AlphaProof Code(deepmind.google)
668 points by ilreb 1 day ago | 195 comments
tl;dr: Google DeepMind used a new agentic framework called AlphaProof Code to formally verify the implementation of Signal's cryptographic protocol, generating over 6 million lines of Lean proofs in four weeks—a task that typically takes expert teams years. The system had humans define specifications while AI agents produced the proofs, flagging only ~1% of the codebase for human review, and successfully detected two reintroduced known bugs. DeepMind is open-sourcing the verification artifacts.
HN Discussion:
  • ↓Skepticism that zero bugs were found suggests theorems may be too narrow or incorrect
  • ~AI-generated Lean proofs can cheat and need adversarial review to be trustworthy
  • ↑This is a promising use of AI that could make software dramatically safer in the future
  • •Curious how this approach compares to other similar AI formal verification efforts
  • •Questions about applicability of AI formal verification beyond cryptographic systems