We use analytics cookies to understand how readers use the site. They only run if you accept. Privacy policy.
Every notable AI system whose training compute has been measured, from Claude Shannon’s maze-solving mouse in 1950 to the frontier of last year. The vertical axis is logarithmic: each gridline is a hundred thousand times the one below it. It has to be, because the record runs from 40 floating-point operations to about ten to the twenty-sixth, and on an ordinary axis everything before roughly 2015 would sit flat on the floor — which would read as “AI began in 2015”, the one thing this chart disproves.
Epoch AI’s notable-models dataset. “Notable” is a selection against stated criteria, not a census: systems never published, never benchmarked, or built in secret are absent by construction. This is the frontier of the public record, which is a narrower claim than the frontier. Training compute is also an input — what was spent, not what was obtained.
The frontier’s doubling time, taken from its own endpoints in each span rather than fitted. The second row is the point of the table: between 1960 and 1987 the frontier of the public record advanced 1.6 orders of magnitude — a doubling every five years, against every five months in the decade before it. Funding collapsed twice in that stretch.
| Era | Orders gained | Years | Doubling time |
|---|---|---|---|
| 1950–1960 | 7.3 | 10 | 5 months |
| 1960–1987 · The winter | 1.6 | 27 | 61.2 months |
| 1987–2010 | 7.7 | 23 | 9.4 months |
| 2010–2025 | 8.5 | 15 | 7.6 months |
The usual lesson drawn from the winter is that the field was wrong. The lesson the record supports is that it was early — the techniques of the 1980s revival were mostly published during the winter and could not be run at the scale that makes them work.