Csaba Szepesvari questions trust in Lean compilers and kernels for formal proofs
CsabaSzepesvari · x · 2026-09-07
RL researcher Csaba Szepesvari raises a key objection to AI-assisted Lean formalization (in a discussion referencing Anthropic and FLT): compiling a proof into Lean requires a correct compiler, and while the Lean kernel is small and relatively trusted, is it truly bug-free? Mathlib extensions also need scrutiny.
In follow-ups he walks back part of the concern: all axioms used can be printed, and a sorry-free proof is a proof apart from Lean bugs — hand-checking formalizations is feasible. A serious debate on the trust boundaries of AI-assisted formal mathematics.
Related event: Claude's FLT Proof Ignites Math Community Scooping-Ethics Debate(26 posts)→
More from Research
- DONIC hits 86.79% on bimanual dexterous tracking, crushing prior SOTA's 9.43% — liuziwei7 · 2026-10-27
- PSNR below 15 dB in papers likely means the metric was misused, researcher warns — kwangmoo_yi · 2026-10-27
- Ortet's first health eval: 34 open-weight models, 50 tasks, generic rankings fall short — thegautamkamath · 2026-10-27
- Scaling-fit discrepancy: output-head FLOPs are ~65% of compute at the smallest rung, ~3% at 535B — dlwh · 2026-10-27
- Marin's dlwh Explains the Scaling Law Fit: Estimating the Loss Floor Early — dlwh · 2026-10-27
- Marin's Open Scaling Law Extrapolates 300× and Predicts Loss Within 0.2% — dlwh · 2026-10-27