Lean creator Leo de Moura on the Collatz kernel exploit and AI-era formal verification

Machine Learning Street Talk · rss · 2026-09-30

Long MLST interview with Leonardo de Moura (creator of Lean, co-creator of Z3) covering: Lean's minimal trusted kernel and independent checkers; the Collatz incident where a purported proof was accepted by both Lean's kernel and nanoda via different bugs in each; reward hacking and safety-by-transparency; Kim Morrison's zlib proof with Claude; AlphaProof and why certificates still matter; Lean 4, dependent types, and Mathlib as infrastructure. 75 min with full timestamps and references.

Original post →

More from AGI Musings

AGI Musings channel →