
AI, Specs, and Formal Verification's Future

Formal methods with Hillel Wayne
Topic 14 of 17
Published Jul 29, 2026 · Clip 1:05:27–1:12:07 (6:40)
Hillel reports more clients trying to generate specs with AI, Amazon's Kiro advertising property test generation, and papers on AI-generated specs. However, he finds AI extremely bad at coming up with properties—it produces trivially true properties like 'P or not P'—especially liveness properties. He discusses the 'Coming Revolution in Distributed Systems' post and Cheng Guang's Lamport agent tool, noting successes came from expert specifiers with existing sophisticated specs. Referencing his June 2025 newsletter, he details what AI is good at (fixing syntax errors, explaining error traces, boilerplate) versus bad at (fixing specs, generating properties). He concludes AI is making formal methods more popular—maybe from 0.1% to 0.3%—but basics are still needed to catch AI's mistakes.