AI & Scientific Verification
Anthropic reports Claude produced a computer-checked proof of Fermat’s Last Theorem
A multi-agent workflow generated a large Lean formalisation, showing how machine-verifiable artifacts can help review AI-assisted research without replacing human explanation.
What happened
Anthropic says a Claude Code-based multi-agent system, coordinated through the Prove2Me platform, produced an end-to-end Lean formalisation of Fermat’s Last Theorem in 11 days. The company reports roughly 13 million lines of Lean, 30,300 computer-verifiable theorems and about six billion output tokens. Anthropic published the proof repository and says Lean checked the result, while mathematician Kevin Buzzard reviewed the proof and its dependency on standard axioms.
Why it matters for AI
AI-generated scientific work can create more material than human reviewers can inspect line by line. A proof assistant changes the review surface: the model still proposes the artifact, but a deterministic checker verifies whether the formal logical steps compile against declared axioms and dependencies.
AIRunsThis View
The reusable pattern is artifact plus independent verifier, not a claim that model output becomes self-validating. Formal checks can strongly test logical consistency inside a defined system, while humans still need to assess whether the formal statement matches the intended question, whether dependencies are acceptable and whether the result is understandable and useful.
What this could mean for us
Relevant to AIRunsThis QA: wherever possible, consequential agent output should be converted into something a deterministic tool can check—tests, schemas, checksums or formal constraints—while keeping interpretation and approval separate from generation.
Source and method
This item is based on a primary source. AIRunsThis separates the source’s factual announcement from its own interpretation. Vendor performance claims have not been independently reproduced unless stated.