Going off the believable citations in the wiki page here, it seems more complicated than "who solved it", as there's a mathematician, Anthropic, and OpenAI having a priority dispute.
For my money, the most “AI does my research” mistake I’ve seen in this whole mess is someone saying “the original researchers did not solve Navier-Stokes, while OpenAI did”. It has all the hallmarks of a classic AI blunder - partially eliding the name of the concept while oversimplifying the concept, a dramatic burst of confidence because it got to express its conclusion as a contradiction, etc. You can practically hear the original AI sentence (“One party lands the Navier-Stokes solve — not the original researchers, but OpenAI.”) before a few words were shuffled around to wash off some of the slop.
OpenAI should put their internal model to solve the rest of Millenium Prize problems too (assuming their difficulty is in the same ballpark). That would be a proof the model did it without using existing work on the problem.
LLMs seem to be much better at finding counterexamples than other types of proof. It’s unlikely the other millennium problems will be solved with counter examples
Finding counter examples might be easier, but AI's also useful for assisting in creating proofs. For example, Anthropic just published a formalization of Fermat's last theorem a few days ago, something that human researchers have been working on for decades.
They did do this, but it seems that Navier-Stokes is the only one that was successful (unless, for some strange reason, they have proofs for the other problems but haven’t released them yet).
Their claim is that they put it to the task after hearing about rumors that Anthropic had gotten there. So not a coincidence in any telling of the tale.
The researchers were also using AI so we're really just arguing over how much AI is allowed. But somehow a lean-verified proof that came first doesn't count because reasons.
If you have been involved in academia in any way, you know this happens all the time because the field is nearing a breakthrough, not a single person.
The evidence is a combination of two things: Tristan insinuating they did (note, not claiming but suggesting or implying), and their own statement saying "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models ."
The latter is being used by everyone as a sign of guilt, but it's clear that they are making a legally-safe statement since promising a forensic level guarantee that nothing Tristan has every typed into ChatGPT has ever made its way into any training data is a massive claim.
It's also a ridiculous expectation since tons of researchers use ChatGPT or Codex and many have "Use this data to improve" setting on.
>and their own statement saying "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models ."
Legally safe? Forensic level guarantee? Since when do they make those kinds of statements? Talk about moving the goalposts. Those are clear weasel-words from OpenAI, I'm shocked by the people who don't realize this.
And a third thing: OpenAI's thuggish behavior. Threatening a researcher with reputational ruin and demanding one of their competitors be denied proper credit is unacceptable, and we know OpenAI did both (unless Buckmaster is simply lying. which I find unlikely).
>many have "Use this data to improve" setting on.
Quite irrelevant. We are talking about scientific misconduct, not IP law. It would be misconduct for OpenAI to publish without proper reference to prior art even if that prior art had been explicitly committed to the public domain 1,000 years ago.
> And a third thing: OpenAI's thuggish behavior. Threatening a researcher with reputational ruin and demanding one of their competitors be denied proper credit is unacceptable, and we know OpenAI did both (unless Buckmaster is simply lying. which I find unlikely).
yeah, its not "we know" but "Buckmaster said this".
People lie often, not sure why we would presume unconditional innocence of one side in this case.
>It is in that context that I said “it would be simpler if Levent was not an Anthropic employee” because I felt it would be inappropriate for an Anthropic employee to author OpenAI’s work.
>I was confused why one would turn an incredible source for celebration (of their achievements!) into such bickering, which is when I said that I did not understand why one would risk their career [over unfounded accusations].
Maybe Buckmaster exaggerated the threatening tone and maybe Levent doesn't actually deserve authorship. But he wasn't lying.
This article is completely mistaken. The original researchers did not solve Navier-Stokes, while OpenAI did.
Ironically, this seems like the type of mistake you'd make if you used AI to do your research.
https://en.wikipedia.org/wiki/Navier%E2%80%93Stokes_existenc...
Going off the believable citations in the wiki page here, it seems more complicated than "who solved it", as there's a mathematician, Anthropic, and OpenAI having a priority dispute.
"Is OpenAI taking everyone for fools?"
Author proceeds to take their audience for fools
For my money, the most “AI does my research” mistake I’ve seen in this whole mess is someone saying “the original researchers did not solve Navier-Stokes, while OpenAI did”. It has all the hallmarks of a classic AI blunder - partially eliding the name of the concept while oversimplifying the concept, a dramatic burst of confidence because it got to express its conclusion as a contradiction, etc. You can practically hear the original AI sentence (“One party lands the Navier-Stokes solve — not the original researchers, but OpenAI.”) before a few words were shuffled around to wash off some of the slop.
OpenAI should put their internal model to solve the rest of Millenium Prize problems too (assuming their difficulty is in the same ballpark). That would be a proof the model did it without using existing work on the problem.
LLMs seem to be much better at finding counterexamples than other types of proof. It’s unlikely the other millennium problems will be solved with counter examples
Finding counter examples might be easier, but AI's also useful for assisting in creating proofs. For example, Anthropic just published a formalization of Fermat's last theorem a few days ago, something that human researchers have been working on for decades.
I haven't been keeping up with all the math proofs. Does anyone know if there have been any LLM proofs that are NOT counterexamples?
likely there are some, the question if its "millennial" type problems.
Also, lots of hype around "millennial" grading, it is kinda funny if we solve all "millennial" problems in first 30 yeas of millenia.
They did do this, but it seems that Navier-Stokes is the only one that was successful (unless, for some strange reason, they have proofs for the other problems but haven’t released them yet).
Though they claim the contrary, it looks that OpenAI would get nowhere if it did not have information about a recent breakthrough.
Comment that will age poorly
why not just solve world hunger while we're blowing smoke.,
OpenAI claims to have solved a decades old mathematical problem just as researchers are about to publish about their solution. Coincidence?
Their claim is that they put it to the task after hearing about rumors that Anthropic had gotten there. So not a coincidence in any telling of the tale.
It's just a desperate attempt to prove how grandiose their new model is
Yeah how desperate to literally solve a Millenium problem? They sure look stupid now. You’ve caught them!
They do look pretty stupid right now, actually.
If they want to desperately solve the remaining Millennium Prize problems, I won't object.
AI derangement syndrome is amazing.
The researchers were also using AI so we're really just arguing over how much AI is allowed. But somehow a lean-verified proof that came first doesn't count because reasons.
If you have been involved in academia in any way, you know this happens all the time because the field is nearing a breakthrough, not a single person.
No, we're arguing over whether or not openai stole the other researcher's work. The evidence is pretty damning.
https://techcrunch.com/2026/09/08/openai-fought-dirty-on-car...
Actually the evidence is not damning at all.
The evidence is a combination of two things: Tristan insinuating they did (note, not claiming but suggesting or implying), and their own statement saying "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models ."
The latter is being used by everyone as a sign of guilt, but it's clear that they are making a legally-safe statement since promising a forensic level guarantee that nothing Tristan has every typed into ChatGPT has ever made its way into any training data is a massive claim.
It's also a ridiculous expectation since tons of researchers use ChatGPT or Codex and many have "Use this data to improve" setting on.
>and their own statement saying "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models ."
Legally safe? Forensic level guarantee? Since when do they make those kinds of statements? Talk about moving the goalposts. Those are clear weasel-words from OpenAI, I'm shocked by the people who don't realize this.
>The evidence is a combination of two things
And a third thing: OpenAI's thuggish behavior. Threatening a researcher with reputational ruin and demanding one of their competitors be denied proper credit is unacceptable, and we know OpenAI did both (unless Buckmaster is simply lying. which I find unlikely).
>many have "Use this data to improve" setting on.
Quite irrelevant. We are talking about scientific misconduct, not IP law. It would be misconduct for OpenAI to publish without proper reference to prior art even if that prior art had been explicitly committed to the public domain 1,000 years ago.
> And a third thing: OpenAI's thuggish behavior. Threatening a researcher with reputational ruin and demanding one of their competitors be denied proper credit is unacceptable, and we know OpenAI did both (unless Buckmaster is simply lying. which I find unlikely).
yeah, its not "we know" but "Buckmaster said this".
People lie often, not sure why we would presume unconditional innocence of one side in this case.
https://xcancel.com/sebastienbubeck/status/20973794116915163...
>It is in that context that I said “it would be simpler if Levent was not an Anthropic employee” because I felt it would be inappropriate for an Anthropic employee to author OpenAI’s work.
>I was confused why one would turn an incredible source for celebration (of their achievements!) into such bickering, which is when I said that I did not understand why one would risk their career [over unfounded accusations].
Maybe Buckmaster exaggerated the threatening tone and maybe Levent doesn't actually deserve authorship. But he wasn't lying.
> But he wasn't lying.
and we know this how?
Discussions:
https://news.ycombinator.com/item?id=49613262
https://news.ycombinator.com/item?id=49605915