AI Atlas 3D

An immersive, interactive 3D website that explains how AI works: from a single neuron to trillion-parameter models, from one token to a fleet of agents.

Nothing on the site is a video. Every model runs live in the browser: a real 1M-parameter GPT predicts tokens, neural networks train while you watch, a CNN reads the digit you draw, particles are denoised into shapes, agents learn a maze, and the performance comparisons are measured on your own machine.

What is inside

Chapter	Route	What you do there
Universe	#/	The hub: fly to any chapter.
Families	#/families	Every kind of AI on one map, as a family tree, a timeline from 1943 or by learning signal.
Labs	#/labs	Seven working models you can take apart (below).
Problems	#/problems	Eight levels of the AI stack, the problems at each level, the competing solutions, and a "requirement lens" that re-ranks them as your priorities change.
Data Flow	#/flow	Follow data through ten pipelines (LLM request, training, RAG, tool calls over MCP, diffusion, vision, speech, serving, agent-to-agent, federated learning): its shape, an example payload and the wire format at every stage.
Performance	#/perf	Quantization, the KV cache, batching, a model bake-off and a hall of sourced records.
Ecosystem	#/ecosystem	A galaxy of the projects that run AI: what each does, its core algorithm, its major process and where it deploys.
Deploy	#/deploy	Ten deployment modes as animated topologies, from a hosted API to a browser tab.
Hot 10	#/hot	The ten hottest issues in AI as of November 2026, each with its own 3D animation and dated sources.

The labs

Lab	Route	The real thing underneath
Inside a GPT	#/labs/gpt	A GPT with 1,071,872 parameters trained on TinyStories: BPE tokens, embeddings, attention maps, the logit lens, sampling, generation.
Neural Network Lab	#/labs/neural	A multilayer perceptron and backpropagation written from scratch, training on toy datasets.
Optimizer Arena	#/labs/optim	SGD, Momentum, Nesterov, RMSProp and Adam racing on five loss landscapes.
How a CNN Sees	#/labs/cnn	A convolutional network trained on MNIST; draw a digit and watch every feature map.
Diffusion Lab	#/labs/diffusion	DDPM, DDIM and flow-matching samplers with the exact denoiser, plus classifier-free guidance.
Plan vs Learn	#/labs/world	BFS, Dijkstra, greedy and A* against Q-learning, SARSA and Dyna-Q in one maze.
Embedding Space	#/labs/embed	Real sentence embeddings, BM25 / dense / hybrid retrieval, and a race between vector indexes.

Run it

npm ci
npm run dev        # http://127.0.0.1:5173
npm run build      # static site in dist/
npm run preview    # serve dist/ on http://127.0.0.1:4173
npm run test:ml    # check the JavaScript models against PyTorch reference values

Requirements: Node 22 or newer and a browser with WebGL 2. There is no backend and no run-time network access: the site is plain static files and works from any path.

Publish to GitHub Pages

.github/workflows/pages.yml builds the site on every push and deploys it from the default branch.

One-time setup: in the repository open Settings → Pages → Build and deployment and set Source to GitHub Actions. Then push (or run the workflow from the Actions tab). The site appears at https://<owner>.github.io/<repository>/.

The workflow deploys from the default branch and from main. GitHub also has to allow the branch in Settings → Environments → github-pages → Deployment branches and tags; a Deploy job that fails within seconds, before running any step, was rejected by that rule.

How it is built

One WebGL canvas, one camera and one render loop; chapters are lazy-loaded modules mounted into the scene one at a time; a glass HUD made of plain DOM floats on top. three.js is the only run-time dependency. The algorithms live in src/ml as plain JavaScript with no rendering code, so they can be tested in Node.

src/core      engine: renderer, camera, picking, HUD, charts, 3D helpers
src/chapters  the top-level chapters          src/labs   the mechanism labs
src/perf      the performance experiments     src/ml     the algorithms themselves
src/data      sourced content                 public/    model weights and data
tools/train   PyTorch scripts that produce the small models

See docs/ARCHITECTURE.md for the chapter contract, the helper APIs and the visual rules, and tools/train/README.md to retrain the models.

About the numbers

• Measured values (timings, accuracies, perplexities, recall) are computed live by the code in src/ml on the visitor's device.
• Quoted values (prices, benchmark scores, model sizes, the hot-issue facts) live in src/data with a source URL and a date. The landscape snapshot was taken on 12 November 2026.
• Ratings on a 1 to 5 scale are editorial judgments and are labeled as such.

Credits

Built with three.js and Vite. Typefaces: Inter, Space Grotesk and JetBrains Mono (SIL Open Font License). The tiny GPT is trained on TinyStories, the digit classifier on MNIST, and the embeddings come from all-MiniLM-L6-v2.