Who verifies the verifier? Notes on DeepMind's formal proof-search paper
DeepMind has developed an AI-driven formal proof search that autonomously resolves mathematical problems. This new approach allows for proofs to be generated and verified using a language model and a compiler, significantly reducing the need for human verification. While this advancement is promising, it does not yet address the challenge of formalizing complex proofs that are currently beyond the capabilities of existing tools.
- ▪DeepMind's AI resolved 9 of 353 open Erdős problems and proved 44 of 492 open conjectures.
- ▪The AI generates proofs in Lean, which are then verified by a compiler, making the process more scalable.
- ▪The current limitations of the AI include its inability to tackle proofs that require deeper mathematical frameworks not yet supported in Lean.
Opening excerpt (first ~120 words) tap to expand
May 28, 2026 · 14 min read Who Verifies the Verifier aimathematicsverificationlean Contents Who Verifies the Verifier Two frontiers Inside problem #125 The machine tried to cheat The frontier that’s still open Quis custodiet Who Verifies the Verifier Three days ago I published an essay that ended on a problem I couldn’t see a solution to. The short version: an AI had disproved one of Erdős’s conjectures, the result was real, and it was verified — but verified in the only way we currently know how, which is that nine of the world’s relevant experts — one of them a Fields medalist — read a hundred-page argument over a weekend and put their names on it. That doesn’t scale.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at Korbonits.
Discussion
0 commentsMore from Korbonits
