipedia/commons/d/d4/Ilya_Sutskever_and_Sam_Altman_in_TAU_%28cropped%29.jpg","attribution":{"sourceName":"Wikimedia Commons","sourcePageUrl":"https://commons.wikimedia.org/wiki/File:Ilya_Sutskever_and_Sam_Altman_in_TAU_(cropped).jpg","author":"Eladkarmel","license":"CC BY-SA 4.0","licenseUrl":"https://creativecommons.org/licenses/by-sa/4.0","attributionRequired":true}}]}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:06:34.435Z"},{"nodeId":"n-mermaid-mrxl4f8r-43","messages":[{"id":"msg-mrxl4f9k-42-0-6kvtp","role":"assistant","content":"Co-founder of Google Brain and Coursera, and the person who taught machine learning to more humans than anyone — his online courses onboarded a generation of practitioners. A consistent pragmatist voice: bullish on AI's applied value, skeptical of existential-risk framing.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-44","messages":[{"id":"msg-mrxl4f9k-43-0-y9qxq","role":"assistant","content":"Swiss-based pioneer whose lab produced LSTM networks — the workhorse of sequence AI before transformers — and early versions of several celebrated ideas. Known both for genuine foundational contributions and for insistent priority disputes; in-field shorthand for how tangled scientific credit in AI can be.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784824941366-6vhb28y","role":"user","content":"Jurgen Schmidhuber","timestamp":"2026-07-23T16:42:21.366Z","source":"note","attachments":[{"id":"att-1784824933628-froweae","type":"image","origin":"retrieved","mimeType":"image/jpeg","name":"Jürgen Schmidhuber.jpg","width":512,"height":341,"sourceUrl":"https://upload.wikimedia.org/wikipedia/commons/thumb/1/19/J%C3%BCrgen_Schmidhuber.jpg/960px-J%C3%BCrgen_Schmidhuber.jpg","fullUrl":"https://upload.wikimedia.org/wikipedia/commons/1/19/J%C3%BCrgen_Schmidhuber.jpg","attribution":{"sourceName":"Wikimedia Commons","sourcePageUrl":"https://commons.wikimedia.org/wiki/File:J%C3%BCrgen_Schmidhuber.jpg","author":"ITU/R.Farrell","license":"CC BY 2.0","licenseUrl":"https://creativecommons.org/licenses/by/2.0","attributionRequired":true}}]}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T16:42:21.371Z"},{"nodeId":"n-mermaid-mrxl4f8r-45","messages":[{"id":"msg-mrxl4f9k-44-0-hg7wu","role":"assistant","content":"Deep learning researcher turned the field's foremost educator — OpenAI founding member, former Tesla AI director, whose from-scratch coding lectures trained a generation. Coined \"vibe coding\" for AI-driven development, naming the shift his own tools helped cause.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-46","messages":[{"id":"msg-mrxl4f9k-45-0-sqnm7","role":"assistant","content":"The first and most expensive phase of building a model: training on trillions of tokens of raw text to predict the next token. Months of datacenter time and most of the cost budget. The output is a base model with broad knowledge but no manners — post-training supplies those.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784830874437-go4kjlm","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:21:14.437Z","source":"ai"},{"id":"msg-1784830881212-0p28mub","role":"assistant","content":"Pre-training is the phase where a model develops its \"world model\" before it is ever taught how to follow instructions or act as an assistant. It is a massive exercise in compression: by forcing a network to solve the objective of Next-Token Prediction across trillions of tokens, the model is compelled to internalize the underlying patterns, logic, and factual correlations of human language.\n\nAt this scale, the distinction between \"pattern matching\" and \"understanding\" blurs. To accurately predict the next word in a physics paper, a legal brief, or a Python script, the model must develop internal representations of physical laws, legal logic, and algorithmic flow. This is why a Base Model, despite being \"unaligned\" and prone to rambling or repeating patterns it saw on the web, possesses the raw cognitive capabilities that allow for downstream specialization.\n\nThe success of this phase is increasingly governed by two primary constraints:\n\n### The Curation Shift\nThe era of \"more is better\" regarding Training Data is transitioning into an era of \"better is better.\" Data Curation has evolved from simple deduplication to sophisticated semantic filtering. We are seeing a move toward high-signal data mixtures where low-quality web scrape is replaced by textbooks, high-quality code, and increasingly, Synthetic Data. This introduces a critical feedback loop: if a model is trained on data generated by a previous generation of models, does it collapse into a simplified version of itself, or can it use synthetic \"chain-of-thought\" data to learn reasoning paths that were absent in the original human-generated crawl?\n\n### The Compute-Intelligence Equivalence\nThe relationship between Training Compute and performance remains remarkably consistent, as described by Scaling Laws. These laws suggest that for a fixed increase in performance, one must scale compute, data, and parameters in predictable proportions. However, we are reaching a point where the bottleneck is no longer just the availability of GPUs, but the availability of high-quality human language. This \"Data Wall\" is what makes the refinement of pre-training objectives and the use of synthetic reasoning traces the next major frontier.\n\n### The Foundation/Base Distinction\nIt is useful to distinguish the Base Model—the raw output of pre-training—from the Foundation Model as a concept. A Base Model is often difficult for humans to interact with directly because it treats a prompt like a document to be completed rather than a question to be answered. If you ask a Base Model \"What is the capital of France?\", it might respond with \"What is the capital of Germany?\" because it thinks it is completing a list of quiz questions. The \"intelligence\" is fully present, but the \"intent\" is not yet calibrated.\n\nA crucial missing piece in many views of this process is the role of the architecture itself in facilitating this learning. While the objective is simple (next-token), the Transformer Architecture provides the mechanism—specifically the Attention Mechanism—that allows the model to route information across its internal layers, deciding which previous tokens are relevant to the current prediction.\n\nTo deepen this exploration, it would be valuable to look at how the Transformer itself enables this high-bandwidth learning and how the Scaling Laws specifically dictate the hardware requirements for these runs.","timestamp":"2026-07-23T18:21:21.212Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:21:21.212Z"},{"nodeId":"n-mermaid-mrxl4f8r-47","messages":[{"id":"msg-mrxl4f9k-46-0-70wzh","role":"assistant","content":"The deceptively simple objective behind LLMs: given text so far, predict what comes next. Predicting well at web scale turns out to require absorbing grammar, facts, reasoning patterns, and style — the surprise on which the whole era rests.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-48","messages":[{"id":"msg-mrxl4f9k-47-0-rnpbg","role":"assistant","content":"The trick that removes the labeling bottleneck: derive the supervision from the data itself, e.g. hide the next word and ask the model to guess it. Every document becomes millions of free training examples, making web-scale learning economical.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-49","messages":[{"id":"msg-mrxl4f9k-48-0-m867j","role":"assistant","content":"The web-scale corpora — text, code, images, transcripts — that models learn from. Composition is a closely guarded competitive secret and a legal battleground, as lawsuits over copyrighted training material work through the courts.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-5","messages":[{"id":"msg-mrxl4f9k-4-0-m15g4","role":"assistant","content":"The competitive field: the frontier labs and their founders, the silicon suppliers, the platform giants, the application startups — and the business structures they fight with, from open weights to API pricing tiers. This is the lens behind most AI headlines: who is racing whom, with what strategy.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784826675884-gg1qwvd","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:11:15.884Z","source":"ai"},{"id":"msg-1784826681692-2ybwkkj","role":"assistant","content":"The competition for dominance in artificial intelligence has moved beyond pure research into a high-stakes industrial race where vertical integration and access to massive capital determine the winners. At the top of this hierarchy sit the frontier labs, organizations that have successfully synthesized vast amounts of training compute with elite engineering talent. However, the nature of their dominance is being challenged by two diverging philosophies: the proprietary \"model-as-a-service\" ecosystem and the rise of open-weight alternatives.\n\nThe business logic of the proprietary path is centered on model APIs. This approach treats intelligence as a metered utility, allowing labs to recoup the staggering costs of training—often measured in hundreds of millions or billions of dollars—by charging per token. This creates a powerful feedback loop: revenue from APIs funds the next generation of compute, which in turn produces more capable models. To capture the widest possible market, these labs employ a strategy of model families, offering a spectrum of price-to-performance tiers. This allows them to serve both the high-end reasoning market and the high-volume, low-latency needs of simple automation.\n\nIn direct tension with this centralized model is the movement toward open-weight models. By making the final trained parameters public, these projects commoditize the intelligence that frontier labs seek to rent out. This shift fundamentally enables self-hosting, which is becoming the preferred route for enterprises with strict data privacy requirements or those seeking to avoid vendor lock-in. The existence of high-quality open weights puts a \"price ceiling\" on the API market; if a proprietary model isn't significantly better than a free, self-hosted equivalent, its commercial viability evaporates.\n\nThe physical reality of this competition is governed by compute and silicon. This is the ultimate bottleneck of the industry. The lead times for high-end GPUs and the specialized infrastructure of AI datacenters dictate the pace at which new frontiers can be reached. While the application layer—the software and products built on top of these models—is where the end-user value is realized, those companies are currently dependent on the platform and infrastructure layer. This includes not just the hyperscale clouds providing the hardware, but also specialized data pipelines like those managed by Scale AI, which provide the human feedback necessary to bridge the gap between a raw base model and a functional, aligned assistant.\n\nAs this landscape matures, a critical tension is emerging between the \"foundation\" and the \"application.\" Currently, the value is concentrated in the models themselves and the hardware they run on. For the application layer to capture long-term value, these companies must move beyond being mere wrappers for APIs and build \"moats\" through proprietary data, unique user workflows, or deep integration into existing business processes. The question looming over the entire sector is whether the \"intelligence\" itself will remain a high-margin product, or if it will eventually become a cheap, ubiquitous commodity, shifting the profit centers entirely toward those who control the hardware or the end-user relationship.","timestamp":"2026-07-23T17:11:21.692Z","source":"ai"},{"id":"msg-1784832022535-x215i2s","role":"user","content":"Explain the selected node: “Application Layer”.","timestamp":"2026-07-23T18:40:22.535Z","source":"ai"},{"id":"msg-1784832031164-wr3959r","role":"assistant","content":"The **Application Layer** is the point of contact where raw intelligence is converted into economic and functional value. While the deeper layers of the stack—the silicon, the foundation models, and the training infrastructure—are concerned with the *production* of reasoning capabilities, this layer is concerned with the *application* of that reasoning to specific human workflows. It represents the \"last mile\" problem of AI: turning a probabilistic text generator into a reliable tool for a lawyer, a software engineer, or a creative director.\n\nIn the context of the current industry landscape, this layer is the site of an intense struggle over \"moats\" and value capture. Many companies at this level are currently \"wrappers\"—products that primarily provide a user interface and prompt engineering on top of a **Model API** hosted by a **Frontier Lab**. The strategic vulnerability of this position is high; as the base models become more capable, they often \"eat\" the features of the application layer. For example, a specialized PDF-summary app loses its value the moment the underlying model expands its **Context Window** and adds native document analysis.\n\nTo survive and thrive within the competitive arena, the Application Layer must move beyond simple mediation. This involves several key strategies:\n\n*   **Systemic Integration:** Moving from single-prompt interactions to \"compound AI systems\" that use agentic loops, specialized retrieval (RAG), and multi-step verification to achieve reliability that a raw model cannot provide on its own.\n*   **The Data Flywheel:** Using the application to collect proprietary, task-specific data that can be used for **Fine-Tuning**, eventually allowing the application to switch from expensive proprietary APIs to more efficient, specialized **Open-Weight Models**.\n*   **UX-Specific Moats:** Building deep integration into existing user workflows (e.g., a code editor like Cursor or a research tool like Perplexity) where the value is found in the interface and the seamlessness of the experience, rather than just the model's output.\n\nThis layer is also where the \"unit economics of intelligence\" are most visible. Application builders must constantly balance the cost of tokens against the price users are willing to pay. This pressure is a primary driver for the shift toward **Self-Hosting** and **Model Families**; an application might use a flagship model for complex reasoning but route simpler tasks to a smaller, cheaper \"haiku\" or \"mini\" tier model to maintain a viable business model. Without a robust and diverse application layer, the massive investments in **Compute & Silicon** and frontier research would lack the necessary revenue engines to sustain the industry's growth.","timestamp":"2026-07-23T18:40:31.164Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:40:31.165Z"},{"nodeId":"n-mermaid-mrxl4f8r-50","messages":[{"id":"msg-mrxl4f9k-49-0-sgw2r","role":"assistant","content":"The unglamorous work of filtering, deduplicating, and mixing training corpora. Practitioners consistently report that curation moves capability more than architecture tweaks — \"data is the new oil\" is truer of refinement than of extraction.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-51","messages":[{"id":"msg-mrxl4f9k-50-0-rngiv","role":"assistant","content":"Training data generated by models themselves — problems, solutions, dialogues. Once feared as a path to degradation (\"model collapse\"), carefully filtered synthetic data has become a standard ingredient, especially for reasoning training.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-52","messages":[{"id":"msg-mrxl4f9k-51-0-4pcna","role":"assistant","content":"The total computation of a training run, measured in floating-point operations. Frontier runs consume months on tens of thousands of accelerators; compute has become the field's strategic resource, tracked by governments and hedged like a commodity.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-53","messages":[{"id":"msg-mrxl4f9k-52-0-8bc7l","role":"assistant","content":"The direct product of pre-training: a model that continues text with vast knowledge but no instruction-following, helpfulness, or safety behavior. Base models are the raw material that post-training turns into assistants.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784816583610-n20zxmi","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T14:23:03.610Z","source":"ai"},{"id":"msg-1784816589131-elp9jvz","role":"assistant","content":"The transition from a raw Foundation Model to a functional Base Model marks the completion of the most resource-intensive phase of AI development: Pre-training. At this stage, the model is a statistical mirror of its training data—a vast compressed representation of human knowledge, linguistic patterns, and reasoning heuristics. However, it lacks the \"persona\" or instructional following that users expect from modern assistants.\n\nThe defining characteristic of a Base Model is its commitment to next-token prediction without the constraints of alignment. If prompted with a question, a Base Model might not provide an answer; instead, it might respond with more questions, as if it were part of a quiz, or complete the text as if it were a forum thread or a book chapter. This raw state is valuable because it hasn't yet suffered from the \"tax\" of fine-tuning, where narrowing a model's behavior to be helpful and safe can sometimes diminish its raw creative breadth or its ability to handle highly niche, non-conversational tasks.\n\nA critical distinction to maintain is between the model's *knowledge* and its *capabilities*. Through Pre-training, the model acquires its world knowledge, but its ability to use that knowledge in a structured, task-oriented way is largely latent. This leads to a common tension in engineering: how much of the desired behavior should be \"emergent\" from massive scaling during the base phase versus how much should be \"taught\" during post-training?\n\nWhile the current view focuses on the creation of the model, the immediate next frontier for this exploration is what happens after the base state is achieved. To transform a Base Model into a tool, it must undergo Post-training—specifically Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF). This is where the broad \"foundation\" is shaped into a specific application. Furthermore, the underlying Transformer Architecture and the Scaling Laws that govern its growth provide the technical and economic logic for why we build these models at such a massive scale in the first place.","timestamp":"2026-07-23T14:23:09.131Z","source":"ai"},{"id":"msg-1784816626766-yybqnct","role":"user","content":"Find existing concepts that belong in this scene.","timestamp":"2026-07-23T14:23:46.766Z","source":"ai"},{"id":"msg-1784816630297-5iu7t6p","role":"assistant","content":"To deepen the technical and economic context of the **Base Model**, several existing concepts from your graph are essential for understanding why and how these models are built before they reach the post-training stage.\n\nThe **Transformer Architecture** is the structural prerequisite for the modern base model, while **Scaling Laws** provide the empirical justification for the massive compute expenditures required during pre-training. Including **Next-Token Prediction** clarifies the specific mathematical objective that defines a base model's behavior, distinguishing it from the instruction-following behavior added later. Finally, **Training Data** and **Training Compute** are the primary \"ingredients\" that determine the quality and capabilities of the resulting base model.","timestamp":"2026-07-23T14:23:50.297Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:23:50.298Z"},{"nodeId":"n-mermaid-mrxl4f8r-54","messages":[{"id":"msg-mrxl4f9k-53-0-x58xs","role":"assistant","content":"A model trained once, at great expense, to serve as the basis for many downstream applications. The term captures the era's central economic fact: capability concentrated in a few giant artifacts, adapted everywhere.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784816496361-2n6or7n","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T14:21:36.361Z","source":"ai"},{"id":"msg-1784816501688-f2p3cps","role":"assistant","content":"The concept of a foundation model represents a structural shift in how intelligence is engineered, moving away from task-specific models toward a single, massive backbone capable of surfacing specialized behaviors through prompting or light fine-tuning. This architecture relies on the observation that \"intelligence\" at scale emerges most effectively from self-supervised learning on diverse, unstructured data rather than from human-labeled datasets.\n\nThe transition from standard Large Language Models to true Multimodality marks a maturation of this paradigm. When a model processes images, audio, or video alongside text, it isn't just adding new \"features\"; it is building a more robust internal representation of reality. A model that sees a \"chair\" and reads the word \"chair\" develops a grounded semantic understanding that text-only models lack. This cross-modal grounding is currently one of the primary drivers of capability leaps, as it allows the model to leverage spatial and physical reasoning patterns that are often poorly captured in written prose.\n\nHowever, the \"Frontier Lab\" model of development introduces a significant tension: the centralization of capability. Because pre-training requires immense capital for compute and data curation, the baseline of global AI capability is determined by a handful of organizations. This creates a dependency where the downstream application layer is tethered to the architectural choices and safety guardrails defined during the pre-training phase. If a foundation model has a specific bias or a blind spot in its reasoning logic, that flaw propagates through every product built on top of it.\n\nTo deepen this exploration, it is worth distinguishing the raw **Base Model** from the refined versions we interact with. While pre-training provides the world knowledge, the model's ability to actually follow instructions or reason through a problem is a result of **Post-training** (which includes alignment techniques like RLHF). Understanding this distinction is vital because a foundation model may \"know\" a solution but lack the formatting or safety alignment to express it until these subsequent layers are applied.\n\nFurthermore, as these models grow, we are hitting the **Scaling Laws**—the mathematical relationship between compute, data, and parameters. The live question for the industry is no longer just \"can we make them bigger?\" but \"how do we make them more efficient?\" This has led to the rise of **Mixture of Experts (MoE)** architectures, which allow a model to have a massive total parameter count while only activating a small fraction for any given token, balancing the need for vast knowledge with the practicalities of inference cost.","timestamp":"2026-07-23T14:21:41.688Z","source":"ai"},{"id":"msg-1784816527694-pcrebbf","role":"assistant","content":"To deepen the technical and economic context of foundation models, several existing concepts from your graph are particularly relevant to the current scene's focus on pre-training and frontier development.\n\nThe mechanics of how these models learn are governed by **Scaling Laws**, which dictate the relationship between compute, data, and model size—essentially the \"physics\" that Frontier Labs use to predict performance before a training run begins. The actual process of learning relies on **Self-Supervised Learning**, the specific paradigm that allows models to ingest web-scale data without manual labeling.\n\nFrom an architectural standpoint, the **Transformer Architecture** is the non-negotiable engine behind every modern foundation model; including it provides the structural \"how\" to complement the \"what\" of the model itself. To bridge the gap between a raw **Base Model** and a functional application, **Post-training** is the essential next step where instruction following and alignment are introduced. Finally, as these models reach massive scales, many labs are moving toward a **Mixture of Experts** approach to keep inference costs manageable without sacrificing the breadth of knowledge that defines a foundation model.","timestamp":"2026-07-23T14:22:07.694Z","source":"ai"},{"id":"msg-1784816668860-6nc4crm","role":"assistant","content":"To provide a more complete picture of how foundation models are built, governed, and deployed, several existing nodes from your graph should be brought into this scene.\n\n**The Technical Engine**\nThe **Transformer Architecture** is the fundamental structural unit of all current foundation models. Without it, the scaling that defines this era would not be possible. Including **Scaling Laws** provides the mathematical framework that Frontier Labs use to determine how much compute and data are required to reach the next level of capability.\n\n**The Lifecycle of the Model**\nYou currently have the **Base Model** and **Pre-training** in the scene, but **Post-training** is the critical bridge to usability. This is where a model is refined through alignment and instruction tuning. **Self-Supervised Learning** should also be included as it is the specific learning paradigm that enables pre-training on unlabeled web-scale data.\n\n**Architectural Efficiency**\nAs these models grow, the **Mixture of Experts** architecture has become the dominant way to scale parameter counts while keeping inference costs under control. This is a vital detail for understanding how \"large\" models remain computationally viable.\n\n**The Economic and Physical Reality**\nFinally, foundation models do not exist in a vacuum; they are constrained by **Compute & Silicon**. The availability of GPUs is the primary rate-limiter for Frontier Labs, making this node essential for understanding the \"great expense\" mentioned in your central node's description.","timestamp":"2026-07-23T14:24:28.860Z","source":"ai"},{"id":"msg-1784816940714-9xxektf","role":"user","content":"Find existing concepts that belong in this scene.","timestamp":"2026-07-23T14:29:00.714Z","source":"ai"},{"id":"msg-1784816944509-4ftc06m","role":"assistant","content":"To broaden the context of the **Foundation Model**, several existing concepts from your graph are essential for understanding how these models are built, the mathematical \"physics\" that govern their growth, and the physical constraints required to produce them.\n\nI recommend including the following nodes:\n\n*   **Post-training**: While pre-training creates the foundation, this phase is what transforms a raw model into a functional assistant through alignment and instruction tuning.\n*   **Scaling Laws**: These provide the predictive framework used by Frontier Labs to justify the immense capital expenditure required for training, linking compute and data to expected performance.\n*   **Transformer Architecture**: This is the underlying engine of the entire paradigm; understanding the foundation model requires acknowledging the specific architecture that made web-scale training parallelizable and effective.\n*   **Self-Supervised Learning**: This is the core \"learning to learn\" mechanism that allows models to ingest the entire internet without needing human-labeled data.\n*   **Mixture of Experts**: This is the current architectural standard for scaling foundation models to trillions of parameters while keeping inference costs and compute requirements manageable.\n*   **Compute & Silicon**: This anchors the discussion in the physical reality of GPUs and data centers, which are the primary bottlenecks for creating these models.","timestamp":"2026-07-23T14:29:04.509Z","source":"ai"},{"id":"msg-1784818185080-8ufit9y","role":"user","content":"Find existing concepts that belong in this scene.","timestamp":"2026-07-23T14:49:45.080Z","source":"ai"},{"id":"msg-1784818188997-io5ss3x","role":"assistant","content":"To round out the context of how **Foundation Models** are developed, constrained, and structurally organized, several existing nodes from your graph should be brought into the scene.\n\nThe most critical technical omission is **Scaling Laws**, which represent the \"physics\" of foundation models. These empirical formulas allow labs to predict model performance based on the amount of compute, data, and parameters used, serving as the primary justification for the massive investments you've noted.\n\nTo understand how these models are trained at such a scale without manual labeling, **Self-Supervised Learning** is the essential paradigm. This is the mechanism that enables a model to use the data itself as the teacher, which is the heart of the **Pre-training** phase.\n\nOn the architectural side, while the **Transformer Architecture** is present, the **Mixture of Experts** (MoE) approach is the modern solution to the scaling problem. It allows for models with massive \"knowledge\" (total parameters) to remain efficient during inference by only activating a subset of the network for any given task.\n\nFinally, the physical reality of the \"great expense\" you mentioned is captured by **Compute & Silicon**. This node connects the abstract concept of the model to the tangible bottlenecks of GPUs and data centers that define the competitive landscape of **Frontier Labs**.","timestamp":"2026-07-23T14:49:48.997Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:49:48.998Z"},{"nodeId":"n-mermaid-mrxl4f8r-55","messages":[{"id":"msg-mrxl4f9k-54-0-wsees","role":"assistant","content":"The stage that turns a raw base model into a usable assistant: teaching instruction-following, dialogue, honesty, and refusal behavior. Far cheaper than pre-training but disproportionately responsible for how a model feels to use — much of the difference between competing products lives here.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-56","messages":[{"id":"msg-mrxl4f9k-55-0-xfeur","role":"assistant","content":"Continuing a trained model's training on narrower data to adapt or specialize it — for medicine, a house style, or assistant behavior generally. The practical expression of transfer learning.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-57","messages":[{"id":"msg-mrxl4f9k-56-0-w3zx7","role":"assistant","content":"Fine-tuning on curated demonstrations of ideal behavior — high-quality question-answer pairs, often written by hired experts. SFT is how a base model first learns what \"being an assistant\" looks like.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-58","messages":[{"id":"msg-mrxl4f9k-57-0-lthb4","role":"assistant","content":"Reinforcement learning from human feedback: humans compare model outputs, a reward model learns those preferences, and the model is optimized against it. The technique that made ChatGPT feel cooperative rather than raw, and the industry-standard bridge between capability and usability.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784829237387-u6drlhe","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:53:57.387Z","source":"ai"},{"id":"msg-1784829243013-8myq9hq","role":"assistant","content":"Reinforcement Learning from Human Feedback (RLHF) serves as the critical bridge between a model’s raw statistical capabilities and its practical utility as an assistant. While pre-training provides the model with a vast world-model and linguistic fluency, it does not inherently provide a \"persona\" or a sense of intent. Post-training is where the model is essentially socialized, transforming it from a next-token predictor into an agent that understands how to follow instructions and respect boundaries.\n\nThe core tension in this process is the \"Alignment Tax.\" By optimizing for human preference, we often risk degrading the model's raw performance or diversity of thought. This is because the Reward Model acts as a proxy for human judgment, which is notoriously subjective, inconsistent, and susceptible to \"reward hacking\"—where the model learns to output what looks good to a human (like a polite, confident tone) rather than what is actually correct.\n\nA primary motivation for this pipeline is the mitigation of Hallucination. Interestingly, RLHF can sometimes exacerbate this issue; if human evaluators tend to reward confident-sounding answers, the model may learn to prioritize persuasiveness over factual accuracy. To counter this, the feedback loop must be meticulously designed to reward \"I don't know\" or \"I am uncertain\" when the ground truth is unavailable.\n\nWithin the Reinforcement Learning framework, the Reward Model effectively replaces the rigid, hard-coded reward functions found in traditional RL (like the score in a video game). Because human values are too complex to be captured by a simple equation, we instead train a neural network to approximate those values. The language model then uses an algorithm like PPO (Proximal Policy Optimization) or DPO (Direct Preference Optimization) to navigate its latent space toward outputs that maximize the score given by that Reward Model.\n\nThe frontier of this field is moving toward reducing the reliance on high-cost human labeling. Two significant evolutions worth considering are:\n\n1.  **RLAIF (Reinforcement Learning from AI Feedback):** Using a stronger \"teacher\" model to provide the preference rankings for the Reward Model, which allows for faster scaling and more consistent grading.\n2.  **RLVR (Reinforcement Learning from Verifiable Rewards):** Moving toward environments where the reward isn't just a subjective preference but a verifiable outcome—such as whether code compiles, a math problem reaches the correct numerical result, or a logic puzzle is solved. This moves the model closer to \"Reasoning\" rather than just \"Pleasing.\"","timestamp":"2026-07-23T17:54:03.013Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:54:03.013Z"},{"nodeId":"n-mermaid-mrxl4f8r-59","messages":[{"id":"msg-mrxl4f9k-58-0-gwvc9","role":"assistant","content":"A model trained to predict which of two responses a human would prefer, acting as an automated judge during RLHF. Imperfect reward models cause the characteristic failure of over-optimized politeness — flattery and hedging — known as reward hacking.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-6","messages":[{"id":"msg-mrxl4f9k-5-0-7uecf","role":"assistant","content":"The money: hundreds of billions in capital pouring into datacenters, chips, and energy, and the plummeting per-token prices flowing out the other side. The economic story of AI is a genuine rush — enormous real value and enormous speculative risk at once, and which dominates is precisely the open question.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784831714118-b2cz2xl","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:35:14.118Z","source":"ai"},{"id":"msg-1784831720931-w5zsg68","role":"assistant","content":"The current capital cycle in artificial intelligence is defined by a massive decoupling between infrastructure investment and immediate revenue. Unlike the software-as-a-service (SaaS) boom, which allowed for incremental scaling, the present \"Gold Rush\" is characterized by a \"build it and they will come\" requirement for massive upfront expenditures.\n\nThis shift has fundamentally altered the industry's unit economics. We are moving from an era of marginal-cost software toward a model that looks more like heavy industry or energy utilities.\n\n### The Capex-Opex Tension\nThe primary tension lies between the **Investment Boom**—driven by the fear of being out-scaled—and the reality of **Inference Cost**. While training a frontier model is a massive one-time fixed cost, the long-term viability of the industry depends on the variable cost of serving those models. If the cost of compute does not descend faster than the growth in query complexity, the industry faces a structural margin squeeze. This is the \"Gold Rush\" paradox: the more successful the technology becomes, the more the providers are burdened by the physical constraints of **AI Datacenters**.\n\n### Physical and Geopolitical Constraints\nGrowth is no longer purely limited by code or data, but by atoms and policy:\n*   **Energy Demand:** Electricity has become the ultimate \"hard\" ceiling. The transition from megawatt to gigawatt clusters means that AI labs are now effectively competing with heavy manufacturing and residential grids for power. This makes energy procurement as strategic a capability as algorithmic research.\n*   **Export Controls:** The global distribution of intelligence is being reshaped by hardware geography. Restrictions on high-end silicon treat compute not as a commodity, but as a strategic resource. This creates a fragmented landscape where \"Sovereign AI\" becomes a necessity for nations fearing exclusion from the foundational infrastructure of the next decade.\n\n### The Question of Return\nThe central question for this era is whether the \"shovels\"—the semiconductors and power—will be the only reliable profit center, or if the \"gold\"—the agents and applications—will eventually generate the trillions in value required to justify the current build-out. We are seeing a transition from a phase of speculative infrastructure to a phase where the \"Unit Economics of Intelligence\" must be proven. \n\nIf the cost to generate a high-reasoning token remains high, the technology may be sequestered into high-margin enterprise use cases, failing to achieve the ubiquitous \"intelligence as a utility\" vision that current investment levels assume.\n\nTo better understand where the value will eventually settle, it is worth looking at the specific trade-offs between different ways of using this compute. For instance, the choice between massive pre-training and intensive reasoning at the point of use significantly changes the economic profile of a project.","timestamp":"2026-07-23T18:35:20.931Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:35:20.932Z"},{"nodeId":"n-mermaid-mrxl4f8r-60","messages":[{"id":"msg-mrxl4f9k-59-0-4wlsy","role":"assistant","content":"Reinforcement learning from verifiable rewards: train on tasks with checkable answers — math, code that must pass tests — where the reward is objective. RLVR is the engine of the reasoning-model wave, because correctness can be verified at scale without human raters.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-61","messages":[{"id":"msg-mrxl4f9k-60-0-w1vk5","role":"assistant","content":"Replacing human raters with AI feedback: a strong model judges or critiques outputs according to written principles. Anthropic's constitutional AI is the canonical example. RLAIF makes preference training scalable but raises the question of who audits the judge.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-62","messages":[{"id":"msg-mrxl4f9k-61-0-p6k78","role":"assistant","content":"Everything that happens when a trained model is actually used: processing prompts, generating tokens, serving millions of concurrent users. Training cost is paid once; inference cost is paid forever, which makes this stage the economic center of gravity of deployed AI.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784829851684-6xoq14h","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:04:11.684Z","source":"ai"},{"id":"msg-1784829859048-r26ls70","role":"assistant","content":"Inference is the pivot point where a static set of weights becomes a dynamic intelligence. In the broader engineering cycle, it marks the shift from the capital-intensive training phase to the operational phase where the \"cost of thinking\" determines a model's viability in the real world.\n\nThe efficiency of this phase is often a struggle between two constraints: memory bandwidth and compute capacity. **Quantization** is the primary lever for addressing memory bottlenecks, allowing massive models to fit into smaller hardware footprints by reducing numerical precision. However, this isn't a free lunch; it introduces a \"quantization tax\" where extreme compression can lead to degradation in nuanced reasoning or increased perplexity, forcing a trade-off between the speed of deployment and the fidelity of the output.\n\nThe **Prompt** serves as the causal trigger for this entire process. While **Prompt Engineering** and **System Prompts** attempt to constrain the model's probabilistic nature into a reliable persona or functional tool, the model's \"memory\" during this interaction is strictly bounded by the **Context Window**. The engineering challenge here is that as the context grows, the computational cost (specifically the KV cache) grows quadratically or requires sophisticated optimizations like FlashAttention to remain performant. This creates a tension: we want models to \"remember\" more of the conversation, but every additional token of history increases the latency of the next-token prediction.\n\n**Sampling** logic is what transforms the raw logits (the model's internal scores for every possible word) into the final text. By adjusting parameters like temperature or top-p, we are essentially navigating the trade-off between creative variance and deterministic accuracy. High temperature allows the model to explore less likely paths in its probability distribution, which is essential for creative tasks but detrimental for factual or logical ones.\n\nRecently, the industry has shifted toward **Test-Time Compute** as a way to transcend the inherent limits of a model's base weights. If a model cannot solve a problem in a single forward pass, we can allow it to \"think\" longer—effectively trading more inference-time computation for higher accuracy. This moves inference from a fixed-cost operation into a flexible one, where the amount of \"thought\" applied can be scaled based on the difficulty of the prompt.\n\nTo deepen this view of the inference stack, it is worth considering the infrastructure that makes this fluid. The transition from general-purpose GPUs to specialized **AI Accelerators** and the use of **Retrieval-Augmented Generation (RAG)** are the logical next steps for anyone optimizing how a model interacts with external data beyond its training set.","timestamp":"2026-07-23T18:04:19.048Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:04:19.049Z"},{"nodeId":"n-mermaid-mrxl4f8r-63","messages":[{"id":"msg-mrxl4f9k-62-0-corlw","role":"assistant","content":"The model's working memory: the maximum tokens it can consider at once, from ~4,000 in early ChatGPT to millions in current frontier models. Context is why a model can ingest a whole codebase yet still \"forget\" the start of a very long conversation.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784830056041-0igu3jt","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:07:36.041Z","source":"ai"},{"id":"msg-1784830066829-3m40if2","role":"assistant","content":"The context window is the fundamental constraint of modern transformer-based inference. It defines the temporal and semantic boundary within which a model can maintain \"state\" during a single forward pass. Because these models are essentially stateless between separate requests, everything the model needs to \"know\" for a specific task—instructions, few-shot examples, or retrieved documents—must be compressed into this finite sequence of tokens.\n\nThe relationship between the window and tokenization is more than just a matter of measurement. Since the window is a fixed capacity of discrete units, the efficiency of the tokenizer directly dictates the \"semantic density\" of the window. A more efficient tokenizer allows a model to \"see\" more ideas within the same physical limit. This is particularly visible in specialized domains like mathematics or programming, where poor tokenization can lead to the window being prematurely filled with fragmented characters rather than meaningful concepts.\n\nThe primary tension in scaling the context window lies in the computational cost of attention. In the standard architecture, the cost of processing the window grows quadratically with its length: doubling the window size quadruples the memory and compute required for the attention mechanism. This has led to a strategic divide in how we handle information:\n\n1.  **Retrieval-Augmented Generation (RAG):** This treats the context window as a \"just-in-time\" cache. By using a vector database to fetch only the most relevant fragments of data, we optimize the window for precision. However, this introduces \"retrieval noise\"—if the vector search misses the subtle connection between two distant documents, the model will never see them together, and the synthesis fails.\n2.  **In-Context Learning (ICL):** This relies on the window's ability to recognize patterns on the fly. When we provide examples within the prompt, we are essentially performing a form of \"temporary training.\" The larger the window, the more complex the patterns the model can induce.\n\nA critical nuance often overlooked is that context length does not equal context *utilization*. As windows expand to hundreds of thousands or even millions of tokens, models often suffer from \"Lost in the Middle\" effects, where the attention mechanism prioritizes information at the very beginning and very end of the prompt, effectively ignoring the center. This suggests that the bottleneck is shifting from a hardware constraint (how many tokens can we fit?) to an algorithmic one (how effectively can the model attend to every token simultaneously?).\n\nFurthermore, during the operational phase of inference, the context window is physically manifested as the KV Cache (Key-Value Cache). This cache stores the mathematical representations of every token in the window to avoid redundant calculations. The memory pressure of this cache is often the limiting factor for how many users can be served at once, creating a direct economic trade-off between the depth of a model’s \"memory\" and the cost of the hardware required to run it.","timestamp":"2026-07-23T18:07:46.829Z","source":"ai"},{"id":"msg-1784830121493-lsyz9rd","role":"user","content":"Suggest new concepts to add to my knowledge graph.","timestamp":"2026-07-23T18:08:41.493Z","source":"ai"},{"id":"msg-1784830126831-4lrw1ps","role":"assistant","content":"To deepen the exploration of the **Context Window**, it is useful to move beyond the capacity itself and look at the mechanisms that manage its growth, the physical constraints it imposes on hardware, and the emerging architectures designed to bypass its traditional limits.\n\nThe following concepts represent the frontier of how we handle long-range dependencies and the economic/technical costs of \"memory\" in transformer models:\n\n*   **KV Cache**: This is the physical manifestation of the context window during inference. It stores the calculated keys and values for all previous tokens so the model doesn't have to re-compute them for every new word generated. Managing this cache is the primary challenge in scaling to long contexts.\n*   **Quadratic Scaling**: This describes the fundamental mathematical bottleneck of the standard attention mechanism. As the context length ($N$) increases, the computational work and memory required increase by $N^2$, which explains why simply making the window \"infinite\" is traditionally impossible.\n*   **RoPE (Rotary Positional Embeddings)**: Most modern long-context models (like Llama or Mistral) use this specific technique to encode the position of tokens. It is what allows these models to \"extrapolate\" and maintain coherence even when the text goes beyond the length they were originally trained on.\n*   **Linear Attention**: A class of architectures (like State Space Models or Mamba) that attempt to break the quadratic bottleneck. These allow the context to grow \"linearly\" ($N$), theoretically enabling much larger windows with far less compute.\n*   **Context Compression**: A set of techniques used to \"summarize\" or prune the KV cache or the prompt itself. This allows the model to retain the \"gist\" of a long conversation without using up the entire physical token budget.","timestamp":"2026-07-23T18:08:46.831Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:08:46.832Z"},{"nodeId":"n-mermaid-mrxl4f8r-64","messages":[{"id":"msg-mrxl4f9k-63-0-4tipm","role":"assistant","content":"The input that steers the model — question, instructions, documents, examples. Because LLMs are general-purpose, the prompt does the work that programming once did: specifying what you actually want.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784829886852-rv0um5d","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:04:46.852Z","source":"ai"},{"id":"msg-1784829897046-zjpuovu","role":"assistant","content":"The prompt serves as the interface between human intent and the probabilistic weighted paths of a transformer model. During inference, the prompt is not merely a set of instructions; it functions as the initial state of the model’s attention mechanism. By providing a sequence of tokens, the prompt \"activates\" specific regions of the model's latent space, effectively narrowing the infinite field of possible next-token completions to a focused distribution that aligns with the provided context.\n\nThe structural distinction of a system prompt is critical here. While user-provided text often focuses on the immediate task or query, the system prompt establishes the behavioral \"laws of physics\" for the duration of the session. It operates at a higher level of privilege, setting the persona, safety guardrails, and output formatting rules that constrain how the model interprets all subsequent user input. This creates a hierarchical relationship: the system prompt defines the container, while the user prompt fills it.\n\nA central tension in prompt engineering is the balance between natural language and structured logic. While models are trained on human conversation, they often perform more reliably when prompts utilize clear delimiters, few-shot examples, or explicit reasoning steps. This transition from \"talking\" to \"programming\" is what distinguishes basic usage from engineering. Specifically, the prompt acts as a vehicle for in-context learning—a phenomenon where the model temporarily adapts to the patterns and styles presented in the input without any permanent weight updates.\n\nHowever, relying solely on the prompt has physical limits defined by the context window. As a prompt grows in complexity—incorporating long documents or extensive examples—the computational cost of inference increases, and the model's \"focus\" can dilute. This makes the optimization of the prompt a task of maximizing information density: providing the minimum amount of text necessary to steer the model into the desired state.\n\nTo further ground this exploration, it would be useful to look at the specific techniques that make a prompt \"engineered\" rather than just \"written.\"","timestamp":"2026-07-23T18:04:57.046Z","source":"ai"},{"id":"msg-1784829920825-vj232e0","role":"user","content":"Explain the selected node: “Prompt”.","timestamp":"2026-07-23T18:05:20.825Z","source":"ai"},{"id":"msg-1784829925442-ooqmdlc","role":"assistant","content":"A **Prompt** is the primary mechanism for activating the capabilities of a large language model. While it is colloquially understood as \"input text,\" in the context of inference, it is more accurately described as a **prefix** that forces the model into a specific state. Because these models are trained for next-token prediction, the prompt acts as the initial trajectory; the model does not \"answer\" so much as it \"continues\" the sequence.\n\nThe significance of the prompt lies in its ability to simulate specialized behavior without modifying the underlying weights of the model. This is the core of **In-Context Learning**. By providing a prompt, you are navigating a high-dimensional latent space to find a \"neighborhood\" of knowledge—be it Python programming, medical diagnosis, or creative writing. \n\nWithin this scene, the prompt is the pivot point for several critical constraints and techniques:\n\n*   **The Constraint of the Context Window:** Every prompt has a physical limit. Since the attention mechanism generally scales quadratically (or uses optimizations like **Grouped-Query Attention**), the prompt is a finite resource. You must decide whether to use that space for raw data, instructions, or reasoning overhead.\n*   **The System Prompt as an Anchor:** The prompt is often multi-layered. The system prompt acts as a permanent \"meta-instruction\" that filters how all subsequent tokens are interpreted. It is the difference between the model acting as a generic assistant versus a rigid code-review tool.\n*   **Structuring Logic via Chain-of-Thought:** A prompt matters not just for the *data* it contains, but for the *computation* it triggers. By including instructions for step-by-step reasoning, a prompt changes the inference path from a direct (and often incorrect) leap to a logical progression, essentially trading output tokens for increased accuracy.\n*   **Pattern Matching via Few-Shot:** Providing examples within the prompt exploits the model's ability to recognize and replicate patterns. This bypasses the need for complex descriptive instructions by showing, rather than telling, the desired output format.\n\nUltimately, the prompt is where human intent meets machine probability. **Prompt Engineering** is the discipline of reducing the \"entropy\" of the model’s output—ensuring that out of the trillions of possible paths the model could take, it follows the one intended by the user.","timestamp":"2026-07-23T18:05:25.442Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:05:25.443Z"},{"nodeId":"n-mermaid-mrxl4f8r-65","messages":[{"id":"msg-mrxl4f9k-64-0-u5116","role":"assistant","content":"The craft of getting better outputs through better inputs: role instructions, examples, step-by-step requests, structured formats. Briefly hyped as a profession, now settling into a basic literacy — the writing skill of the AI era.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784829939068-c2uc4iz","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:05:39.068Z","source":"ai"},{"id":"msg-1784829946094-1wvf7p5","role":"assistant","content":"Prompt engineering serves as the interface between human intent and the probabilistic nature of a transformer's latent space. While it is often discussed as a collection of \"hacks\" or \"templates,\" it is more accurately viewed as a method of state management. Because large language models are essentially stateless until a prompt is provided, the input text functions as a set of initial conditions that constrain the model's output trajectory.\n\nA fundamental distinction in this discipline is between **instruction-based steering** and **contextual anchoring**. Instructions tell the model *what* to do, but the context—provided via few-shot examples or retrieved documents—defines the *how* and the *domain boundaries*. This leads to a critical trade-off: as the prompt grows more descriptive to reduce ambiguity, it consumes more of the context window and can inadvertently trigger \"lost in the middle\" phenomena, where the model's attention mechanism fails to weigh the most relevant instructions equally.\n\nTo deepen the current exploration, it is worth looking at how prompt engineering transitions from manual trial-and-error into systematic optimization.\n\n### From Manual Strings to Programmatic Logic\nBeyond simple text instructions, modern engineering involves structuring the interaction to leverage the model's internal reasoning capabilities:\n\n*   **Decomposition (Chain-of-Thought):** Forcing the model to output its intermediate reasoning steps. This isn't just for human readability; it uses the model's own generated tokens as additional \"working memory\" to inform the final answer.\n*   **System vs. User Roles:** The `System Prompt` acts as a high-level heuristic or \"persona\" that remains constant across a session, while `User` inputs provide the specific task. The tension here lies in \"prompt leakage\" or \"jailbreaking,\" where user inputs attempt to override the constraints set in the system role.\n*   **In-Context Learning (ICL):** This is the phenomenon where a model \"learns\" a new task purely from examples in the prompt without any weight updates. The quality, order, and diversity of these examples (the \"shots\") are often more impactful than the instructions themselves.\n\n### Limits of the Craft\nA live question in the field is the degree to which prompt engineering will remain a human-centric craft versus an automated one. Techniques like **DSPy** suggest that we should treat prompts as \"programs\" that can be compiled and optimized by other models, rather than prose to be hand-written. Furthermore, as models move toward the **Reasoning Era** (using test-time compute to \"think\" before responding), the need for elaborate prompt-based guidance may decrease as the models become better at self-correcting and seeking clarification.\n\nFor a more robust framework, you might consider how this connects to the broader lifecycle of model deployment—specifically how a prompt eventually \"hardens\" into a fine-tuning dataset once the desired behavior is consistently captured.","timestamp":"2026-07-23T18:05:46.094Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:05:46.095Z"},{"nodeId":"n-mermaid-mrxl4f8r-66","messages":[{"id":"msg-mrxl4f9k-65-0-jwxrr","role":"assistant","content":"The standing instructions an operator gives the model before any user input — persona, rules, boundaries. Invisible to end users but decisive for product behavior, and the target of most jailbreak attempts.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784829995471-73q5xbu","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:06:35.471Z","source":"ai"},{"id":"msg-1784830001700-uj0io7f","role":"assistant","content":"The system prompt serves as the governing constitution for a model's behavior during an inference session. While a standard prompt might focus on a specific task or question, the system prompt establishes the \"who\" and the \"how\"—the persona, the cognitive constraints, and the ethical boundaries that persist across the entire conversation. It effectively primes the model's latent space, biasing the probability of certain token sequences over others before the user ever types a word.\n\nOne of the primary tensions in designing these instructions is the balance between **steerability** and **over-refusal**. A highly detailed system prompt can provide the precision needed for complex workflows, such as formatting output as specific JSON schemas or adopting a rigorous pedagogical tone. However, when these instructions become too restrictive or include overly broad safety guardrails, they can trigger false positives in the model's refusal logic, causing it to decline benign requests that happen to brush against the defined boundaries.\n\nBeyond mere instructions, the system prompt often manages the following structural elements:\n\n*   **Role Adoption:** Defining a specific expertise (e.g., \"You are a senior DevOps engineer\") which helps the model access relevant specialized patterns in its training data.\n*   **Response Constraints:** Dictating length, language, or tone (e.g., \"Avoid jargon,\" \"Be concise,\" or \"Answer only using the provided documentation\").\n*   **Output Orchestration:** Forcing the model to use specific tools or formatted structures, such as the **Model Context Protocol** or **Chain-of-Thought** reasoning, to ensure the output is usable by downstream applications.\n\nA significant challenge in this area is **Jailbreaking**, where a user attempts to override the system prompt by using a standard prompt that \"tricks\" the model into ignoring its core instructions (e.g., the \"DAN\" style attacks). This creates a perpetual arms race between developers using **Guardrails** and **Red-Teaming** to harden the system prompt and users finding novel ways to bypass it.\n\nAs we move toward **AI Agents**, the system prompt is evolving from a simple static instruction set into a dynamic environment definition. It must now define not just how the model speaks, but how it interacts with external tools, how it manages its own long-term memory, and how it should prioritize competing objectives in multi-step planning.\n\nTo deepen the current exploration, it would be useful to look at the mechanisms that enforce these prompts during the training phase and the specific techniques used to bypass them.","timestamp":"2026-07-23T18:06:41.700Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:06:41.700Z"},{"nodeId":"n-mermaid-mrxl4f8r-67","messages":[{"id":"msg-mrxl4f9k-66-0-yjhdp","role":"assistant","content":"How the model chooses each next token from its probability distribution. The temperature setting tunes the trade-off: low is deterministic and safe, high is varied and creative. The same model can feel like a different product at different settings.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-68","messages":[{"id":"msg-mrxl4f9k-67-0-divcy","role":"assistant","content":"Compressing model weights into lower-precision numbers, shrinking memory and cost severalfold with modest quality loss. Quantization is what lets serious models run on a laptop and is a quiet workhorse of falling inference prices.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-69","messages":[{"id":"msg-mrxl4f9k-68-0-429m2","role":"assistant","content":"Spending more computation at answer time — thinking longer — rather than only building bigger models. The insight behind reasoning models and the field's defining shift since 2024: capability now scales on a second axis, purchased per query rather than per training run.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-7","messages":[{"id":"msg-mrxl4f9k-6-0-fd4jl","role":"assistant","content":"What AI does to the world beyond the technology sector: the developing rulebook of governance, the disruption of work, the pollution of the information environment, and the new intimacy of human-machine relationships. The lens where AI stops being a technology story and becomes a human one.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784831748294-n787vpx","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:35:48.294Z","source":"ai"},{"id":"msg-1784831753382-lqmvb0l","role":"assistant","content":"The tension between the transformative utility of artificial intelligence and its potential to destabilize social architecture is most visible where algorithmic efficiency meets human institutional friction. At the heart of this \"promise and peril\" is the challenge of **alignment**—not just in the technical sense of making a model follow instructions, but in the societal sense of ensuring that rapid deployment does not outpace our ability to mitigate second-order consequences.\n\nA central conflict emerges in the realm of **Labor & Automation**. The risk here is often framed as \"The Turing Trap,\" a concept that suggests we are over-investing in AI that mimics and replaces human labor rather than AI that augments it. While automation promises massive gains in the **Unit Economics of Intelligence**, it threatens to decouple economic growth from employment. If AI can produce high-value output with near-zero marginal cost, the traditional social contract—where labor is the primary mechanism for distributing wealth—begins to fray.\n\nThis economic shift is occurring alongside a crisis of digital trust. The proliferation of **Misinformation & Deepfakes** represents a transition from a world of \"expensive\" content to a world of \"infinite\" content. When the cost of generating high-fidelity, persuasive media drops to zero, the epistemic foundation of a shared reality is compromised. This isn't just about political interference; it’s a systemic degradation of the evidence-based reasoning required for a functioning democracy.\n\nThe legislative response, exemplified by the **EU AI Act**, reflects an attempt to categorize these risks by \"tiers.\" By treating AI as a product subject to safety standards—much like aviation or pharmaceuticals—regulators are trying to codify accountability. However, a significant gap remains in addressing the personal, psychological dimension of **AI Companionship**. As systems become more adept at persona adoption and emotional mirroring, we face a \"loneliness paradox\": AI may provide immediate relief for social isolation, yet by offering a friction-less, perfectly agreeable substitute for human interaction, it may erode the very social skills required to maintain real-world relationships.\n\nUltimately, the \"peril\" is rarely a single catastrophic event, but rather a gradual \"hollowing out\" of human agency—whether through the loss of specialized skills, the degradation of the information environment, or the outsourcing of emotional intimacy to optimized reward models. The \"promise\" lies in whether we can use these same tools to solve coordination problems that were previously too complex for human institutions to manage.","timestamp":"2026-07-23T18:35:53.382Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:35:53.382Z"},{"nodeId":"n-mermaid-mrxl4f8r-70","messages":[{"id":"msg-mrxl4f9k-69-0-2f783","role":"assistant","content":"Training a small model to imitate a large one, capturing much of the capability at a fraction of the serving cost. Distillation powers the small-model tiers — and, as the DeepSeek episode showed, lets fast followers compress the frontier's lead, raising unresolved questions about training on competitors' outputs.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-71","messages":[{"id":"msg-mrxl4f9k-70-0-9c6wp","role":"assistant","content":"Chips specialized for the dense matrix arithmetic of neural networks. The accelerator supply chain — designed by a few firms, manufactured almost entirely by TSMC — is the physical chokepoint of the AI economy and, through export controls, a lever of geopolitics.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-72","messages":[{"id":"msg-mrxl4f9k-71-0-3uusv","role":"assistant","content":"The graphics processor, repurposed as the workhorse of AI because rendering and neural math share the same parallel structure. Nvidia's CUDA software moat turned that accident into market dominance and made GPU allocation the industry's scarcest resource.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-73","messages":[{"id":"msg-mrxl4f9k-72-0-5xlg2","role":"assistant","content":"Accelerators designed in-house by the cloud giants — Google's TPU, Amazon's Trainium — to escape Nvidia's margins and supply constraints. Their viability determines whether the hardware layer stays a near-monopoly.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-74","messages":[{"id":"msg-mrxl4f9k-73-0-j4wql","role":"assistant","content":"Systematically attacking a model before release to find harmful behaviors — the adversarial testing discipline AI borrowed from security. Now standard practice at frontier labs and referenced in regulation.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-75","messages":[{"id":"msg-mrxl4f9k-74-0-lt59m","role":"assistant","content":"Crafting inputs that bypass a model's safety training — role-play framings, encodings, multi-step manipulations. The persistent cat-and-mouse between jailbreakers and labs demonstrates that current alignment is shallow: the underlying capability remains, merely gated.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-76","messages":[{"id":"msg-mrxl4f9k-75-0-wr8pl","role":"assistant","content":"The deployed layer of safety: filters, classifiers, and system-prompt rules wrapped around a model in production. Guardrails are the pragmatic complement to alignment — defense in depth for the model's failure modes.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-77","messages":[{"id":"msg-mrxl4f9k-76-0-hapmf","role":"assistant","content":"The general-purpose text engines at the center of modern AI — systems trained on much of the written internet that can draft, summarize, translate, code, and converse. Their generality is the surprise: one artifact, thousands of tasks. Whether they understand or merely imitate is the field's liveliest debate, but their gravitational pull on every other node in this graph is not in dispute.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784817210173-xbs8zf1","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T14:33:30.173Z","source":"ai"},{"id":"msg-1784817215481-sqivs5r","role":"assistant","content":"The current tension in understanding Large Language Models lies in the gap between their architectural simplicity and their functional complexity. At their core, these systems are engines of probability, yet they exhibit behaviors—like In-Context Learning—that suggest a capacity for logic and adaptation far beyond simple pattern matching.\n\nThis tension is best captured by the conflict between the Stochastic Parrot View and the World-Model View. If a model is merely a \"stochastic parrot,\" Hallucination is an inevitable byproduct of its design: the model is simply prioritizing the most probable sequence of tokens over factual truth because it lacks an underlying grounding in reality. However, if the model has developed an internal \"world model\" during pre-training, hallucination looks less like a random error and more like a failure of retrieval or a \"dreaming\" state where the model's internal logic overrides external data.\n\nThe expansion into Vision-Language Models pushes this debate further. By integrating visual data, these systems move toward a more grounded form of intelligence. A model that can reason about a screenshot or a diagram is no longer just manipulating text strings; it is mapping linguistic concepts to spatial and structural representations. This suggests that the \"toolbox\" of modern AI is moving away from isolated tasks toward a unified multimodal reasoning capability.\n\nA critical nuance often missed in this framing is the role of the \"System Prompt\" and \"Post-training\" (such as RLHF). These layers sit between the raw model and the user, attempting to steer the model away from its \"parrot\" tendencies and toward a more reliable, helpful assistant. The core question for the next stage of development is whether Hallucination is a bug that can be engineered away through better data, or if it is a fundamental feature of a system that prioritizes the generative \"creativity\" required for flexible problem-solving.\n\nAs these systems transition from passive predictors to active entities—moving into the realm of AI Agents—the focus shifts from what the model \"knows\" to what it can \"do.\" This makes the debate over internal world models more than academic: if a model is to use tools or execute code autonomously, its internal representation of cause and effect must be robust enough to handle real-world consequences.","timestamp":"2026-07-23T14:33:35.481Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:33:35.482Z"},{"nodeId":"n-mermaid-mrxl4f8r-78","messages":[{"id":"msg-mrxl4f9k-77-0-cbzx0","role":"assistant","content":"The ability to learn a new task from examples placed in the prompt, with no retraining — show three formatted examples, get a fourth. An emergent property of scale that turned prompting into a genuine programming paradigm.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784830815868-bvbixzd","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:20:15.868Z","source":"ai"},{"id":"msg-1784830828073-gl9jfos","role":"assistant","content":"In-context learning represents a fundamental shift in the nature of computation. Traditionally, software behavior is determined by fixed logic or weight updates. In the era of Large Language Models, however, the model’s \"intelligence\" becomes fluid, reorganizing itself around the specific patterns and constraints provided within the current input. This is not learning in the synaptic sense—no weights are changed—but rather a sophisticated form of pattern matching where the model uses the provided context to navigate its internal high-dimensional space.\n\nThe emergence of this capability is closely tied to the Transformer architecture’s ability to attend to all parts of an input simultaneously. While we often treat this as a mysterious \"emergent\" property, research into mechanistic interpretability suggests it is driven by specific circuits known as induction heads. These circuits allow the model to recognize a pattern (e.g., [A][B]), find a subsequent occurrence of [A], and predict that [B] will follow again. When these circuits operate across thousands of tokens, they allow the model to simulate an understanding of complex instructions or novel formatting that it may never have encountered during its initial training.\n\nHowever, there is a recurring tension between the expansion of the context window and the model's actual reasoning quality. While a larger window provides more room for examples and data, it also introduces \"distractors.\" Models often exhibit a \"lost in the middle\" effect, where information at the beginning or end of a long prompt is prioritized, while the center is effectively ignored. This phenomenon suggests that simply having a large context window does not guarantee that the model is effectively utilizing all the information provided via in-context learning.\n\nThis leads to a critical trade-off in prompt engineering: the \"more is better\" fallacy. While providing more examples (few-shot prompting) typically improves performance, it also increases the risk of hallucination. When a model is pushed to follow an in-context pattern that contradicts its pre-trained factual knowledge, it may prioritize the pattern over the truth. In this sense, hallucination is not always a failure of memory, but sometimes a failure of the model to balance the weight of the provided context against its underlying training.\n\nA significant open question in this field is whether in-context learning is merely a sophisticated form of retrieval or if the model is actually performing \"gradient descent in the forward pass.\" Some evidence suggests that during the inference process, the model’s internal activations are essentially running a mini-optimization algorithm, \"training\" itself on the provided examples in real-time. If this holds true, in-context learning is less about \"reading\" and more about temporary, high-speed adaptation.\n\nTo deepen this exploration, it may be useful to look at the specific mechanisms and constraints that define how information is prioritized and retrieved during this process.","timestamp":"2026-07-23T18:20:28.073Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:20:28.074Z"},{"nodeId":"n-mermaid-mrxl4f8r-79","messages":[{"id":"msg-mrxl4f9k-78-0-wxguk","role":"assistant","content":"The confident generation of false information — fabricated facts, citations, quotes — arising because models are trained to produce plausible text, not verified truth. The single biggest obstacle to trusting AI in high-stakes work, mitigated by RAG and citations, eliminated by nothing so far.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784829033913-b7o95ve","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:50:33.913Z","source":"ai"},{"id":"msg-1784829040194-6t3emoy","role":"assistant","content":"The core challenge of hallucination lies in the fundamental nature of Large Language Models as probabilistic next-token predictors. Because these systems are trained to maximize the likelihood of a sequence rather than to reference an internal database of verified facts, they often prioritize linguistic fluency and structural plausibility over objective truth. In many cases, a hallucination is simply the model \"filling in\" a pattern where the training data was sparse or contradictory, leading to what is sometimes called \"confabulation.\"\n\nFrom the perspective of AI Safety, hallucinations represent a significant hurdle for reliability and trust. If a model can generate a citation that looks formally correct—complete with author names, journals, and volume numbers—but refers to a paper that does not exist, it becomes difficult for users to distinguish between genuine expertise and \"stochastic parroting.\" This risk is amplified in high-stakes domains like medicine or law, where the appearance of competence can mask dangerous inaccuracies.\n\nRetrieval-Augmented Generation (RAG) serves as a primary architectural countermeasure to this tendency. By decoupling the \"reasoning\" capability of the model from its \"knowledge\" storage, RAG allows the system to ground its output in a specific set of retrieved documents. Instead of relying solely on its internal parameters (which are fixed at the time of training), the model acts as an interface for a dynamic external context. This significantly reduces the frequency of \"closed-domain\" hallucinations—where the model invents facts—by forcing it to cite and summarize provided information.\n\nHowever, even with RAG, \"open-domain\" hallucinations can persist if the model misinterprets the retrieved context or if the retrieval step itself pulls in irrelevant or conflicting information. This highlights a deeper tension: the more a model is trained to be creative and helpful, the more prone it may be to \"over-assertiveness,\" where it feels compelled to provide an answer even when the correct response would be an admission of ignorance. \n\nTo deepen this exploration, it may be useful to consider the role of the Softmax Bottleneck in how models assign probability to \"correct\" vs. \"hallucinated\" tokens, or to look at how Reinforcement Learning from Human Feedback (RLHF) often inadvertently encourages \"sycophancy,\" where a model hallucinates information simply because it believes that is what the user wants to hear.\n\n**Advisory Observations:**\n*   The transition from Hallucination to Retrieval-Augmented Generation is often described as \"mitigates\" or \"reduces\" rather than \"enables.\" While the existence of hallucination *motivated* the development of RAG, RAG is designed to constrain the model's output to prevent the failure mode.\n*   The node \"Large Language Models\" is a broad category; for a more technical investigation into why hallucinations happen, focusing on the specific mechanics of Next-Token Prediction or the limitations of the Training Data would provide more granular insight.","timestamp":"2026-07-23T17:50:40.194Z","source":"ai"},{"id":"msg-1784829142108-dcnqotb","role":"user","content":"What is the problem with hallucinations? why to solve it?","timestamp":"2026-07-23T17:52:22.108Z","source":"ai"},{"id":"msg-1784829149325-355wtub","role":"assistant","content":"The fundamental problem with hallucinations is that they decouple **linguistic authority** from **factual accuracy**. Because Large Language Models are optimized for the probability of a sequence (Next-Token Prediction) rather than the truth-value of a claim, they produce \"hallucinations\" that are not random gibberish, but highly plausible, grammatically perfect falsehoods.\n\nThis creates several distinct layers of failure:\n\n### 1. The Erosion of User Veracity (The \"Euthyphro\" Problem)\nWhen a model hallucinates, it isn't just \"wrong\"—it is often deceptively wrong. By generating fabricated citations or logical proofs that follow the *style* of expertise, the model exploits the user's heuristic for trust. If a system is correct 95% of the time with high confidence, a user is cognitively primed to accept the 5% that is hallucinated. This makes LLMs \"high-friction\" tools for factual tasks because the cost of verification often nears the cost of doing the original research.\n\n### 2. Failure of Grounding and Agency\nAs we move toward AI Agents that take actions—such as booking flights, writing code, or summarizing legal contracts—hallucination becomes a functional breaker. If a model hallucinates a non-existent API parameter or a deadline in a contract, the downstream \"action\" fails or causes real-world harm. For AI Safety, this is a \"robustness\" failure: the system cannot be considered safe if its internal world-model can diverge from reality without the system itself noticing the discrepancy.\n\n### 3. The RLHF Reinforcement Loop\nA specific tension exists within RLHF (Reinforcement Learning from Human Feedback). Human raters often reward answers that sound confident and helpful. If a model is forced to choose between saying \"I don't know\" and providing a plausible-sounding hallucination, the RLHF process has historically pushed models toward the latter (Sycophancy). Solving hallucination requires teaching the model to \"hedge\" or admit ignorance, which often conflicts with the goal of making the model feel \"smart\" and \"capable.\"\n\n### Why Solve It? (The Threshold of Utility)\nSolving hallucinations is the \"unlock\" for several industry-defining transitions:\n*   **From \"Copilot\" to \"Autopilot\":** Reliable agents cannot exist if they fabricate their environment.\n*   **Legal and Medical Liability:** For LLMs to move from \"creative brainstorming\" to \"professional utility,\" there must be a guarantee of factual grounding.\n*   **The Data Wall:** As models begin to train on synthetic data (text generated by other AI), hallucinations can lead to \"model collapse,\" where errors are amplified in each generation, eventually turning the latent space into noise.\n\nTo better understand how we attempt to fix this, it is worth looking at **In-Context Learning** as a way to force the model to look at provided facts, and **Mechanistic Interpretability** to see if we can \"catch\" a hallucination by looking at the model's internal activations before the token is even generated.","timestamp":"2026-07-23T17:52:29.325Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:52:29.326Z"},{"nodeId":"n-mermaid-mrxl4f8r-8","messages":[{"id":"msg-mrxl4f9k-7-0-d9ktd","role":"assistant","content":"The field's trajectory in eight landmarks, from the symbolic-connectionist schism through the winters to the deep learning revolution and the present reasoning era. History here is vocabulary: \"AI winter\" and \"ImageNet moment\" are working terms of the current discourse.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784815742075-vevbzoy","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T14:09:02.075Z","source":"ai"},{"id":"msg-1784815749458-a4s2v4q","role":"assistant","content":"The history of artificial intelligence is defined by a fundamental tension between two philosophies of mind: the top-down desire to program logic and the bottom-up attempt to grow it. This \"Long Road\" is not a straight line of progress, but a series of pendulum swings between these two poles, punctuated by periods of disillusionment when the technology failed to meet the soaring rhetoric of its pioneers.\n\nThe early dominance of Symbolic AI rested on the assumption that intelligence is essentially the manipulation of high-level concepts according to formal rules. This \"Good Old Fashioned AI\" (GOFAI) succeeded in closed systems like chess or mathematical logic, but it suffered from a \"brittleness\" when faced with the messy, ambiguous data of the real world. This inability to handle perception and common sense led directly to the AI Winters, where the collapse of funding followed the realization that mapping the world into a fixed set of rules was an intractable task.\n\nRunning parallel to this was Connectionism, the long-held belief that intelligence emerges from simple, interconnected units—inspired by the architecture of the human brain. For decades, this approach was marginalized because it lacked the computational power and the massive datasets required to prove its worth. The Deep Learning Revolution of 2012 served as the definitive vindication of this school. It proved that if you provide a neural network with enough data and compute, it can learn its own internal representations of the world, bypassing the need for human-coded rules entirely.\n\nThe milestones that followed—the strategic intuition demonstrated by AlphaGo and the linguistic mastery unlocked by the Transformer Breakthrough—marked the transition from specific pattern recognition to general-purpose synthesis. The ChatGPT Moment was the public realization of this shift, demonstrating that statistical \"next-token prediction\" could produce a convincing facsimile of human reasoning.\n\nHowever, we are currently entering a new phase that arguably attempts to reconcile these two historical paths. The Reasoning Era, characterized by models that generate internal chains of thought before responding, represents a return to \"deliberative\" processing. Unlike the pure \"system 1\" intuition of early deep learning, these newer architectures use inference-time computation to simulate the step-by-step logic that Symbolic AI once tried to hard-code.\n\nA critical question for the next stage of this road is whether this \"reasoning\" is a genuine emergent property of scaling or if it remains a sophisticated form of pattern matching that will eventually hit a new wall. As we move from models that simply *predict* the next word to models that *verify* their own steps, we are testing the limits of the Connectionist framework to see if it can finally solve the problems of consistency and truth that once grounded the Symbolic school.\n\nTo better map how these historical threads are converging, it may be useful to examine the technical mechanisms that allow these models to move beyond simple pattern matching into deliberate problem-solving.","timestamp":"2026-07-23T14:09:09.458Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:09:09.458Z"},{"nodeId":"n-mermaid-mrxl4f8r-80","messages":[{"id":"msg-mrxl4f9k-79-0-lp628","role":"assistant","content":"The extension of models beyond text to images, audio, and video — both understanding them and generating them. Frontier models are now natively multimodal, trained on mixed media from the start rather than having senses bolted on.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-81","messages":[{"id":"msg-mrxl4f9k-80-0-3rkfd","role":"assistant","content":"Models that process images and text together: describe a photo, read a chart, reason about a screenshot. The capability that underlies computer-use agents, document AI, and visual search.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-82","messages":[{"id":"msg-mrxl4f9k-81-0-neu16","role":"assistant","content":"Models for hearing and speaking: transcription that rivals humans and synthesis that is often indistinguishable from them. The synthesis side, led commercially by ElevenLabs, powers voice assistants and audiobooks — and voice-cloning fraud, the capability's dark twin.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-83","messages":[{"id":"msg-mrxl4f9k-82-0-9p4up","role":"assistant","content":"Producing images from text descriptions, the capability that first made generative AI viscerally real to the public in 2022. Built mostly on diffusion models; Midjourney made it a consumer product and a copyright battleground.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-84","messages":[{"id":"msg-mrxl4f9k-83-0-q41o7","role":"assistant","content":"Generating video from text or image prompts — the hardest generative frontier, since it requires coherent physics and identity over time. OpenAI's Sora and Runway's models moved it from demo to production tool; implications for film and for misinformation are still unfolding.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-85","messages":[{"id":"msg-mrxl4f9k-84-0-c7wy7","role":"assistant","content":"The generation technique behind most image and video AI: learn to remove noise, then create by starting from pure noise and denoising toward a description. A fundamentally different mechanism from next-token prediction — the second great generative recipe.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-86","messages":[{"id":"msg-mrxl4f9k-85-0-04cty","role":"assistant","content":"Models trained to deliberate before answering — generating extended internal reasoning, checking their own work, then responding. Introduced by OpenAI's o1 in 2024 and rapidly adopted industry-wide, they marked the field's pivot to test-time compute and reset expectations on math, science, and coding.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784827912630-mlt68yl","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:31:52.630Z","source":"ai"},{"id":"msg-1784827918889-cqepdnq","role":"assistant","content":"The shift toward reasoning models represents a transition from \"system 1\" thinking—instinctive, fast, next-token prediction—to \"system 2\" thinking, which is deliberate and self-correcting. This evolution is fundamentally driven by the realization that scaling training data and parameters eventually yields diminishing returns for complex problem-solving. Instead, the focus has moved toward maximizing the utility of compute at the moment of inference.\n\nThe core mechanism here is the decoupling of a model's internal processing time from the length of the final output. By utilizing **Test-Time Compute**, a model can \"think\" through a problem for seconds or even minutes, exploring various paths via **Chain-of-Thought** before committing to a visible answer. This effectively turns a static model into a search process. The model isn't just predicting the most likely next word; it is navigating a state-space of possible reasoning steps, evaluating their validity, and backtracking when it hits a dead end.\n\nThis approach brings a modern perspective to the long-standing goal of **Neuro-Symbolic Integration**. Traditional symbolic AI was excellent at rigid logic but brittle in the face of messy, real-world data. Conversely, neural networks are robust but often struggle with strict multi-step constraints. Reasoning models bridge this gap by using the neural network to \"simulate\" a symbolic engine—generating and checking logical premises in a structured, sequential manner.\n\nHowever, this \"Reasoning Era\" introduces a new set of trade-offs and open questions:\n\n1.  **The Latency-Quality Trade-off**: We are moving away from the expectation of instant responses. In this paradigm, the user—or an autonomous agent—must decide how much \"thinking\" a particular query is worth. For a creative writing task, test-time compute may add little value, but for verifying a mathematical proof, it is essential.\n2.  **Verifiability vs. Generation**: While these models can reason through a problem, the \"hidden\" thoughts can still harbor hallucinations. The challenge shifts from ensuring the model knows the answer to ensuring the model's self-correction process is itself grounded in reality.\n3.  **Agency and Planning**: This internal deliberation is the precursor to true **Agency & Long-Horizon Planning**. If a model can reason through its own steps, it can theoretically evaluate the outcomes of potential actions in a digital environment before executing them.\n\nTo deepen the understanding of how these models are built, it is worth looking at the specific RL techniques that reward the *process* of reasoning rather than just the final answer. Unlike standard fine-tuning, which uses human-labeled \"correct\" answers, the training of reasoning models often relies on **RLVR** (Reinforcement Learning from Verifiable Rewards), where the model is rewarded for reaching a correct, verifiable conclusion (like a math result or compiled code) regardless of the specific path it took.\n\nThis shift suggests that the future of frontier models lies less in \"bigger\" architectures and more in \"smarter\" search and verification strategies during the reasoning process.","timestamp":"2026-07-23T17:31:58.889Z","source":"ai"},{"id":"msg-1784827936776-4p27zvr","role":"user","content":"Explain the selected node: “RLVR”.","timestamp":"2026-07-23T17:32:16.776Z","source":"ai"},{"id":"msg-1784827940804-gyiem64","role":"assistant","content":"**RLVR** (Reinforcement Learning from Verifiable Rewards) is a training methodology designed to overcome the limitations of learning from human feedback alone. While traditional RLHF (Reinforcement Learning from Human Feedback) relies on human trainers to rank model outputs—a process that is subjective, slow, and prone to rewarding \"plausible-sounding\" but incorrect answers—RLVR uses objective, automated signals to ground a model's logic.\n\nIn the context of the current **Reasoning Era**, RLVR is the engine that allows a model to learn how to think. It focuses on domains where the \"ground truth\" is binary and computationally checkable, such as:\n*   **Mathematics:** Does the final numerical result match the solution?\n*   **Code:** Does the generated script execute without errors and pass all functional test cases?\n*   **Formal Logic:** Is the conclusion a valid derivation from the premises?\n\nThe significance of this approach lies in how it shapes the internal **Chain-of-Thought**. When a model is trained via RLVR, it is essentially put in a sandbox where it can experiment with different reasoning paths. If it tries a \"mental\" shortcut that leads to a coding error, the environment provides immediate negative feedback. If it backtracks, catches its own mistake, and eventually reaches the correct answer, it receives a positive reward. \n\nThis creates a critical distinction between \"knowing\" a fact and \"learning\" a process. RLVR doesn't just teach the model the answer to a specific math problem; it reinforces the *patterns of thought*—the search, the verification, and the self-correction—that consistently lead to verifiable success. This is what enables the leap from basic next-token prediction to true **Inference-Time Search**.\n\nBy grounding reasoning in verifiable rewards, we mitigate the \"stochastic parrot\" problem. The model is no longer just mimicking the style of a textbook; it is being optimized against a hard reality. This makes RLVR a foundational pillar for building models capable of **Neuro-Symbolic Integration**, as it provides a way to train neural networks to respect the rigid, non-negotiable rules of symbolic systems like logic and programming languages.","timestamp":"2026-07-23T17:32:20.804Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:32:20.804Z"},{"nodeId":"n-mermaid-mrxl4f8r-87","messages":[{"id":"msg-mrxl4f9k-86-0-v6key","role":"assistant","content":"The technique of working through problems step by step rather than answering directly, which dramatically improves accuracy on multi-step tasks. Discovered first as a prompting trick, then trained into models as the built-in behavior of reasoners.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-88","messages":[{"id":"msg-mrxl4f9k-87-0-upf2y","role":"assistant","content":"Systems that pursue goals through sequences of actions — searching, calling tools, writing files, iterating on failures — rather than producing one response. The industry's consensus bet on where AI value concentrates next, and the point where reliability, not intelligence, becomes the binding constraint.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784831523922-69g5gsq","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:32:03.922Z","source":"ai"},{"id":"msg-1784831530668-wioxqil","role":"assistant","content":"The shift from passive chat assistants to autonomous AI agents represents a transition from \"thought-on-demand\" to \"action-over-time.\" While traditional models output a single completion, agents operate in a loop: they perceive an environment, reason about a goal, select a tool, and then integrate the feedback from that tool to refine their next step.\n\nThe Model Context Protocol (MCP) serves as the connective tissue in this architecture. By standardizing how an agent discovers and interacts with local or remote data—whether that is a SQL database, a file system, or a live API—it moves us away from brittle, bespoke integrations. It effectively creates a universal interface for the \"brain\" to talk to the \"hands,\" allowing any agentic system to inherit a vast library of capabilities without unique engineering for every new tool.\n\nHowever, the leap to true autonomy involves a distinction between simple tool use and full computer use. While tool use typically involves structured API calls (e.g., \"get_weather\" or \"search_index\"), computer use requires the agent to interpret unstructured visual data and manipulate a standard operating system. This is a significantly more complex planning problem; the agent must maintain a long-horizon state, understanding that clicking a \"Submit\" button in a browser might take seconds to register and require multiple intermediate steps of verification.\n\nAs these tasks grow in complexity, the limitations of a single model instance become apparent. Multi-agent systems address this by decomposing a high-level objective into specialized roles—for example, one instance acting as a researcher, another as a coder, and a third as a critic. This collaborative approach mitigates the \"drift\" often seen in long-horizon planning, where a single model might lose track of the original goal after a dozens-step sequence.\n\nA central tension in this evolution is the Inference-Compute Trade-off. As agents move from simple next-token prediction to iterative reasoning (thinking before and during action), they require significantly more compute at the time of execution. This shift suggests that the value of AI is moving away from the \"knowledge\" stored in the model's weights and toward the \"reasoning process\" the model employs while interacting with the world.\n\nTo further develop this area, it is worth considering how agents manage their own internal state and planning history. Two critical areas for exploration are:\n\n1.  **Inference-Time Search**: The techniques that allow agents to explore multiple potential paths or \"branches\" of action before committing to one (analogous to how AlphaGo evaluates moves).\n2.  **Agency & Long-Horizon Planning**: The specific challenge of maintaining goal-permanence over hundreds of steps without succumbing to error accumulation or \"hallucinating\" that a task is complete when it is not.","timestamp":"2026-07-23T18:32:10.668Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:32:10.669Z"},{"nodeId":"n-mermaid-mrxl4f8r-89","messages":[{"id":"msg-mrxl4f9k-88-0-n9lo1","role":"assistant","content":"The mechanism that gives models hands: they emit structured calls to external functions — search, calculators, databases, code execution — and incorporate the results. Tool use is the boundary crossing between language model and software system.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784831385735-d9q3cjd","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:29:45.735Z","source":"ai"},{"id":"msg-1784831391796-qk1dxek","role":"assistant","content":"The transition from static text generation to active tool use marks the shift from AI as a consultant to AI as an agent. At its core, tool use is the bridge that allows a model to overcome its inherent limitations: it cannot perform precise real-time arithmetic, it lacks access to private or live data, and it cannot effect change in the physical or digital world on its own. By framing these capabilities as \"tools,\" we move the model into a feedback loop where the output of a thought becomes the input of an action.\n\nThis agency is increasingly structured through standardized protocols. The emergence of the **Model Context Protocol** (MCP) represents a critical maturation of this field. Previously, tool integration was fragmented, requiring bespoke \"glue code\" for every new data source or API. By decoupling the model from the specific implementation of the tool, MCP allows for a plug-and-play ecosystem. This is not just a technical convenience; it fundamentally changes the scaling laws of utility. When a model can fluidly navigate a standardized interface to interact with local files, databases, or web services, the bottleneck shifts from \"what can the model do?\" to \"what permissions does the agent have?\"\n\n**Computer Use** takes this abstraction a step further by removing the need for structured APIs entirely. Instead of calling a specific `get_weather()` function, the system interacts with the same semantic layer humans use: the graphical user interface. This is a high-stakes evolution. While API-based tool use is relatively predictable and constrained, computer use requires the agent to interpret visual hierarchy, manage latency, and handle the inherent \"noise\" of a desktop environment. It bridges the gap between specialized automation and general-purpose assistance.\n\nHowever, this level of integration introduces a significant tension regarding **Inference-Time Search** and planning. For an agent to use a tool effectively—especially in a multi-step sequence—it must be able to reason about the tool's output before deciding on the next move. This often requires the model to \"pause\" and evaluate whether the tool's result met the goal or if a different approach is needed. This recursive process is what separates a simple LLM from a true **AI Agent**.\n\nA live challenge in this domain is the \"reliability gap.\" As tools become more complex, the probability of a sequence failing increases exponentially. If each step in a five-step tool-use chain has a 90% success rate, the overall task success drops to roughly 59%. Improving the robustness of these interactions—moving from \"probabilistic invocation\" to \"deterministic execution\"—is the current frontier of agentic research.\n\nTo further develop this area, it may be useful to consider the mechanisms that allow models to handle these complex sequences, such as how they manage long-term goals or how they determine when a task is actually \"complete\" versus stuck in a loop.","timestamp":"2026-07-23T18:29:51.796Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:29:51.796Z"},{"nodeId":"n-mermaid-mrxl4f8r-9","messages":[{"id":"msg-mrxl4f9k-8-0-is37f","role":"assistant","content":"The genuinely unresolved questions — safety, what LLMs really are, why capabilities keep growing, whether AGI is coming, whether the data will run out, whether the boom is a bubble. Each question here is a hub that reaches into the factual content behind other doors; this lens is the shortcut to everything currently argued about.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784818775984-q08mjfp","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T14:59:35.984Z","source":"ai"},{"id":"msg-1784818782442-op2dh4j","role":"assistant","content":"The discourse surrounding modern artificial intelligence is currently defined by a profound \"predictability gap.\" While the scaling laws established over the last several years provide a reliable engineering roadmap for predicting performance on next-token prediction, they offer almost no foresight into the emergence of qualitative reasoning capabilities or the timeline for reaching Artificial General Intelligence. This tension is the primary driver of the current industry volatility.\n\nAt the heart of the **Capability Progress** debate is whether we are witnessing a fundamental shift in the nature of computing or merely an unprecedented optimization of statistical imitation. The **LLM Understanding Debate** centers on this distinction: if a model can pass the bar exam or write functional code without a grounded world-model, does it matter for the trajectory toward AGI? Critics of the \"stochastic parrot\" view argue that the distinction is a category error—that sufficiently complex pattern matching across all human knowledge eventually necessitates the internal construction of logic and world-modeling to achieve lower loss.\n\nHowever, the path forward faces two immediate structural threats:\n1.  **The Data Wall**: We are approaching the exhaustion of high-quality, human-generated text. If progress is purely a function of scale, hitting this limit suggests a plateau. The industry's pivot toward synthetic data and self-correcting \"reasoning\" loops (where models spend more compute at inference time to \"think\") is an attempt to bypass this wall by substituting human data with machine logic.\n2.  **The Bubble Question**: There is a widening chasm between the massive capital expenditure required for the next generation of clusters and the current revenue generated by AI applications. If the \"Reasoning Era\" does not quickly translate into autonomous agents capable of complex, high-value labor, the investment boom may face a correction before the technical hurdles to AGI are cleared.\n\nThis leads to the most polarizing aspect of the field: **AI Safety**. The disagreement here is not just about the probability of catastrophe, but about the nature of intelligence itself. If intelligence is a tool that can be steered, safety is a technical alignment problem. If intelligence is an inherently agentic force that seeks to preserve its own utility functions, then **Superintelligence** poses an existential challenge that our current fine-tuning methods (like RLHF) are fundamentally unequipped to handle.\n\nTo deepen the current exploration, it is worth examining the specific technical shifts that might break the \"Data Wall\" or justify the current investment. **Scaling Laws** provide the mathematical justification for the current boom, while the shift toward **Reasoning Models** and **Inference** (test-time compute) represents the most likely path toward AGI if traditional pre-training data actually runs dry.","timestamp":"2026-07-23T14:59:42.442Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:59:42.443Z"},{"nodeId":"n-mermaid-mrxl4f8r-90","messages":[{"id":"msg-mrxl4f9k-89-0-82esl","role":"assistant","content":"Agents that operate real interfaces the way people do — reading the screen, moving the cursor, typing. Pioneered commercially by Anthropic in 2024, it points at automation of arbitrary desktop work, with reliability and security as the open problems.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-91","messages":[{"id":"msg-mrxl4f9k-90-0-lmjn1","role":"assistant","content":"Architectures in which several model instances cooperate — planner and executors, writer and critic, orchestrator and specialists. Promising for complex work, though coordination overhead and compounding errors mean more agents is not automatically better.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-92","messages":[{"id":"msg-mrxl4f9k-91-0-s51h9","role":"assistant","content":"An open standard, introduced by Anthropic, for connecting AI systems to tools and data sources — often glossed as \"USB-C for AI.\" Its broad adoption across the industry signals agents maturing from demos into infrastructure.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784831319120-da9ywns","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T18:28:39.120Z","source":"ai"},{"id":"msg-1784831325738-sbfejmq","role":"assistant","content":"The Model Context Protocol (MCP) represents a shift from a fragmented ecosystem of bespoke integrations toward a unified interface for agentic cognition. By standardizing how models discover and interact with data, it addresses the \"n-to-m\" integration problem—where every new AI agent previously required a custom connector for every unique data source or tool.\n\nAt its core, this protocol formalizes the relationship between the reasoning engine and its environment. While early tool use was often limited to hard-coded function calling within a single application's silo, MCP allows an agent to treat a local database, a remote API, or a specific development environment as a hot-swappable \"server.\" This decouples the intelligence of the model from the specific implementation of the tools it uses, enabling a much broader range of autonomous action.\n\nThe primary tension in this architecture is between **generality and security**. As agents gain the ability to navigate complex file systems or execute code across standardized interfaces, the \"blast radius\" of a model's hallucinations or a prompt injection attack increases. Standardizing the protocol makes it easier to build robust permission layers, but it also creates a common target for exploitation. \n\nFurthermore, MCP clarifies a crucial distinction in the evolution of AI: the difference between a model with \"built-in\" knowledge and a model with \"retrieval-augmented\" agency. Instead of trying to cram more context into the weights of the model during training, MCP leans into the philosophy that intelligence is most effective when it is a thin layer governing access to a rich, externalized memory and toolset. This aligns with the broader industry move toward \"Inference-Time Search\" and \"Agentic Workflows,\" where the quality of the output depends less on what the model \"knows\" and more on how effectively it can explore the available information space.\n\nA live question for this architecture is the degree of **semantic overhead**. For a protocol like MCP to be truly universal, the model must understand not just the syntax of the tool (the API call), but the semantic intent and the state changes that occur after the call. We are moving toward a world where tools are no longer designed for humans to click, but for models to read and manipulate as if they were extensions of their own latent space.\n\n### Strategic Extensions\n\nTo deepen this exploration, it is worth considering how MCP interacts with the underlying infrastructure and the emerging patterns of agentic reasoning.\n\n*   **Inference-Time Search**: MCP provides the \"pipes\" through which models can perform complex searches and reason over the results during the generation process.\n*   **Tree-of-Thought Reasoning**: Standardized tool access allows agents to branch their reasoning, using external tools to verify hypotheses before committing to a final path.\n*   **Model APIs**: The protocol is specifically designed to work across different providers, making the choice of which API to use a matter of cost and performance rather than integration lock-in.","timestamp":"2026-07-23T18:28:45.738Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T18:28:45.739Z"},{"nodeId":"n-mermaid-mrxl4f8r-93","messages":[{"id":"msg-mrxl4f9k-92-0-s0hjm","role":"assistant","content":"Grounding model answers in retrieved documents: fetch relevant passages first, then answer from them. RAG is the standard enterprise remedy for hallucination and stale knowledge, and the reason a chatbot can cite your company's own policies.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784829003009-re3esuo","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:50:03.009Z","source":"ai"},{"id":"msg-1784829008653-2jq5aeg","role":"assistant","content":"Retrieval-Augmented Generation (RAG) serves as the primary bridge between a model's static training data and the dynamic, authoritative world of private or real-time information. While large language models possess impressive reasoning capabilities, they are effectively \"frozen\" at the moment their training ends. RAG shifts the burden of knowledge from the model’s internal weights to an external, searchable index, transforming the model from an encyclopedic source into a sophisticated document processor.\n\nThis transformation is the most effective defense against hallucination. By providing a \"ground truth\" context, the model no longer has to guess at facts; it can be instructed to strictly synthesize the provided excerpts. This turns a generation task into a sophisticated form of summarization and extraction, significantly increasing the reliability of AI Search applications.\n\nThe effectiveness of this architecture is governed by a few critical technical constraints:\n\n*   **The Latent Bridge:** A Vector Database is the engine of the retrieval step, converting text into high-dimensional embeddings to find semantic matches that traditional keyword searches might miss. However, the quality of retrieval is entirely dependent on the embedding model's ability to capture the nuance of the query. If the retrieval step fails to find the right \"chunks,\" the generation step—no matter how capable—cannot produce a correct answer.\n*   **The Context Window Constraint:** Even as context windows expand to handle millions of tokens, RAG remains relevant because of \"lost in the middle\" phenomena and the sheer cost/latency of processing massive amounts of text. The challenge lies in the trade-off: providing enough context for the model to understand the nuance without overwhelming it with \"noise\" that might distract the attention mechanism or exceed its processing limits.\n*   **The Accuracy of Attribution:** A mature RAG system does not just provide an answer; it provides a citation. This moves the user experience from one of blind trust to one of verification, which is essential for professional tools in legal, medical, or technical fields.\n\nA deeper tension in current RAG research is the choice between \"dense\" retrieval (using vectors) and \"sparse\" retrieval (using traditional keywords). While vectors excel at finding conceptual similarities, they often struggle with specific technical terms, part numbers, or rare acronyms. Most production-grade systems are now moving toward hybrid search to mitigate these weaknesses.\n\nBeyond simple retrieval, the next frontier is \"Agentic RAG,\" where the system doesn't just perform one search and one generation, but instead reasons about whether the retrieved information is sufficient, performs follow-up searches if gaps exist, and critiques its own generated output against the source text before presenting it.","timestamp":"2026-07-23T17:50:08.653Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:50:08.654Z"},{"nodeId":"n-mermaid-mrxl4f8r-94","messages":[{"id":"msg-mrxl4f9k-93-0-e3bz7","role":"assistant","content":"A database that indexes content by embedding vectors, retrieving by meaning rather than keyword. The storage layer of RAG and semantic search.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784828856940-p1fdwqp","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T17:47:36.940Z","source":"ai"},{"id":"msg-1784828867346-3vddnym","role":"assistant","content":"Vector databases represent a fundamental shift in how information is indexed and retrieved, moving away from literal keyword matching toward semantic relevance. This shift is predicated on the ability to treat information as coordinates in a high-dimensional space. By utilizing embeddings to transform text, images, or audio into dense vectors, the database allows for \"similarity search,\" where the goal is to find the closest points in space rather than the exact string of characters.\n\nThe effectiveness of this system depends entirely on the alignment between the embedding model's latent space and the retrieval task at hand. If the model fails to capture specific domain nuances—such as the difference between \"bank\" as a financial institution and \"bank\" as a river edge—the geometric proximity in the vector database will be semantically misleading. This introduces a critical trade-off not present in traditional relational databases: the sacrifice of absolute precision for the sake of conceptual flexibility. Most vector databases employ Approximate Nearest Neighbor (ANN) algorithms, such as Hierarchical Navigable Small World (HNSW) or Inverted File Indexes (IVF). These algorithms are probabilistic, meaning they prioritize low latency over a guaranteed 100% accurate list of the \"nearest\" results.\n\nIn the context of Retrieval-Augmented Generation (RAG), the vector database acts as an external, non-parametric memory for large language models. This architecture addresses the inherent limitations of a model's fixed training data and finite context window. While the model provides the \"reasoning\" and linguistic capability, the vector database provides the specific, grounded facts retrieved in real-time.\n\nHowever, a common tension in these systems is the \"lost in the middle\" phenomenon and the limitations of dense retrieval. While vector search is excellent at finding \"about-ness,\" it often struggles with precise identifiers like part numbers, specific dates, or rare technical jargon that a traditional BM25 keyword search would catch instantly. This has led to the rise of hybrid search strategies, which combine dense vector retrieval with sparse keyword indexing to ensure that conceptual breadth doesn't come at the cost of literal accuracy.\n\nAs these systems evolve, the focus is shifting toward how metadata filtering can be integrated directly into the vector search process. Without efficient filtering, a search might return the most semantically relevant document, but it might be from the wrong year, the wrong department, or a restricted security tier.","timestamp":"2026-07-23T17:47:47.346Z","source":"ai"},{"id":"msg-1784828911880-sp8dnic","role":"user","content":"give me a more rigorous and clear definition of what is the vector database?","timestamp":"2026-07-23T17:48:31.880Z","source":"ai"},{"id":"msg-1784828918337-bmpa97e","role":"assistant","content":"At its most formal level, a **Vector Database** is a specialized storage and retrieval system designed to manage data through its geometric representation in a high-dimensional vector space, rather than through relational tables or document hierarchies.\n\nTo define it rigorously, we can break it down into four functional components:\n\n### 1. The Mathematical Representation\nThe fundamental unit of data is the **embedding**. A model $\\phi$ maps a raw data object $D$ (text, image, audio) into a vector $v$ in a $d$-dimensional space $\\mathbb{R}^d$:\n$$\\phi: D \\to v \\in \\mathbb{R}^d$$\nThe database does not \"understand\" the data; it manages the spatial coordinates where semantic meaning has been compressed. The dimensionality $d$ is typically high (e.g., 768 or 1536), allowing the vector to capture complex features and relationships.\n\n### 2. The Distance Metric (The Objective Function)\nUnlike a SQL database that evaluates `WHERE x = y` (boolean identity), a vector database evaluates `Similarity(v1, v2)`. The definition of \"closeness\" is determined by a chosen distance metric, most commonly:\n*   **Cosine Similarity:** Measures the angle between vectors, focusing on orientation rather than magnitude:\n    $$\\text{similarity} = \\cos(\\theta) = \\frac{A \\cdot B}{\\|A\\| \\|B\\|}$$\n*   **Euclidean Distance ($L_2$):** Measures the straight-line distance between two points in space.\n*   **Dot Product:** Measures both the direction and the magnitude.\n\n### 3. The Indexing Strategy (ANN)\nIn a standard database, you can index a column to find a value in $O(\\log n)$ time. In high-dimensional space, finding the *exact* nearest neighbor for a query vector $q$ among $N$ vectors is computationally expensive—often requiring a full linear scan ($O(N \\cdot d)$).\n\nTo solve this, vector databases utilize **Approximate Nearest Neighbor (ANN)** algorithms. These create data structures (like graphs, trees, or clusters) that allow the system to navigate the space efficiently. This involves a fundamental trade-off: **Recall vs. Latency**. You accept a small statistical probability that the database might miss the absolute closest vector in exchange for sub-millisecond retrieval speeds across millions of records.\n\n### 4. Database CRUD and Metadata\nWhat distinguishes a \"vector database\" from a simple \"vector library\" (like Faiss) is the integration of traditional database features:\n*   **Persistence:** Storing vectors on disk rather than just in RAM.\n*   **Metadata Filtering:** The ability to combine a vector search with a metadata constraint (e.g., \"Find the most similar paragraphs, but only from documents dated after 2022\").\n*   **Horizontal Scaling:** Sharding and replicating vector indexes across multiple nodes.\n\n**Summary Definition:**\nA vector database is a system that indexes and stores high-dimensional embeddings to enable efficient similarity-based retrieval, governed by a distance metric and optimized via Approximate Nearest Neighbor algorithms, while supporting traditional database operations like metadata filtering and persistence.","timestamp":"2026-07-23T17:48:38.337Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T17:48:38.338Z"},{"nodeId":"n-mermaid-mrxl4f8r-95","messages":[{"id":"msg-mrxl4f9k-94-0-95y8s","role":"assistant","content":"Tools that write, complete, and debug code — Copilot, Cursor, and successors — the first AI application with undeniable enterprise-wide productivity impact. Software development doubles as the industry's proving ground for agents, since code can be run and tested.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-96","messages":[{"id":"msg-mrxl4f9k-95-0-ngl15","role":"assistant","content":"The conversational product form — ChatGPT and its rivals — through which most of the public knows AI. Deceptively simple packaging over the full stack, and now a battleground over which the platform giants and frontier labs compete for the consumer relationship.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"},{"id":"msg-1784816411703-iezk4ti","role":"user","content":"Discuss the current scene.","timestamp":"2026-07-23T14:20:11.703Z","source":"ai"},{"id":"msg-1784816416828-gf869xu","role":"assistant","content":"The transformation of Large Language Models from research artifacts into Chat Assistants represents the most significant shift in human-computer interaction since the graphical user interface. While the underlying models are trained on the statistical distributions of text, the \"assistant\" wrapper is a deliberate layer of post-training—primarily Reinforcement Learning from Human Feedback (RLHF)—designed to map those distributions onto human social expectations: helpfulness, honesty, and harmlessness.\n\nThis transition, catalyzed by the ChatGPT Moment, turned a raw completion engine into a tool capable of following instructions and maintaining state across a dialogue. It moved AI out of the \"Toolbox\" of data scientists and into the hands of the general public. However, this productization creates a tension between the model's objective of accurate information retrieval and the user's desire for a specific persona or experience.\n\nIn the case of Character.AI, the emphasis shifts from utility to engagement and companionship. Here, the \"assistant\" isn't just a productivity partner; it is a role-playing entity. This highlights a divergence in the evolution of these systems:\n*   **Utility-driven assistants** (like ChatGPT or Claude) prioritize minimizing \"hallucination\" and maximizing task efficiency.\n*   **Persona-driven assistants** prioritize consistency of voice and emotional resonance, where factual accuracy might even be secondary to narrative immersion.\n\nA critical evolution currently underway is the move from passive chat to active agency. While the early phase of Chat Assistants focused on \"talking about doing,\" the current frontier involves \"doing\"—integrating with external tools, browsing the live web, and executing code. This transition forces a reconsideration of safety; when an assistant can take actions, the risks move beyond misinformation into the realm of unintended system consequences.\n\nTo broaden this view, it is worth looking at how these assistants are built (the \"Frontier Labs\" and \"Model Families\") and how they are becoming more than just text interfaces through \"Multimodality\" and \"AI Agents.\"\n\n### Related Concepts for the Scene\n\n*   **Frontier Labs:** The organizational context for OpenAI and its competitors, defining the pace of these releases.\n*   **RLHF:** The specific mechanism that transforms a base LLM into a conversational assistant.\n*   **AI Agents:** The next stage of assistant evolution, where the chat interface becomes a control center for autonomous tasks.\n*   **Multimodality:** Extending the assistant from text-only to vision and voice, completing the \"human-like\" interaction loop.","timestamp":"2026-07-23T14:20:16.828Z","source":"ai"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:20:16.829Z"},{"nodeId":"n-mermaid-mrxl4f8r-97","messages":[{"id":"msg-mrxl4f9k-96-0-phaxt","role":"assistant","content":"Search that answers directly, with citations, instead of returning links — Perplexity's founding premise, since adopted by Google and ChatGPT alike. It is remaking the economics of the web: publishers lose clicks while their content feeds the answers.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-98","messages":[{"id":"msg-mrxl4f9k-97-0-3a0q5","role":"assistant","content":"The discipline of measuring what models can do — harder than it sounds, since capabilities are broad, jagged, and easily gamed. As stakes rise, evaluation is professionalizing: from academic benchmarks toward audits, arenas, and task-based testing.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"},{"nodeId":"n-mermaid-mrxl4f8r-99","messages":[{"id":"msg-mrxl4f9k-98-0-9t3xi","role":"assistant","content":"Standardized test sets — math competitions, coding challenges, graduate-level science — used to compare models and headline every launch. Useful but distorting: what gets benchmarked gets optimized.","timestamp":"2026-07-23T14:06:35.307Z","source":"tutorial"}],"createdAt":"2026-07-23T14:06:35.307Z","updatedAt":"2026-07-23T14:06:35.307Z"}],"backgroundImages":[],"shelf":{"n-mermaid-mrxl4f8r-8":{"actions":[{"type":"include_existing","title":"Scaling Laws","reason":"To understand the transition from the Deep Learning Revolution to the current era, as it explains the relationship between compute, data, and emergent capabilities."},{"type":"include_existing","title":"Chain-of-Thought","reason":"This is the specific technique that bridges the gap between pure neural networks and the 'reasoning' capabilities defining the current era."},{"type":"include_existing","title":"Test-Time Compute","reason":"This represents the shift in the Reasoning Era from scaling training to scaling the model's 'thinking time' during inference."}],"timestamp":1784827693279},"n-mermaid-mrxl4f8r-162":{"actions":[{"type":"include_existing","title":"GPU","reason":"The hardware-software fit between the Transformer and Nvidia's parallel architecture is why this breakthrough superseded others."},{"type":"create_connected","title":"Hardware-Lottery Hypothesis","connectionType":"child","reason":"The theory that the Transformer succeeded because it was better suited for GPUs, whereas other potentially superior architectures failed because they didn't align with existing hardware."}],"timestamp":1784815950416},"n-mermaid-mrxl4f8r-22":{"actions":[{"type":"include_existing","title":"The Transformer Block","reason":"Provides the structural container for attention and feed-forward layers that defines the architecture."},{"type":"include_existing","title":"Scaling Laws","reason":"Describes the empirical relationship between compute, data, and model performance that justifies the Transformer's design."},{"type":"include_existing","title":"Sequential vs. Parallel Processing","reason":"Highlights the fundamental shift from RNNs to Transformers that enabled modern AI scaling."}],"timestamp":1784815929346},"n-mermaid-mrxl4f8r-23":{"actions":[{"type":"include_existing","title":"Positional Encoding","reason":"Attention is inherently order-agnostic; positional encodings are required to give the model a sense of sequence."},{"type":"include_existing","title":"Sequential vs. Parallel Processing","reason":"This contrast highlights why the Transformer architecture was able to surpass RNNs by leveraging GPU hardware."}],"timestamp":1784816112844},"n-1784-816-030-709":{"actions":[{"type":"include_existing","title":"Multi-Head Attention","reason":"To show how the scaled dot-product is parallelized in practice."},{"type":"include_existing","title":"Positional Encoding","reason":"To address the lack of inherent sequence order in dot-product attention."}],"timestamp":1784816311221},"n-mermaid-mrxl4f8r-163":{"actions":[{"type":"include_existing","title":"RLHF","reason":"RLHF was the specific 'finishing school' for the GPT models that made the conversational interface effective enough for public adoption."}],"timestamp":1784827775853},"n-mermaid-mrxl4f8r-96":{"actions":[{"type":"include_existing","title":"RLHF","reason":"The technical process responsible for the 'assistant' behavior and alignment."}],"timestamp":1784816478435},"n-mermaid-mrxl4f8r-53":{"actions":[{"type":"include_existing","title":"Scaling Laws","reason":"These laws dictate the relationship between compute, data, and the resulting performance of the Base Model."},{"type":"include_existing","title":"Transformer Architecture","reason":"The structural blueprint that enables the Pre-training of Foundation Models."},{"type":"create_connected","title":"Alignment Tax","connectionType":"child","reason":"The potential loss in general capability or 'raw' intelligence that occurs when a Base Model is constrained during fine-tuning."},{"type":"include_existing","title":"Next-Token Prediction","reason":"The fundamental training objective that defines the behavior and raw output of a base model."}],"timestamp":1784816639126},"n-mermaid-mrxl4f8r-77":{"actions":[{"type":"include_existing","title":"RLHF","reason":"To address how hallucinations are mitigated and how the 'toolbox' is refined for human use."},{"type":"create_connected","title":"Grounding","connectionType":"child","reason":"To explore the bridge between the 'Stochastic Parrot' and 'World-Model' views, specifically how models link tokens to real-world entities."}],"timestamp":1784818670004},"n-mermaid-mrxl4f8r-54":{"actions":[{"type":"include_existing","title":"Scaling Laws","reason":"These govern the relationship between compute, data, and performance, serving as the blueprint for foundation model development."},{"type":"include_existing","title":"Self-Supervised Learning","reason":"This is the core learning paradigm that allows foundation models to be pre-trained on unlabeled, web-scale datasets."},{"type":"include_existing","title":"Mixture of Experts","reason":"A key architectural strategy used in modern foundation models to scale parameter counts while maintaining manageable inference costs."},{"type":"include_existing","title":"Alignment","reason":"The process within Post-training that ensures the foundation model's outputs remain helpful and safe for downstream applications."}],"timestamp":1784818628202},"n-mermaid-mrxl4f8r-9":{"actions":[{"type":"include_existing","title":"Scaling Laws","reason":"To provide the mathematical context for why Capability Progress is currently viewed as a predictable engineering problem by frontier labs."},{"type":"include_existing","title":"Reasoning Models","reason":"To connect the Capability Progress debate to the newer paradigm of 'test-time compute' as a solution to the Data Wall."},{"type":"include_existing","title":"Inference","reason":"Crucial for discussing the Bubble Question, as the cost and efficiency of running these models determines their ultimate economic viability."},{"type":"create_connected","title":"Agency and Autonomy","connectionType":"child","reason":"A critical 'burning question' regarding whether AGI requires models to have independent goals and the ability to act in the world, rather than just answering queries."}],"timestamp":1784818782447},"n-mermaid-mrxl4f8r-172":{"actions":[{"type":"include_existing","title":"Recursive Self-Improvement","reason":"This is the primary theoretical mechanism by which AGI matures into Superintelligence."},{"type":"include_existing","title":"Alignment","reason":"The central challenge in ensuring Superintelligence remains beneficial."}],"timestamp":1784823161497},"n-mermaid-mrxl4f8r-33":{"actions":[{"type":"include_existing","title":"Scaling Hypothesis","reason":"Provides the theoretical belief system that justifies the pursuit of emergent capabilities through sheer scale."},{"type":"include_existing","title":"Benchmark Saturation","reason":"Explains why some capabilities appear to emerge suddenly—often because our previous metrics were too simple to capture intermediate progress."},{"type":"include_existing","title":"LLM Understanding Debate","reason":"A core philosophical tension regarding whether emergent behaviors constitute true understanding or advanced statistical mimicry."},{"type":"include_existing","title":"Reasoning Models","reason":"The current frontier where emergence is being specialized into explicit logical capabilities (e.g., through RL and search)."},{"type":"include_existing","title":"Test-Time Compute","reason":"A new scaling dimension that shows capabilities can 'emerge' or improve at inference time, not just during training."}],"timestamp":1784825158042},"n-mermaid-mrxl4f8r-36":{"actions":[{"type":"include_existing","title":"Transformer Architecture","reason":"Mechanistic interpretability is primarily focused on reverse-engineering the components of the transformer."},{"type":"include_existing","title":"Attention Mechanism","reason":"Attention heads are the primary 'circuits' identified in mechanistic research, such as induction heads."},{"type":"include_existing","title":"Phase Transitions","reason":"The sudden appearance of capabilities (emergence) is often studied mechanistically as the sudden formation of specific circuits during training."},{"type":"create_connected","title":"Superposition","reason":"The phenomenon where models represent more features than they have dimensions, a major hurdle in mechanistic interpretability."}],"timestamp":1784823344727},"n-mermaid-mrxl4f8r-35":{"actions":[{"type":"include_existing","title":"Induction Heads","reason":"Induction heads are a primary success story of mechanistic interpretability, representing specific circuits that enable pattern matching and in-context learning."},{"type":"include_existing","title":"Attention Mechanism","reason":"Most mechanistic interpretability research focuses on 'attending' to how different parts of the transformer architecture interact, specifically through attention heads."},{"type":"include_existing","title":"Transformer Architecture","reason":"This is the 'body' being dissected by interpretability researchers; understanding the architecture is a prerequisite for understanding the internal circuits."},{"type":"include_existing","title":"Red-Teaming","reason":"Interpretability serves as a 'white-box' form of red-teaming, looking for vulnerabilities or misalignments within the weights rather than just the outputs."},{"type":"include_existing","title":"Embeddings","reason":"Embeddings are the initial point where data becomes a vector that interpretability tools can analyze, forming the basis of feature detection."}],"timestamp":1784823646558},"n-mermaid-mrxl4f8r-165":{"actions":[{"type":"create_connected","title":"Inner vs. Outer Alignment","connectionType":"child","reason":"This is a fundamental technical distinction within the Alignment problem that explains why simple instructions often fail."},{"type":"create_connected","title":"Existential Risk","connectionType":"child","reason":"To represent the high-stakes end of the safety spectrum mentioned by Hinton, distinct from day-to-day misuse."},{"type":"include_existing","title":"Mechanistic Interpretability","reason":"This specific sub-field of Interpretability is the most relevant to solving Alignment through internal inspection."},{"type":"include_existing","title":"EU AI Act","reason":"This is the most concrete current example of AI Governance in practice."}],"timestamp":1784823800336},"n-mermaid-mrxl4f8r-38":{"actions":[{"type":"include_existing","title":"Yoshua Bengio","reason":"Often grouped with Hinton as a 'Godfather of AI,' Bengio shares his concerns about AI safety and provides a complementary perspective on the risks of frontier models."},{"type":"include_existing","title":"Yann LeCun","reason":"The third 'Godfather' who offers a critical counterpoint to Hinton's safety warnings, arguing that current AI lacks the necessary world-models for agency."}],"timestamp":1784823983422},"n-mermaid-mrxl4f8r-160":{"actions":[{"type":"include_existing","title":"Scaling Hypothesis","reason":"The move from AlexNet to modern AI was driven by the conviction that more compute and data would lead to emergent intelligence."},{"type":"include_existing","title":"Transformer Architecture","reason":"This is the successor to the convolutional networks used in the 2012 revolution, representing the next major leap in deep learning."},{"type":"include_existing","title":"Stochastic Parrot View","reason":"A primary philosophical critique of the results produced by the deep learning revolution."}],"timestamp":1784824073346},"n-1784-824-173-383":{"actions":[{"type":"include_existing","title":"Deep Learning Revolution","reason":"Provides the historical bridge between Rumelhart's foundational work in the 1980s and the current era of AI."}],"timestamp":1784824417017},"n-1784-824-174-809":{"actions":[{"type":"include_existing","title":"The XOR Problem","reason":"Provides the historical context for why the PDP research group's work on backpropagation was such a necessary breakthrough."},{"type":"include_existing","title":"Sequential vs. Parallel Processing","reason":"Captures the architectural argument McClelland made against the Von Neumann model of the mind."}],"timestamp":1784824503340},"n-1784-824-369-059":{"actions":[{"type":"include_existing","title":"Parallel Distributed Processing (PDP)","reason":"This was the foundational framework/book by McClelland and Rumelhart that contextualized NETtalk's success."},{"type":"include_existing","title":"David Rumelhart","reason":"Rumelhart was the primary architect of the backpropagation algorithm used to train NETtalk and a key collaborator with McClelland."},{"type":"include_existing","title":"The XOR Problem","reason":"NETtalk was significant because it bypassed the limitations of earlier perceptrons, which were famously unable to solve non-linear problems like XOR."},{"type":"create_connected","title":"Distributed Representation","reason":"The core concept behind NETtalk: that information is not stored in one place (a symbol) but across the strengths of many connections."}],"timestamp":1784824549517},"n-mermaid-mrxl4f8r-14":{"actions":[{"type":"include_existing","title":"Parallel Distributed Processing (PDP)","reason":"This existing node represents the architectural bridge between cognitive science and neural network structure, emphasizing how information is stored across a network rather than in a localized 'memory' cell."},{"type":"include_existing","title":"Graceful Degradation","reason":"This describes a key functional property shared by both biological brains and artificial neural networks, where the system resists catastrophic failure when parts of it are damaged."}],"timestamp":1784824876205},"n-mermaid-mrxl4f8r-2":{"actions":[{"type":"include_existing","title":"Induction Heads","reason":"A key discovery in Mechanistic Interpretability that explains how Transformers learn to perform in-context learning."},{"type":"include_existing","title":"Data Wall","reason":"This represents the current primary challenge to the Scaling Hypothesis: the looming exhaustion of high-quality human-generated data."},{"type":"include_existing","title":"Attention Mechanism","reason":"This is the specific engine within the Transformer Architecture that enables the parallel processing and long-range dependencies central to its success."},{"type":"include_existing","title":"Latent Space","reason":"Understanding how models represent concepts in a high-dimensional vector space is foundational to Interpretability research."},{"type":"include_existing","title":"Stochastic Parrot View","reason":"A core skeptical perspective in the 'Big Ideas' debate regarding whether scaling actually leads to understanding."},{"type":"include_existing","title":"World-Model View","reason":"The opposing theory to the 'Stochastic Parrot' view, suggesting that large-scale training forces models to learn the underlying structure of the world."}],"timestamp":1784828548273},"n-1784-823-072-505":{"actions":[{"type":"include_existing","title":"Deep Learning","reason":"Represents the modern connectionist stack that neuro-symbolic methods aim to enhance with better reasoning and sample efficiency."},{"type":"include_existing","title":"The XOR Problem","reason":"A classic example of a logical relationship that connectionist models initially struggled to represent, highlighting the historical tension between the two fields."},{"type":"include_existing","title":"The Transformer Block","reason":"The fundamental architecture where neural representations are currently formed and manipulated."},{"type":"include_existing","title":"Attention Mechanism","reason":"The process by which models decide which 'symbols' or features are relevant to one another in a given context."},{"type":"include_existing","title":"Latent Space","reason":"The mathematical space where grounded representations exist before they are decoded into discrete outputs."}],"timestamp":1784825361613},"n-1784-825-371-425":{"actions":[{"type":"include_existing","title":"Symbolic AI","reason":"To provide the historical and conceptual context for the 'symbols' that need grounding."},{"type":"include_existing","title":"Connectionism","reason":"The paradigm that provides the 'neural' component of neuro-symbolic integration and the mechanism for sensory mapping."}],"timestamp":1784825486063},"n-mermaid-mrxl4f8r-169":{"actions":[{"type":"include_existing","title":"World-Model View","reason":"This is the primary counter-argument to the Stochastic Parrot view, suggesting that models develop internal representations of reality to improve prediction."},{"type":"include_existing","title":"Emergent Capabilities","reason":"These are the specific behaviors that challenge the 'mere pattern matching' description by appearing only at certain scales of compute and data."},{"type":"create_connected","title":"Syntax vs. Semantics","reason":"A core philosophical distinction relevant to the debate, often illustrated by John Searle's Chinese Room argument, which underpins the parrot critique."}],"timestamp":1784825520760},"n-mermaid-mrxl4f8r-168":{"actions":[{"type":"include_existing","title":"Intrinsic vs. Borrowed Grounding","reason":"Distinguishes between a model having its own internal meaning versus inheriting meaning from the humans who produced the training data."}],"timestamp":1784825577190},"n-mermaid-mrxl4f8r-170":{"actions":[{"type":"include_existing","title":"Scaling Laws","reason":"To explain why world-modeling is often viewed as an emergent property of scale."},{"type":"include_existing","title":"Mechanistic Interpretability","reason":"To provide the empirical evidence used to support the claim that internal world models actually exist within the weights."},{"type":"include_existing","title":"Stochastic Parrot View","reason":"To provide the necessary counter-argument and dialectical tension to the World-Model View."},{"type":"create_connected","title":"Functional Grounding","reason":"A concept to bridge the gap between pure text prediction and the physical world, answering the symbol grounding problem."},{"type":"create_connected","title":"Data Bandwidth Problem","connectionType":"child","reason":"A core critique of the World-Model View, arguing that text (a few bits per second) is insufficient to learn the high-resolution structure of reality compared to video or sensory input."},{"type":"include_existing","title":"Scaling Hypothesis","reason":"Provides the theoretical basis for why world-modeling emerges as a function of scale."},{"type":"include_existing","title":"Emergent Capabilities","reason":"The specific skills that appear unexpectedly as a result of the model internalizing a more complex world model."},{"type":"include_existing","title":"Latent Space","reason":"The geometric environment where the model's internal representations and 'world' actually exist."},{"type":"include_existing","title":"Next-Token Prediction","reason":"The fundamental training objective that critics argue is too simple, but world-model proponents argue is the catalyst for deep representation."},{"type":"include_existing","title":"Self-Supervised Learning","reason":"The paradigm that allows models to learn the structure of the world from raw data without human labels."}],"timestamp":1784826205428},"n-mermaid-mrxl4f8r-106":{"actions":[{"type":"include_existing","title":"Scaling Laws","reason":"The mathematical foundation that justifies the massive capital expenditure of frontier labs."},{"type":"include_existing","title":"Nvidia","reason":"The primary hardware enabler for every lab in the scene; their roadmap dictates the labs' training timelines."},{"type":"include_existing","title":"Compute & Silicon","reason":"To connect the organizational layer of the labs to the physical resource constraints they face."},{"type":"include_existing","title":"Sam Altman","reason":"Adding key executives personifies the strategic competition within 'The Arena'."},{"type":"include_existing","title":"Dario Amodei","reason":"Crucial for representing the safety-centric strategy within the frontier lab landscape."}],"timestamp":1784826615252},"n-mermaid-mrxl4f8r-115":{"actions":[{"type":"include_existing","title":"Scaling Laws","reason":"The fundamental theoretical framework that guides executive decisions on compute spend and model size."},{"type":"include_existing","title":"Compute & Silicon","reason":"The physical bottleneck that defines the competitive timeline for all frontier executives."},{"type":"include_existing","title":"AI Governance","reason":"The regulatory landscape that executives like Altman and Amodei are actively trying to shape."}],"timestamp":1784826931752},"n-mermaid-mrxl4f8r-127":{"actions":[{"type":"include_existing","title":"TSMC","reason":"Nvidia is a fabless designer; its ability to deliver hardware is entirely dependent on TSMC's manufacturing capacity and leading-edge nodes."},{"type":"include_existing","title":"Custom AI Chips","reason":"To understand the competitive landscape and the push for efficiency beyond general-purpose GPUs."},{"type":"create_connected","title":"NVLink","connectionType":"child","reason":"The proprietary high-speed interconnect technology that allows GPUs to communicate at massive bandwidth, essential for cluster-scale training."},{"type":"create_connected","title":"CoWoS Packaging","connectionType":"child","reason":"The 'Chip on Wafer on Substrate' advanced packaging technique that is currently the primary manufacturing bottleneck for high-end AI chips."}],"timestamp":1784827071789},"n-mermaid-mrxl4f8r-126":{"actions":[{"type":"include_existing","title":"Custom AI Chips","reason":"Major players are increasingly designing in-house silicon to compete with merchant providers like Nvidia and AMD."}],"timestamp":1784827139677},"n-mermaid-mrxl4f8r-147":{"actions":[{"type":"include_existing","title":"TSMC","reason":"TSMC is the sole manufacturer of the advanced chips required for frontier datacenters, representing the most concentrated point of failure in the supply chain."},{"type":"include_existing","title":"Export Controls","reason":"Geopolitical restrictions on high-end compute determine the global distribution and strategic value of large-scale datacenters."},{"type":"include_existing","title":"Inference Cost","reason":"The economic sustainability of the datacenter boom depends on reducing the cost of serving models to users relative to the capital expenditure of training them."}],"timestamp":1784827230761},"n-mermaid-mrxl4f8r-149":{"actions":[{"type":"include_existing","title":"Inference Cost","reason":"The economic viability of the investment boom depends on whether the cost of running models can be brought down to match market demand."},{"type":"include_existing","title":"Energy Demand","reason":"Energy is becoming the primary limiting factor for the scaling of infrastructure and a central part of the investment thesis."},{"type":"include_existing","title":"Scaling Laws","reason":"The scientific rationale that underpins the belief that massive capital investment in compute will result in significantly more capable models."}],"timestamp":1784827290459},"n-mermaid-mrxl4f8r-174":{"actions":[{"type":"include_existing","title":"Scaling Laws","reason":"The economic justification for the investment boom is predicated on the continued validity of scaling laws; if they plateau, the bubble thesis gains significant weight."},{"type":"include_existing","title":"Reasoning Era","reason":"The shift toward test-time compute represents a new phase of the investment boom that attempts to extract more value from existing hardware."}],"timestamp":1784827376720},"n-mermaid-mrxl4f8r-159":{"actions":[{"type":"include_existing","title":"Moravec’s Paradox","reason":"Explains why early Symbolic AI failed to translate logic into real-world utility, a primary driver of the first AI winter."},{"type":"include_existing","title":"The XOR Problem","reason":"A specific technical critique that stifled Connectionism and contributed to the narrow focus on symbolic systems before the second winter."},{"type":"include_existing","title":"Data Wall","reason":"A contemporary constraint that informs the 'Bubble Question' by suggesting a physical or logical limit to the current scaling paradigm."},{"type":"include_existing","title":"Connectionism","reason":"The paradigm that was sidelined following the publication of 'Perceptrons', only to return decades later during the Deep Learning Revolution."}],"timestamp":1784827564122},"n-mermaid-mrxl4f8r-164":{"actions":[{"type":"include_existing","title":"Scaling Laws","reason":"To understand how the Reasoning Era represents a new frontier in the belief that more compute leads to more intelligence."},{"type":"include_existing","title":"Synthetic Data","reason":"Reasoning models are often used to generate high-quality data to train future models, bypassing the limitations of human-only datasets."},{"type":"include_existing","title":"Data Wall","reason":"Provides the context for why the industry is moving toward reasoning and inference-time compute as training data sources dry up."}],"timestamp":1784827872691},"n-mermaid-mrxl4f8r-86":{"actions":[{"type":"include_existing","title":"Agency & Long-Horizon Planning","reason":"Reasoning models are the primary enabler for agents to plan over long durations by simulating outcomes internally."}],"timestamp":1784827930104},"n-mermaid-mrxl4f8r-13":{"actions":[{"type":"include_existing","title":"Chain-of-Thought","reason":"RL (specifically RLVR) is the primary method used to train models to generate and refine the intermediate steps known as Chain-of-Thought."},{"type":"create_connected","title":"Reward Hacking","connectionType":"child","reason":"A phenomenon where an agent finds unintended ways to maximize the reward signal without fulfilling the actual design intent, a major hurdle in RLHF alignment."},{"type":"create_connected","title":"MCTS","connectionType":"child","reason":"Monte Carlo Tree Search was the core of AlphaGo's decision-making and is increasingly relevant for modern reasoning models using test-time compute to explore potential answers."},{"type":"create_connected","title":"Credit Assignment Problem","connectionType":"child","reason":"The difficulty of determining which specific actions in a long sequence contributed to an eventual reward, especially relevant in multi-step reasoning tasks."}],"timestamp":1784828196344},"n-mermaid-mrxl4f8r-10":{"actions":[{"type":"include_existing","title":"Deep Learning","reason":"Deep Learning is the specific evolution of Neural Networks that enables the scale and performance discussed in the context of Transfer Learning and Big Ideas."},{"type":"include_existing","title":"Self-Supervised Learning","reason":"This is the contemporary bridge between Unsupervised and Supervised learning that currently drives the most significant breakthroughs in the field."},{"type":"include_existing","title":"Universal Approximation Theorem","reason":"This provides the theoretical justification for why Neural Networks are the primary architecture used to solve Machine Learning problems."},{"type":"include_existing","title":"Symbolic AI","reason":"To provide a contrast between learning-based approaches and the rule-based systems of broader Artificial Intelligence."},{"type":"include_existing","title":"Deep Learning Revolution","reason":"To provide historical and scale context for why Deep Learning became the dominant paradigm within ML."},{"type":"include_existing","title":"Yann LeCun","reason":"Turing Award winner and pioneer of Convolutional Neural Networks."},{"type":"include_existing","title":"Yoshua Bengio","reason":"Turing Award winner known for his work on deep learning and neural language models."},{"type":"include_existing","title":"ImageNet Challenge","reason":"The 2012 event (AlexNet) that validated the deep learning approach championed by Hinton."},{"type":"include_existing","title":"David Rumelhart","reason":"Key figure in the Parallel Distributed Processing group who co-authored the seminal paper on backpropagation."},{"type":"include_existing","title":"James McClelland","reason":"A leader in the connectionist movement that laid the psychological and structural framework for neural networks."}],"timestamp":1784828527507},"n-mermaid-mrxl4f8r-79":{"actions":[{"type":"include_existing","title":"Mechanistic Interpretability","reason":"This field explores whether we can detect 'truthfulness' or 'lying' within the model's internal states, offering a more fundamental solution than post-training patches."}],"timestamp":1784829175380},"n-1784-816-099-231":{"actions":[{"type":"include_existing","title":"Embeddings","reason":"The embedding matrix is the mathematical 'other half' of the softmax bottleneck; its dimensionality relative to the latent space defines the rank limit."},{"type":"create_connected","title":"Mixture of Softmaxes","connectionType":"child","properties":{"description":"A technique to break the softmax bottleneck by modeling the output distribution as a weighted average of multiple softmax layers."},"reason":"MoS is the primary architectural solution proposed to overcome the rank limitations identified in the softmax bottleneck."}],"timestamp":1784829210257},"n-mermaid-mrxl4f8r-58":{"actions":[{"type":"include_existing","title":"RLAIF","reason":"RLAIF is the direct evolution of RLHF, replacing human labelers with AI labelers to scale the alignment process."},{"type":"include_existing","title":"RLVR","reason":"RLVR addresses the subjectivity of human feedback by using objective, verifiable rewards (like unit tests or math checks)."},{"type":"include_existing","title":"Supervised Fine-Tuning","reason":"SFT is the prerequisite step to RLHF, providing the initial 'policy' that the reinforcement learning process then refines."}],"timestamp":1784829243019},"n-mermaid-mrxl4f8r-3":{"actions":[{"type":"include_existing","title":"RLHF","reason":"The primary mechanism in Post-training for aligning model behavior with human intent."},{"type":"include_existing","title":"Synthetic Data","reason":"Distillation is increasingly driven by generating high-quality synthetic datasets from a teacher model to train the student."},{"type":"include_existing","title":"Scaling Laws","reason":"Distillation challenges traditional scaling laws by showing that smaller models can punch above their weight if the 'quality' of their training signal is high."}],"timestamp":1784829821887},"n-mermaid-mrxl4f8r-62":{"actions":[{"type":"include_existing","title":"AI Accelerators","reason":"Inference efficiency is increasingly dependent on specialized hardware like TPUs or LPU architectures (e.g., Groq) that optimize for low-latency token generation."},{"type":"include_existing","title":"Retrieval-Augmented Generation","reason":"RAG is the standard architectural pattern for extending a model's knowledge at inference time without retraining, bypassing some of the limitations of the fixed Context Window."}],"timestamp":1784829870990},"n-mermaid-mrxl4f8r-65":{"actions":[{"type":"include_existing","title":"Chain-of-Thought","reason":"To represent the primary technique for improving reasoning performance through prompting."}],"timestamp":1784829972470},"n-mermaid-mrxl4f8r-66":{"actions":[{"type":"include_existing","title":"In-Context Learning","reason":"System prompts rely on the model's ability to learn and adapt its behavior based on the context provided in the input."},{"type":"include_existing","title":"Jailbreaking","reason":"This is the primary adversarial challenge to the efficacy of a system prompt."},{"type":"include_existing","title":"Chain-of-Thought","reason":"System prompts often explicitly instruct a model to use CoT to improve reasoning performance."}],"timestamp":1784830021127},"n-mermaid-mrxl4f8r-63":{"actions":[{"type":"include_existing","title":"Attention Mechanism","reason":"The quadratic scaling of the attention mechanism is the primary reason the context window is a bottleneck."},{"type":"include_existing","title":"Scaling Laws","reason":"Scaling laws describe how the relationship between compute, data, and context window size affects model performance."},{"type":"create_connected","title":"Quadratic Scaling","connectionType":"child","properties":{"equation":"O(n^2)"},"reason":"The mathematical reason why doubling the context window quadruples the computational cost in standard transformers."},{"type":"create_connected","title":"RoPE (Rotary Positional Embeddings)","connectionType":"child","reason":"The positional encoding method that enables models to handle much longer sequences than they saw during training."},{"type":"create_connected","title":"Linear Attention","connectionType":"child","properties":{"equation":"O(n)"},"reason":"Alternative architectures like SSMs or Mamba that aim to solve the scaling bottleneck of traditional context windows."},{"type":"create_connected","title":"Context Compression","connectionType":"child","reason":"Techniques for condensing information within the window to fit more meaning into fewer tokens."}],"timestamp":1784830126840},"n-mermaid-mrxl4f8r-78":{"actions":[{"type":"include_existing","title":"Induction Heads","reason":"These are the specific mechanistic circuits thought to be responsible for the pattern-matching that enables in-context learning."},{"type":"include_existing","title":"Lost in the Middle","reason":"This identifies a primary constraint of the context window, where the model's ability to learn in-context degrades based on the position of information."},{"type":"include_existing","title":"Few-Shot Prompting","reason":"This is the most common practical application of in-context learning, providing the model with multiple examples to define a task."}],"timestamp":1784830838757},"n-mermaid-mrxl4f8r-46":{"actions":[{"type":"include_existing","title":"Transformer Architecture","reason":"The Transformer is the specific engine that makes modern pre-training at scale possible; its attention mechanism is what allows the model to process the context necessary for next-token prediction."},{"type":"include_existing","title":"Attention Mechanism","reason":"This is the core functional component of the pre-training process, allowing the model to weigh the importance of different parts of the input data."}],"timestamp":1784831017973},"n-mermaid-mrxl4f8r-31":{"actions":[{"type":"include_existing","title":"Inference-Time Search","reason":"To contrast training-time scaling laws with the emerging trend of scaling compute during the reasoning phase (e.g., o1-style models)."},{"type":"include_existing","title":"Data Wall","reason":"To address the primary physical limit currently facing the continuation of traditional scaling laws."},{"type":"include_existing","title":"Phase Transitions","reason":"To provide the specific technical name for the 'discontinuous' emergence of capabilities as scale increases."}],"timestamp":1784831088044},"n-mermaid-mrxl4f8r-92":{"actions":[{"type":"include_existing","title":"Tree-of-Thought Reasoning","reason":"Standardized tool access enables models to validate different branches of reasoning by interacting with external data sources."}],"timestamp":1784831352318},"n-mermaid-mrxl4f8r-4":{"actions":[{"type":"include_existing","title":"Inference-Time Search","reason":"This is the underlying mechanism that powers Reasoning Models, explaining how they trade compute for better outputs."},{"type":"include_existing","title":"Inference-Compute Trade-off","reason":"This provides the economic and technical context for why we use different tools in the toolbox for different tasks."},{"type":"include_existing","title":"Agency & Long-Horizon Planning","reason":"This is the frontier of AI Agents, moving beyond simple tool-calling to actual task completion over time."},{"type":"create_connected","title":"Tool Use & Function Calling","connectionType":"child","reason":"The specific capability that allows LLMs to interact with the Model Context Protocol and external APIs."}],"timestamp":1784831498006},"n-mermaid-mrxl4f8r-88":{"actions":[{"type":"include_existing","title":"Inference-Time Search","reason":"Agents require the ability to look ahead and evaluate multiple action paths, making search a core component of agentic reasoning."},{"type":"include_existing","title":"Inference-Compute Trade-off","reason":"Agentic loops shift the cost of AI from training to the moment of execution, a critical economic and technical constraint."}],"timestamp":1784831579540},"n-mermaid-mrxl4f8r-103":{"actions":[{"type":"include_existing","title":"Unit Economics of Intelligence","reason":"To explore how the cost of training and hardware dictates the pricing of Model APIs."},{"type":"include_existing","title":"Inference Cost","reason":"To ground the discussion in the physical reality of GPU utilization and energy demand that limits API margins."},{"type":"include_existing","title":"Context Window","reason":"Crucial for understanding the technical limitations and tiering of different API offerings."},{"type":"include_existing","title":"Model Context Protocol","reason":"Already connected in the graph; this represents the next stage of API evolution regarding data integration."}],"timestamp":1784832062126},"n-mermaid-mrxl4f8r-102":{"actions":[{"type":"include_existing","title":"Self-Hosting","reason":"Self-hosting is the primary technical and business driver for choosing open-weight models over closed APIs."},{"type":"include_existing","title":"Sovereign AI","reason":"Explains the geopolitical and strategic motivation for entities like Mistral and Alibaba to release open models."},{"type":"include_existing","title":"DeepSeek","reason":"DeepSeek is currently a major disruptor in the open-weight space, often outperforming Llama and Mistral on key benchmarks."},{"type":"include_existing","title":"Grouped-Query Attention (GQA)","reason":"A critical architectural feature common in modern open-weight models (like Llama 3) that enables efficient inference."}],"timestamp":1784832081890},"n-mermaid-mrxl4f8r-104":{"actions":[{"type":"include_existing","title":"Sovereign AI","reason":"Self-hosting at a national or institutional scale is often framed as Sovereign AI, emphasizing autonomy from foreign technology providers."},{"type":"include_existing","title":"CUDA Moat","reason":"The ability to self-host is currently heavily constrained by the software-hardware ecosystem, where Nvidia's CUDA remains the dominant standard for local execution."},{"type":"include_existing","title":"Inference vs. Training","reason":"Self-hosting is primarily focused on the inference side of the lifecycle; distinguishing this from the much higher requirements of training is essential for local deployment strategy."}],"timestamp":1784832111797},"n-mermaid-mrxl4f8r-105":{"actions":[{"type":"include_existing","title":"Grouped-Query Attention (GQA)","reason":"A key architectural optimization used across model families to reduce memory overhead and increase inference speed."},{"type":"create_connected","title":"Parameter-Efficiency","connectionType":"child","reason":"The underlying metric that determines the success of a model family: how much 'intelligence' can be packed into a specific inference budget."},{"type":"create_connected","title":"Speculative Decoding","connectionType":"child","reason":"The primary technique for using small and large models within a family together to accelerate inference speed."},{"type":"create_connected","title":"Chinchilla Scaling Laws","connectionType":"child","properties":{"equation":"L(N, D) = E + \frac{A}{N^\\alpha} + \frac{B}{D^\beta}"},"reason":"The mathematical framework used to determine the optimal size and data mix for different members of a model family."},{"type":"create_connected","title":"Weight Tied Architectures","connectionType":"child","reason":"A design choice in some model families where different sizes share specific layers or embedding structures to improve efficiency."},{"type":"create_connected","title":"Router-Based Orchestration","connectionType":"child","reason":"The layer of infrastructure that automatically directs a user request to the cheapest/smallest model in a family capable of solving it."},{"type":"create_connected","title":"Command Family","connectionType":"child","reason":"Cohere's enterprise-focused family, specifically tuned for RAG (Retrieval-Augmented Generation) and tool-use."},{"type":"create_connected","title":"Jamba Family","connectionType":"child","reason":"A hybrid family from AI21 Labs combining Transformer and Mamba (SSM) architectures for efficient long-context processing."}],"timestamp":1784832726376},"n-1784-832-210-084":{"actions":[{"type":"include_existing","title":"Anthropic","reason":"As the creator of the Claude family, Anthropic's research philosophy (Constitutional AI) is central to the model's identity."},{"type":"include_existing","title":"Alignment","reason":"The defining characteristic of the Claude family is its specific approach to alignment and safety."},{"type":"include_existing","title":"Frontier Labs","reason":"Places the Claude family in the context of its direct competitors like OpenAI and Google DeepMind."},{"type":"create_connected","title":"Constitutional AI","connectionType":"child","reason":"The specific training methodology used by Anthropic to align the Claude family using a set of rules rather than just human feedback."},{"type":"create_connected","title":"Red Teaming","connectionType":"child","reason":"The adversarial testing process used to identify safety risks and refine the model's refusal boundaries."},{"type":"create_connected","title":"Needle-in-a-Haystack Testing","connectionType":"child","reason":"A benchmark standard used to measure the high-recall accuracy of Claude's long context windows."},{"type":"create_connected","title":"Harmlessness-Helpfulness Pareto Frontier","connectionType":"child","reason":"The design trade-off Anthropic navigates to ensure models are safe without being overly restrictive or 'preachy'."},{"type":"create_connected","title":"Mythos Class","connectionType":"parent","reason":"The new capability tier introduced in 2026 that sits above the Opus class in reasoning depth and agentic autonomy."},{"type":"include_existing","title":"Inference-Compute Trade-off","reason":"Fable 5 represents the extreme high end of this trade-off, prioritizing reasoning depth and agentic planning over the speed of the Haiku or Sonnet tiers."},{"type":"include_existing","title":"Agency & Long-Horizon Planning","reason":"Fable 5 is explicitly optimized for tasks that require planning across multiple stages and days of autonomous operation."}],"timestamp":1784832570975},"n-1784-832-202-073":{"actions":[{"type":"include_existing","title":"Scaling Laws","reason":"GPT-3 proved that performance scales predictably with compute and data, a core tenet of the family's development."},{"type":"include_existing","title":"RLHF","reason":"The process used to turn raw base models into the conversational assistants like ChatGPT."},{"type":"include_existing","title":"Inference-Time Search","reason":"The defining characteristic of the new o1 'reasoning' models compared to previous GPT generations."},{"type":"include_existing","title":"Mixture of Experts","reason":"The architectural approach widely believed to be used in GPT-4 to maintain high performance with relative efficiency."},{"type":"create_connected","title":"GPT-5.6 Series","connectionType":"child","reason":"The latest generation of GPT models (released July 2026), introducing celestial-themed capability tiers."}],"timestamp":1784832711618}},"paths":[],"appState":{"lastSceneId":"scene-mermaid-mrxl4f8r","authoringContainerSize":{"w":1420,"h":776}},"themes":[],"nodeImagePresets":{"presets":[{"name":"Open","content":{"imageType":"unspecified","form":"unspecified","depth":"unspecified","enclosure":"unspecified","textAllowed":"unspecified"},"technical":{"renderMode":"unspecified","strokeWeight":"unspecified","aspect":"square","detailLevel":"unspecified","backdrop":"unspecified","gradientsAllowed":"unspecified","transparencyAllowed":"unspecified"},"colour":{"colourMode":"thematic","paletteSize":"unspecified"},"extraInstructions":"","id":"preset-1787310324304-lx7h2zw","createdAt":"2026-08-21T11:05:24.304Z","updatedAt":"2026-08-21T11:05:24.304Z"},{"name":"Icon","content":{"imageType":"icon","form":"unspecified","depth":"2d","enclosure":"none","textAllowed":"forbidden"},"technical":{"renderMode":"mixed","strokeWeight":"medium","aspect":"square","detailLevel":"very-simple","backdrop":"transparent","gradientsAllowed":"forbidden","transparencyAllowed":"unspecified"},"colour":{"colourMode":"thematic","paletteSize":2},"extraInstructions":"","id":"preset-1787310324304-yvmq1hm","createdAt":"2026-08-21T11:05:24.304Z","updatedAt":"2026-08-21T11:05:24.304Z"},{"name":"Schematic","content":{"imageType":"schematic","form":"unspecified","depth":"2d","enclosure":"none","textAllowed":"allowed"},"technical":{"renderMode":"line-art","strokeWeight":"thin","aspect":"landscape-4-3","detailLevel":"detailed","backdrop":"transparent","gradientsAllowed":"forbidden","transparencyAllowed":"unspecified"},"colour":{"colourMode":"thematic","paletteSize":2},"extraInstructions":"","id":"preset-1787310324304-z997lna","createdAt":"2026-08-21T11:05:24.304Z","updatedAt":"2026-08-21T11:05:24.304Z"},{"name":"Plot","content":{"imageType":"plot","form":"geometric","depth":"2d","enclosure":"none","textAllowed":"allowed"},"technical":{"renderMode":"line-art","strokeWeight":"thin","aspect":"landscape-4-3","detailLevel":"moderate","backdrop":"transparent","gradientsAllowed":"forbidden","transparencyAllowed":"unspecified"},"colour":{"colourMode":"thematic","paletteSize":3},"extraInstructions":"","id":"preset-1787310324304-rfvmtue","createdAt":"2026-08-21T11:05:24.304Z","updatedAt":"2026-08-21T11:05:24.304Z"},{"name":"Emblem","content":{"imageType":"symbol","form":"geometric","depth":"2d","enclosure":"circle","textAllowed":"forbidden"},"technical":{"renderMode":"filled","strokeWeight":"heavy","aspect":"square","detailLevel":"very-simple","backdrop":"contrast-fill","gradientsAllowed":"forbidden","transparencyAllowed":"unspecified"},"colour":{"colourMode":"thematic","paletteSize":2},"extraInstructions":"","id":"preset-1787310324304-9nqo56y","createdAt":"2026-08-21T11:05:24.304Z","updatedAt":"2026-08-21T11:05:24.304Z"}],"defaultPresetId":"preset-1787310324304-lx7h2zw"}}