Dozens of agents used a process called autoformalization to turn advanced human mathematics into proofs that software can check line by line. The resulting output verifies a theorem conjectured over 350 years ago and first proven by Sir Andrew Wiles in 1995. Anthropic's Claude AI completed the formalization in 11 days, producing a 13 million line Lean proof, which is the largest ever written, including 29,500 intermediate theorems across geometry and number theory.

The company stated the project marks a step toward firming up the core of mathematical knowledge and reducing the labor involved in refereeing proofs. Researchers said the system proves that AI autoformalization artifacts are robust enough to be built upon, allowing for the verification of complex, multi layered reasoning without assumptions beyond basic axioms.

Sign in to suggest edits

Key sources

  1. SOURCE@anthropicai“the largest Lean proof ever written”x.com
  2. SUPPORT@scaling01“proved 29,500 intermediate theorems over 11 days”x.com
  3. SUPPORT@sammcallister“no assumptions other than the axioms of mathematics”x.com
  4. SUPPORT@rohanpaul_ai“dozens of Claude agents took the existing Wiles-based proof”x.com
  5. DISCUSSIONni5argalobste.rs
  6. SOURCElobstersanthropic.com
  7. SOURCE@blackhc“Checking that a major mathematical proof is correct can take years”x.com
  8. SUPPORT@andytng28“over 29,000 other theorems that the proof requires”x.com
Markdown