Big Tech

Anthropic Researcher Claims Riemann Hypothesis Proof

Anthropic's Levent Alpoge used Claude to claim a proof of the $1M Riemann Hypothesis, publishing 175,187 Lean 4 files MIT has verified.

Share:XLinkedIn

Key Takeaways

  • 175,187 Lean 4 files, 61.5 GB unpacked: Levent Alpoge released the largest formal mathematics archive ever associated with a Millennium Prize Problem, covering both the classical Riemann Hypothesis and the Generalized form for all Dirichlet L-functions.
  • MIT verified all 175,187 modules with zero errors: Andrew Sutherland ran five independent kernel checks using 100+ GB RAM and 27 minutes of compute, every check passing with no failures.
  • Claude supplied the decisive sqrt(X) transformation insight: the precise count that makes Stepanov's method work on the Riemann Hypothesis came from Claude, not from the human mathematician guiding the high-level approach.
  • A competing all-human proof by Alessandro Masullo failed within 8 hours: problems were found by other researchers the same day, with some identified using AI tools, leaving Alpoge's archive as the only surviving candidate from November 20.
  • $1 million Clay Prize requires 2 more years of community review: the Clay Institute mandates peer-reviewed publication plus a two-year scrutiny window before any prize consideration begins, putting the earliest possible prize date at late 2028.

Levent Alpöge did not wait for peer review. The Anthropic mathematician and Harvard Society of Fellows junior fellow announced on November 20, 2026, that he and Claude had produced a formal proof of the Riemann Hypothesis and its Generalized form covering all Dirichlet L-functions, one of the seven Millennium Prize Problems, each worth $1 million from the Clay Mathematics Institute. The claim, backed by 175,187 Lean 4 source files that MIT has machine-verified without errors, is unconfirmed as mathematics. That tension is exactly the story.

What Alpöge and Claude Published and Why It Weighs 61 Gigabytes

The Riemann Hypothesis was proposed by Bernhard Riemann in 1859. It says every nontrivial zero of the Riemann zeta function lies on the vertical line in the complex plane whose real part is exactly one-half. No one has proved it in 167 years, and the Clay Mathematics Institute has listed it as one of seven Millennium Prize Problems since 2000. Alpöge's four-page written statement, posted on his personal website, says: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." The claim extends to the Generalized Riemann Hypothesis, which covers all Dirichlet L-functions, a broader class of mathematical objects connected to prime number distribution and to some of the deepest open questions in analytic number theory.

The release is a 14-part Lean 4 archive. Compressed, it weighs 239.3 megabytes. Decompressed, it expands to more than 61.5 gigabytes containing 175,187 source files, according to CASRAI's technical review. If the full argument were printed on paper, the resulting document would run to more than 250,000 pages, per The Decoder. The core statement appears in a file called Main.lean, formalized against Mathlib, the standard mathematics library for Lean 4. The axioms used are propext, Classical.choice, and Quot.sound, the same three standard axioms every Mathlib proof relies on. The release contains no sorry or admit placeholders, which are the Lean equivalents of "proof left to the reader."

This announcement comes at the end of a research trajectory. In August 2026, Alpöge and Anthropic colleague Ralph Furman published a peer-reviewed paper on arXiv showing that at least 67.25007 percent of the nontrivial zeros of the Riemann zeta function are simple and lie on the critical line, with the proof autonomously discovered by Claude and independently Lean-verified. That was the highest-ever confirmed lower bound on zeros and drew external review from number theorists Brian Conrey and Dan Goldston. The November 20 claim goes further: it asserts the full statement, with no zeros permitted off the critical line at all, extending to all Dirichlet L-functions under the Generalized form.

Stay Ahead

Get daily AI signals before the market moves.

Join founders, investors, and operators reading TechFastForward.

How Stepanov's Method Became a Path to the Full Proof

The technical approach traces back to Sergei Stepanov, who in the early 1970s solved a related problem over finite fields by constructing a polynomial function whose roots encode the information needed. Stepanov's method produces exact bounds rather than approximations, which makes it unusually powerful compared to the sieve methods and density arguments that most previous progress on the Riemann Hypothesis relied on. The challenge is that the original method was designed for finite fields, not for the complex numbers where the Riemann zeta function lives. Standard derivatives fail on integers, which blocked the obvious extension for decades and left Stepanov's approach as a curiosity that number theorists assumed would not transfer.

Alpöge's approach maps the problem into a higher-dimensional setting using automorphic forms and a construction called the theta lift, which moves mathematical objects between different spaces in a controlled way. The lift is applied across totally definite unitary groups, drawing simultaneously on arithmetic algebraic geometry, automorphic forms on higher-rank Lie groups, Iwasawa theory, and spectral analysis. These are four distinct research programs that rarely appear in a single argument. The fields involved are deep enough that Alpöge told The Decoder he has not "fully mastered" some of the surrounding mathematics and plans a human-readable version over the coming months.

The key numerical insight came from Claude, not from Alpöge. For a sequence of length X, you need exactly √X transformations to make the Stepanov method work in this setting. That is not an approximation or a heuristic: it is a precise count that determines whether the construction produces the bounds the proof needs. Alpöge says he directed the high-level architecture, moving ideas from toy models to higher-rank automorphic forms. Claude then handled most of the algebraic manipulation, trace formula computations, and technical lemmas in the final stages. Scientific American reported Alpöge's statement that Claude "moved beyond his own command" of some parts of the mathematics.

The timeline as described in alpo.ge/remark.pdf places the final pieces coming together on the night of Sunday, November 15. The entry for November 16 is labeled "GRH," marking when the Generalized Riemann Hypothesis extension was completed. Alpöge then consulted with Anthropic CEO Dario Amodei and colleagues before releasing on Friday, November 20, one day after Anthropic's stock market debut. He denied the timing was deliberate, saying he sought clearance from leadership before proceeding and rejected any suggestion of a publicity stunt.

MIT's Five-Check Verification Table Passes at Zero Errors

Andrew Sutherland, a number theorist at MIT, independently downloaded Alpöge's 14-part archive and ran it through five separate formal verification tools. The results are unambiguous in one narrow sense: the code is internally consistent. All 175,187 Lean 4 modules compiled without errors. A replay using the leanchecker tool accepted every one of the 175,187 declarations. The nanoda Rust kernel, running on 80 threads, exited with code zero in 27 minutes and 42 seconds. A con-leche comparison matched 51,692 challenge declarations against the specification. The lean4lean tool confirmed all 175,187 modules.

Sutherland said afterward the checks required more than 100 gigabytes of RAM, hundreds of gigabytes of disk space, and dozens of processor cores to complete. He described the moment as "historic" and said the most important step was "building a machine that can think." He called it an event "worth celebrating" and said a full understanding of the proof "is going to take some time." These statements confirm that Sutherland endorses the significance of the moment while stopping well short of endorsing the proof as mathematically understood.

Senior number theorists cited by CASRAI are careful to distinguish machine verification from mathematical proof. The Lean kernel confirms the code is internally consistent, but the kernel cannot determine whether the Mathlib definitions in Main.lean faithfully represent the actual mathematical objects of number theory. Those definitions are human-made. If the Mathlib formulation of the Riemann zeta function, or of the Dirichlet L-functions at the core of the GRH claim, differs in any way from the standard definition, the proof would pass every kernel check and still not prove what it claims. This is not a theoretical concern: subtle definition mismatches have been found in formal mathematics projects before, and no number theorist outside the verification team has yet publicly reviewed the definitions used here.

The Stanford Counter-Proof That Collapsed by Nightfall

Within hours of Alpöge's announcement, Stanford mathematics professor Ravi Vakil, who also serves as president of the American Mathematical Society, posted a competing draft. The proof was written by Alessandro Masullo, an independent researcher. Stanford's Brian Conrad had reviewed the work for months before the posting. "It is all human, with no AI," Vakil wrote, in what read as a direct contrast to Alpöge's approach. The competing proof offered the mathematical community what many people wanted: a human-readable argument they could examine in the traditional way, with techniques that mathematicians could check line by line without compiling gigabytes of code.

Within approximately eight hours, Vakil posted an update. Several researchers had identified problems in the Masullo draft. Notably, some of those problems were found with the help of AI models, which added a pointed dimension to Vakil's earlier emphasis on human authorship. Vakil wrote: "all issues need to be resolved to the satisfaction of the community before this proof would be accepted as true."

The Masullo reversal matters for two reasons. First, it demonstrates how fragile unverified proofs of hard problems tend to be, even from credible researchers with months of expert review behind them. Second, it left Alpöge's Lean archive as the only surviving candidate from November 20, which shifts pressure toward the Lean verification community to find the first error, if one exists. The fact that a human-readable proof failed faster than a machine-verified one is not proof that formal mathematics outperforms traditional argument, but it is evidence that the check the Lean kernel provides is not easily bypassed.

What Clay's Prize Rules Mean for a 61-Gigabyte Lean File

The Clay Mathematics Institute's rules are specific: a proof must be published in a peer-reviewed journal of international standing and then survive at least two years of community scrutiny before the prize committee considers awarding the $1 million. Alpöge's release does not currently qualify for prize consideration on either count. There is no peer-reviewed paper yet, only a four-page written statement and a 61.5-gigabyte Lean archive that has passed machine checks but not human expert review. The two-year clock, if it starts, would not expire until late 2028 at the earliest.

The bear case is straightforward. A formal proof can be internally consistent while proving a subtly different statement from the one the mathematical community cares about. The question is whether the Mathlib definition of the Riemann zeta function, or of the Dirichlet L-functions at the heart of the GRH claim, encodes the correct mathematical objects at the correct level of generality. If they do not, the proof would be a valid theorem about something slightly different from what it claims to be, and the error would be invisible to the Lean kernel. Number theorists who have not been involved in building Mathlib's mathematical library would need to audit those definitions carefully before endorsing the result, and that review has not yet happened.

Hector Pasten, a number theorist at the Pontifical Catholic University of Chile, told Interesting Engineering that if the proof is correct and understood, it "completes a dictionary between different branches of math," comparable to a Rosetta Stone. Carlo Pagano, a researcher at Concordia University and a visiting researcher at Google DeepMind, said the new ideas could open new avenues in number theory even if further review identifies gaps. What to watch in the next 90 days: whether any number theorist publicly endorses the correctness of the Mathlib definitions; whether Lean experts on the community Zulip channel flag a semantic gap between the code and the actual hypothesis; and whether Alpöge's promised human-readable draft appears before mid-February 2027. These three signals will tell observers more about the proof's prospects than the kernel checks already have.

What It Means That the Decisive Insight Came From Claude

Alpöge's disclosure that the √X transformation count came from Claude, not from him, is not a footnote. It is the most consequential claim in the announcement, independent of whether the proof is eventually verified. For 167 years, the Riemann Hypothesis resisted every human attempt, and the attempts were made by mathematicians including Riemann himself, Hilbert, Hardy, Littlewood, Selberg, and many others with decades of specialized training. The √X count is not the kind of result that comes from pattern matching or from recombining known techniques: it is a structural insight into how many times a specific mathematical operation must be applied to produce a desired bound in a novel setting.

Andrew Sutherland made a remark at an October 2024 workshop that an AI proving the Riemann Hypothesis would make mathematicians' jobs "the least of our problems." That comment reflected a widespread assumption that the Riemann Hypothesis was structurally out of reach for AI systems, requiring a form of deep mathematical creativity that large language models were assumed to lack. Alpöge's archive is the first empirical test of that assumption at the actual problem. At minimum, it shows that Claude operating in a formal mathematics environment can contribute specific, identifiable mathematical structure to a claimed proof of a Millennium Prize Problem, even when the human collaborator has not fully mastered all of the surrounding mathematics.

What comes next has no precedent. The mathematical community has never reviewed a 250,000-page formal proof of a Millennium Problem before. Traditional peer review involves human experts reading and checking each step against known results. For a Lean archive of this size, the review process will likely require mathematicians who specialize in both number theory and formal verification, a combination that is rare. The Clay Institute has not publicly commented on how its prize rules apply to a machine-verified result with no accompanying human-readable paper, which is itself a question the mathematical community will need to resolve before the prize conversation can even start. If the proof survives the next 24 months of scrutiny, the prize rules will need to be tested against a scenario their authors almost certainly did not anticipate.

Claude found the √X insight that eluded every human mathematician for 167 years. Whether the proof holds, that fact does not go away.


Key Takeaways

  • 175,187 Lean 4 files, 61.5 GB unpacked: Levent Alpöge released the largest formal mathematics archive ever associated with a Millennium Prize Problem, covering both the classical Riemann Hypothesis and the Generalized form for all Dirichlet L-functions.
  • MIT verified all 175,187 modules with zero errors: Andrew Sutherland ran five independent kernel checks using 100+ GB RAM and 27 minutes of compute, every check passing with no failures.
  • Claude supplied the decisive √X transformation insight: the precise count that makes Stepanov's method work on the Riemann Hypothesis came from Claude, not from the human mathematician guiding the high-level approach.
  • A competing all-human proof by Alessandro Masullo failed within 8 hours: problems were found by other researchers the same day, with some identified using AI tools, leaving Alpöge's archive as the only surviving candidate from November 20.
  • $1 million Clay Prize requires 2 more years of community review: the Clay Institute mandates peer-reviewed publication plus a two-year scrutiny window before any prize consideration begins, putting the earliest possible prize date at late 2028.

Questions Worth Asking

  1. If the Lean kernel confirms every step but no human mathematician can read the full 250,000-page argument, does the mathematical community have a standard for accepting or rejecting it, and who decides what that standard is?
  2. The competing human proof failed within hours partly because AI found errors in it. What does it say about the future of mathematical proof that AI is simultaneously the prover and the reviewer for the same class of problem?
  3. If Claude contributed the decisive √X insight, what does authorship of a mathematical theorem mean when the key structural idea came from a model trained on human-generated mathematics?

Sources

  1. Alpöge's four-page written statement alpo.ge
  2. CASRAI's technical review casrai.org
  3. The Decoder the-decoder.com
  4. a peer-reviewed paper on arXiv arxiv.org
  5. Scientific American scientificamerican.com
  6. Interesting Engineering interestingengineering.com

Read Next

Cognition Weighs $75B Round as Devin Nears $1B Revenue

Senate Passes Permitting Reform to Fast-Track AI Data Centers

RayNeo Air 4 Pro Undercuts Rivals at $299 With HDR10 Display

Gemini API Cuts Token Overhead for Agents With Many Tools

Newsletter

Enjoyed this analysis? Join the list for the email edition.

Daily AI signals. No noise. Built for founders, investors, and operators.

Share:XLinkedIn
</> Embed this article

Copy the iframe code below to embed on your site:

<iframe src="https://techfastforward.com/embed/anthropic-researcher-claims-riemann-hypothesis-proof" width="480" height="260" frameborder="0" style="border-radius:16px;max-width:100%;" loading="lazy"></iframe>