AIsaAIsaStudio English 打开 AIsa Studio

Anthropic 数学家宣布 Claude 协助证明黎曼猜想,斯坦福随后发布纯人类竞争草稿

原标题:Hours after Anthropic claims a Claude proof of the Riemann hypothesis, Stanford releases an all-human competing draft

Anthropic 数学家 Levent Alpöge 宣布与 Claude 合作完成黎曼猜想及广义黎曼猜想(GRH)的证明草稿,附完整代码,证明以 Lean 语言发布、打印将超过 25 万页。数小时后,斯坦福 AMS 主席 Ravi Vakil 发布了数学家 Alessandro Masullo 历经 Brian Conrad 数月审阅的纯人类证明草稿,但约 8 小时后已有人在 AI 协助下发现其中问题,Vakil 表示所有问题需经学界确认后证明才能成立。

为什么值得看

原文梳理了 AI 证明与人类证明的竞争细节和验证方式,读者可以据此了解两份草稿各自的依据和后续审查流程。

阅读原文 ↗11月22日 AI 日报

本页的摘要、推荐理由与解读出自 AIsa Studio 的 AI 风向智能体,原文版权归原作者所有。

原文

Levent Alpöge, a mathematician at Anthropic, said Friday he used a large language model to solve math's most famous open problem.

Alpöge posted a paper and the accompanying computer code on his website, according to Scientific American and other US outlets. "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis," he wrote in his announcement.

According to Alpöge, his method goes beyond the original prime-number problem and also resolves the Generalized Riemann Hypothesis (GRH), which extends that ordering principle to a large family of related equations. Given a bit more time, he expects to solve the even broader Extended Riemann Hypothesis as well.

Alpöge had treated the problem as a secret lifelong goal since college, expecting it to take decades. Over the past few months, he worked closely with Anthropic's AI model, setting the direction and combining two approaches on a simplified toy model. Once they moved to the actual problem, the AI took over and left him far behind near the finish line.

To crack the puzzle, the pair extended an idea from mathematician Stepanov, who had solved a similar question in another mathematical system. That method relies on a tool similar to taking derivatives of smooth curves in high school calculus, a trick that normally fails on whole numbers because individual values cannot be differentiated like a continuous line.

Alpöge and Claude got around that obstacle by chaining special mathematical transformations together like links to mimic the effect of a derivative. The language model supplied the key insight, finding that for a number sequence of length X, you need to chain exactly as many transformations as the square root of X. Ideas like that alone should leave fellow mathematicians stunned, Alpöge says.

Stanford mathematicians rush out an all-human competing proof

Hours after Alpöge's announcement, the story took a dramatic turn. Just before a board meeting, Stanford math professor and American Mathematical Society (AMS) president Ravi Vakil posted a "hurried announcement" on his blog, sharing an independent draft proof of the Generalized Riemann Hypothesis written by mathematician Alessandro Masullo.

"It is all human, with no AI," Vakil wrote. Renowned Stanford number theorist Brian Conrad had quietly reviewed Masullo's work for months, offering extensive feedback that is now being incorporated, Vakil said.

Masullo was originally supposed to spend a year at Stanford refining the draft before going public, but Anthropic's sudden AI announcement forced the researchers to release their current version right away.

About eight hours after his initial announcement, Vakil updated his post, saying several people had contacted him to report problems in the proposed proof, identified with the assistance of different AI models. Several mathematicians in particular were first to flag the issues, he said. "As is usual, all issues need to be resolved to the satisfaction of the community before this proof would be accepted as true. (I make no claim as to the correctness of the proof strategy.)," Vakil wrote.

Why both proofs could hold up and what happens next

Mathematicians are taking both claims seriously right away because of the groundwork behind them. Masullo's traditional proof comes with months of vetting by Brian Conrad, while Alpöge backs his AI proof with full computer verification.

Printed out, Claude's argument would run over 250,000 pages, Alpöge says, because the AI spells out even basic rules, such as a number other than one-half being either larger or smaller than one-half. To spare mathematicians that wall of text, Alpöge released the proof only in the Lean programming language, split into 14 data packages.

Lean acts like a digital referee, checking step by step whether every logical deduction follows from math's basic rules without errors or unproven assumptions. Human error is still possible if the original question was translated into code incorrectly at the start, though checking whether Alpöge coded the starting problem correctly should not take experts long, according to Scientific American. Vakil, meanwhile, urged mathematicians to quickly review Masullo's draft, translate it into systems like Lean, and close any small gaps.

Alpöge admits Claude was so far ahead of him in some complex subfields that he is presenting a result whose surrounding math he has not fully mastered yet. He expects the massive AI proof can be shortened quickly and plans to write a human-readable version over the coming months.

Anthropic's announcement also lands in a high-stakes business week, just a day after the company's major IPO on Thursday. Alpöge says he made the breakthrough late Sunday night and then contacted Anthropic CEO Dario Amodei for clearance. He calls the timing right after the public debut extremely unfortunate, but rejects any suggestion of a PR stunt, saying waiting any longer to publish simply was not an option.

This caps a fast-moving 2026 in which Claude has already solved several open math problems. In July, Alpöge published a Claude-generated solution to the well-known Jacobian conjecture. September brought a fierce dispute over the Navier-Stokes equations, which describe how fluids and gases flow, when NYU mathematician Tristan Buckmaster suggested that OpenAI had used prior work by him and Alpöge. In October, OpenAI took an important first step on prime numbers with a major release of results. Alpöge explicitly praised the OpenAI researchers' problem selection, saying it gave him key ideas for his own proof.

Riemann's 1859 hypothesis seeks hidden order in prime numbers

Journalist Joseph Howlett traced the core of the problem in a detailed Scientific American analysis. Since Bernhard Riemann proposed his hypothesis in 1859, it has served as number theory's north star. David Hilbert placed it on his 1900 list of math's most important problems, and since 2000 it has carried a $1 million prize as one of the Clay Mathematics Institute's seven "Millennium Problems."

At the center are prime numbers like 2, 3, 5, 7, and 17, which divide evenly only by one and themselves. They act as the indivisible atoms of math because every other whole number is a unique product of primes, like 15 from 3 times 5. Stanford mathematician Brian Conrad compares their role in number theory to forces in physics.

Around 300 BC, Euclid proved there are infinitely many primes, though their exact positions on the number line look random. They do grow rarer as numbers get larger, dropping from 25 primes between 0 and 100 to just 16 between 1,000 and 1,100. As a teenager in the late 18th century, Carl Friedrich Gauss used long tables to find a formula that roughly estimates how many primes appear in a given range, yet real primes keep drifting unpredictably from that estimate.

Riemann, a student of Gauss in Göttingen, looked for a way to calculate that error exactly and pinpoint every single prime. His tool is the Riemann zeta function, which works like a calculating machine over a two-dimensional map with horizontal and vertical axes. Feed the machine a point's coordinates, and it outputs a new value. At certain points the result is exactly zero. Mathematicians call these points zeros.

Riemann saw that these zeros explain the apparent chaos of primes. Much like breaking a musical note into individual sound waves or overtones, the zeros correct Gauss's estimation error. Each zero acts like an instrument in an orchestra, with its vertical position setting the pitch and its horizontal position setting the volume.

Because calculating infinitely many points one by one was impossible, Riemann conjectured in 1859 that all the key zeros, known as non-trivial zeros, lie on a single vertical line at the horizontal value of one-half. "The Riemann hypothesis says all these different instruments in the orchestra of the primes play at exactly the same volume," Rutgers University's Alex Kontorovich explains. Apparent randomness turns into strict harmony.

A confirmed proof would validate hundreds of math theories and reshape AI's role

If at least one of the two proofs holds up, number theorists would gain perfect control over how primes are distributed across the entire number line. Because a proof seemed unreachable for so long, mathematicians worked ahead for decades. Hundreds of papers assume the Riemann hypothesis is true, according to Kontorovich, and all of those theories would become established theorems overnight. Since both Alpöge's AI method and Masullo's work cover the Generalized Riemann Hypothesis, the breakthrough would also extend to fields ranging from simple equations to high-dimensional geometry.

The links also reach into computer science and physics. Prime numbers form the foundation of modern digital encryption. In 1972, physicist Freeman Dyson accidentally discovered over tea with a mathematician that the spacing of Riemann's zeros matches the energy levels inside atomic nuclei. That same formula keeps showing up in random particle motion, chaos theory, and black hole calculations.

Before this sudden double breakthrough, the field had hit a wall. Aside from quiet efforts like Masullo and Conrad's, almost no top mathematicians worked directly on the problem because they lacked a starting point. Two years ago, Oxford's James Maynard and MIT's Larry Guth showed that not too many points can stray far from Riemann's line. Though the biggest advance in decades, it struck Maynard himself and colleagues like Andrew Granville of the University of Montreal as only a minor improvement and stopgap.

That made the prospect of an AI cracking the puzzle all the more daunting. At a Harvard University workshop in October 2024, MIT number theorist Andrew Sutherland summed up the stakes. "An AI that can prove the Riemann hypothesis is not one I'd want to meet. If that happens, mathematicians having jobs will be the least of our problems."

Even if Masullo's work saves human pride in this head-to-head race, models like Claude have changed fast. In just a year, they have gone from unreliable tools to systems that break through deep roadblocks and push the math community into an identity crisis. Alpöge still feels a mix of sadness at losing a 150-year-old guiding star and optimism about what comes next. "But what I saw during the course of progress so amazed me that I believe it is one dawn across science of something far greater."

原文版权归 The Decoder 及原作者所有,本站转载仅供阅读。如需删除,请联系 support@aisa.one。

事件综述

Anthropic 数学家称与 Claude 合作证明黎曼猜想,同周现人类竞争证明

2026 年 11 月 20 日,Anthropic 数学家 Levent Alpöge 公开声称在与公司 AI 模型 Claude 的反复交互中完成了黎曼猜想(1859 年提出、悬赏的千禧年大奖难题)的证明,据称证明以 Lean 语言形式化且无额外假设,文件已公开,打印将超过 25 万页。据 X 上流传的说法,该方法用迭代 theta lift 替代整数上不可用的求导来叠加零点重数,长度为 X 的和约需 √X 轮,据称该轮数由 Claude 提出;Alpöge 称方法可推广至广义黎曼猜想(GRH),甚至预期也能解决扩展黎猜想。报道 2 与报道 8 在完成范围上存在差异:早先报道称 Claude 已解决广义黎曼猜想,后续报道称其与 Claude 完成的是黎曼猜想及 GRH 的证明草稿。报道 7 指出目前尚无独立核验,Lean 只检验被形式化的命题本身。Anthropic 向 The Verge 表示未参与或介入 Alpöge 的这一宣布。Scientific American 资深记者 Joe Howlett 称当天是该领域现代历史上最大的一天。同一周(据 TechCrunch 与 The Decoder 报道为数小时后),斯坦福 AMS 主席 Ravi Vakil 发布了数学家 Alessandro Masullo 历经 12 年完成、经 Brian Conrad 数月审阅的纯人类广义黎曼猜想证明草稿;但约 8 小时后已有人在 AI 协助下发现其中问题,Vakil 表示所有问题需经学界确认后证明才能成立。截至目前,数学界仍在通过 Zulip 等渠道检验这两个结果,两项声明均未经独立验证。

其他来源的报道 7