About this record
AI So Far is a chronological record of artificial intelligence from January 2020 to the present. It exists because the period moved fast enough that most people, including people working in it, only ever saw fragments.
What an entry is
Every entry is a short brief — between 100 and 500 words — describing what happened, what was claimed, what was disputed, and why it mattered. Briefs are written to be neutral: they attribute claims rather than assert them, name disagreements as disagreements, and avoid adjectives that do argumentative work. Where a figure is contested, both the figure and the contest are stated.
Every entry links to its sources, marked primary where the link is to the thing itself — the paper, the announcement, the statute, the filing — rather than to coverage of it. There are currently 1398 entries carrying 2373 sources; 1398 of 1398 have at least one.
Significance
Each entry is rated 1–5. The rating drives the detail control on the timeline, so it is applied strictly — inflating it would make the top level useless.
- 5 · Era-defining
- A retrospective of the whole period is wrong without it. Changed the trajectory of the field or the public's understanding of it.
- 4 · Major
- Front-page tech news at the time; still cited a year later; shifted what labs or governments did next.
- 3 · Notable
- Belongs in any serious retrospective of its quarter. A real development, not just an announcement.
- 2 · Minor
- Interesting and worth recording, but a well-informed reader could have missed it without losing the plot.
- 1 · Colour
- Texture, trivia, and the odd moments that make the period feel like a period.
Tracks
Entries carry one to three tracks. They are deliberately broad; the point is to let you follow one kind of development through time, not to classify precisely.
- Models & capabilities — Model releases, capability jumps, new modalities, product launches that matter because of what the model can do.
- Benchmarks & progress — Evaluation results, benchmark releases and saturation, competition wins, measured-progress reports, scaling and forecasting analyses.
- Labs & people — Frontier-lab organisational news: founding, restructuring, leadership, departures, strategy shifts, notable hires and splits.
- Safety & alignment — Alignment research, interpretability, safety frameworks and policies, model behaviour and welfare, evaluations for dangerous capability.
- Security & misuse — Jailbreaks, attacks, cyber-offence capability, state-actor misuse, agentic-harm incidents, model theft and infrastructure compromise.
- Government & policy — Legislation, regulation, executive orders, export controls, safety institutes, international summits and treaties.
- Courts & copyright — Litigation, copyright and IP disputes, settlements, rulings, regulatory enforcement actions.
- Money & business — Funding rounds, valuations, revenue milestones, acquisitions, corporate restructurings, market moves, the talent market.
- Compute & infrastructure — Chips, accelerators, datacentres, energy, supply chain, cluster buildouts, inference cost curves.
- Ideas & essays — Landmark papers, influential essays, open letters, public statements, arguments that changed how the field thinks.
- Open weights & ecosystem — Open-weight releases, licences, the open-vs-closed argument, tooling and ecosystem infrastructure, community projects.
- Culture & impact — Public reaction, notable failures and embarrassments, deepfakes, labour and jobs, education, art, harms to individuals, discourse.
What it isn't
It is not comprehensive, and it does not try to be. Thousands of models were released in this period; the ones here are the ones that changed something. It is also not neutral about inclusion — deciding what counts as a development is an editorial judgement, and a different editor would have made a different list.
Corrections are welcome and dates are the most likely thing to be wrong. Where an entry's date is uncertain, it is shown to the month or quarter rather than given a false precision.