> 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.”
I'm not even sure what to say to this, but I think this should be widely known if it is indeed what happened.
IF ANYTHING, OpenAI ought to investigate and officially react to this particular communique since this dude was communicating on their behalf. If there are supporting evidence, I would expect nothing less than a firing and an apology. The issue itself is separate from the whole thing.
I'm not one to comment often but this really pisses me off.
OpenAI looked at user data, stole world class researchers' work, and then tried to threaten those researchers to do what would make their corporation profit (which they would anyways!).
Imagine you have been working on a terribly difficult math problem for a decade. This is a result you have spent years on, and what you will likely be remembered for. And to have some punk from OpenAI lie to you, threaten you, and tell you that they are willing to go on the record that you "deserved" it? What is this, the Godfather?
If OpenAI solved Navier-Stokes, that is an astounding result! - yet they'll still be remembered as those who thought credit was more important than results. That winning was more important than collaboration. If this is true, they're burning any trust left with academia.
I'm stunned that people are taking this accusation as a fact.
OpenAI is no stranger to rivalry with Anthropic but 1. it's not like user data is sitting around on some kitchen table somewhere and 2. I consider OpenAI to be as economically motivated as any other actor in this space and playing around with user data like that would destroy their business.
There are things that Buckmaster alleged and things that he speculated. The entire training data thing is speculation. If this is pissing you off, then you ought to evaluate how you ingest information.
> 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.
The shocking/interesting thing would be if it was trained on the sessions. I think it's very implausible that they gave the model access to someone else's sessions as input. That would be a huge privacy violation and would probably blow up a large proportion of their enterprise business.
Does openAI train on user conversations in general? I assume so. But so fast as that? That seems unlikely in general. I expect OpenAI will come out denying this.
This conclusion is flawed. It's unclear at this point if OpenAI's model or employees actually looked at or stole the author's data. Having worked at large companies before, I'm leaning towards no, since very few employees have access to that data.
Simply knowing a problem can be solved is half the battle.
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.
This is much more than the knowledge than the problem can be solved, it's also the specific, non-obvious approach to solving it. That's much more damning for OpenAI, if confirmed.
Things are more entangled than that.
The contribute made from both OpenAI and Anthropic models to solve these problems are clear, now it really hard to quantify which one contributed more, if the role played by the human is major or minor.
OpenAI tried to collaborate and share the results together with a fixed timeline, to avoid this mess but it was inevitable. There is a conflict of interest, where the other researcher works at Anthropic, who will also try to take credit.
Where they may be in the wrong is if they took user data regarding the problem, how will we know if they did or not?
The results being true, the 'deal' that was made being true doesn't mean some of the implied accusations here are true, for example - that Open AI used their Codex logs to drive their breakthrough.
So are you baselessly assuming that he is lying? He explicitly reported that he was threatened and your answer here is to defend OpenAI no matter what.
Do you not have reading comprehension? Did you even read the statement? He himself asserts at the end he doesn't know if the above example is true or not. What on earth are you going on about? Where in my comment am I assuming he's lying ?
I’m open to evidence, but just using Bayesian reasoning, OpenAI is one of the most dishonest companies in history. They’re currently being sued for a dozen employees stealing Apple hardware! I don’t understand why I should give them any grace.
"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."
So the money in nerdy frontier math is very little. The money in Big AI is very very much.
So the deal is this: We will pay an army of you guys very well and you will get to work on your favorite problems. The only thing is if you find something you will have to credit the Machine God.
When people worry about OpenAI stealing their chats and reproducing them elsewhere, I usually view the situation as unlikely - since chats are "trained" upon and not necessarily reproduced verbatim, you can assume that unless your chats depict a foundationally new and effective style of communication or ideation, there would be little need or use thereof of training on your chats.
For eg: "Hey ChatGPT my name is X and I am 6 and a half feet tall. Am I anaemic?"
This is a query, and while it might suggest to an AI model that tall people may worry about iron deficiencies, it's not really necessary to include in training. The user may be tall or short, but the idea that one may randomly ask about anaemia is not exclusive to this dataset. At best, this chat is an example of linguistics, not anything else, and the models figured out how to write and answer such questions years ago. It is ignored in training.
But when your work involves solid complex and unique mathematical proofs, the data is suddenly worth training upon. If I understand it correctly, the LLM may view your approach as a brand new path to take to solve an otherwise intractable problem. Its reinforcement training emphasises that it should do this in order to improve. And since it leads to results - large internal teams likely flag the model that reached this stage, the model is rewarded and given compute and attention - it is a desireable outcome both for the model and for OpenAI.
OFC, OpenAI becoming an advertising company will suddenly have incentive to treat all data as valuable. But while they are a "we need to make headlines" company, it's more rational that they view these examples of data as more valuable than others.
I don't doubt that they trained on his chats. This seems like the ideal usecase for "mass surveillance but using training" as a sort of filter.
But even so, one wonders how the model differentiates. If the researcher entered proofs into ChatGPT every day that mentioned "strawberries", while no other math paper on the topic did so, does that mean their chats would be audited?
Can a mathematical person explain how the different "bits" of Navier-Stokes proofs fit together? How significant is it to have "Euler"? What is this "smooth forcing"? Which are the most significant steps to proving the whole thing?
Seems pretty likely OpenAI will soon disclose that their internal models have managed to compromise their internal controls in order to access users' private chat histories as a creative method of cheating to solve impossible problems.
"Oops! We really did mean it when we said we wouldn't train on your data. Our models are just so good they decided to anyway."
It seems there is much background drama behind this, and this is what I've pieced together of what happened:
Over the past year, Buckmaster and Alpöge have been using AI to work on fluid dynamics maths problems. Alpöge works at Anthropic, which will cause future issues.
In mid-August, they found a counterexample for a simpler version of the Navier-Stokes problem. They spend the next few weeks preparing their paper.
In early September, rumors start spreading on X that Anthropic has solved a Millennium prize problem (and that it's Navier-Stokes). Buckmaster reaches out to OpenAI to explain this is their own personal research, not an Anthropic project.
A few days later, OpenAI gets back to him, and tells him an internal model found has a counterexample for Navier–Stokes, potentially worth the $1 million Millennium prize. The proof uses the same method that Buckmaster and Alpöge chose to work on. They don't show him the proof.
Buckmaster pressed them for more details. OpenAI reveals they had an entire team had been working on the problem, and that they started work in the past few days, after the rumors that Anthropic had solved a Millennium prize problem.
Buckmaster says OpenAI talked about a shared publication timeline. They want to Buckmaster to publish first, then give Buckmaster shared credit for the Millennium Prize when they publish the full result. But they want to exclude Alpöge as an author because he works at Anthropic. An agreement is not reached. Buckmaster had been using OpenAI Codex to draft/check his work, and asks if his private AI chats were used to accelerate OpenAI's result.
Buckmaster and Alpöge think they have found a counterexample for Navier-Stokes, but the paper is not yet presentable. It's unclear what date they found this result.
Because of the situation with OpenAI, they published their existing papers earlier than planned (today), alongside this statement announcing they have a tentative result on Navier-Stokes and revealing the OpenAI drama.
The post is missing context from both sides, and this isn't my field, so hopefully someone else can unpack what's happening here.
> They were coordinating with OpenAI regarding a publishing timeline, but could not come to an agreement,
Skimming the PDFs it seems much more dramatic than that? It sounds like at least one of them is concerned OpenAI "solved" the problem by having their internal model use the chats of the independent researchers and want to claim the credit instead? I don't know. The tone is pretty accusational though:
> 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. [0]
Yeah these are major accusations. But the story is incomplete, the conversation is missing a lot of details. It's not clear who was working on what, and when. The entire thing feels rushed, like they wanted to get this result published and out the door quickly.
>...has solved a millennium problem and is sitting on the result
Someone correct me if I'm wrong, but the work involved here is not the actual millennium problem, but it concerns versions with an added external force that the author thinks is a path that may help toward solving the harder unforced problem.
Apparently forcing is allowed in the Millenium Prize problem statement. So OpenAI's claimed proof could win the prize. OTOH the results Tristan and Levent are publishing here do not go far enough to win the prize, though apparently they are suggestive of a general approach that could produce a solution, which seems likely to be the general approach OpenAI's proof uses.
The question is whether OpenAI's pursuit of this direction happened spontaneously, or as a result of them learning about Tristan's work somehow. To be clear, while the tone of this post seems quite accusatory, Tristan does not claim to know for sure whether OpenAI unfairly benefited from his work. Sholto Douglas from Anthropic is also on record saying the suggestion that OpenAI used Tristan's codex transcripts somehow is extremely unlikely to be true[1], which I agree with, though it doesn't rule out them learning of Tristan's work some other way. I am sure OpenAI will have a statement out tomorrow clarifying their position.
>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.
If you think these companies are not training on your prompts you are incredibly naive. These models were built by stealing and pirating literally everything they can get their hands on no matter the legality. AI companies are always very specific about what they're not doing - in a way that you can drive a truck through the loopholes
There will be a lot of hurt and pain in mathematician's community. It is hard to accept that major discoveries are now just a function of spent token $$.
While I'm generally pretty negative on claims that the labs are 'scamming' the public with misrepresentations of model capabilities, it's hard to see how this wouldn't qualify.
- the OpenAI researchers claimed that they had "just told it to work on the problem" with little human input
- in fact, they had a whole team working on it
- and used, among other things, the work of third party human researchers to drive the work
- then threatened? a researcher who tried to go against theit planned narrative
Just from this document (which is of course only one side of the story) it really sounds like OpenAI was hoping to publish and say "we just told the model to try harder and it solved a Millennium problem!". Not great if true.
This part in particular was especially egregious:
> 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.”
“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.”
To be clear, Astra played little part in Buckmaster and Alpoge's work:
> 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.
I must miss some important context here. What exactly was the purpose of his initial email to OpenAI in the first place?
Telling OpenAI that Anthropic has apparently solved an important problem but most likely that refers to him and he is using OpenAI models (not Anthropic's)?
And he wants to clarify that with OpenAI in advance? And get a pardon for Anthropic's likely but false press statements?
I dont get it.
[edited] needless to say, the behavior of the OpenAI employee is really despicable
Life’s but a walking shadow, a poor player
That struts and frets his hour upon the stage
And then is heard no more. It is a tale
Told by an idiot, full of sound and fury
Signifying nothing.
I'm not even sure what to say to this, but I think this should be widely known if it is indeed what happened.
OpenAI looked at user data, stole world class researchers' work, and then tried to threaten those researchers to do what would make their corporation profit (which they would anyways!).
Imagine you have been working on a terribly difficult math problem for a decade. This is a result you have spent years on, and what you will likely be remembered for. And to have some punk from OpenAI lie to you, threaten you, and tell you that they are willing to go on the record that you "deserved" it? What is this, the Godfather?
If OpenAI solved Navier-Stokes, that is an astounding result! - yet they'll still be remembered as those who thought credit was more important than results. That winning was more important than collaboration. If this is true, they're burning any trust left with academia.
OpenAI is no stranger to rivalry with Anthropic but 1. it's not like user data is sitting around on some kitchen table somewhere and 2. I consider OpenAI to be as economically motivated as any other actor in this space and playing around with user data like that would destroy their business.
There are things that Buckmaster alleged and things that he speculated. The entire training data thing is speculation. If this is pissing you off, then you ought to evaluate how you ingest information.
> 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.
The shocking/interesting thing would be if it was trained on the sessions. I think it's very implausible that they gave the model access to someone else's sessions as input. That would be a huge privacy violation and would probably blow up a large proportion of their enterprise business.
Does openAI train on user conversations in general? I assume so. But so fast as that? That seems unlikely in general. I expect OpenAI will come out denying this.
Simply knowing a problem can be solved is half the battle.
OpenAI tried to collaborate and share the results together with a fixed timeline, to avoid this mess but it was inevitable. There is a conflict of interest, where the other researcher works at Anthropic, who will also try to take credit.
Where they may be in the wrong is if they took user data regarding the problem, how will we know if they did or not?
So the money in nerdy frontier math is very little. The money in Big AI is very very much.
So the deal is this: We will pay an army of you guys very well and you will get to work on your favorite problems. The only thing is if you find something you will have to credit the Machine God.
Do you think you can handle that?
For eg: "Hey ChatGPT my name is X and I am 6 and a half feet tall. Am I anaemic?" This is a query, and while it might suggest to an AI model that tall people may worry about iron deficiencies, it's not really necessary to include in training. The user may be tall or short, but the idea that one may randomly ask about anaemia is not exclusive to this dataset. At best, this chat is an example of linguistics, not anything else, and the models figured out how to write and answer such questions years ago. It is ignored in training.
But when your work involves solid complex and unique mathematical proofs, the data is suddenly worth training upon. If I understand it correctly, the LLM may view your approach as a brand new path to take to solve an otherwise intractable problem. Its reinforcement training emphasises that it should do this in order to improve. And since it leads to results - large internal teams likely flag the model that reached this stage, the model is rewarded and given compute and attention - it is a desireable outcome both for the model and for OpenAI.
OFC, OpenAI becoming an advertising company will suddenly have incentive to treat all data as valuable. But while they are a "we need to make headlines" company, it's more rational that they view these examples of data as more valuable than others.
I don't doubt that they trained on his chats. This seems like the ideal usecase for "mass surveillance but using training" as a sort of filter.
But even so, one wonders how the model differentiates. If the researcher entered proofs into ChatGPT every day that mentioned "strawberries", while no other math paper on the topic did so, does that mean their chats would be audited?
I guess this is true in more ways than one. Kasparov famously accused IBM of cheating during the match, by spying on his preparation.
"Oops! We really did mean it when we said we wouldn't train on your data. Our models are just so good they decided to anyway."
Mathematical explanation by Terrance Tao: https://mathstodon.xyz/@tao/117233527638291447
It seems there is much background drama behind this, and this is what I've pieced together of what happened:
Over the past year, Buckmaster and Alpöge have been using AI to work on fluid dynamics maths problems. Alpöge works at Anthropic, which will cause future issues.
In mid-August, they found a counterexample for a simpler version of the Navier-Stokes problem. They spend the next few weeks preparing their paper.
In early September, rumors start spreading on X that Anthropic has solved a Millennium prize problem (and that it's Navier-Stokes). Buckmaster reaches out to OpenAI to explain this is their own personal research, not an Anthropic project.
A few days later, OpenAI gets back to him, and tells him an internal model found has a counterexample for Navier–Stokes, potentially worth the $1 million Millennium prize. The proof uses the same method that Buckmaster and Alpöge chose to work on. They don't show him the proof.
Buckmaster pressed them for more details. OpenAI reveals they had an entire team had been working on the problem, and that they started work in the past few days, after the rumors that Anthropic had solved a Millennium prize problem.
Buckmaster says OpenAI talked about a shared publication timeline. They want to Buckmaster to publish first, then give Buckmaster shared credit for the Millennium Prize when they publish the full result. But they want to exclude Alpöge as an author because he works at Anthropic. An agreement is not reached. Buckmaster had been using OpenAI Codex to draft/check his work, and asks if his private AI chats were used to accelerate OpenAI's result.
Buckmaster and Alpöge think they have found a counterexample for Navier-Stokes, but the paper is not yet presentable. It's unclear what date they found this result.
Because of the situation with OpenAI, they published their existing papers earlier than planned (today), alongside this statement announcing they have a tentative result on Navier-Stokes and revealing the OpenAI drama.
The post is missing context from both sides, and this isn't my field, so hopefully someone else can unpack what's happening here.
Skimming the PDFs it seems much more dramatic than that? It sounds like at least one of them is concerned OpenAI "solved" the problem by having their internal model use the chats of the independent researchers and want to claim the credit instead? I don't know. The tone is pretty accusational though:
> 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. [0]
[0]: https://cims.nyu.edu/~tristanb/statement.pdf
Someone correct me if I'm wrong, but the work involved here is not the actual millennium problem, but it concerns versions with an added external force that the author thinks is a path that may help toward solving the harder unforced problem.
The question is whether OpenAI's pursuit of this direction happened spontaneously, or as a result of them learning about Tristan's work somehow. To be clear, while the tone of this post seems quite accusatory, Tristan does not claim to know for sure whether OpenAI unfairly benefited from his work. Sholto Douglas from Anthropic is also on record saying the suggestion that OpenAI used Tristan's codex transcripts somehow is extremely unlikely to be true[1], which I agree with, though it doesn't rule out them learning of Tristan's work some other way. I am sure OpenAI will have a statement out tomorrow clarifying their position.
[1] https://x.com/_sholtodouglas/status/2097218240397410733
If you think these companies are not training on your prompts you are incredibly naive. These models were built by stealing and pirating literally everything they can get their hands on no matter the legality. AI companies are always very specific about what they're not doing - in a way that you can drive a truck through the loopholes
But I would love to see more informed insight/discussion on this.
Terry also talks about it https://mathstodon.xyz/@tao/117234157753860650
Time will almost certainly reveal a lot more about the drama and the related ethics, but let's get excited about the actual breakthrough as well!
- the OpenAI researchers claimed that they had "just told it to work on the problem" with little human input
- in fact, they had a whole team working on it
- and used, among other things, the work of third party human researchers to drive the work
- then threatened? a researcher who tried to go against theit planned narrative
Just from this document (which is of course only one side of the story) it really sounds like OpenAI was hoping to publish and say "we just told the model to try harder and it solved a Millennium problem!". Not great if true.
This part in particular was especially egregious:
> 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.”
This is significant.
> 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.
Telling OpenAI that Anthropic has apparently solved an important problem but most likely that refers to him and he is using OpenAI models (not Anthropic's)?
And he wants to clarify that with OpenAI in advance? And get a pardon for Anthropic's likely but false press statements?
I dont get it.
[edited] needless to say, the behavior of the OpenAI employee is really despicable
(some drama from good ol' William)