今天,我和 Levent Alp¨oge 公开了三项成果:不可压缩多孔介质、Boussinesq 方程以及三维不可压缩 Euler 方程在光滑强迫下的有限时间爆破。我们相信对于 hypo-dissipative Navier-Stokes 方程也能得到爆破结果,但今天没有发布那篇论文:与上述工作不同,它的 Lean 验证尚未完成。我们目前也还没有任何接近可展示的文稿。我提到这一点,是因为它暗示了一条通往无强迫 Euler 方程的道路。这个研究纲领并非由我们发起,也不是大语言模型提出的。该纲领基本思想的功劳归于 Diego C´ordoba 和 Luis Mart´ınez-Zoroa,他们多年来一直在探索强迫爆破的构造。我们以他们的工作为起点,利用大语言模型将他们的纲领推进到完成。具体来说,Levent 和我所做的是,以 C´ordoba 和 Mart´ınez-Zoroa 的纲领(该纲领在粗糙强迫下取得了爆破结果)为出发点,在大语言模型的大量帮助下,将其推进到光滑强迫情形,并推广到不可压缩 Euler 方程。使这条研究路线成为可能的思想归功于 C´ordoba 和 Mart´ınez-Zoroa。请允许我明确说出我私下对同事们讲过的话:鉴于这一系列工作,我认为 Luis Mart´ınez-Zoroa 理应获得菲尔兹奖。我与 Levent 的合作纯粹是个人性质的,不涉及任何机构协议,也不涉及我们双方雇主的任何官方参与。我们在整个过程中使用了多个大语言模型:Anthropic 的 Claude、OpenAI 的 Codex,尤其是 GPT-5.6 Sol,以及最近使用的 Astra。后者仅用于文稿撰写和论证审计。在过去一年的大部分时间里,进展缓慢。我们梳理了文献,升级了各种初步结果,直到获得了不可压缩多孔介质方程(带光滑强迫)的有限时间爆破。直到大约一个月前,我们才取得了实质性进展:8 月 15 日,我们获得了 Boussinesq 方程和 Euler 方程在光滑强迫下的爆破结果。可以说,Levent 发给我的第一个 LLM 生成的证明是我读过的最糟糕的;我们于 8 月 22 日在 Lean 上验证了它。从那时起,我们夜以继日地工作,试图理解这个证明并将其转化为可读的内容。这个故事还有另一部分,坦白说,我非常希望自己不必为此操心。我对这些论文的呈现质量并不满意。理想情况下,我们更愿意花上数周时间,从第一页起就把 LLM 生成的证明改写成可读的形式。这种程度的用心,才是这些问题以及致力于解决这些问题的社区所应得的。尤其是那些 Boussinesq 和 Euler 的文稿,它们更接近模型在人类指导下产出的结果,而非人类撰写的论文。特别是 Euler 的文稿,只能被形容为“AI 垃圾”。对此我感到抱歉。原因如下,涉及我们受到外部因素的压力。我这么说不是找借口,而是解释。我原本计划在宣布我们的工作时说,结果本身并不重要;重要的是,一位数学家和 LLM 模型现在能在一个月内完成所有这些工作,这一事实的意义——对于如何培养学生、分配功劳、评审论文、判断什么值得一个人毕生关注——怎么强调都不为过。这是一个“深蓝”对卡斯帕罗夫的时刻。社区需要严肃且从容地讨论接下来该何去何从。然而,我发现自己却在写别的东西,而不是这些极其重要的发展。我想尽可能直白地陈述发生了什么。9 月 3 日(星期四),有传言称 Anthropic 解决了一个重大的开放问题,而 Levent 也收到消息说我们进展的信息已泄露给 OpenAI,于是我写信给 OpenAI 的一位著名数学家。我完整引用我的邮件,因为我宁愿大家读全文,而不是我的概括:“你好[……]],我们未曾谋面,但我是[…]上的一位演讲者。我写信是因为有一个传言似乎正在迅速传播,称 Anthropic 已经解决了一个重大的开放问题。我不确定传言所指的就是我,但我在 Courant 的一位同事昨天通过邮件告诉了我这件事——他是从英国一位分析师那里听说的,而那位分析师又是从更上游的某处得知的——所以似乎可以放心地假定传言指的就是我。我了解到,将 Levent Alpöge 与某个解决方案联系起来的说法也在科技界流传。不过,关于具体解决了什么问题,似乎存在很多混淆。我还想强调,这并非机构层面的努力。这严格来说是我们两人之间的个人合作,背后没有任何正式协议。我用自己的研究经费支付我团队使用的工具费用,包括向 OpenAI 支付一大笔账单。我以前与 DeepMind 有过行业合作,那是有正式机构安排的。我们确实在这一领域有我们确信的工作成果,我们很快就会发布,论文和形式化证明会一起发布。我们有意决定不仓促地在未经打磨的预印本旁边发布一份 Lean 证书。我强烈认为,人们读到的第一样东西应该是以正常方式呈现的数学论证,而不仅仅是一份形式化证书。我直接写信给你,而不是公开发表任何言论,以便你掌握事实,在你那边处理此事。此致敬意,Tristan” 他当天就回复了:“如果你愿意提供任何细节,那将有助于避免我们这边的重复工作,而且总的来说,每当数学家利用我们的模型取得进展时,我们总是感到非常高兴。此外,如果 OpenAI 这边在算力方面有任何需要,我们很乐意提供。” 我要求下周再谈。9 月 4 日星期五,有人问我当天能否见面;我再次说下周。9 月 6 日星期日中午 12:45,有人问我能否“今天任何时间”见面。Sebastien Bubeck 也加入了。那天下午我们三人谈了两次。Levent 没有参加这些电话会议。我被告知,OpenAI 内部的一个模型已经为受迫的纳维-斯托克斯方程生成了有限时间爆破的证明。当 Levent 通过短信询问精确表述时,得到的回答是:“在 R3 中存在受迫爆破,且
T 3”,并提到“强制函数在 Fefferman 的表述中是光滑的选项 c 和 d”。我被告知证明大约有 100 页。我没有看过。我应当在此说明我为何如此解读他们的表述——我将在下文讨论这一解读。通过光滑力(即 Fefferman 问题陈述中的选项 c 和 d)来攻克 Clay 问题的路径,是 Luis 和 Diego 开辟的,也是 Levent 和我悄悄选择的攻击方向。据我所知,几乎没有人在这条路上工作。这不是把问题陈述丢给模型几天就能走出来的方向。当我听到“强制”一词时,那是一个醒目的危险信号。有人给我看了一段提示词,并告诉我这个内部研究模型只是被直接给了问题陈述。Sebastien 曾告诉 Levent,只用了“极少的人工输入”。事实证明这并非实情。在通话过程中,随着他们团队成员通过内部聊天向 Sebastien 发送更正和细节,逐渐浮现出以下事实:有一个完整团队一直在研究这个问题;这只是众多尝试之一;工作是从非强制问题开始的;团队先让模型解决更简单的问题,包括 Euler 问题;甚至给我看的那段提示词也是通过提示 Codex 写出来的;而且使用了极其庞大的计算资源。我问他们第一次发送提示词是什么时候。OpenAI 在一段时间内没有直接回答这个问题。最终他们承认提示词是在过去几天内发送的,是在关于我们工作的信息传到 OpenAI 之后。我问模型是否在训练中使用了或能够访问我们在 Codex 中的会话——我们整个项目期间的所有草稿都放在那里。我被告知模型不会查阅用户数据。我再次询问关于训练的问题,但没有得到回答。他们向我提出了两个方案。第一个方案是:我们发布 Euler 结果,OpenAI 第二天发布其 Navier-Stokes 结果。第二点是,在发布欧拉论文之后,我将独自撰写一篇展示纳维-斯托克斯结果的论文,并承认是 OpenAI 内部的一个模型解决了这个问题。塞巴斯蒂安两次坚称他希望将莱文特从作者名单中移除,并表示如果不是因为莱文特在 Anthropic 工作这件事,一切都会很简单,而这件事既令人恼火又让人无奈。他还说,如果 OpenAI 在我们之后发布,他们会说我们理应获得克莱奖,并且我们是“最接近这个问题的人类”。我拒绝了这两个提议。我说,如果 OpenAI 按照提议的方式发布其结果,我会公开所发生的事情。得到的回答是:“你为什么要毁掉自己的职业生涯?”我回答说,我是一名学者,并问他为什么认为公开此事会毁掉我的职业生涯。回答是:“如果你不想让我客气,那我也不必客气。”一段时间后,莱文特收到一条短信,提议他和塞巴斯蒂安单独谈谈,短信中说:“我不确定特里斯坦现在是否完全理性。”莱文特拒绝了,并表示谈话应该和我进行。塞巴斯蒂安当晚发了一封后续邮件,请求在 9 月 37 日(周一)与我通话。我没有回复。我想明确说明我并没有在指控什么。我没有看过 OpenAI 的证明。我不知道他们的模型做了什么,也不知道是如何做到的。我不知道我们的数据是否被使用。我没有指控任何人任何事。我只是陈述我所被告知的内容、时间以及向我提出的提议。我之所以陈述这些,是因为如果不这样做,就等于任由一系列公告说出一些我知道是虚假的事情。如果确实是 OpenAI 的一个模型弥合了纳维-斯托克斯问题的差距,那是一件了不起的事情,应该由他们大声说出来,并完整保留这段历史。我更愿意谈论数学、路易斯和迭戈的想法,以及这一切对我们其他人意味着什么。最后,我想感谢在过去 24 小时里给予我巨大支持的整个数学界。4
Today, Levent Alp¨ oge and I have made public three results: finite-time blowup with smooth forcing for incompressible porous media, for Boussinesq, and for 3d incompressible Euler. We believe we also have blowup for hypo-dissipative Navier-Stokes. We are not releasing that paper today: unlike the above, the Lean verification has not yet finished. We do not yet have anything resembling a presentable writeup. I mention it because it is suggestive of a path to unforced Euler. The program this fits into was not started by us nor was it proposed by a Large Language Model. The credit for the basic idea of this program goes to Diego C´ ordoba and Luis Mart´ ınez-Zoroa, who for several years have been exploring the construction of forced blow ups. We took their work as a starting point, using Large Language Models to push their program to completion. Concretely, what Levent and I did was to take the C´ ordoba and Mart´ ınez-Zoroa program, which achieved blowup results with rough forcing, and, with a great deal of help from LLMs, push it to smooth forcing and to the incompress-ible Euler equations. The ideas making this line of attack possible are due to C´ ordoba and Mart´ ınez-Zoroa. Let me make plain what I have said to colleagues in private: in view of this body of work, I believe Luis Mart´ ınez-Zoroa deserves a Fields Medal. My work with Levent has been a purely personal collaboration, free of any institutional agreements or official involvement by either of our employers. We used several LLMs throughout: Anthropic’s Claude, OpenAI’s Codex, especially with GPT-5 .6 Sol and, more recently, Astra. The latter was only used for writeups and auditing our arguments. For most of the past year progress was slow. We worked through the lit-erature and upgraded various preliminary results, up to obtaining finite time blow up for the Incompressible Porous Media equation (with smooth forcing). This was until about a month ago, when we had real progress: on August 15th, we obtained the blow up results, with smooth forcing, for both Boussinesq and Euler. I can say the first LLM generated proof Levent sent me was the most horrendous I have ever read; we verified it on Lean on August 22nd. Since this point, we have been working around the clock to understand this proof and turn it into something readable. There is another part of this story, and one that, honestly, I very much wish I did not have to be concerned with. I am not happy about the presentation quality in these papers. Ideally, we would have preferred to spend weeks turning the LLM generated proofs into something readable from the very first page. This level of care is what these problems and the community devoted to these problems deserves. The various Boussinesq and Euler write-ups in particular are much closer to what models produce under human direction than to a paper written by a person. The Euler writeup, in particular, can only be described as AI slop. I am sorry for this. The reasons are below, and they involve our being pressured by outside factors. I say this not as an excuse but as an explanation. I had planned to say on announcing our work that the results are not the important thing. Rather the important thing is instead the significance that a mathematician and an LLM model can now do all this work in a month. The 1significance of this with respect to the way we train students, assign credit, referee, and decide what is worth one human life’s attention cannot be under-stated. This is a a Deep Blue-Kasparov moment. The community needs to have serious and unhurried discussion about where to go from here. Instead of these incredibly important developments, I find myself writing about something else. I want to set out what happened as plainly as I can. On Thursday, September 3rd, with a rumor circulating that Anthropic had resolved a major open problem, and with Levent having received tips that infor-mation about our progress had been passed to OpenAI, I wrote to a prominent mathematician at OpenAI. I am quoting my email in full because I would rather the full text be read rather than my summary of it: “Hi [...], We have not met in person, but I was one of the speakers at the [...]. I am writing because a rumor seems to be spreading quickly that Anthropic has resolved a major open problem. I cannot be certain I am the person it attaches to, but a colleague at Courant emailed me about it yesterday – who heard it from an analyst in the UK, who had it from somewhere further upstream – so it seems safe to assume I am. I gather versions linking Levent Alp¨ oge to a solution are going around in tech as well. There however appears to be a lot of confusion with regards to the exact problem solved. I should also emphasize that this is not an institutional effort. It is a strictly personal collaboration between the two of us, and there is no formal agreement behind it. I pay for the tools my group uses out of my own research funds, including footing a large bill to OpenAI. I have had an industry collaboration before, with DeepMind, which had a formal institutional arrangement. We do have work in this area that we are confident in, and we will post it shortly, the paper and the formalization together. We intentionally decided against rushing out a Lean certificate alongside an unpolished preprint. I feel strongly that the first thing anyone reads should be a mathematical argument presented in the normal manner, rather than just a formal certificate. I am writing to you directly rather than saying anything publicly so that you have the facts to address this on your end. Best wishes, Tristan” He replied the same day: “If you are willing to give any details it would be useful to avoid competing here and in general we are always thrilled when mathematician make progress with our models. Additionally if there is anything in terms of compute from OpenAI’s end we would be happy to provide it.” I asked to speak the following week. On Friday, September 4th, I was asked whether I could meet that day; I again said the following week. At 12:45 on Sunday, September 6th, I was asked whether I could meet “at any point today.” Sebastien Bubeck joined. The three of us spoke twice that afternoon. Levent was not on the calls. I was told that an internal OpenAI model had produced a proof of finite time blowup for the forced Navier-Stokes equations. When Levent asked by text for 2the precise statement, the answer was: “Existence of forced blowup in R3 and
T 3,” with “the forcing function is smooth option c and d in Fefferman.” I was told the proof is about 100 pages. I have not seen it. I should say here why I interpreted their statement the way I did, the in-terpretation I will discuss below. The route to the Clay problem through a smooth force, options c and d in Fefferman’s statement of the problem, is the route Luis and Diego opened and the one Levent and I had quietly chosen to attack. Almost nobody else I know of was working on it. It is not the direction one arrives at in a few days by giving a model the problem statement. When I heard “forced,” it was a bright red flag. I was shown a prompt and told the internal research model had simply been given the problem statement. Levent had been told by Sebastien “very little human input” had been used. This turned out not to be true. Over the course of the call, as members of their team sent Sebastien corrections and details over their internal chat, it emerged that an entire team had been working on the problem, that this was one of a number of things that was tried, that work had started on the unforced problem, that the team first set the model on easier problems, including Euler, that even the prompt that had been shown to me had been written by prompting Codex, and that an insane amount of compute had been used. I asked when the first prompt had been sent by them. This question was not answered directly by OpenAI for some time. Eventually it was agreed that it had been sent in the past few days, after information about our work had reached OpenAI. 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. I was told the model did not look up user data. I asked again, about training, and I did not get an answer. Two proposals were offered to me. The first was that we post our Euler result, and that OpenAI post its Navier-Stokes result the next day. The second was that, after posting Euler, I alone write a paper presenting the Navier-Stokes result, acknowledging that an internal OpenAI model had resolved it. Sebastien twice asserted that he wanted Levent removed from authorship, and said it would all be simple if only it were not the case that, and it was so annoying that, Levent works at Anthropic. It was also said that if OpenAI posted after us, they would say that we deserved the Clay Prize, and that we were the “closest humans to the problem”. I declined both offers. I said that if OpenAI released its result in the way proposed I would go public with what happened. The reply was, “Why would you ruin your career?” I replied that I am an academic, and asked why he thought going public would ruin my career. The reply was, “If you don’t want me to be nice, then I don’t have to be nice.” Some time later Levent received a text proposing that he and Sebastien speak one on one, saying, “I don’t know if Tristan is being fully rational right now.” Levent declined and said conversations should be with me. Sebastien sent a follow-up email that night requesting to speak with me on Monday, September 37th. I did not respond. I would like to be clear about what I am not claiming. I have not seen OpenAI’s proof. I do not know what their model did, or how. I do not know whether our data was used. I am not accusing anyone of anything. I am stating what I was told, when, and what was proposed to me. I am stating it because the alternative is to let a sequence of announcements say something I know to be false. If indeed an OpenAI model did close the gap to Navier-Stokes, that is a remarkable thing and it should be said loudly, by them, with the history intact. I would much rather be talking about mathematics, Luis and Diego’s ideas, and what this all means for the rest of us. Lastly, I would like to thank the entire mathematics community that have been so supportive of me over the last 24 hours. 4