Every term, timeline station, API feature, model and event in the apps below on one map. An open-weight embedding model read each one, and things that read alike sit together, whichever app they come from.
. Each item’s title and text went through Qwen3-Embedding-8B, an open-weight model run on a Mac, and UMAP flattened the result to two dimensions; the topics are regions of the map found by k-means, each named by a local model (Qwen 3.6) from the titles nearest its middle, and the ringed dots are the dictionary’s own terms. Distances are rough: two dots close together read alike, but the gap between far-apart clusters means little. Scroll to zoom, drag to pan, click a dot to open it. Built by ai/scripts/build_atlas.py.
The whole picture
Everything since ChatGPT on one clock, the question it keeps raising, and where the labs say the road leads.
Five timelines on one clock, drawn as a transit map: what AI achieved, what the models learned, the harness built around them, the research behind both, and the compute it all runs on. Breakthroughs mark where an improvement on one line met something on another that was ready to use it, and two panels underneath follow METR’s time horizons and the Artificial Analysis Intelligence Index of the best model and the best open-weight one.
Eleven published definitions of AGI, from lab charters to the coffee test, each held up against independent measurements of the frontier: which are met, which are partly met, and why reliability is what divides them. A change log records every review, including the ones that move nothing.
From each definition’s source, and the published measurements: ARC Prize, METR, OSWorld, FrontierMath · sources
→
Superintelligence · 1965 to the White House Accord
Sixty years of a word, on four lanes: the ideas, the labs that made it their goal, the rules, and what the machines did. It marks 29 September 2026, when six AI company heads signed the White House Accord on Super Intelligence. The page reads the accord’s four layers closely, redraws who sat at the table, sets it beside the 2023 and Seoul pledges, and lists the deadlines still ahead. A scoreboard built from Frontier’s data shows which lab leads on the Intelligence Index, and every time the lead changed hands.
From the accord, the executive orders, each lab’s own posts and the reporting on the lunch · sources
The words every page here leans on, from token and attention to harness, MCP and superintelligence, in plain English. Read it in order as a short course or search one term at a time; hover any linked term for its one-line meaning, and follow a graph of how each term connects to the rest. Each entry quotes my blog where I’ve explained it and points to where it shows up in these apps.
From the definitions in these apps, my posts and the paper or post that introduced each term · start at token
A running weekly paper for these apps: each issue tells what happened in AI that week, from new models and agents to mathematics and policy, and what it changed in the apps, with links into each one. Back issues stay up, corrected in place but never rewritten.
From the apps’ own entries, so every figure is one they already cite · sources
→
The models
How far the open-weight models trail the closed ones, what intelligence costs, what it takes to run them yourself, which ones people actually download, and what’s inside them.
Open weights against proprietary · release by release
The best open-weight model against the best proprietary one on the Artificial Analysis Intelligence Index, at every release that moved either line, and how far behind the open side runs, in points and in months. Beside it: what one task costs, which models fit on your Mac, how to set them up, and boards for text to speech and speech to text.
Hold a bar still, the score GPT-4, o1 or Claude Opus 4.8 earned, and watch the cheapest model to clear it step down, release by release: GPT-4’s bar went from $37.50 per million tokens to about four cents. Against it, what the best model costs, which barely moves, and what one task costs once reasoning tokens are counted.
From Artificial Analysis’s prices, GPQA Diamond and Intelligence Index scores, re-read model by model · sources
Every model Docker publishes for Docker Model Runner, from Gemma and Qwen to GLM, Kimi and the embedding models, with how often each is pulled, how big it is, which formats it ships in, and how well it scores. The pulls go to small models a laptop can run; the best-scoring ones are hundreds of gigabytes and barely downloaded.
From the Docker Hub API, each repo’s README, the Artificial Analysis Intelligence Index for chat models and the MTEB leaderboard for embedding models · sources
The most-downloaded open models on Hugging Face, from chat and embeddings to speech, image generation and decision models, plus the benchmark leaders fewer people download, with every GGUF, MLX and other quantized re-upload counted once under the model it copies. Beside each: file sizes, likes, license and the score its kind is measured on, matched by exact repository.
From the Hugging Face API and its base-model tags, the Artificial Analysis Intelligence Index and Speech Arena, the MTEB leaderboard, the Open ASR Leaderboard and the community Decision Index · sources
The open-weight flagships from LLaMA on, read from the config file each release ships with: the attention each uses (GQA, MLA, sliding windows, linear attention), mixture of experts, multi-token prediction, precision and context length, and which release brought each idea to open weights first.
From each release’s config.json on Hugging Face · sources
→
The platforms
An agent is a model calling tools in a loop. Each lab’s API has taken over more of that loop from your code: three timelines follow how, and a fourth puts the labs side by side.
The OpenAI API from Completions to the Agents API, feature by feature, with request shapes and the models released alongside. A table follows which parts of an agent your code ran and which OpenAI ran, from the conversation history and the tool loop to the computer the agent works on.
Built three times from one prompt: by Claude Opus 5.5 in Claude Code, and by GPT-6.1 Sol and GPT-6 Astra in Codex. The switch in the corner moves between them.
From OpenAI’s announcements, API changelog and docs · sources
Anthropic’s platform in three columns: the Claude API on one side, the Claude Code harness release by release on the other, and the models between them. Tables follow which parts of an agent your code ran and which Anthropic ran, and how Claude Code took on the customization points of the agent frameworks before it.
Also built three times from one prompt, by the same three models.
From Anthropic’s announcements and docs, and the Claude Code changelog · sources
→
Gemini, Google Cloud & Antigravity · 2023 to today
Google builds every layer under the agent loop itself, from TPU chips to the Gemini models. Its APIs run down one side and its harnesses (ADK, Gemini CLI, Jules, Antigravity) down the other, with which layers it opened, from Gemma’s weights to an open-source CLI, and which it kept.
From Google’s blogs, release notes and GitHub repositories · sources
The same capabilities across all three platforms, from a chat API and function calling to hosted agents: the day each lab first shipped each one, who led, and how long the others took to follow.
From the three timelines above, card by card, so the dates cannot drift · sources
→
The harnesses
The loops built on top of the models: what one does with a single prompt, year by year; the coding agents, which all converged on the same features; and the personal agents that carried them into your messages.
One request, “fix the failing test”, animated through the agent harness of five years: GPT-3 through an API, ChatGPT and its code sandbox, the editor agent that calls tools in a loop, the terminal agent with its instructions file, subagents and compaction, and the cloud agent with teammates and a reviewer model. Pause at any step to see what fills the context window and who approves each action.
Claude Code, Codex, Gemini CLI, Cursor, GitHub Copilot, Cline, OpenHands, OpenCode, Aider and Pi against the features that define a coding agent (instruction files, MCP, plan mode, subagents, hooks, skills, sandboxes, cloud agents): the day each shipped each one, who had it first, and how widely it was copied.
From each harness’s changelogs, releases and package registries · sources
Agents that work for one person, stay on after the chat window closes, remember you and answer by text: from OpenClaw, a weekend project that wired a coding agent into WhatsApp, to every large lab’s own. Each launch, rename, feature, incident and price sits on one clock, in three bands: open source on your machine, proprietary in their cloud, or built yourself from a coding agent and a repo.
From GitHub releases, the projects’ and companies’ own announcements and release notes, and press where no primary page exists · sources
→
The tests
Benchmarks built from data these apps already hold, each with its dataset to download: what models know about who believes what, and how well search models find a sermon from its passage. The quiz on AI’s own history has moved to the Evals page.
Every stance in the doctrine apps cites something the theologian published, so each one becomes a test question: does a model know where Piper, Ladd or N.T. Wright stands on the millennium, election or justification? Claude models, Qwen and Llama scored, with where each one goes wrong.
From the cited stances in the shared theology database, with every model reply kept · sources
Give a search model a Bible passage and ask for the sermons preached on it, with titles and spoken references hidden so it can’t just match strings. BM25 against five embedding models on a thousand of Spurgeon’s sermons and on transcribed church sermons.
From the Spurgeon Archive’s index of texts and the sermon archive’s titles · sources