On November 20, Levent Alpöge, a mathematician at Anthropic, posted a statement on his personal website: together with Claude, he says, he has found a proof of the Riemann hypothesis — and the same method, he claims, extends to the generalized Riemann hypothesis. The Riemann hypothesis is one of the Millennium Prize Problems, formulated by Riemann in 1859, and it describes the distribution of prime numbers. Right now this is a claim, not an accepted result.
The publication isn't a press release — it's working material for verification: 4 pages of text and a formal proof in Lean, packaged into 14 archives. The mechanics Alpöge describes go like this: he chains theta lifts into a sequence, mimicking the "multiplicity" from classical methods in number theory. The key idea came from Claude: the length of this chain of transformations should grow as the square root of the summation range. According to Alpöge, in the later stages Claude was moving faster than he was, and he's still preparing a human-readable version of the proof.
Why now: over the past two years, LLMs have stopped being just "text assistants" and moved into the territory of formal proofs — Lean, Coq, Isabelle. Here the model isn't reasoning in a vacuum; it produces code that a machine checks step by step. Our take: that's the real reason an announcement like this is even possible — it's not that "AI got smarter," it's that a language now exists where errors are caught mechanically.
What it changes: if the Lean files pass independent verification, mathematics gets a precedent where AI doesn't just suggest a conjecture but closes out one of the seven Millennium Problems. That will shake up how grants are allocated and how the contribution of an AI co-author is assessed. If they don't pass, we're left with a tidy but loud case study: "AI generated a mountain of code nobody has digested yet."
Who wins and who loses: Anthropic gets ammunition in the argument over reasoning models, and Lean and formal methods get attention — and probably fresh funding. The losers, by definition, are everyone who bet on "AI can't do real math" — but only after verification. Terence Tao wrote on his blog on November 20 that Alpöge had produced a large-scale proof of the GRH only as Lean code, and that understanding it will take time; he also noted that at least there's a living author you can actually talk to.
The big unknown: whether the 14 archives pass independent Lean verification, and whether the formalized theorem actually proves the Riemann hypothesis itself rather than something that merely looks like it. The word "proof" here still belongs to the author of the claim, not to peer review. Tao's response is not a stamp of approval — it's a polite "this needs to be looked at." There's exactly one way to check it yourself: open the files and run them through Lean — if you have weeks to spare.
Forecast: within the quarter, at least one public analysis of the Lean files by outside mathematicians will appear — concluding either "the formalization is correct" or "the theorem doesn't match the hypothesis."
Disclaimer: not officially confirmed, no independent verification; the facts come from the author's public statement and his blog.
Source: (ifanr.com)
Source: www.ifanr.com · post in Telegram