Lean Community Completes Formal Verification: Confidence That Anthropic Mathematician's Riemann Hypothesis Proof Is Correct

Deep News
8 hours ago

On November 21, the open-source formal verification community Lean and Mathlib issued a joint statement confirming that the proof of the Riemann Hypothesis by Anthropic mathematician Levent Alpöge is valid. The two organizations jointly endorsed it, stating:

"We are fully confident that Alpöge's proof of the Riemann Hypothesis is correct."

On November 20, the day after Anthropic's IPO, Alpöge announced that he had worked with the AI model Claude to find a proof of the Riemann Hypothesis and the Generalized Riemann Hypothesis, and publicly released over 17 million lines of Lean source code. Because the text version of the proof exceeded 250,000 pages and was difficult to read, it was not made public, and combined with the timing of the announcement being so close to the IPO, outsiders initially questioned its authenticity and commercial motives. Odds on prediction platforms also reflected market disagreement.

Twenty-four hours later, the Lean community provided a computer-level response using quadruple formal verification. But machine verification and human comprehension of a proof are two different things — a kernel written by Claude even appeared in the verification chain, some final tests have not yet concluded, and claiming the Clay Mathematics Institute's $1 million prize would still require at least two years of public scrutiny.

24 hours to verify 17 million lines of code: How Lean validates, and where the boundaries lie

Traditional mathematical breakthroughs rely on peers reviewing papers one by one and judging whether each step of reasoning holds. But faced with a proof draft exceeding 250,000 pages that no one could read through in the short term, the formal verification software Lean became the fastest "notary office" in the mathematics world.

Lean was founded by Brazilian computer scientist Leonardo de Moura in 2013, initially developed at Microsoft Research, later continued at Amazon Web Services AWS, and in 2021 a rewritten version Lean 4 was released, now supported by Lean FRO. In this programming language and proof assistant, mathematicians must write every definition, prerequisite conclusion, and reasoning step as computer-readable code, with the program verifying whether the logic is rigorous step by step.

Its community-shared mathematics library Mathlib, started in 2017, has exceeded 1 million lines, and its documentation even includes a footnote about the Riemann Hypothesis, saying that constructing its proof is "worth a million dollars." Lean was used in 2021 to verify Peter Scholze's condensed mathematics proof, in 2023 assisted Terence Tao in formalizing the proof of the polynomial Freiman-Ruzsa conjecture, is also the underlying tool for Google DeepMind's AlphaProof, and has been used for Amazon's Cedar policy language and Rust code verification.

But code compiling successfully does not mean the Riemann Hypothesis has been conquered. Mathematician Kit Yates pointed out in a Substack column that although Lean verification is far more rigorous than having another AI judge, experts still must additionally confirm three things: that the code proves the Riemann Hypothesis and not a weaker proposition, that the definitions faithfully reproduce standard objects such as the zeta function, and that no additional assumptions were smuggled into the reasoning. The Lean language reference manual also warns that the system cannot check whether new axioms are self-consistent, and one must be especially wary of the placeholder "sorry" that can "prove everything."

The Lean community dared to make a decision within one day because Alpöge's codebase passed all four hard criteria proposed by Imperial College London mathematician Kevin Buzzard on November 12.

The first hurdle was verifying the "question statement." The codebase directly depends only on Mathlib, and the theorem it proves points to Mathlib's built-in standard definition of RiemannHypothesis — that among all complex numbers making the zeta function zero, apart from trivial zeros such as -2, -4, -6, the real part is always 1/2. This definition had already been repeatedly verified by Lean and number theory experts.

The second hurdle was independent reproduction. The codebase, split into 14 compressed files covering 175,187 modules, has extremely high computational requirements; a complete verification by a single person requires approximately 100GB of memory and 70GB of hard drive space. MIT mathematician Andrew Sutherland independently completed an error-free compilation using Lean and Mathlib v4.33.1, and the Mathlib files matched the official cache. The joint statement showed that dozens of people had reproduced the compilation and generated completely identical raw proof export files, totaling 61.6GB.

The third hurdle concerned "hidden manipulation." The community separately created an isolated file that imports only Mathlib, and used two tools — comparator and con-leche --compare — to compare, both confirming that the proposition proved by Alpöge exactly matches Mathlib's original definition. On axiom review, Sutherland ran the #print axioms command on five theorems in the main file Main.lean, and the results showed they depend only on three core axioms out of Lean's seven standard axioms — propext, Classical.choice, and Quot.sound — contain no sorryAx, and all originate from the core library. "The probability that there's something fishy going on here is approximately zero," Sutherland told Scientific American.

The final hurdle targeted the verification software itself. Lean's official kernel has historically had reliability vulnerabilities, which were fixed within 4 hours of discovery, and mathematician Elliot Glazer estimated the probability of Lean itself having vulnerabilities at below 1%. To isolate potential malicious code in the source code and circumvent single-kernel flaws, the community skipped the source code and directly handed the 61.6GB lazy export file to multiple independent kernels in different languages for cross-verification.

In addition to Lean's official kernel and leanchecker and lean4lean, which fully pass all 175,187 modules, the third-party kernel nanoda written in Rust took 27 minutes and 42 seconds to accept the complete export; con-leche took only 3 minutes and 15 seconds to review 1.468 million declarations with none rejected; con-ron also passed. con-leche and con-ron themselves also come with formal self-consistency proofs relative to set theory, mathematically and logically ensuring they cannot accept an incorrect proof.

The effect of passing verification has already manifested in specific mathematical corollaries. Combined with existing formalization results, Sutherland's page confirms that the Robin inequality and Lagarias inequality concerning the sum of divisors of integers have been derived from this proof, also depending only on the three standard axioms. In 1984, French mathematician Guy Robin proved that for any integer n greater than 5040, the sum of all its divisors is less than [expression] — this intuitive proposition is completely equivalent to the Riemann Hypothesis — where γ is the Euler constant and e^γ is approximately equal to 1.78.

However, comparing the Lean joint statement with Sutherland's MIT verification page, this verification still has several technical boundaries that have not been completely closed.

Although the Lean joint statement announced that both comparator and the official kernel have passed, it did not publish specific tool versions and detailed output logs. As of 12:55 UTC on November 21, Sutherland's own end-to-end comparator run on Lean v4.36.0 had completed proposition comparison and axiom checking, while the kernel phase was still continuing, and Lean's default and strict kernels had not yet produced results; his earlier build and export were not completed in a sandbox and still await comparison with sandbox build results. As for the older versions of con-leche and con-ron bundled with Lean v4.36.0, which returned exit code 2 because they do not support a certain type in the proof, Sutherland's page noted that this does not mean the proof was rejected.

There is also a discrepancy in the scope of the proof. Elliot Glazer stated on November 20 that the formalized proposition is actually the Generalized Riemann Hypothesis, while Sutherland's page is more cautious: the main theorem is the Riemann Hypothesis as defined by Mathlib, and the other four theorems target Dirichlet L-functions, with the page mentioning "GRH" only in this limited sense. Sutherland also reminded that even if the axiom check is correct, the final conclusion still presupposes that Mathlib's definitions are faithful to mathematical intent and that the verification kernel itself is reliable.

From proof to "grading"

Looking behind the scenes, Claude's presence runs through multiple stages from proposing the proof to participating in "grading."

According to the research timeline attached by Alpöge, the work began on October 6 with the "Stepanov determinant" and was completed in the late night of November 15 to the early morning of November 16 local time, with only the three letters "GRH" remaining on the November 16 record. This timeline was reportedly organized by Claude. Alpöge stated that he was responsible for indicating the direction, and Claude then proposed the breakthrough step — a "theta lifting" chain whose length is approximately equal to the square root of the defined prime, with each lifting corresponding to the "multiplicity" that classical function field arguments depend on but is extremely difficult to find over the integers. "Near the end, Claude far surpassed me. No, Claude was simply flying, dancing, and shining near the finish line," he wrote in the statement.

At the verification stage, Claude was still present. The signature at the top of Sutherland's verification page shows that the entire set of checks was run through Claude Code; and con-leche, which reviewed 1.468 million declarations in 3 minutes and 15 seconds, was introduced by Lean founder Leonardo de Moura in a blog post last month as the first kernel written in Lean and proven self-consistent — both the code and the self-consistency proof were completed by Claude, with strict oversight by Lean FRO's Joachim Breitner. Because of the formal self-consistency proof as a backstop and the final conclusion being consistent with multiple third-party independent kernels such as the Rust kernel nanoda, verification credit does not depend on a single tool.

This pace of progress exceeded Anthropic's own expectations from three months ago. In August of this year, the company published research results using an unreleased version of Claude with Lean, pushing the lower bound on the proportion of critical-line zeros from 41.6% to 67.2%, and passing comparator verification. At the time, Anthropic explicitly stated that it did not expect these techniques to ultimately prove the Riemann Hypothesis.

A leap achieved just three months later happened to coincide with the sensitive window after the November 19 IPO. In his statement, Alpöge proactively admitted that "the precise timing is very bad," revealing that after completing the work he had asked Anthropic CEO Dario Amodei whether it could be made public, and "ultimately we figured out that there were indeed restrictions that needed to be observed." He denied negative speculation about the release timing, saying that this "obviously cynical view" was wrong, and frankly stated that he "simply could not wait any longer," and was willing to take full responsibility.

Public frenzy and academic restraint

At 23:12 Beijing time on November 21, the Lean official announcement was posted on the social platform X, quickly sparking discussion in the tech world and mainstream academia.

The AI community quickly grew excited: AI researcher Yao Shunyu had already posted on the evening of the 20th, exclaiming "OMG, Riemann Hypothesis?!" After the official statement came out, AI news writer Rohan Paul summarized it as "After 167 years, the Riemann Hypothesis has finally been completed by AI," while the commentary account Dr Singularity marveled, "What a time to be alive." Netizens also dug up an old post from 2022 in which OpenAI CEO Sam Altman half-jokingly asked on X, "When will the Riemann Hypothesis be solved?" and joked that "Anthropic beat you to it."

Mainstream scholars were noticeably more cautious in their wording. On the morning of November 21, before the Lean official joint statement was published, Columbia University professor of mathematics and physics Brian Greene posted introducing that AI "may have conquered the Riemann Hypothesis, which stands above all difficult problems," but his wording left room and was not his own independent verification. Although Rutgers University mathematics department chair Alex Kontorovich grew increasingly optimistic about the proof's success, his action on X that day was merely to repost Quanta Magazine's report on Alpöge's statement.

On the question of credit, academia also refused to simplify this into "AI solving the problem alone." A mathematics and AI researcher at Princeton University emphasized that human participation remains the best way to use large language models to solve mathematical problems, and Alpöge should receive equal recognition alongside Claude. After independent compilation passed, Sutherland also called it "a victory for humanity."

The more difficult reality is that the machine has verified it, but humans still cannot understand it. University of Manchester mathematician Thomas Bloom said bluntly on X that he could not tell whether what shocked him more was that AI claimed to have proved the Riemann Hypothesis, or that AI generated a proof that, when converted to human-readable text, exceeded 250,000 pages and was logically coherent.

2006 Fields Medal winner Terence Tao offered a preliminary judgment on his personal blog: the proof appears to comprehensively use modern research results from the Langlands program and arithmetic geometry, and verification and digestion will still take time. "Fortunately, this time, at least there is a human author to communicate with."

Next, breaking down and simplifying this massive proof will become a core task for the mathematics community. The Lean and Mathlib communities stated at the end of their announcement that they look forward to the mathematics community's subsequent work explaining the new ideas in the proof. Alpöge himself also predicted that this astonishingly large proof "is very likely to be greatly simplified, and possibly quite soon."

Even though the computer has given an affirmative answer, formal coronation by the mathematics community must still cross a long institutional threshold. According to the rules of the Clay Mathematics Institute, the $1 million prize for the Millennium Prize Problems requires that the solution be formally published, undergo at least two years of public scrutiny, and receive widespread recognition from the international mathematics community.

For Alpöge, what he calls "a dawn of something greater in the scientific field" is still several years away from truly being settled.

Disclaimer: Investing carries risk. This is not financial advice. The above content should not be regarded as an offer, recommendation, or solicitation on acquiring or disposing of any financial products, any associated discussions, comments, or posts by author or other users should not be considered as such either. It is solely for general information purpose only, which does not consider your own investment objectives, financial situations or needs. TTM assumes no responsibility or warranty for the accuracy and completeness of the information, investors should do their own research and may seek professional advice before investing.

Most Discussed

  1. 1
     
     
     
     
  2. 2
     
     
     
     
  3. 3
     
     
     
     
  4. 4
     
     
     
     
  5. 5
     
     
     
     
  6. 6
     
     
     
     
  7. 7
     
     
     
     
  8. 8
     
     
     
     
  9. 9
     
     
     
     
  10. 10