OpenAI 传奇数学里程碑陷入争议漩涡

据 The New York Times 和 Wired 此前报道，OpenAI 表示，它已找到一项重大数学问题的解法，该问题约 90 年来一直未获解决。OpenAI 在周二的一篇博客文章中宣布，公司借助一款性能强于新近发布的 GPT-6 Astra 的内部 AI 模型，以及 1 万个并发 Agent，发现了 Navier-Stokes 问题的解法。Navier-Stokes 问题涉及液体和气体的流动，是七个千禧年大奖难题之一；每道难题的解答均可获得 100 万美元奖金。

OpenAI 表示，公司于 8 月 28 日开始训练这款内部 AI 模型，该模型“在我们的基准测试中展现出前所未有的表现，其中包括数学”。这项解法对数学界而言是一项重大突破，但也并非毫无争议。

就在 OpenAI 宣布的前一天，纽约大学数学教授 Tristan Buckmaster 与 Anthropic 研究员 Levent Alpöge 合作，针对一个相关问题发布了研究结果。Buckmaster 在公布这些发现时称，在得知 OpenAI 已听说他们取得的进展后，他曾联系 OpenAI。不过，Buckmaster 发现，OpenAI 已经通过一条他与 Alpöge 一直在使用 OpenAI 的 Codex 和 Anthropic 的 Claude 探索的路径，为 Navier-Stokes 方程给出了证明。

Buckmaster 在声明中提出质疑：OpenAI 是否访问了他们的 Codex 数据，从而更接近 Navier-Stokes 问题的解法。“我询问该模型是否接受过我们在 Codex 中进行的会话数据训练，或是否能够访问这些会话；在整个项目中，我们一直把所有草稿都放进这些会话里，”Buckmaster 写道。“对方告诉我，模型不会查阅用户数据。我再次询问训练数据的问题，但没有得到答复。”

OpenAI 目前正试图通过周二的公告消除这些疑虑，称“为解决这一问题，没有访问任何特定用户数据”。公司还补充说，“虽然可能性不大，但我们无法排除源自他们使用我们产品的去标识化数据曾帮助改进我们模型的可能性”。在被要求置评时，OpenAI 向 The Verge 提供了其在 X 上发布的声明，该声明与其博客文章的表述一致。

OpenAI 技术团队成员 Sebastien Bubeck 也表示：“直到昨晚他们 [Buckmaster 和 Alpöge] 公开发布研究成果后，我们才看到他们的任何工作。事后可以看出，我们的证明存在显著差异，甚至所证明的具体结果也不相同。”

与此同时，Buckmaster 在 Mastodon 的一篇帖子中作出回应，声称 OpenAI“公开承认，他们使用了我们发现结果之后一段时期的数据进行训练”。

OpenAI 表示，公司不计划领取这 100 万美元奖金。

OpenAI says it found a solution to a major math problem that has remained unsolved for around 90 years, as reported earlier by The New York Times and Wired. In a blog post on Tuesday, OpenAI announced that it discovered a solution to the Navier-Stokes problem — which relates to the flow of liquid and gas — using an internal AI model more powerful than the newly released GPT-6 Astra alongside 10,000 concurrent agents. The Navier-Stokes problem is one of seven Millennium Prize Problems, each of which comes with a $1 million reward for solving.

OpenAI says it started training the internal AI model on August 28th, which has “exhibited unprecedented performance in our benchmarks, including mathematics.” The solution is a big breakthrough for the mathematics community, but it doesn’t come without controversy.

Just one day before OpenAI’s announcement, New York University mathematics professor Tristan Buckmaster published findings on a related problem in partnership with Levent Alpöge, a researcher at Anthropic. When announcing these findings, Buckmaster claims he contacted OpenAI after learning the company had heard about their progress. However, Buckmaster found that OpenAI had produced a proof for the Navier-Stokes equation using a route he and Alpöge had been working on with OpenAI’s Codex and Anthropic’s Claude.

In his statement, Buckmaster raises concerns about whether OpenAI had accessed their Codex data to get closer to the Navier-Stokes solution. “I asked whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project,” Buckmaster writes. “I was told the model did not look up user data. I asked again, about training, and I did not get an answer.”

OpenAI is now attempting to squash these suspicions with its Tuesday announcement, saying “no specific user data was accessed in order to solve this problem.” It adds that “while unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models⁠.” When reached for comment, OpenAI pointed The Verge to its statement on X, which echoes its blog post.

Sebastien Bubeck, a member of technical staff at OpenAI, similarly said: “We did not see any of their [Buckmaster and Alpöge’s] work until they released it publicly last night. One can in hindsight see that our proofs differ significantly and even the precise results proved are different.”

Meanwhile, Buckmaster responded to this in a post on Mastodon, claiming that OpenAI is “openly admitting they used training data from a period after we found our result.”

OpenAI says it doesn’t plan on taking the $1 million prize.