2026-07-04 / Signal #1
Mistral Leanstral 1.5 (formal proof agent in Lean 4)
“One open model just proved hundreds of grad-level math problems and caught bugs humans missed in production code—for pocket change.”
Why It Matters
AI making "proof abundance" practical changes builder behavior in software verification, math, and trust institutions—agent loops turn formal methods from elite/academic to routine, with real bug-catching as concrete artifact (not generic hype).
Evidence
Mistral official blog (released ~July 3, 2026), HN top post (269 pts), widespread X discussion (e.g., threads with 1K+ likes detailing benchmarks). Saturates miniF2F (100%), solves 587/672 PutnamBench problems (~$4/problem with test-time scaling), SOTA on FATE-H (87%)/FATE-X (34%), trained as agent with compiler feedback + RL; found 5 previously unknown bugs (e.g., integer overflow) across 57 real OSS repos via Aeneas/Lean pipeline. Apache-2.0, weights on HF, free API.
Signal Read
Source Trail
Daily scan: 2026-07-04