What it means when AI compresses years of expert work into 11 days
Claude finished a mathematical proof in 11 days that experts estimated would take several years — and the lesson for knowledge work is bigger than the math.
- 11 days
- to complete what experts expected to take years
- 13 million
- lines of computer-verified Lean code written
- 29,500
- intermediate theorems proved along the way
- 350 years
- since Fermat wrote the problem down in 1637
Last week, Anthropic's AI model Claude completed one of the hardest jobs in mathematics. The target was a proof called Fermat's Last Theorem — a problem that sat open for 350 years after Pierre de Fermat wrote it in the margin of a book in 1637. A human mathematician, Andrew Wiles, finally solved it in 1995, in 129 dense pages that took years to finish. Claude finished its version in 11 days.
To be precise, Claude wasn't solving the theorem from scratch — Wiles already did that. The task was to "formalize" the proof: translate the human-written math into computer code that a machine checks step by step, with no room for error. Experts had estimated this would take a team of specialists several years. Claude ran dozens of agents in parallel, all coordinating automatically, and produced 13 million lines of verified code — five times bigger than any math library before.
Along the way, the agents proved 29,500 smaller theorems and consumed about 6 billion tokens — the small chunks of text an AI processes. A separate follow-up experiment then formalized another famous theorem in just three days, using consumer plans anyone can buy today.
The real story here isn't mathematics — it's time. If AI can collapse a multi-year expert project into 11 days, the same pattern could apply to legal document review, scientific literature checks, compliance analysis, or code auditing. Any high-stakes, expert-heavy work that takes time to verify is now a candidate for this: months of expert work compressed into days.
One honest note: this worked because the right infrastructure was already in place — community-built math libraries, an open-source coordination platform, and humans setting direction at key moments. AI didn't replace the experts; it worked fast alongside them. The practical question for any team: where does slow, expert-heavy verification hold everything else back?