François Fleuret: 'The model made a prediction which Lean confirmed' makes no sense
francoisfleuret · x · 2026-10-13
AI researcher François Fleuret pushed back on the claim that 'the model made a prediction which Lean confirmed,' calling it meaningless: Lean checking a model's prediction gives no more guarantee than using it to verify a stock-market prediction Python script. Formal correctness, he argues, is simply not the problem in such scenarios — a substantive point in debates over AI-for-science safety arguments.
Related event: Fleuret: Formal Math Cannot Guarantee Model Predictions(3 posts)→
More from AGI Musings
- Veteran Unix admin: LLMs cured my decades-long dread of system administration — esrtweet · 2026-10-13
- Computational biology shifts as AI builds 'better predictors' in under a day — jmschreiber91 · 2026-10-13
- halvarflake: most production code wasn't production quality even before AI — TheKanter · 2026-10-13
- Academia vs Anthropic: scholars defend university research independence in AI — sethlazar · 2026-10-13
- Zhejiang University paper: LLMs match human average choices but adjust along divergent paths — ValerioCapraro · 2026-10-13
- AI Lab Insiders Insist on Writing Themselves: 'Writing Is Thinking' and Outsourcing It Kills Your Thinking — rseroter · 2026-10-13