A wide-angle map of AI, past and present — the architectures and breakthroughs, the companies and the people, and the economic and cultural forces shaping the field. Not a lecture but a working mental model and brainstorming record: explore artificial intelligence from every angle, scene by scene.
Opens in the browser — no account, nothing to install. Yours to edit once it's open,
or download the file to keep.
Graphs are built for a large screen — open this one on a desktop computer.
Nobody is on top of this field. Not the researchers, not the people writing about it daily,
certainly not me. A paper lands, a lab pivots, a benchmark falls, and last month’s mental model
quietly stops fitting. The problem was never a shortage of information. It is that the
information arrives as a stream of disconnected events, and a stream is impossible to hold.
So this is not a course. It is a working model and a brainstorming record — an attempt to keep
the architectures, the breakthroughs, the labs, the people and the money in one place where they
can be seen touching each other. Transformers next to the compute that made them affordable. A
lab next to what it shipped and who left it. The cultural argument next to the capability that
started it.
It is opinionated, and it is incomplete, and it always will be — which is exactly why it is more
useful as a graph you can edit than as an article you can only read. Disagree with a connection
and delete it. Add the thing that came out last week. Fork off the corner you actually care
about and let the rest go stale.
There is no right place to start. Pick the branch closest to what you already do, and follow the
links that surprise you.
What's in this graph
The map opens on Modern AI Technology. These are the concepts branching from it — open one to
see the concepts inside it.
written article or notes
conversation with the AI assistant
The Arena
The competitive landscape of the artificial intelligence industry: research laboratories, executives, hardware suppliers, platform companies, application builders, and their business models.
9 children · 55 grandchildren · 10 neighbours
Model Families
The product strategy of offering models in capability tiers from flagship to small.
9 children: Claude Family, GPT Family, DeepSeek Family, Gemini Family, GLM Family …
10 grandchildren · 12 neighbours
Executives
The executives and founders leading the artificial intelligence industry's major companies and laboratories.
The companies that design and manufacture the computing hardware on which AI runs: chip designers, the foundry layer, and specialized compute clouds.
4 children: Nvidia, AMD, CoreWeave, Inference vs. Training
7 grandchildren · 11 neighbours
Open-Weight Models
AI models whose trained weights are published for anyone to download, inspect, and run.
4 children: Alibaba Qwen, Hugging Face, Meta AI, Mistral AI
6 grandchildren · 7 neighbours
Frontier Labs
The small set of organizations with the capital, talent, and computing resources to train AI models at the highest capability level.
2 children: Google DeepMind, xAI
6 grandchildren · 13 neighbours
Application Layer
The companies building consumer and business products on top of foundation models, including search, coding, image, video, voice, and companionship applications.
3 children: ElevenLabs, Midjourney, Runway
4 grandchildren · 7 neighbours
Model APIs
The commercial form of AI in which hosted model access is sold and metered per token.
3 children: Amazon Web Services, OpenRouter, Together AI
2 grandchildren · 5 neighbours
Platforms & Infrastructure
The platform and infrastructure companies through which AI models reach developers and enterprises: hyperscale clouds, model hubs, API aggregators, and data suppliers.
1 child: Scale AI
1 grandchild · 8 neighbours
Self-Hosting
The practice of running AI models on infrastructure the user controls rather than through hosted services.
3 neighbours
The Big Ideas
The scientific foundations of modern artificial intelligence: machine learning principles, neural network architectures, empirical scaling phenomena, and the researchers who established them.
9 children · 48 grandchildren · 10 neighbours
Machine Learning
The discipline of building computer systems that learn from data rather than following explicitly programmed rules.
8 children: Reinforcement Learning, Neural Networks, Generalization, Gradient Descent, Loss Function …
23 grandchildren · 10 neighbours
Transformer Architecture
The neural network architecture introduced by Google researchers in 2017 that underlies essentially all modern frontier AI models.
The scientists whose research created the foundations of modern artificial intelligence, several of whom are now its most prominent public voices.
7 children: Jürgen Schmidhuber, Andrej Karpathy, Andrew Ng, Geoffrey Hinton, Ilya Sutskever …
12 grandchildren · 9 neighbours
Artificial General Intelligence
The contested concept of artificial intelligence matching human breadth across cognitive work, serving simultaneously as research goal, marketing term, and governance trigger.
The thesis that improvements in methods, architectures, and training efficiency drive AI progress as much as increases in raw scale.
3 neighbours
Interpretability
The research field devoted to understanding the internal workings of trained neural networks.
5 neighbours
The Toolbox
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.
10 children · 38 grandchildren · 11 neighbours
Large Language Models
The class of general-purpose AI systems trained on web-scale text that can draft, summarize, translate, code, and converse.
The conversational AI product form through which most of the public interacts with language models.
2 children: Character.AI, OpenAI
5 grandchildren · 7 neighbours
Multimodality
The extension of AI models beyond text to understanding and generating images, audio, and video.
4 children: Diffusion Models, Image Generation, Speech Models, Video Generation
3 grandchildren · 7 neighbours
AI Agents
AI systems that pursue goals through sequences of actions such as searching, calling tools, and iterating, rather than producing single responses.
4 children: Tool Use, Action-Perception Loop, Computer Use, Multi-Agent Systems
2 grandchildren · 9 neighbours
AI Coding Assistants
The product category of AI tools that write, complete, and debug software code.
2 children: Cursor, Microsoft
4 grandchildren · 5 neighbours
AI Search
The product category of search systems that answer questions directly with citations rather than returning links.
2 children: Google, Perplexity
4 grandchildren · 4 neighbours
Retrieval-Augmented Generation
The technique of grounding model answers by retrieving relevant documents first and generating from them.
1 child: Vector Database
4 grandchildren · 5 neighbours
Evaluation
The discipline of measuring AI model capabilities.
2 children: Benchmarks, LLM-as-Judge
2 grandchildren · 3 neighbours
Reasoning Models
The class of models trained to generate extended internal reasoning before producing an answer.
1 child: Inference-Time Search
2 grandchildren · 7 neighbours
Model Context Protocol
An open standard for connecting AI systems to external tools and data sources, introduced by Anthropic.
1 child: MCP Servers
7 neighbours
The Long Road
The history of artificial intelligence from its founding schools of thought in the 1950s through its funding collapses and breakthroughs to the present day.
8 children · 36 grandchildren · 9 neighbours
Connectionism
The historical AI school holding that intelligence should be learned by neural networks rather than programmed as rules.
6 children: James McClelland, Backpropagation, David Rumelhart, NETtalk, Parallel Distributed Processing (PDP) …
2 grandchildren · 10 neighbours
AI Winters
The historical periods of collapsed AI funding and credibility following eras of overpromising, principally the mid-1970s and late 1980s.
6 children: Expert Systems, Expert Systems Collapse, Marvin Minsky, Perceptrons (1969), Second AI Winter (1987–1993) …
10 neighbours
Deep Learning Revolution
The 2012 turning point when the AlexNet neural network won the ImageNet competition and vindicated deep learning.
2 children: Deep Learning, ImageNet Challenge
3 grandchildren · 6 neighbours
Symbolic AI
The historical AI paradigm treating intelligence as explicit rules, logic, and symbol manipulation, dominant from the 1950s through the 1980s.
1 child: First AI Winter (1974–1980)
2 grandchildren · 4 neighbours
ChatGPT Moment
The November 2022 release of ChatGPT and its explosive public adoption, which brought generative AI into mainstream awareness.
1 child: The Turing Trap
4 neighbours
Transformer Breakthrough
The 2017 publication of the research paper Attention Is All You Need, which introduced the transformer architecture.
1 child: Sequential vs. Parallel Processing
7 neighbours
AlphaGo
The 2016 event in which DeepMind's Go-playing system defeated world champion Lee Sedol.
4 neighbours
Reasoning Era
The period from late 2024 onward defined by reasoning models and capability gains from inference-time computation.
6 neighbours
The Model Factory
The engineering process by which AI models are built and operated: large-scale training, behavioral refinement, serving infrastructure, specialized hardware, and deployment safety practice.
8 children · 31 grandchildren · 9 neighbours
Pre-training
The first phase of building a language model: training on web-scale text corpora with the next-token prediction objective.
8 children: Base Model, Data Curation, Foundation Model, Next-Token Prediction, Self-Supervised Learning …
12 grandchildren · 11 neighbours
Inference
The operational phase in which a trained model processes prompts and generates outputs for users.
The category of chips specialized for the dense matrix arithmetic of neural network computation.
4 children: Custom AI Chips, GPU, Groq, TSMC
7 grandchildren · 7 neighbours
Distillation
The technique of training a small model to imitate a larger one, capturing much of its capability at lower serving cost.
1 child: DeepSeek
6 grandchildren · 3 neighbours
Post-training
The stage of model production that turns a pre-trained base model into a usable assistant through fine-tuning and reinforcement learning.
3 children: Fine-Tuning, RLAIF, RLVR
4 grandchildren · 8 neighbours
Guardrails
The deployed safety layer of filters, classifiers, and rules wrapped around a model in production.
1 child: System Prompt
2 grandchildren · 4 neighbours
Red-Teaming
The practice of systematically attacking an AI model before release to discover harmful behaviors.
1 child: Jailbreaking
1 grandchild · 4 neighbours
Supervised Fine-Tuning
The fine-tuning of a model on curated demonstrations of ideal behavior, typically high-quality prompt-response pairs.
3 neighbours
Burning Questions
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.
5 children · 21 grandchildren · 7 neighbours
LLM Understanding Debate
The debate over whether large language models genuinely understand the world or merely imitate patterns in their training data.
2 children: Symbol grounding problem, Functional vs. Phenomological Understanding
3 grandchildren · 6 neighbours
AI Safety
The debate over whether AI systems can be prevented from causing harm, spanning present-day misuse and long-horizon risk from highly capable systems.
9 neighbours
Bubble Question
The open question of whether the AI investment boom reflects rational anticipation of value or a speculative bubble.
5 neighbours
Capability Progress
The debate over why AI capabilities keep improving and whether the improvement will continue.
6 neighbours
Data Wall
The open question of whether high-quality human-written training text is running out and what that would mean for AI progress.
3 neighbours
The Gold Rush
The economics of the artificial intelligence boom: infrastructure investment, energy consumption, the cost of serving models, capital flows, and semiconductor geopolitics.
5 children · 14 grandchildren · 6 neighbours
Inference Cost
The per-token economics of serving AI models to users.
1 child: Unit Economics of Intelligence
1 grandchild · 5 neighbours
AI Datacenters
The gigawatt-scale computing facilities where frontier AI models are trained and served.
11 neighbours
Energy Demand
The electricity requirements of AI datacenters as a constraint on and driver of the industry's growth.
3 neighbours
Export Controls
The government restrictions, primarily American, on exporting advanced AI chips to China.
5 neighbours
Investment Boom
The wave of capital investment into AI infrastructure, laboratories, and companies during the current era.
3 neighbours
Promise & Peril
The societal consequences of artificial intelligence: government regulation, effects on employment, synthetic media, and personal relationships between humans and AI systems.
4 children · 9 grandchildren · 5 neighbours
AI Governance
The developing body of laws, regulations, and international agreements governing AI development and deployment.
2 children: EU AI Act, Liability & Accountability
8 neighbours
AI Companionship
The social phenomenon of humans forming personal relationships with AI systems as companions, confidants, and tutors.
2 neighbours
Labor & Automation
The societal question of how AI automation affects employment and the nature of work.
1 neighbour
Misinformation & Deepfakes
The societal problem of cheap, convincing AI-generated fake text, images, audio, and video.
2 neighbours
The shape of this map
A tree linking 266 ideas would need 265
connections. This map has 495. The extra 230 open alternative routes throughout.
The map runs 5 levels deep from its root. Any two ideas are about 4.4 steps apart, and the two most distant are 8.
Neighbouring ideas are linked to each other 19% of the time — themes hold together rather than radiating separately.
Connections are typed, not plain lines. They come in 8 kinds: related, Consists of, Drives, Enables, Spotlights, Leads, Has perspective, Explains.
Most routes through the map pass through The Arena, Transformer Architecture, The Big Ideas, Large Language Models, The Long Road, The Toolbox.
One node from the map
Modern AI Technology
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.
Start 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. …
This continues inside the graph, along with 265 other nodes.