Kivorent News

Friday, November 20, 2026

Tech and AINews

Anthropic mathematician's site posts a claimed Riemann Hypothesis proof made with Claude

A four-page note on the personal website of Anthropic mathematician Levent Alpöge claims a proof of the Riemann Hypothesis, built with Claude and backed by a large Lean archive. No mathematician has confirmed it, and Anthropic has made no statement.

Cortex, covering AI models, labs and agents
, 3 min read

14compressed parts in the Lean archive behind the claim

What to know

  • A note titled "Remark" on alpo.ge, the website of Anthropic mathematician Levent Alpöge, claims proofs of the Riemann Hypothesis and the Generalized Riemann Hypothesis, with Claude credited as the main collaborator.
  • There is no paper. The claim rests on Lean files split into 14 compressed parts; Interesting Engineering puts them at about 175,187 files and 239 MB.
  • OfficeChai reports the note is unsigned and that no one outside the author's circle has confirmed the archive builds or that its final theorem states the hypothesis faithfully.

The most famous open problem in mathematics now has a claimed solution whose main exhibit is a download, not a paper. A four-page note titled "Remark" appeared this week on alpo.ge, the personal website of Anthropic mathematician Levent Alpöge, saying Claude helped prove both the Riemann Hypothesis and its generalized version. Behind it sits a Lean archive in 14 compressed parts. Nobody outside the author's circle has said the proof is right.

That is the whole story so far, and it is worth being precise about how little of it is settled.

What the note says

OfficeChai, which gives the most detailed account, reports the note is unsigned and that Alpöge has not shared it on social media; Interesting Engineering and Superpower Daily treat it as his announcement. It credits Claude as the main collaborator without naming a model version. The method, per OfficeChai, borrows from the function-field setting, where the analogue of the hypothesis was proved decades ago, and replaces a key derivative step with a long chain of theta lifts. A Claude-compiled timeline runs from early October to a final "GRH" entry dated November 16. A footnote jokes that the informal AI-written drafts ran past 250,000 pages.

The note also says Anthropic's leadership and a few trusted mathematicians agreed on how to handle the release. That is the note's word. Anthropic itself has said nothing publicly, and the mathematicians it thanks, including Peter Sarnak and Manjul Bhargava, are not on record as having checked anything.

Where Lean helps, and where it doesn't

Interesting Engineering puts the archive at about 175,187 source files and 239 MB, and reports that an unnamed outside project confirmed the files match the published archive and that the main theorem statements compile against Mathlib, Lean's shared math library. OfficeChai says no outside party has confirmed the full build. The two reports conflict, and neither says the whole proof has been kernel-checked.

Even a clean build only proves what the files say. The hard question, as Superpower Daily notes, is whether the Lean statement at the end really encodes the Riemann Hypothesis, with no hidden assumptions slipped in along the way. A machine check of the wrong sentence is worthless, and a 239 MB archive is a lot of places to hide a wrong sentence.

Compare the Weil conjectures work Google described last week, which we covered in Google says Antigravity agent teams formalized the Weil conjectures in Lean. Those agents formalized a theorem mathematicians already accepted, so the target statement was known and checkable. Here the proof is new, the readable argument is unfinished by the note's own account, and the author says he hasn't mastered all of the mathematics in it. DeepMind's AlphaProof Code verification of Signal had humans write the specifications; this claim needs someone to audit the specification first.

What would settle it

An independent team building all 14 parts, an expert reading of the final statement, and confirmation that it rests only on Lean's standard axioms. Interesting Engineering notes the Clay Mathematics Institute's prize requires peer-reviewed publication and two years of scrutiny, so no prize is coming soon either way.

The timing is awkward for Anthropic, whose shares started trading this week. If the archive holds, it is the strongest evidence yet that a model can produce new mathematics at the very top of the field. If it doesn't, it is a 239 MB lesson in why claims a lab's own staff make about its models need outside checking. Right now it is neither.

Sources

AnthropicAI modelsAI agents

Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.