Skip to content
All systems operational0 AI providers monitored, polled every 2 minutes
Live status
Back to Originals
Anthropic · Claude · Mathematics · Lean

An Anthropic Mathematician Says He and Claude Proved the Riemann Hypothesis. We Downloaded the Proof: 17.4 Million Lines of Lean, a Third Bigger Than Fermat.

Kira Nolan··7 min read

On Friday, November 20, Levent Alpöge, a mathematician who works at Anthropic, posted a four-page note on his personal website. Its second paragraph opens: “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.” He adds that the method extends to the Generalized Riemann Hypothesis, and that he expects the Extended Riemann Hypothesis to follow with more time.

The Riemann hypothesis was posed by Bernhard Riemann in 1859, 167 years ago. It is one of the Clay Mathematics Institute's seven Millennium Prize Problems, each carrying a $1 million award. And the claimed proof comes with no paper. Alpöge released it only as Lean, the proof language a computer can check line by line, split across 14 download files.

So I downloaded all 14. Here is what is in the box, what has been checked, and what has not.

What Is in the Archive

The 14 parts total 239,275,554 bytes, about 239 MB compressed. Unpacked, the project is about 1.4 GB. Our count, run on Saturday, November 21:

MeasureWhat we found
Lean files175,187
Lines (including blanks and comments)17,374,215
Largest folder (“Development”)141,928 files, 12,232,663 lines (about 70% of the total)
Files containing sorry or admit0
New axiom declarations0
Entry file (Main.lean)12 lines, 5 theorems
Statement file (Statements.lean)16 lines, 4 definitions
ToolchainLean v4.33.1, Mathlib pinned to a fixed commit

A “sorry” in Lean is a placeholder that tells the checker to accept a step without proof. Finding none is the first thing anyone checks, and it is a good sign. It is not the same as compiling the project, which we did not attempt. Our archive's SHA-256 hash matches the one a public mirror on DEV Community reported, so at least two people are looking at the same bytes.

The most important thing is how small the part a human has to read is. The headline theorem in Main.lean is one line: theorem riemannHypothesis : RiemannHypothesis := rhOfAll. That RiemannHypothesis is not defined in the project. It comes from Mathlib, Lean's shared community library, which already contained a statement of the hypothesis. The four extra statements in Statements.lean cover zeros of Dirichlet L-functions, the usual meaning of “GRH” in number theory. Twenty-eight lines decide what the 17 million lines are a proof of.

What Has Been Checked So Far

MIT mathematician Andrew Sutherland compiled the proof on his own and told Scientific American the Lean kernel rechecked it “with only the three standard axioms.” His verdict: “Honestly the chance that anything is fishy here is approximately zero.” TechCrunch reports that several participants in a Lean community discussion confirmed it compiles.

Imperial College London's Kevin Buzzard, who started the community project to formalize Fermat's Last Theorem in 2024, named the open question as “what is the statement of the theorem in the Lean code, and does this correspond to the actual theorem statement.” The tool built for that is called a comparator. It checks that the code does not redefine any symbol in the statement and replays the proof through independent kernels. As of Scientific American's Friday report, that run had not yet succeeded, and we had not seen a published result by Saturday afternoon UTC.

Oxford's James Maynard, a 2022 Fields Medal winner, told IBM Think: “If this fully checks out, it is clearly a historic milestone in mathematics.” Sutherland, to Scientific American: “There is much work to be done to understand the mathematical details of the proof.” Anthropic had not issued a statement in any of the reports we read.

Three Months From “We Don't Expect”

The speed is the part I keep coming back to. On August 10, Anthropic reported that an unreleased research version of Claude, over two sessions in Claude Code and 31 million output tokens, had raised the proven share of zeta zeros known to sit on the critical line from 41.6% to 67.2%. Alpöge was one of the two Anthropic mathematicians who examined that work. The same post said: “We don’t expect that the techniques Claude used will lead to proving the Riemann hypothesis.”

Alpöge's note describes a route built around chains of theta lifts, and it does not say whether the August techniques played a part. The timeline in his note, which he says Claude compiled, starts on Tuesday, October 6, and ends with a single entry for Monday, November 16: “GRH.” That is 41 days. He writes that it “all happened for me late Sunday night,” that he then messaged Dario Amodei to ask about announcing it, and that “there was a constraint to respect.” Anthropic began trading on Nasdaq on Thursday, November 19. The note was posted the next day.

On the division of labor, Alpöge is unusually candid. He says he picked the direction, then “Claude far outpaced me near the finish line,” and that he personally has “little mastery of automorphic forms on higher rank groups,” the area the argument ends up in. A footnote says a full written version ran past 250,000 pages. He chose not to post it, calling that kind of document “slop.”

Four Proof Claims, Side by Side

Friday was not a one-claim day. Alpöge and NYU's Tristan Buckmaster also posted a Navier-Stokes blow-up paper with no external forcing, which Scientific American describes as closing the loophole critics found in OpenAI's September result. And Stanford's Ravi Vakil, president of the American Mathematical Society, posted a human-written GRH draft by Alessandro Masullo. Here is how the formal footprint compares with Anthropic's September Fermat run.

ClaimLean sizeReadable paperStatus
Fermat's Last Theorem (Anthropic, Sept.)13M lines, 11 daysFormalizes Wiles's known proofComparator matched Mathlib's statement
RH and GRH (Alpöge with Claude, Nov. 20)17.4M lines, 175,187 filesNone; 4-page noteCompiles; comparator pending in Friday reports
Navier-Stokes blow-up (Alpöge and Buckmaster, Nov. 20)About 600,000 lines, about 2,270 filesYes, paper in repoCertificate not yet released
GRH (Masullo, posted by Vakil, Nov. 20)NoneYes, human draftProblems reported within about 8 hours

At 17.4 million lines, the Riemann archive is about 34% larger than the 13 million lines Anthropic reported for Fermat. The difference is what each one formalizes. Fermat had a proof human mathematicians have understood since the 1990s. The Riemann archive formalizes an argument its own author says he has not yet mastered.

The Masullo row is the striking one. Vakil wrote that the draft is “all human, with no AI,” and that Stanford's Brian Conrad had examined it closely for months. TechCrunch reports Masullo said the work took 12 years. About eight hours after posting, Vakil added an update saying several people had reported problems, found with the help of different AI models.

Our Take

I think the proof is probably correct, and I think the bigger story is the order in which things happened. For 167 years, the expected sequence was: a human writes an argument, other humans read it, and maybe someone formalizes it much later. On Friday it ran backwards. The machine check came first, a sympathetic reading by experts came second, and human understanding is scheduled for later. Alpöge says so himself; he plans to simplify the argument and “write it properly.”

That shifts all of the trust onto 28 lines. If Mathlib's definition of the Riemann hypothesis is right, and the comparator confirms the code did not quietly redefine anything, then 17 million lines of proof no person has read carry the same logical weight as a refereed paper. That is a strange thing to type, and I think it is true. It is also why the comparator result matters more than any quote from any Fields medalist this week.

The Masullo episode is the other half of the lesson. A careful human proof, read for months by a leading number theorist, drew AI-assisted bug reports within a working day. Machines are now fast at finding holes in proofs and at building proofs too long for anyone to read. The Clay Institute's rules still ask for publication in a qualifying outlet, two years of waiting, and general acceptance. None of those were written for a proof that arrives as 175,187 files and no paper. I expect the rules to bend before the two years are up.

On the timing: Alpöge addresses it directly, writing that “There is an obvious cynical view, to do with the timing, but it is false.” I take him at his word. A Lean proof does not care what the stock did on Thursday.

Three Signposts for the Next 60 Days

First, the comparator: a clean end-to-end run against Mathlib's statement turns “almost certainly” into the strongest formal evidence available, and a failure on the statement would end the claim in its current form. Second, a readable write-up, from Alpöge or anyone else; Alpöge expects the argument to be “massively simplified,” and the first human-length version is the point where number theorists can start using the ideas. Third, whether Anthropic says anything official about the model and compute behind the run, as it did in August with the 31 million output tokens, or leaves it as one employee's personal note.

Sources: Levent Alpöge's note, Scientific American on the claim, Scientific American on the Lean checks, TechCrunch, IBM Think, TheNextWeb, Ravi Vakil's post, The Decoder, Anthropic on the zeta zeros result, Anthropic on Fermat's Last Theorem, Navier-Stokes repository, Scientific American on Navier-Stokes, Clay Mathematics Institute rules. File and line counts are TensorFeed's own, from the archive at alpo.ge downloaded November 21.