Levent Alpöge claims a proof of the Riemann Hypothesis via Claude, but the artifact is Lean code, not a paper. That shifts the trust boundary for anyone building with AI. You stop parsing dense English text and start compiling logic. If a model outputs formal structure, you verify it mechanically. That is a better workflow than reading a 250,000-page document. I would treat the Lean code as the source of truth; the natural language explanation is secondary. For self-hosted tooling, verification pipelines matter more than generation pipelines. If the code compiles, the logic holds. Do not trust the prose. Trust the compiler.
The story — Levent Alpöge, an Anthropic employee, announced a possible proof of the Riemann Hypothesis developed with Claude. The proof is formalized in the Lean proof assistant, resulting in extensive code rather than a traditional manuscript. Alpöge withheld a 250,000-page document titled “GRH book” because he found it unreadable. Independent checks confirm the Lean code compiles and is accepted by two independent kernels. However, whether the formal statement exactly matches the Riemann Hypothesis remains under expert review. No final consensus has been reached yet. (Source)