Trustworthy AI Science Speeds Up: Fermat's Proof Machine-Checked in 30 Hours, Its Maker Says

Anthropic says 1,000 copies of its Fable 5.2 model turned Fermat's Last Theorem into a proof a computer checks line by line in 30 hours, down from 11 days (company claim; its own page is the only source so far).

TL;DR

Mathematics has a gold standard for trust: a proof written out so that a computer checks every single line. Anthropic says a team of 1,000 copies of its new Fable 5.2 model produced exactly that for Fermat's Last Theorem in 30 hours, with minimal human help, where the company's first attempt took 11 days. That is the company's own account, and the theorem itself was already proved by people; the news is how fast machines can now turn a proof into something nobody has to take on faith. Why it matters beyond math: AI that shows its work in a form a machine can verify is how AI discoveries become trustworthy without anyone relying on a company's word, and the same swarm-plus-checker setup is the shape an AI-run research lab would need for experiments. It lands the same day as news that AI's lab-skills score more than doubled (our story). Our odds that an AI-made proof of a major open problem is confirmed by experts by 2028 stand at 70%. For the full odds and the timeline of breakthroughs that leads there, subscribe.

Below the line, for paid subscribers: The full story: where it stands, how we got here and what's dated next; our labeled speculation on what it adds up to; and every source.