Deep Dive — Claude's Riemann Hypothesis proof: 17.4M lines of Lean

By Saturday morning the machine had all but signed off on the most famous open problem in mathematics. A Lean archive of 175,187 modules compiles with zero errors, replays through Lean's kernel, and is accepted by two independently written kernels, one of them a Rust checker that shares no code with the official one. The five headline theorems use only Lean's three standard axioms. As of 08:45 UTC today, the last remaining check — a full comparator run and a paranoid-kernel replay of a 62 GB export — was still grinding through single-threaded stages expected to take hours.
No human being has read the proof.
That is the shape of the strangest 36 hours in the history of mathematics. On Friday morning, Levent Alpöge, a mathematician employed by Anthropic, posted a four-page note titled "Remark" on his personal site and a fourteen-part tarball of Lean code. "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis," it opens, according to the note itself. It goes further: the method "easily extends" to the Generalized Riemann Hypothesis, and he expects the Extended version "with more time."
What was actually released
Start with the arithmetic of the release, because it is the argument. There is no paper. Alpöge says the informal draft the model produced ran past 250,000 pages — "I of course amused myself, and hit 250,000+ pages," he writes, noting that "Claude cutely called it the GRH … 'book'." Rather than post that, he released only machine-checkable code, on the reasoning that "it being obviously inappropriate to attempt to post some massive unreadable travesty of 'slop', I decided to only present Lean."
The source archive is 239 MB across fourteen parts. Unpacked, it is 175,187 Lean modules built against Lean and Mathlib version 4.33.1. Scott Aaronson, who read the repository the morning it landed, put the line count at 17.4 million lines of Lean — a figure that is less a measure of the proof than of how much scaffolding a proof this size needs before a kernel will accept it.
The mathematics underneath is a descendant of Stepanov's method, which proves the Riemann hypothesis for curves over finite fields. Stepanov's argument uses a divisor with multiplicity; over the integers there is no obvious multiplicity to use, because producing one would be something like differentiating whole numbers. Alpöge's account of the breakthrough is specific enough to check: the pair worked in a toy setting with new Fourier-analytic ideas, then needed the same trick to recur "in a natively automorphic manner," which led to chaining theta lifts — iterated transitions between unitary groups, each one standing in for a unit of multiplicity.
"The line: for the von Mangoldt sum of length X, Claude's suggested length of chain of theta lifts, in view of each mimicking multiplicity, was √X," he writes. "(And yet, what else could it be?)" Around that spine sit a relative trace formula, an Iwasawa tower, Bessel periods, and a step he describes as fighting for "a significant weakening of Vandiver." His timeline runs from October 6, when the Iwasawa tower analogy appears as entry 11, to November 16, entry 20, labeled simply "GRH." Forty-one days.
He is candid to the point of discomfort about what he does and does not understand. "I am presenting the world with a complete argument before having mastered a massive neighbourhood around its mathematics," he writes. Where Claude was the expert and he was not, his descriptions paraphrase a summary he calls "slop," and he apologizes for it.
The machine has already agreed
The verification, unusually, is public and timestamped. Andrew Sutherland at MIT rebuilt the archive on his own machines and posted a running log of every check, most of it completed within eight hours of the announcement.
All 175,187 modules compiled with zero errors by 16:43 UTC. An axiom audit on the five theorems in the submission's main file returned only propext, Classical.choice and Quot.sound — the three axioms Lean's own mathematics rests on, with nothing declared outside the core. Every module was replayed through leanchecker with zero rejections. The independent nanoda kernel accepted the full exported proof in 27 minutes 42 seconds on 80 threads. con-leche, the kernel written in Lean and proved consistent, accepted 1,468,418 declarations, and its statement-comparison mode confirmed all five theorems against the challenge files. lean4lean passed all 175,187 modules, five of them only after a step limit was raised.
The statement itself is the part that matters most. The theorem is riemannHypothesis : RiemannHypothesis, where RiemannHypothesis is Mathlib's own definition, not one written for this release: for every complex number that is a zero of the zeta function, is not a trivial zero at a negative even integer, and is not 1, the real part is one half. "GRH" here means the analogous statement for Dirichlet L-functions.
Then there is the check that deserves more attention than it has received. Mathlib's definition of the zeta function is a definition like any other; a proof could be internally perfect and still be about the wrong object. So Sutherland combined Alpöge's proof with an independently written formalization of Robin's and Lagarias's criteria — two statements equivalent to the Riemann hypothesis that talk only about divisor sums — and got two inequalities out the other side that mention no zeta function at all. If the arithmetic statements hold, the object being proved about is the right object. The combined export is 62 GB and its own comparator run is still going.
"The Lean proof has been compiled and independently re-checked by the Lean kernel with only the three standard axioms," Sutherland told Scientific American. "Honestly the chance that anything is fishy here is approximately zero."
What is not done is the comparator run in paranoid mode, which Kevin Buzzard of Imperial College described to Scientific American as the check that the code does not "do anything nefarious like redefine the meaning of any of the symbols in the statement," and that it compiles under several independent kernels rather than exploiting an unknown bug in Lean's official one. That is not a formality. Lean shipped a kernel soundness bug in August, patched and documented by Leonardo de Moura; the paranoid mode exists because the checker is software, and software has defects.
Why "compiles" is not "true"
Formalization claims die at exactly one place: the gap between the theorem a human meant and the theorem the code states. Buzzard wrote the field's audit protocol on November 12, eight days before this release, after reviewing OpenAI's dump of 722 mathematical manuscripts — about 45 percent of which came with Lean code, and three of which, in the non-formalized portion, have since been retracted for a sign error. His first question is always the same: what is the statement in the file, and does it correspond to the actual theorem.
The second failure mode is history. Michael Atiyah announced a five-page proof of the Riemann hypothesis in 2018 and it did not survive a week of scrutiny. Alpöge has made large claims before: Scientific American notes he posted a Claude-generated solution to the Jacobian conjecture in July.
So the honest reading of Saturday morning is narrower and more interesting than either headline. The probability that the Lean artifact means what it says — that a kernel-checked, axiom-audited, statement-matched derivation of the Riemann hypothesis for Dirichlet L-functions exists — is now very high. The probability that mathematics has absorbed a proof is zero, because absorption has not begun. Clay Mathematics Institute rules require publication in a qualifying outlet, two years of elapsed time, and "general acceptance in the global mathematics community." Even a paper this month puts the earliest possible award in late 2028, and Clay still lists the problem as unsolved.
The other half of the story is why this keeps happening to the credit machinery rather than to the mathematics. We traced that breakdown in Math's credit system was built for humans. AI just broke it: priority, peer review and the two-year wait all assume a human pace and a single author, and none of them has an answer for a proof that arrives as a repository. In October, OpenAI published 722 math manuscripts from a model nobody can run, and the reading problem it created — thousands of pages nobody had time to check — is now the same problem in a single artifact, at 17.4 million lines.
The timing, and the sales line
The note landed one day after Anthropic's IPO opened at $111.50 and closed at $97.17 and a day after the stock gave back most of its first-day gain. Alpöge addresses it head-on rather than letting others do it: "The precise timing is horrible. If I could have, I'd have finished this all much earlier." He says the result came together late Sunday night, that he messaged chief executive Dario Amodei asking "Hypothetically, if I were to have…", and that "there was a constraint to respect." His verdict on the cynical reading: "There is an obvious cynical view, to do with the timing, but it is false."
Andrew Granville of the University of Montreal told Nature that Alpöge should have prepared a clearer write-up first, and that "what he released is garbled," Nature reported. James Maynard, a Fields Medalist at Oxford, gave the conditional version: "If this fully checks out, it is clearly a historic milestone in mathematics."
The commercial logic does not need a conspiracy. A lab selling reasoning to enterprises now has a headline no competitor can match: the most famous unsolved problem in mathematics, solved on its model, verified by strangers within a day. Whether or not that was the intent, it is the asset. It also arrives with an awkward footnote from the company's own record. Anthropic's August write-up of Claude's earlier zeta work says the model "did take a real stab" and "didn't succeed," and states plainly: "We don't expect that the techniques Claude used will lead to proving the Riemann hypothesis." That was true of those techniques. It was written twelve weeks before a different set of techniques closed the problem, which is the more useful fact about the pace.
The other proof broke in eight hours
The same Friday, a second claimed proof of the Generalized Riemann Hypothesis appeared, and it is the mirror image of the first. Ravi Vakil, president of the American Mathematical Society, posted a 940-page manuscript by Stanford postdoc Alessandro Masullo, writing on his blog that Brian Conrad had been examining it for months and giving extensive feedback, and that "it is all human, with no AI."
Eight hours later, Vakil appended an update: several people had reported problems in the proposed proof, "identified with the assistance of different AI models."
Read those two paragraphs together and the actual state of the field is visible. The machine-generated proof has passed every mechanical test and awaits human comprehension. The human proof has human comprehension built in and was found wanting, in under a day, by machines reading it. Brian Conrey, one of the leading specialists on the zeta function, told AFP he had not yet seen a manuscript, only the remarks, and that "I don't think any human was about to find this" — while predicting "there will be simpler ways to do it."
That last clause is the one to hold onto. A 17.4-million-line proof is not how mathematics ends up knowing something; it is how a search process records a result. Alpöge himself expects the argument to be "massively simplified, likely immediately." The open question is whether simplification produces a proof a number theorist can read, check and teach — or whether the field has to accept a result on the strength of its certificate, the way it accepts a computer-checked enumeration it cannot inspect by hand.
What to watch
Sutherland's page is the scoreboard: the comparator run and the paranoid-kernel replay are the last mechanical gates, and both had passed their statement-comparison and axiom stages. Watch for the readable write-up Alpöge has promised, and for whether the theta-lift-as-multiplicity idea survives contact with a human reader who can generalize it — that, not the line count, is what would make this a method rather than an artifact. Watch whether anyone attacks the parts Alpöge says he does not fully understand himself. And watch the Extended Riemann Hypothesis claim, which is currently a sentence in a note and nothing more.
Meanwhile the Clay prize sits where it has sat since 2000, behind a two-year clock that has not started. The proof may be finished. The mathematics is not.
If a theorem compiles but no mathematician can yet explain it, has it been proved or only certified? Write to us via our contact page.







