Anthropic mathematician Alpöge used Claude to prove the Riemann Hypothesis, formalizing the entire argument into roughly 17 million lines of machine-checkable code that Lean's official community certified within a single day. It marks the first time AI has helped crack mathematics' most famous open problem after 167 years, with Claude proposing a key step in areas beyond the human's own expertise.
Formal Verification
形式化验证
Technique that uses mathematical proof to rigorously establish that software or protocols satisfy their security and functional
Curated Intelligence
Andrew Sutherland of MIT, using Claude Code, has independently verified a Lean formalization of the Riemann Hypothesis: all 175,187 modules compile cleanly, the five headline theorems depend only on the three standard axioms, and two independent kernels, nanoda and con-leche, accept the full exported proof. It marks the first time AI-assisted formal mathematics on a major open problem has reached machine-checkable verification strength.
The Lean and Mathlib communities jointly certified that Levent Alpöge's 17-million-line Lean library, released on November 20, 2026, correctly proves the Riemann Hypothesis, with both statement and proof vetted by experts. This marks a milestone for AI-assisted formal mathematics and machine verification reaching a top-tier open problem.
Anthropic researcher Levent Alpöge released a Claude-driven proof of the Generalized Riemann Hypothesis as roughly 250,000 pages of Lean code, while mathematician Alessandro Masullo published a 940-page fully human proof built over 12 years. The near-simultaneous independent claims put AI-assisted formal mathematics to its most consequential test yet.
As AI-generated code proliferates, low-quality output known as "slop" has become a daily engineering concern. Companies including Harmonic, Theorem, and Axiom Math are pushing formal verification systems, which rely on mathematical proofs and logic, toward practical use for checking the correctness of code delivered by AI agents. AI's improving math capabilities are making this once high-barrier approach viable.
Levent Alpöge claims to have used Anthropic models to prove the Generalized Riemann Hypothesis, generating 17.4 million lines of Lean code now published on GitHub. Coming on the heels of an AI-assisted result bounding nontrivial zeros within 3/8 of the critical line, a verified proof would mark a landmark moment for AI for mathematics.
Anthropic mathematician Levent Alpöge has published an essay claiming a Lean-compilable proof of the Riemann Hypothesis, an open problem since 1859 and one of the Clay Millennium Prize problems. Mathematicians are scrutinizing the result in Lean community channels, while a second researcher has also claimed an independent proof, leaving verification to peer review. If it holds, the result would mark a landmark intersection of mathematics and AI-assisted formal verification.
Mathematician Levent Alpöge publicly claims to have discovered a proof of the Riemann Hypothesis working alongside Claude, saying the method extends to the Generalized Riemann Hypothesis and that the arguments were formalized without assumptions. If verified by peers, this would mark a milestone for AI-assisted mathematics, though it currently remains a personal claim awaiting formal verification and community review.
Harmonic announced that its formal verification tool Aristotle now supports Rust, proving code free of undefined behavior and compliant with its specification before shipping, including inside unsafe blocks. The capability extends Aristotle's existing C and Verilog support to mission-critical domains such as smart contracts and systems infrastructure.
OpenAI's recent dump of 722 mathematics papers included Lean formalization for roughly 45%, yet three non-Leanified papers were already retracted over a sign error. The post argues that Lean code compiling is not proof of correctness, especially when AI-generated, and that additional tests against reward hacking and kernel soundness are needed but have not yet been systematically performed.