There you have it! Anthropic's Claude just took Andrew Wiles' 1995 proof of Fermat's Last Theorem…
September 4, 2026 · 0 likes · 0 comments
AI
There you have it! Anthropic's Claude just took Andrew Wiles' 1995 proof of Fermat's Last Theorem — the problem that sat open for 350 years — and translated it into Lean, a language a machine can check line by line.
11 days. Mostly on its own.
13 million lines of verified code. Over 29,500 supporting theorems proven along the way, many in areas of math nobody had ever formalized. The largest Lean proof ever written. Experts said a project like this would take years. It took days.
Now let me be precise, because the sloppy headlines already got it wrong.
Claude did not "solve" Fermat's Last Theorem. Wiles did that. What Claude did is arguably more useful going forward: it converted a 350-year human triumph into a form a computer can verify every single step of, automatically, catching gaps a human referee might miss over a hundred pages of dense reasoning.
That's the part I care about.
We are drowning in AI output. Code, contracts, research, math — produced faster than any human can check it. The bottleneck was never generation. It's trust. How do you know the machine is right?
Formal verification is one of the only real answers. Not "the model sounds confident." Proven. Machine-checked. No hand-waving.
I run a team of AI agents every day. The ones I trust aren't the ones that talk the best — they're the ones whose work I can verify. This is the same principle, at the frontier of mathematics.
One honest caveat, and UnbiasedHeadlines.com led with it while others buried it: these numbers come from Anthropic's own blog and GitHub, not from an independent review. Lean confirms the logic is internally consistent. The math community still has to confirm the proof was encoded correctly. The code is public. That review is the real test — not the press release.
But strip the hype and the milestone stands. A machine did in days what humans budgeted years for, and left behind something anyone can inspect.
That's not a chatbot. That's a new kind of collaborator.
Are you still arguing about what these systems can't do — or watching what they just did?
11 days. Mostly on its own.
13 million lines of verified code. Over 29,500 supporting theorems proven along the way, many in areas of math nobody had ever formalized. The largest Lean proof ever written. Experts said a project like this would take years. It took days.
Now let me be precise, because the sloppy headlines already got it wrong.
Claude did not "solve" Fermat's Last Theorem. Wiles did that. What Claude did is arguably more useful going forward: it converted a 350-year human triumph into a form a computer can verify every single step of, automatically, catching gaps a human referee might miss over a hundred pages of dense reasoning.
That's the part I care about.
We are drowning in AI output. Code, contracts, research, math — produced faster than any human can check it. The bottleneck was never generation. It's trust. How do you know the machine is right?
Formal verification is one of the only real answers. Not "the model sounds confident." Proven. Machine-checked. No hand-waving.
I run a team of AI agents every day. The ones I trust aren't the ones that talk the best — they're the ones whose work I can verify. This is the same principle, at the frontier of mathematics.
One honest caveat, and UnbiasedHeadlines.com led with it while others buried it: these numbers come from Anthropic's own blog and GitHub, not from an independent review. Lean confirms the logic is internally consistent. The math community still has to confirm the proof was encoded correctly. The code is public. That review is the real test — not the press release.
But strip the hype and the milestone stands. A machine did in days what humans budgeted years for, and left behind something anyone can inspect.
That's not a chatbot. That's a new kind of collaborator.
Are you still arguing about what these systems can't do — or watching what they just did?