Anthropic uses Claude to verify Fermat's Last Theorem proof

The formalization of Wiles's proof is a notable step in using AI for mathematics. Wiles's original proof spanned 129 pages and took months to verify manually. Formalization is notoriously difficult because proofs often omit explanations that computers need, and a single error can invalidate later code. Anthropic's model overcame these challenges by generating 6 billion tokens over 11 days, a task that mathematicians had estimated would take years. The use of Prove2Me appears to have been pivotal, turning an unsuccessful attempt into a breakthrough.

This work follows another Anthropic achievement announced a month earlier, where Claude was used to discover new information about the Riemann zeta function. Rival OpenAI also applied its latest Astra model to solve several Erdos problems last month. Together, these advances suggest a growing trend of using large language models to tackle complex mathematical proofs, though the article does not claim this is a universal solution. The comments from mathematician Kevin Buzzard, whose work was used, indicate that AI formalization artifacts are now robust enough to build upon, lending credibility to the milestone.