OpenAI宣布破解千禧年数学难题引发原创性与数据归属争议OpenAI Claims Solution to Millennium Math Problem Amid Credit Dispute

OpenAI在2026年9月8日发布的声明中表示，其开发的一个未公开发布的内部模型成功破解了纳维-斯托克斯问题。该模型于8月下旬开始训练，性能超过该公司以往发布的所有公开模型。纳维-斯托克斯方程涉及空气与水等流体的运动规律，广泛应用于飞机设计、天气预报及血液流动研究，是克雷数学研究所设立的七大千禧年大奖难题之一。OpenAI的研究称，该工作表明这些方程存在致命缺陷，从而回答了奖项设立的问题。

OpenAI stated in a release on September 8, 2026, that an unreleased internal artificial intelligence model cracked the Navier-Stokes problem, an open question concerning the equations that govern the motion of fluids such as air and water. These mathematical equations are widely applied across aircraft design, weather forecasting, and the study of blood flow, forming one of the seven Millennium Prize Problems established by the Clay Mathematics Institute. OpenAI stated that its findings demonstrate these equations are fatally flawed, directly answering the challenge framed by the prize foundation.

在技术实现方面，OpenAI调动了约10,000个并发运行的独立智能体程序。这些智能体被细分为不同规模的小组，具备组内通信、互相传递信息、读取互联网缓存版本以及运行代码的能力。OpenAI研究员Sebastien Bubeck指出，最终阶段的大规模智能体协同运算耗费了数百万美元算力，支出达到以往数学研究计算成本的约1,000倍。智能体集群在启动约88小时后于9月5日得出了最终方案，公司随后公开了展示解题过程的Lean形式化证明，并表示即使该发现获得证实也不会申请该难题设立的100万美元奖金。

Detailing the technical execution, OpenAI explained that it deployed an internal model trained in late August to coordinate approximately 10,000 concurrent AI agent programs. These agents operated within subdivided groups capable of intra-group communication, passing messages back and forth, reading from cached internet data, and executing code. OpenAI researcher Sebastien Bubeck noted that computing costs reached emphatically into the millions of dollars—roughly 1,000 times the expenditure of earlier mathematical results. The agent network reached its solution on Saturday, September 5, roughly 88 hours after launch, leading OpenAI to publish a Lean proof detailing the resolution and state that it would decline the $1 million Millennium Prize reward if its findings are confirmed.

在OpenAI发布公告的前一天晚上，纽约大学数学教授Tristan Buckmaster与其合作者、Anthropic研究员Levent Alpöge公开发布了两人关于三个相关方程的AI辅助研究成果。Buckmaster指出，他们自去年启动该项目，花费数月时间设计出独特的解题路径，期间使用了包括OpenAI Codex在内的多个大型语言模型。当他们准备公开结果并联系OpenAI研究团队后，得知对方也即将宣布类似突破。

The night before OpenAI's announcement, New York University mathematics professor Tristan Buckmaster and collaborator Levent Alpöge, a researcher at rival firm Anthropic, published their own AI-assisted work addressing three related mathematical equations. Buckmaster noted that their personal collaboration began the prior year and required months to craft a distinctive solution path, during which they utilized a mix of large language models that included OpenAI's Codex tool. As the pair prepared to publicize their results and contacted OpenAI researchers, they learned the company was preparing to disclose a comparable finding.

Buckmaster在网站声明中透露，Alpöge曾收到关于两人研究进展信息被传递给OpenAI的线索。他强调OpenAI采取的解题路径与他们耗时数月构建的方法高度一致，并非仅通过给模型输入题目陈述就能在几天内得出的方向。此外，他还对OpenAI是否在模型训练中使用了他们使用Codex时的会话数据提出了疑问。

In a statement on his website, Buckmaster revealed that Alpöge had received tips indicating information regarding their research progress had been shared with OpenAI. Buckmaster asserted that OpenAI pursued a resolution path remarkably similar to their own, describing it as an approach that a model would not arrive at within days simply by being fed the problem statement. He also raised questions regarding whether OpenAI models had accessed or been trained on data from their interaction sessions within Codex.

针对外界对解题过程与立项动机的疑问，OpenAI在声明中明确了其研发时间线。该公司表示，攻坚纳维-斯托克斯方程的行动始于2026年9月1日，起因是听到网络传闻称有竞争对手正在解决千禧年难题，公司随后将该模型投入到剩余的六个千禧年难题中进行尝试。在发现纳维-斯托克斯方程展现出突破迹象后，研究团队集中算力推进该项目，并在数日后得出了成果。

Addressing questions regarding its timeline and motivation, OpenAI clarified that its targeted initiative on the Navier-Stokes problem began on September 1, 2026. The company stated that after seeing online rumors suggesting a competitor had solved Millennium Prize problems, it directed its unreleased model across all six remaining open questions. Upon observing unexpected promise on Navier-Stokes equations, the research team funneled massive computing power specifically into that effort, leading to the breakthrough several days later.

针对研究数据独立性，OpenAI研究员强调，团队与智能体在对方公开前未通过任何途径接触其研究成果，也没有为解决该问题调用特定用户数据，更未利用两人的提示词或证明来引导智能体。公司最初曾提议与对方联合发布成果并共享信用，但在发现对方尚未完全解题后放弃了该方案。与此同时，OpenAI在社交平台与声明中承认，尽管可能性较低，但无法完全排除来自两人使用其产品时产生的去标识化数据对基础模型改进起到了一定促进作用。

Regarding data independence, OpenAI researchers maintained that neither human researchers nor agent systems accessed the duo's work prior to its public release, emphasizing that no specific user data was accessed to solve the challenge. OpenAI stated it initially offered Buckmaster and Alpöge a concurrent release and full visibility into prompt logs, but moved forward independently after realizing the duo had not fully solved the problem. However, OpenAI stated in a public update that while unlikely, it cannot rule out that de-identified data derived from the researchers' usage of its products helped improve its underlying models.

克雷数学研究所于2000年设立千禧年大奖难题，旨在向公众展示数学前沿领域仍存在重要未解难题，并为每个难题的解答设立100万美元奖金。在七大难题中，此前仅有一个问题得到解决。总部位于马萨诸塞州剑桥的克雷数学研究所目前尚未对OpenAI提出的纳维-斯托克斯解答方案发表正式评论。

The Clay Mathematics Institute of Cambridge, Massachusetts, established the Millennium Prize Problems in 2000 to highlight critical unsolved challenges at the frontier of mathematics, offering a $1 million reward for the resolution of each problem. Only one of the seven designated problems had been solved prior to this development. The institute has not yet issued formal comments regarding the validity of OpenAI's proposed resolution to the Navier-Stokes existence and smoothness problem.

加利福尼亚大学洛杉矶分校数学教授陶哲轩等顶尖学者对AI介入基础数学研究的方式提出了担忧。陶哲轩指出，虽然OpenAI有可能借鉴了人类数学家的既有探索，但当前的局面已演变成狂热竞争。他认为跳过中间过程直接得到结果会失去探索的价值，并批评不加区分地使用AI正在将数学转变为一种生产指标竞赛，对数学学科本身与外部世界均缺乏实质益处。

Prominent mathematicians expressed reservations regarding the growing influence of high-powered AI systems on theoretical research. Terence Tao, a mathematics professor at the University of California, Los Angeles, observed that current developments reflect a frenetic competitive dynamic. He compared jumping directly to AI-generated answers to watching only the start and end of a film, where technical resolution occurs but the essential value of exploration is lost, cautioning that indiscriminate AI deployment risks reducing pure mathematics to a production quota game.