Anthropic数学家Levent Alpöge(莱文特·阿尔珀格)在9月曾与OpenAI就一道奖金问题发生争议,如今他宣称已证明黎曼猜想——数学界最著名的未解难题——并表示是在Anthropic的AI模型Claude的协助下找到的证明。
Levent Alpöge, the Anthropic mathematician at the centre of a September dispute with OpenAI over a prize problem, has claimed a proof of the Riemann Hypothesis, the best-known unsolved question in mathematics, and says he found it working with Anthropic’s AI model, Claude.
Alpöge在个人网站上悄然发布了一份四页声明,未在社交媒体上作任何宣传。他在声明中说,该方法同时也证明了广义黎曼猜想(该问题的更广泛版本),且他与Claude已在Lean中核查了完整论证,“没有任何假设”。Lean是一种验证证明每一步骤的软件。相关Lean文件也发布在同一网站。他未发布任何人类可读的证明,声明亦未提及圈外人士是否已核查这些文件。
In a four-page statement posted quietly on his personal website, without any announcement on social media, Alpöge said the method also proves the Generalized Riemann Hypothesis, a broader version of the problem, and that he and Claude had checked the full argument in Lean, software that verifies every step of a proof, “without assumptions”. The Lean files are posted on the same site. He has published no human-readable proof, and the statement does not say that anyone outside his circle has checked the files.
Bernhard Riemann(伯恩哈德·黎曼)于1859年提出该猜想。它预言素数(仅能被1和自身整除的数)在数轴上的分布尽可能均匀。这是克莱数学研究所2000年设立的七大千禧年奖金难题之一,每题奖金100万美元,目前仅庞加莱猜想被解决。数论的大量内容以及计算领域的部分成果,都是在广义版本成立的条件下得到证明的。
Bernhard Riemann posed the hypothesis in 1859. It predicts that prime numbers, those divisible only by 1 and themselves, are spread along the number line as evenly as they possibly could be. It is one of seven Millennium Prize problems set by the Clay Mathematics Institute in 2000, each carrying a $1 million award, and only one of them, the Poincaré conjecture, has been solved. Large parts of number theory, and some results in computing, are proved on the condition that the generalised version holds.
Alpöge表示,之所以只发布Lean代码,是因为完整书面版本将难以阅读,描述的是一份超过25万页的机器生成草稿。他称打算弄懂该论证,“并将其妥善写出”。
Alpöge said he released only the Lean code because a full written version would be unreadable, describing a machine-generated draft of more than 250,000 pages. He said he intends to understand the argument and “to write it properly”.Lean is a tool that confirms that an argument establishes the claim written at the top of a file.
Lean是一种工具,可确认某论证确立了文件开头所写的断言。但它无法确认该断言就是著名的黎曼猜想。评审人员仍需核对所陈述的定理,并确保无步骤遗漏、无额外假设被引入,这一过程可能需要数周。
Alpöge将关键步骤归功于Claude,他写道,在接近尾声时,Claude“远远超过了我”。
It cannot confirm that the claim is the famous problem. Reviewers will need to check the stated theorem, and make sure no step was left unfinished and no extra assumption was added in a process that could likely take weeks.
他在谢尔盖·斯捷潘诺夫(Sergei Stepanov)1969年一项证明的基础上继续推进。该证明涉及数学中一个更简单分支里的类似命题,这个分支模拟整数。那项证明依赖一种微积分技巧,而这种技巧对整数并不适用。阿尔珀格说,他找到了替代办法,即一连串数学变换,而克劳德算出了这串变换需要多长。
Alpöge credits Claude with the crucial steps, writing that near the end it “far outpaced me”. He built on a 1969 proof by Sergei Stepanov of a similar statement in a simpler branch of mathematics that mirrors whole numbers. That proof relies on a calculus technique that does not work for whole numbers. Alpöge says he found a substitute, a chain of mathematical transformations, and that Claude worked out how long the chain needed to be.
在脚注中,他承认自己尚不精通克劳德推导出的那部分论证,并且他对这些内容的叙述取自一份机器撰写的摘要,“可能完全说不通”。他说,Lean代码应该能消除任何疑问。
In a footnote, he acknowledged that he is not yet expert in the parts of the argument Claude worked out, and that his account of them, drawn from a machine-written summary, “might make zero sense”. He said the Lean code should settle any doubt.
9月,阿尔珀格和纽约大学的特里斯坦·巴克马斯特(Tristan Buckmaster)发表了论文,表明在施加光滑外力时,某些流体方程会在有限时间内崩溃。OpenAI同日宣布,其人工智能代理解决了纳维-斯托克斯问题的受迫版本。巴克马斯特称,OpenAI研究员塞巴斯蒂安·布贝克(Sébastien Bubeck)向他施压,要他取消阿尔珀格的共同作者身份,因为阿尔珀格在Anthropic工作。布贝克称这些指控“虚假且具有煽动性”。
In September, Alpöge and Tristan Buckmaster of New York University posted papers showing that certain fluid equations can break down in finite time when a smooth outside force is applied. OpenAI announced the same day that its AI agents had resolved a forced version of the Navier–Stokes problem. Buckmaster alleged that an OpenAI researcher, Sébastien Bubeck, pressed him to drop Alpöge as a co-author because he worked at Anthropic. Bubeck called the allegations “false and inflammatory”.Alpöge worked on the fluid problem in a personal capacity.
阿尔珀格以个人身份研究这一流体问题。他的新声明则指向公司。他写道,他就自己能否公布一事给“达里奥”发了信息,达里奥显然指Anthropic首席执行官达里奥·阿莫代伊;他写道,“有一项限制需要遵守”,而且“我们的领导层”认同他对如何推进的看法。他称时机“糟糕”,但没有解释原因,并表示对此作愤世嫉俗的解读是错误的。Anthropic未公开评论这一说法。
His new statement points to the company. He wrote that he messaged “Dario”, apparently Anthropic chief executive Dario Amodei, about whether he could announce, that “there was a constraint to respect”, and that “our leadership” shared his view of how to proceed. He called the timing “horrible”, did not explain why, and said a cynical reading of it is false. Anthropic has not commented publicly on the claim.
他还称赞OpenAI在数学上的“品味”,写道其早先的工作让他看到纳维-斯托克斯“将会倒下”。声明感谢了几位数学家,包括普林斯顿的马恩朱尔·巴尔加瓦(Manjul Bhargava)和彼得·萨纳克(Peter Sarnak),但没有说他们中任何人审阅过该证明。阿尔珀格写道,他预计该论证将“被大幅简化,很可能立刻如此”。
He also praised OpenAI’s “good taste” in mathematics, writing that its earlier work showed him Navier–Stokes “was going to fall”.The statement thanks several mathematicians, including Princeton’s Manjul Bhargava and Peter Sarnak, but does not say any of them has reviewed the proof. Alpöge wrote that he expects the argument to be “massively simplified, likely immediately”.