Sutherland's Claude-Assisted Lean Proof of the Riemann Hypothesis Passes Multiple Independent Kernels
- Andrew Sutherland (MIT) reports that the Lean proof of the Riemann Hypothesis compiles with zero errors across all 175,187 modules, and its five headline theorems depend only on the three standard axioms.
- Two independently written kernels, nanoda and con-leche, accept the full exported proof, and lean4lean accepts every compiled module, though five modules needed a raised step limit to pass.
- The checks were run with Claude Code, and as of the last page update the default and paranoid Lean kernel replays and the full lake comparator run were still in progress.
- The theorems cover Mathlib's zeta function and Dirichlet L-functions for all moduli and characters, but not Dedekind, Hecke, or automorphic L-functions.
Hacker News 의견들
Kind of a monoculture here. Claude commits the Lean kernel, the Con Leche kernel, writes the proof, writes the site, then verifies its own proof. Could be malware in the kernels, no?
Are you saying Claude ran a long con? If so, I'd be more impressed by Claude than by the GRH proof itself.
Con Leche is supposed to be correct iff ZFC is, so the real questions are whether the binary is untampered and whether Claude parsed the proof faithfully. I'd put it around 90% not fraudulent, but even a clean result is a little worrying if it erodes why we do math.
The dependency on LLMs could be avoided by checking Con Leche in a pre-2022 Metamath, which has a tiny trusted base.
Trusting trust applies here too. Anyone with a motive will probably check that Claude didn't cheat, and a mathematician who finds a flaw would have a strong incentive to say so.
Someone built a tool to run the verification on your home CPU with a global leaderboard, which is neat.