About

A record of what artificial intelligence has actually achieved, when, and on whose evidence. Every entry is sourced, every date is only as precise as the record allows, and every correction is made in public.

Who compiles it

This is a project of ThinkML, a publication about machine learning, and is edited by the ThinkML research team. It began in 2020 as a personal record and was rebuilt from the ground up in 2026.

How an entry gets here

A milestone is included if it meets at least one of these, and can be dated to at least month precision from a primary source:

• A first of its kind, in a specific and defensible sense
• A capability threshold: a machine does something it demonstrably could not before
• A method that later work was built on, judged by adoption rather than novelty
• An institutional or legal event that shaped the field's direction
• A deployment at a scale that changed public or industrial exposure to AI

Funding rounds, product refreshes, incremental benchmark gains and announcements without a released artefact are not milestones, however widely they were covered at the time.

How claims are checked

• Two independent sources minimum, at least one primary, official or institutional
• Dates recorded only to the precision the evidence supports. 115 of 203 entries carry a year and no month, because that is all the record establishes
• Where sources disagree, the entry says so rather than choosing
• First-of-its-kind claims are challenged rather than repeated. 36 hold only under a stated definition, and 15 were refuted outright with an earlier example named
• Source links are archived, and checked quarterly for rot

Across 203 entries there are 528 citations, and 203 entries carry at least one primary, official or institutional source.

How the writing is done

Research and drafting are assisted by language models, working only from the sources attached to each entry rather than from recall. Every entry is then edited by the ThinkML team, and the site refuses to build one that lacks two sources, that claims a date precision its evidence does not support, or that a model has marked as verified on its own authority.

Saying so is deliberate. Concealing it would be the wrong call quite apart from honesty: a reference work asks to be trusted, and the strongest case for trusting one is a method stated plainly enough to be argued with.

Each entry shows how many sources it rests on and when it was last checked, at the foot of the page. Where sources disagree, the entry says so instead of choosing.

Corrections

Rebuilding this site found 976 factual errors in its previous version. Others will remain. Corrections are checked against primary sources, applied openly with a dated note on the entry, and credited if you would like. Report one here.

Reuse

The structured data is available as JSON under CC BY-NC 4.0, so it can be used freely for research and teaching with attribution. The written text of each entry is not licensed for reuse. For commercial licensing, get in touch through the contact form.