Skip to content

Repository files navigation

The Atlas — an atlas of language models

From the boundaries of a context window to the autonomous loops of an agent —
the whole territory of a modern LLM, charted as one map.

Explore the live site →

License: MIT React 19 TypeScript 5.7 strict Tailwind v4 Tested with Playwright and axe

The Atlas home chart — a clickable route map of seven plates across two volumes, drawn as a nautical survey

The Atlas is a single-page, nautical-chart-themed explainer of LLM internals. It treats a language model as territory to be surveyed: every concept is a numbered chart plate, every process a route, every budget a sounding taken in tokens. One continuously scrolling page with a fixed station rail, hash deep-links (#/plate-iv), keyboard paging (←/k, →/j), and per-plate entrance animations that fully stand down under prefers-reduced-motion. Dark only — dark is the design.


Honesty — read this first

Every number on this site is illustrative. Token counts, similarity scores, attention weights, and the candidate logits behind the interactive decode controls are fixed design values, not measurements of any real model. What is real is the math: Plate IV·Detail runs an exact softmax with temperature scaling and top-p nucleus truncation over those illustrative logits, live, as you move the sliders. The site never claims to visualize a frontier model's internals — providers do not expose them.


The charts

Plate Title Subject
— Home Clickable route map of the whole territory
I The Boundaries of Memory The context window — a 128K-token transect
II Standing Orders The system prompt — read first, every turn
III Bearings from Afar Retrieval/RAG — nearness is similarity
IV The Inference Passage Prompt → tokenize → embed → attend → predict → sample → output
IV·D The Passage, Sounded One prompt through all seven stations, with live temperature / top-p / decode controls
V The Self-Directed Survey Agents — goal, instruments, and the reason–act–observe loop
VI The Scouting Party Subagents — parallel isolated scouts, condensed findings
VII The Circuit Loops — iterations and stop conditions
App. Gazetteer of Terms 17 definitions, each cross-linked to its plate
00 About the Atlas Colophon and how to read these charts

Plate IV·Detail — the interactive sounding

Move the temperature and top-p sliders or flip sampling/greedy and watch the prediction bars and nucleus cutoff recompute instantly — an exact softmax running in the page.

Plate IV Detail — the prompt "What is a token?" charted across seven stations with live temperature, top-p, and decode controls

Plate VI — the scouting party

A lead agent fans dispatch routes out to three isolated subagents; their condensed findings converge back to a single chart.

Plate VI — a lead-agent card dispatching dashed gold routes to three subagent cards, whose blue return routes converge back


Under the hood

  • No routing library. One scrolling document. The active section is the last one whose top passed the viewport midpoint (a rAF-throttled scroll listener), mirrored to #/{id} via replaceState; deep links land the right plate under the header on a fresh load.
  • A fixed survey sheet. Every screen is a 1280×960 chart sheet; CSS zoom scales the whole stage to the viewport, so plate geometry is authored once in absolute coordinates and never reflows.
  • Real math over illustrative inputs. src/plates/nextToken.ts implements the decode pipeline exactly — softmax with temperature clamped to ≥ 0.05, descending sort, cum < top-p nucleus rule, greedy ⇒ rank 1 — with unit tests proving the distribution.
  • A fixed motion grammar. Each plate enters once — sheet rises, routes dash-draw, stations pop — then gold routes flow and hollow halos pulse quietly. Entrances run on the motion library, ambient loops on the Web Animations API, and all of it is skipped under prefers-reduced-motion with content fully visible.
  • Self-hosted typography. Fraunces, Hanken Grotesk, and Spline Sans Mono ship as latin variable woff2 from public/fonts/ — no third-party font requests.
  • Accessibility as a release gate. Every section is a uniquely labelled landmark; keyboard paging ignores form controls so the Plate IV·D sliders stay usable. axe runs twice — per plate in jsdom (vitest-axe) and across all 11 live sections in Chromium (@axe-core/playwright).
  • Cookieless analytics with a hard ceiling. Umami, capped at ≤ 15 named events (currently 5); click events are declarative data-umami-event attributes, and the site's own track() wrapper no-ops when the script hasn't loaded.

Develop

nvm use            # Node 20
npm install
npm run dev        # localhost:5173
Command What it does
npm run build Typecheck + production build to dist/
npm test Vitest unit + a11y-unit suites
npm run e2e Playwright E2E (Chromium)
npm run a11y / npm run a11y:e2e axe scans — per plate (jsdom) and across all 11 live sections
npm run lint / npm run format ESLint (zero-warning policy) / Prettier
src/
├── app/          # nav context + cross-cutting hooks (scroll-spy, hash sync, keyboard, stage zoom)
├── components/   # chart-sheet frame primitives, header, station rail, section shell
├── plates/       # the eleven screens + plates.config.ts (single source of truth) + decode math
├── motion/       # entrance + ambient animation grammar
└── analytics/    # Umami wrapper, event taxonomy, scene-reach hook

Deployed on Cloudflare Pages from main; every PR gets a preview URL. Per-release notes live in CHANGELOG.md.

License

MIT — see LICENSE.

About

The Atlas is a knowledge base of language models with information on how they work as well as agentic loops and loop engineering!!

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages