OpenAI 纳维-斯托克斯方程证明争议：Buckmaster 指控不当行为，各方回应引出开放科学之问

原文梳理各方公开回应与 Terence Tao 的警告，读者可据此思考 AI 实验室与学术界的信任问题。

数学家 Tristan Buckmaster 指控 OpenAI 施压、拒绝其合作者 Levent Alpöge（任职于 Anthropic）署名，且可能用两人上传到 Codex 的草稿训练模型，称其为绝对学术不端。

在数学家特里斯坦·巴克马斯特指控 OpenAI 存在不当行为之后，各方现已公开发声。此案为开放科学提出了棘手的难题。

围绕 OpenAI 用 AI 生成的纳维-斯托克斯方程证明所引发的争议——该问题属于克雷数学研究所千禧年大奖难题之一，悬赏 100 万美元——所有相关方都已发表公开声明。

数学家 Tristan Buckmaster 此前曾对 OpenAI 提出严重指控。他声称，在他的研究信息泄露后，该公司向他施压，试图将他的合著者 Levent Alpöge 从论文中除名，因为 Alpöge 在 Anthropic 工作，并威胁他的职业生涯会受到损害。此外还有怀疑认为，OpenAI 可能在两位研究者上传到 Codex 的草稿上训练了其模型。

有一点是无可争议的。据 OpenAI 自己称，它听到传闻说 Anthropic 的模型已经解决了一个千禧年大奖难题，随后便将自身资源也指向了同一问题。在其他几个方面，双方说法则存在分歧。

Alpöge 和 Buckmaster 对 OpenAI 所称其自身解决方案与他们的方案存在显著差异的说法提出异议。他们表示曾将类似的方法输入到 OpenAI 的系统中。据 Buckmaster 的理解，该输入最终进入了训练数据。Buckmaster 认为这是“绝对的学术不端行为。”

OpenAI 员工表示，OpenAI 实际在提交的解答上进行训练的可能性很低，尤其是如果这两位研究者没有禁用“将输入排除在训练之外”的选项的话。他们是否这样做了，外界不得而知。OpenAI 员工 Boaz Barak 也反驳了该模型需要外部帮助的观点。“认为这个模型需要这些帮助才能做到，只是一种自我安慰。它实际上从一开始就证明了比他们更强的结论。任何见过这个模型工作的人都不会认为它需要‘提示’。”

OpenAI 在其官方博客文章中承认了这种不确定性。"虽然可能性不大，但我们不能排除用户使用我们产品所产生的去标识化数据帮助改进了我们模型的可能性。"

OpenAI CEO Sam Altman 在 X 上发文支持 Bubeck，称团队"始终以正直和慷慨的态度行事"。他表示，曾有人建议将奖项授予 Buckmaster 和 Alpöge，但团队却遭遇了"毫无根据的抄袭指控"。

Altman 还写道，这项工作的起因是"上周互联网上有传言称 Anthropic 的模型解决了一个千禧年难题，我们很好奇我们的模型是否也能做到。"

Alpöge 在一个关键点上与 Altman 的说法相矛盾。Altman 写道，很难向 Alpöge 提出同样的合作邀约，因为他"本来就不愿意与我们交谈或协调"。Alpöge 则表示，他本来很乐意与 OpenAI 合作，而且署名问题对他来说并不重要。"我也喜欢实验室之间合作的想法，尤其是在科学进步上合作更好。真遗憾！"Alpöge 写道。

无论谁在每个细节上是对的，这件事都提出了一个基本问题：AI 实验室与研究社区之间究竟该如何互动。仅仅凭传闻，OpenAI 就足以在短时间内向一个研究问题投入大量资源，争分夺秒地去解决它，并可能抢先发表成果。而且也不能排除该公司在自身系统中输入的数据上进行了训练。

对学术界和公司而言，结论很简单。任何把研究数据输入 OpenAI 系统的人，都有可能被自己的发现反超。通过设置选择退出数据训练，充其量只能提供微弱的保护，尤其是在 AI 实验室在训练和数据实践方面尚未赢得信任的情况下。

数学家陶哲轩（Terence Tao）是在世最具影响力的数学家之一，他在 Mastodon 上警告称，“哪怕只是有人在研究某个问题的传闻，都可能触发大规模的 AI 驱动研究力量去抢先攻克它，而原始研究项目还没来得及充分发挥其潜力。”

“当前的激励机制可能正在指向这样一个方向：不再与更广泛的社区分享任何有前景的研究方向，这将逆转数百年来开放科学的传统，并对该领域的未来造成严重的长期损害，”陶哲轩写道。

Clay Math / 千禧年大奖难题

Alpöge / 声明

Buckmaster / 欺诈指控

Boaz Barak / 训练数据

布贝克 / 道歉

山姆·奥特曼 / 辩护

陶哲轩 / 警告

After mathematician Tristan Buckmaster accused OpenAI of misconduct, all parties have now gone on the record. The case raises hard questions for open science.

The fight over OpenAI's AI-generated proof of the Navier-Stokes equations, one of the Clay Millennium Problems that carries a $1 million prize, has drawn public statements from everyone involved.

Mathematician Tristan Buckmaster had previously leveled serious allegations against OpenAI. He claimed that after information about his research leaked, the company pressured him, tried to remove his co-author Levent Alpöge from the paper because Alpöge works at Anthropic, and threatened him with career consequences. There's also suspicion that OpenAI may have trained its models on drafts the two researchers uploaded to Codex.

One thing is undisputed. By its own account, OpenAI heard rumors that Anthropic's models had solved a Millennium Problem and then pointed its own resources at the same problem. On several other points, the stories diverge.

Alpöge and Buckmaster dispute OpenAI's claim that its own solution differs significantly from theirs. They say they had entered a similar approach into OpenAI's systems. As far as Buckmaster understands, that input ended up in the training data. Buckmaster sees this as "absolute academic malpractice."

OpenAI employees say the chance that OpenAI actually trained on the submitted solutions is low, especially if the two had disabled the option to exclude their inputs from training. Whether they did so isn't publicly known. OpenAI employee Boaz Barak also pushed back on the idea that the model needed outside help at all. "It's just cope to think that the model would have needed this. It actually started off by proving a stronger claim than they did. Anyone who has seen this model at work would not think it needs 'hints.'"

In its official blog post, OpenAI acknowledges the gap in certainty. "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models⁠."

OpenAI CEO Sam Altman backed Bubeck on X, saying the team had "acted with integrity and generosity throughout." He said it had been suggested that Buckmaster and Alpöge should receive the award, but the team was met with "unfounded accusations of plagiarism."

Altman also wrote that the effort started because of "rumors on the internet last week that Anthropic's models had solved a millennium problem and we were curious if ours could do it too."

Alpöge contradicts Altman on one key point. Altman writes that it was difficult to make the same offer to Alpöge because he "who was not willing to talk or coordinate with us anyway." Alpöge says he would have liked to work with OpenAI, and the authorship question didn't matter to him. "I also like the idea of the labs cooperating, and even better on scientific progress. It’s a shame!" Alpöge writes.

Regardless of who is right on every detail, the case raises a basic question about how AI labs interact with the research community. Rumors alone were enough for OpenAI to throw massive resources at a research problem on short notice, racing to solve it and possibly publish first. And it can't be ruled out that the company trained on data fed into its own systems.

For academics and companies alike, the takeaway is simple. Anyone who feeds research data into OpenAI's systems risks being beaten by their own findings. Opting out of data training through settings offers thin protection at best, especially since AI labs haven't earned the benefit of the doubt on training and data practices.

Mathematician Terence Tao, one of the most influential living mathematicians, warns on Mastodon that "even the rumor of someone working on a problem can trigger a massive amount of AI-powered effort to flatten it before the original research project has time to reach its full potential."

"The incentives may now be pointing in the direction of no longer sharing any promising research directions with the broader community, which would reverse centuries of traditions of open science and do serious long-term damage to the future of the field," Tao writes.

Clay Math / Millennium Problems

Alpöge / Statement

Buckmaster / Fraud allegation

Boaz Barak / Training data

Bubeck / Apology

Sam Altman / Defense

Terence Tao / Warning

来源：The Decoder：AI News（RSS）· the-decoder.com