Skip to content
The Shape of Intelligence

sources · method · licence

Sourced, all the way down.

Every event here cites a primary source, names what it built on, and is generated from one file that the build validates. This page is that method, the licence, and the way to correct it.

events
169
turning points
23
citations
266
influences
265
roots
21
notable models
537

Editorial cutoff and review

The contemporary chapter is verified to an editorial cutoff of 2026-06-30; the whole exhibition was last reviewed on 2026-09-12. Events after the cutoff are not in the archive until a review adds them. Every chapter claim is either an event in the archive with its own sources or an entry in the public claim ledger with a source, a quoted passage where the claim is a quotation, and a status of verified, disputed or unverified. Chapters are published only when their claims are verified and their artefacts have a rights record.

Method

The Shape of Intelligence carries an archive of 169 events in the history of artificial intelligence, 1943 to 2026. Each event is one Markdown file with validated frontmatter (title, date and its precision, category, significance, summary, people, organisations, tags, influences, sources) and a body written in three beats: what happened, why it matters, what it led to.

Events were chosen for consequence, not fame. An event is in the archive when something downstream would not have happened the same way without it, and when a primary source can be cited for it. Categories are theory, hardware, model, data, product, policy and culture.

Significance runs 1 to 5. Five is a turning point that changed the direction of the field (23 of them). Four is a major result or release. Three is notable. Two is supporting. One is context that a later event needs. Significance drives the default order of the timeline.

Eras are derived from dates, not assigned by hand: prologue (1943), foundations (1944–1972), the first winter (1973–1980), connection (1981–1987), the second winter (1988–1993), statistics and data (1994–2011), deep learning (2012–2016), transformers (2017–2021), everyone (2022–2024) and agents (2025–2026). The winters deliberately hold the quiet work done in the cold. Each era belongs to one of the exhibition's eight chapters.

The influence graph

Every event names the earlier events it built on directly. Those 265 links form a directed acyclic graph with 21 deliberate roots. The build refuses a cycle, an unknown id, a parent dated after its child, or an orphan that has not been marked as a root on purpose.

A link is editorial: it means direct intellectual or technical descent that the sources support, not mere chronology. Ancestry and descent shown on event pages are the transitive closure of those links, and the exhibition lights the same path when it can.

Primary sources

Every event cites at least one source with a URL, typed as paper, announcement, article, book, dataset, video or archive; 266 citations in all, and every turning point cites at least two. Papers link to the canonical record (arXiv, a journal, a proceedings page); announcements link to the organisation that made them; archives link to the institution holding the document.

Links are checked by script before a build. A source that has moved is replaced with an archived copy rather than dropped.

Dataset and licence

The dataset is generated at build time from the same files that render this site, so it cannot drift from what you read. It is released under Creative Commons Attribution 4.0 (CC BY 4.0). Use it, remix it, train on it; credit "The Shape of Intelligence by Jamie McKaye, https://shapeofintelligence.com".

Notable models

The scale chapter uses a second dataset: Epoch AI's "Data on AI Models" (notable models), CC BY 4.0, trimmed to date, name, organisation, parameters and training compute. The snapshot holds 537 models and was fetched on 2026-09-11.

Machine surfaces

Everything here is readable without a browser. Each content page has a text/markdown mirror under /md/, there is an llms.txt with a link per event and an llms-full.txt with every event inline, JSON-LD describes the site, the dataset and each article, and an MCP server exposes the archive and the chapters as tools: list_events, get_event, lineage, search_events, list_eras, list_chapters and get_chapter.

Corrections

Dates, attributions and lineage are the kind of thing that is wrong in small ways. If you find one, send the correction with a primary source and it will be fixed in the file; the build re-validates the whole archive on every change, and the dataset, the mirrors and the exhibition update together.