AI autonomously formalizes Fermat's Last Theorem proof in Lean after 11 days

The event marks the first time Fermat’s Last Theorem has been fully formalized in a proof assistant. Claude’s output, 13 million lines of Lean code, represents a complete, machine-checkable version of Wiles’ 1995 proof. This removes ambiguity in verification, as Lean’s algorithmic checks leave no room for human oversight errors. For engineers, the takeaway is that AI can now handle the tedious, step-by-step logical work that previously required years of manual effort by mathematicians.

The cost of adoption lies in the tooling and expertise required. Lean and other proof assistants demand specialized knowledge to use effectively, and formalizing proofs at this scale still requires significant computational resources. The 11-day runtime for Claude’s work suggests that while AI accelerates the process, it is not yet instantaneous. Additionally, the proof’s reliance on Lean means it is only as trustworthy as the assistant’s implementation, which itself must be verified.

Where this approach stops working is in proofs that rely on intuition or creative leaps not easily encoded in formal logic. Wiles’ proof, for example, required novel insights in algebraic geometry and modular forms, areas where AI may struggle to replicate human ingenuity. The formalization also does not address whether Fermat’s original conjecture could have been proven with 17th-century methods, as the AI’s work builds entirely on modern techniques.

The broader implication is that AI-driven formalization could democratize verification in mathematics. Historically, verifying a proof like Wiles’ took months of peer review; now, it can be done in days. This could reduce the bottleneck for new mathematical discoveries, as researchers spend less time checking existing work and more time building on it. However, it also raises questions about how much trust to place in AI-generated proofs, especially in fields where human intuition remains critical.