Formal verification advocate: AI suggestions are only OK when you can verify them, Lean proves the point
gerardsans · x · 2026-10-05
Responding to reports of engineers line-by-line rejecting AI code, gerardsans notes math faces the same issue and already has an answer: relying on AI is only acceptable when the problem is well understood and suggestions can be verified with confidence — formal runtimes like Lean do this for proofs. Skipping verification steps and over-relying on a system you no longer control is too risky; responsible engineers should keep results but hold deployment until the verification gap closes.
Related event: Engineers Can't Read AI-Generated Code Anymore(3 posts)→
More from AGI Musings
- Margaret Mitchell: feeding models their own energy signals could cut LLM power use — mmitchell_ai · 2026-10-05
- Ex-OpenAI exec Miles Brundage slams Claude's misuse by sanctioned states — Miles_Brundage · 2026-10-05
- Creator lived 30 days with AI 'Worst Boyfriend Ever' for a docu-drama — teortaxesTex · 2026-10-05
- Brundage: Anthropic keeps deploying Claude despite ransomware and propaganda misuse — Miles_Brundage · 2026-10-05
- LLM-Assisted Math: Mathieu Group M_23 Proven Galois over Q, Video Explains — _sathvikr · 2026-10-05
- Robin Hanson: computers truly compute on symbols, brains do process information — ctjlewis · 2026-10-05