Is this the one that was allegedly based on someone else's actual work & prompts?
https://news.ycombinator.com/item?id=49605915
https://bsky.app/profile/quantian.bsky.social/post/3muyhwbcd...
https://news.ycombinator.com/item?id=49605915
https://bsky.app/profile/quantian.bsky.social/post/3muyhwbcd...
I work at OpenAI, though not on the team that did this, and my understanding is:
- we decided to ask our model for Millenium problem solutions because of two reasons: (a) our new model was looking incredibly good and (b) we heard rumors that some Millenium problems had been solved and were curious if our models could solve them (the goal here was not to scoop any particular individuals and we were looking at many problems beyond these)
- we did not read any private chats (but of course the model was aware of prior research literature published to the internet)
- the proof generated by our model was very different from theirs and also goes far beyond the published literature
- we made an effort to jointly announce rather than immediately scoop (I understand Tristan was unhappy with the conversations; I know zero details here and I hope more is shared today)
Edit: Here's is Sebastian's take: https://x.com/SebastienBubeck/status/2097379411691516310?s=2...
https://news.ycombinator.com/item?id=49605915#49610498
https://x.com/dheeraj_nagaraj/status/2097266146445774924?s=6...
(I can't reply to the below comment, but I was aware this was about Sebastien, I was trying to be charitable by including stuff said about both people)
I dedicate my life to its complete destruction beginning today.
- This, from Tristan Buckmaster's writeup yesterday, indicates to me that there was more than incidental inspiration from Alpoge and Buckmaster.
- "very little human" input feels ambiguous, and if someone spends a few days prompting a model to solve a super hairy problem requiring a 100-page proof, I can understand reasonable people interpreting that as both "very little" and "not very little" human input
- it's all true that a team worked on this, a bunch of compute was burned, and the problem was solved in stages and pieces
I'm not sure how any of this provides evidence that OpenAI took any of their work.
As evidence against, we never looked at any of their ChatGPT conversations and our model's proof is quite different from theirs.
(I work at OpenAI, but not on the team that did this proof.)
Sorry, but the burden of proof lies in the other direction: OpenAI needs to definitively prove that their agents did not look at the existing work that was about to be published. Otherwise OpenAI simply stole the glory and the spotlight (and I'm being charitable here).
Only the CIA knows whether or not they're actively covering up reptilian space aliens exerting control over the US government. Therefore the burden of proof remains on the CIA to prove that they are not actively participating in such a scheme.
I didn't realize you had insider knowledge about their systems. Do please explain for the class.
As I understand it they will only have trained on his data if he consented to it. Do you have evidence that they do otherwise?
If it was enabled, then their work was included in the training dataset.
At least, as an ignorant outsider, that's how it seems to me.
You don't just get to subpoena your neighbor's bank account because "I know he's stealing from me" you need to first present credible evidence that you were stolen from and that he is among the most likely culprits.
Any other argument, fc417fc802?
See Russell's teapot for an explanation https://en.wikipedia.org/wiki/Russell%27s_teapot
This is incorrect, and you invoke Russell's teapot incorrectly too.
It would only apply if the accusation rested solely on the fact that neither of us have evidence against the accusation.
But that's not the case. First, we know that there could be proof, it's just apparently burdensome and expensive to produce. At that point you're not in fallacy land anymore, you just need a way to balance the cost required of someone to prove the accusations against them false.
Second, we have an arguably plausible mechanism of action that OpenAI does not dispute is possible.
This isn't a legal dispute, so no one is going to force OpenAI to do anything here, but it's not unreasonable (and certainly not fallacious) to suggest that Buckmaster's suggestions are plausible enough it's up to OpenAI to stand behind their denial.
First, the conversations are anonymized, so there's no simple way to inspect the training dataset and identify which specific conversations belong to Buckmaster.
Second, OpenAI uses these anonymized chats to generate synthetic training data, i.e. they fabricate new conversations based on specific conversation patterns where the model performs poorly, and uses these synthetic conversations as training data for future models. The synthetic data could potentially contain some of selections of Buckmaster's original chats, but it is unknowable how his specific writing could have influenced these synthetic data sets or what portion belongs to him. This information is untraceable and effectively double anonymized.
Third, OpenAI explicitly uses user feedback (the thumbs up or thumbs down ratings), as RLHF to train models. However, this feedback is anonymized and stripped of user identifiers. It's not possible to trace a specific feedback to Buckmaster, nor do we know if Buckmaster ever used this feature. I doubt Buckmaster recalls or can provide a list of every time he used this feature over the past year. OpenAI doesn't have one.
Note that the first and second only happen if Buckmaster "Improve the model for everyone" setting enabled, which I find unlikely. But that doesn't exclude option three from this list.
You seem to think that it is some "gotcha" that OpenAI refuses to make a blanket denial, but they cannot do so in good faith, because they have a genuine understanding of their own system. They don't know where the data they have came from.
This situation meets the requirement of Russell's teapot, since neither party has enough evidence to prove nor disprove what information is actually in OpenAI's training set.
But at that point you've circled back around to my original objection. That reasoning isn't limited to user data but applies to literally all the training data which at this point (AFAIK) covers the vast majority of everything ever written.
If we accept that position then what do we make of all the other output? Isn't everything it spits out plagiarized? So then is everyone who uses a frontier model to help them in their research effectively laundering plagiarized work? But then the other researchers involved in this controversy were also using the openai model ...
Tristan + Levent: 3D incompressible Euler with forcing
OpenAI: 3D incompressible Euler without forcing
OpenAI: Navier-Stokes with forcing
No one: Navier-Stokes without forcing
Tristan+Levent ticked the weakest case, OpenAI ticked the two next weakest, then the final case is unsolved. Only the last two are eligible for the Millennium Prize. The Navier-Stokes general case remains unsolved.
Tristan and Levent only solved the easiest version of the problem and did not have the key insights to solve the harder versions of the problem required for the Millennium Prize.
The researchers didn't even solve the same problem as OpenAI, so your argument doesn't hold up.
Let's check if an OpenAI employee agrees with you: https://news.ycombinator.com/item?id=49614154
Nope. Expensive, yes, impossible, no.
Moreover, you (and Sanders) aren't asking the more fundamental questions (and getting sloppy with your assumptions). - was Buckmaster's account set to prevent conversations from being trained on? This is easily answerable - Was user feedback ever activated on the account? I would bet this is logged, even if not tied to specific data - it's very easy to de-anonymize data in practice. In this case, there will be uncommon phrases used in material not published online until after the training material of their internal model was generated. Do any of these phrases appear in that material?
This is not sharing medical information. Companies publish postmortems relating to specific customers all the time with permission of those customers.
> You seem to think that it is some "gotcha" that OpenAI refuses to make a blanket denial
You're putting incorrect words into my mouth.
What I think is that this whole situation speaks to the character of OpenAI as a participant in the mathematics community, especially when they put no effort into getting to the bottom of this, and, of course, when the extent of them reaching out and collaborating with their peers involves rushed Sunday night video calls and apparent pressure on authorship and credit.
That's their choice, there doesn't appear to be anything illegal here, but they're going to continue to get called out on this kind of nonsense which can ruin the big moment they were clearly hoping to have. Bummer, but there are consequences.
The comment by tedsanders you linked directly contradicts your claim. From your linked comment:
There's no reason to believe that anything they did in ChatGPT led to our solution; it's just impossible for us to truly prove it.
In other words, yes, they had been using ChatGPT, and yes, ChatGPT could very well have trained on their data. Now that there is evidence, we need an investigation: yes or no, was it the case?
If there's more to the story I'd be interested to hear it.
I don’t think they’re too concerned about appeasing you, enraged_camel.
For most reasonable people, achievement in solving the other Millenium Prize problems at an unprecedented rate will be enough. At some point people will see models are capable of solving hard issues without whatever 0.00001% of the training data coming from irate individuals who believe their sample was the key component of the solution.
https://en.wikipedia.org/wiki/Burden_of_proof_(philosophy)#P...
It's unknowable and not possible to prove if any one specific conversation was the key to solving Navier–Stokes.
If the conversation was in the training set, there's a high likelihood that the small set of conversations related to solving Navier-Stokes was used by the model. I get Astra to still quote some of my friends' books or blogposts nearly verbatim on certain niche issues.
Much more importantly, we _can_ determine whether a conversation was used in the training data. And if it was, it gives us a great idea whether that logic was captured in reasoning for a novel problem never yet solved.
Given that you don't see any of this as below the belt according to your other comments, maybe your contribution here is more for yourself than a fair conversation about attribution.
You might retort that ChatGPT used the training data to copy their approach, but the approach Buckmaster and Alpöge chose was already published by Luis and Diego in 2023 and in every frontier model's training set.
However the way you are conducting yourself in public, while announcing yourself as an OpenAI employee is doing enormous harm to the greater and magnanimous aim of your organisation. Take a step back and read the temperature of the room. Being the smartest guy in the room will never protect you from alienating the rest of the room into a baying mob. Right now you are Icarus flying straight into the sun.
If I were at OpenAI, I'd naturally want to snipe that from them. I am completely unsurprised they formed a crack team to steal Anthropic's glory, and do so in just five days.
All this to say, trust is important, and grounded in social convention. So I do agree with you, but also disagree.
Whenever this is OK or not really depends on how the breakthrough is contextualized, and how there people at play, here, agree to contextualize it.
In my view, in the blog post, there is much discussion about who will be publishing the paper. If instead it was just a blog post that said "oops, we beat you to it, our model is the best", it would have been different.
FWIW, publicly facing OAI docs are very unclear about whether this setting even applies to Codex conversations.
it is exceptionally unlikely that anything they ever did made it into any part of training, and the chances are zero if they have opted out (likely). it would be a terrible precedent to break the the PII-scrubbing boundary to go and round it down to 0, and we won’t do it
We have lots of examples now of their model doing what they say is impossible.
Now we have another example of something that they say is impossible or very unlikely. Do we take their word for it this time? Really?
Many upstart Chinese labs got around the user data issue by just buying copious amounts of Claude and ChatGPT session logs from model routers.
I don't really understand how the quantity of training data/rollouts used in training is relevant to the question of whether or not it was trained on these conversations.
I also don't really believe that whether or not this model was trained on these conversations is unknowable information.
A model being trained on lots of irrelevant information does not mean relevant information was not used.
How many of those trillion conversations were about Navier-Stokes you reckon?
If prompts were submitted earlier than that and training was not opted out, there's a chance they made their way into our training pipeline in some form. But this would be a droplet in an ocean and unlikely to have made any difference, imo.
See: https://www.nytimes.com/2026/09/10/science/tristan-buckmaste...
I'm not coming from a place of distrust here. This should just be definitively answerable given the weight of the claims here. Surely between you, your lawyers, and other members of your team you can just clear this part up.
Your post says “While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models .” We can discuss what it means to “read” things but obviously the issue here isn't whether you did it manually or automatically.
But more importantly, what on earth are you doing threatening real scientists to remove their coauthors, then making fun of them on social media? Does the entire company run on that toxic culture, or did those people run off of some kind of outrageous tangent?
It's completely childish, and not befitting of the weight of the times we're living in.
The question I am interested in is not "did we read private chats", but "was this new model trained using any of Tristan and Levent's chats, regardless of whether they were marked private". Can you comment on that?
No answer is also an answer.
He's a human, like everybody else. Mostly a bunch of hungry animals looking to put bread in our mouths. It's rarely ever something a bit more sophisticated than that.
If they did not opt out, then I don't personally know if training signals came from their chats, and I don't think we'd be able to tell without their cooperation in identifying them. And even if signals were trained on in some manner, I highly doubt it made a difference to a problem as challenging as the NS proof.
Reasons for my doubt:
- I know most of our training recipes
- Our model's proof is very different from theirs
- The proof took a tremendous amount of tokens to derive (it wasn't a recall/lookup type question)
- This unreleased model has beastly performance on many unsolved math problems, not just the Euler solution
I acknowledge that this requires trust, and if you think we'd lie shamelessly about this stuff, then nothing we say can really help our case here.
Reminds me a bit of the Frontier Math fiasco, where people accused us of training on the eval set (we didn't), but it's hard to convince someone if they think you're lying.
If you're convinced we lie and cheat, then nothing I say may help. But if you're not sure, then hopefully providing my perspective is helpful.
Can't you guys just check their account settings so the public knows what was set?
EDIT: Why was this downvoted? I'm genuinely asking because I have no idea. Opting out is just a normal setting in the profile, It's not like I'm asking for their private conversations or PII. If I were the person claiming that they trained on my conversations, I'd make sure to disclose that I had opted out and hadn't given them permission to do so. And if I were the accused party, I'd disclose whether that setting was turned on or off to provide evidence against the accusation.
Per OpenAI's privacy policy, they use de-identified data to improve their products. From Mark Chen's comment, improving products includes improving ChatGPT and Codex in a holistic way. Improving models in a holistic way sounds a lot like training to me.
That’s not inconsistent with what you responded to. They use your data unless you opt out. If the user doesn’t opt out, their de-identified data is used to improve their products.
I think that's quite a leap. Using de-indetified data to improve the products is what everyone has been doing since the dawn of web analytics.
Even better would be more research and tools to help determine the impact of particular training data on models. Right now, proprietary LLM providers get to hide a lot behind "we just train it, we don't know what inputs affect the outputs," and that can be a problem, both because of lack of traceability of factual informaiton as well as lack of traceability of things like this, where the model itself may have had unpublished work in its training set.
Either way, it seems worth having clarity, and I'm a bit surprised OpenAI's stance is just "we can't rule this out, but don't worry about it". OpenAI is, apparently, very happy to use unreleased models to try to scoop big results if they get a whiff that someone else is close (which strikes me as pretty scummy regardless of any issues of training contamination). It seems like people who might want to use OpenAI's models as part of their research would want to be very clear about whether doing so can make them, even in principle, more likely to fall victim to this.
Fall “victim” to what? Having their responses in the training data if they fail to opt out? That is what will happen.
If you’re referring to falling “victim” to OpenAI scooping a problem discussed in training, this also wasn’t the case. They chose the problem based off human-spread rumors.
are opted-out-of-training chats ever paraphrased, and thus "de-identified" (in openai's own words)? what this means is, it would be hard to prove that a "synthetic" (but actually paraphrased) training instance came from a particular chat. except if the chat was about some esoteric math proof, of course.
That is your opinion, but the optics of that should raise for you some flags. OAI could have waited (how long is a task left to the ethics committee) to see how the rumors panned out. Right now the optics look a lot like "we don´t care there is a 1/7 chance we one-up a human researcher by reacting to this rumor immediately, might makes right"
Every company I've worked at has said very clearly not to comment about work things on social media. I would definitely have been in serious trouble for this. I imagine some strongly worded emails from marketing are flying around OpenAI right now.
If someone was speaking on behalf of their employer, they would've used the official channels (such as I dunno an `openai` HN handle, or whatever other channel).
I'm outraged that people think this "opinions are my own" disclaimer is ever necessary.
Turns out it wasn’t actually Anthropic and just a researcher with a single Anthropic guy friend working on it .
Wild times
My guess was it was probably this was more a nerd snipe than any action from OpenAI that was a "massive team" being put on it. Literally someone looking at this and asking "I wonder if our models are good enough yet".
It's easy to assume that having access to massive compute amounts means significant coordination, but this assumes that you're looking at the costs of this sort of thing from an external lens. Internally, tokens are often treated as free and infinite.
If you consider this as a cost to be compared against the question: "What does it take to be able to prove that you have a model that can solve the hardest problems that humans know about?", then spending a some amount of thousands/millions to know the boundaries of that seems not too important in comparison.
You've also got to consider this as compute that's allocated to pushing the frontier of what models can do, so while it's using GPUs that have been paid for etc., it's not like it's a cost that's supposed to be use less of this so that others can have capacity. If you made researchers afraid to use capacity like this, a lot of the things that improve would tend to do so significantly slower. (some may say that's a good thing ;)
A good way to think about this is when tokens are free, you get to choose whether you're optimizing for latency or intelligence rather than having to consider price.
---
Publically, tibo (Codex owner) in Feb this year: https://x.com/thsottiaux/status/2024649339344445825
> OpenAI employees currently get unlimited inference. Usage is now peaking at > XX billion tokens per week for some of them.
Mathew Berman (AI Youtuber) in Jun: https://x.com/MatthewBerman/status/2067270730795134984
> I've used 25 billion tokens in the last 7 days.
I'm hearing two completely conflicting stories. Buckmaster is claiming the approach used by OpenAI is so strikingly similar to the one he used, that mere coincidence is astronomically small. Yet OpenAI is claiming that the methods used are entirely different.
Anyone care to provide primary evidence proving one way or the other?
OpenAI's approach was to copy his work, which is technically a different method of coming up with an approach.
Details may be different, but the use of a very similar tack is very suspicious. Combined with thuggish comments "why would you ruin your career" and "I don't have to be nice" take away pretty much any credibility the OpenAI team's statements might have had.
This really need to be a top-level story on HN..
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.
This whole episode is more horrific than "AI is eating math". We now have a clear and economically damaging (or at least career damaging) example of the "training on customer tokens" problem.
We can't ignore this problem any longer.
It’s fair to give benefit of doubt to Buckmaster given OpenAI is currently being very credibly sued by Apple for openly stealing others’ original work in another context.
In short I think a lot of people are jumping to conclusions without supporting evidence and that's really not helping the situation.
The guy mocked accessing his prior employer's circuit diagrams and was protected by OpenAI until Apple filed suit.
Tabula rasa, sure, we need more evidence. But ignoring the priors should at least be explicitly acknowledged.
> We (the researchers and the agents) did not see any of their work through any means until they released it publicly — in particular, no specific user data was accessed in order to solve this problem. While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models . However, our proofs differ significantly and even the precise results proved are different in the Euler case (forced vs unforced).
That they don't is telling.
It's unknowable and not possible to prove if any one specific conversation contained the insights for solving Navier–Stokes.
We also don't know if the authors unintentionally provided data to OpenAI through alternate means, such as via alternate accounts or model feedback queries.
Knowability and likelihood are almost orthogonal here. If I commit a crime and perfectly destroy the evidence, my deed may be unknowable. That doesn’t make it more or less likely.
> not possible to prove if any one specific conversation contained the insights for solving Navier–Stokes
It may be. We haven’t seen the researchers’ transcripts. We don’t know what Buckmaster or his co-author uploaded to OpenAI or with what permissions (or if OpenAI actually respects those toggles).
The fact they can't make a blanket denial triggers everyone's bullshit detectors, and they're getting eviscerated over it.
Again, see the Apple lawsuit. OpenAI has never been constrained by facts in what it can and can’t say.
To the extent anything is setting off my bullshit detector, it’s in the idea that this time is different (Moreover, the idea that we should assume this divergence without evidence.)
OpenAI doesn’t have the benefit of doubt. They shouldn’t for anyone who’s honest and reasonable. That doesn’t mean they’re automatically at fault. But when the twentieth person comes forward and says a pattern is continuing, I’m giving them the preliminary benefit of doubt. It’s a bit extreme to conclude based on that. But it’s far more baseless to swing to the other side and claim we need to clear the table for a serial offender.
Further, given that this is all in the open now, they can search the training data. No way somebody is using some specific unique cutting edge mathematical approach to solve a fluid dynamic problem 99.9% of people have never heard of and it's not locatable. Considering they spent $15,000,000 already on this, they could afford to grep around to be able to state that their hands are clean.
No, it’s genuinely impossible to know how much of Buckmaster’s Codex data is in OpenAI’s training set.
First, the conversations are anonymized, so there's no simple way to inspect the training dataset and identify which specific conversations belong to Buckmaster.
Second, OpenAI uses these anonymized chats to generate synthetic training data, i.e. they fabricate new conversations based on specific conversation patterns where the model performs poorly, and uses these synthetic conversations as training data for future models. The synthetic data could potentially contain some of selections of Buckmaster's original chats, but it is unknowable how his specific writing could have influenced these synthetic data sets or what portion belongs to him. This information is untraceable and effectively double anonymized.
Third, OpenAI explicitly uses user feedback (the thumbs up or thumbs down ratings), as RLHF to train models. However, this feedback is anonymized and stripped of user identifiers. It's not possible to trace a specific feedback to Buckmaster, nor do we know if Buckmaster ever used this feature. I doubt Buckmaster recalls or can provide a list of every time he used this feature over the past year. OpenAI doesn't have one.
Note that the first and second only happen if Buckmaster "Improve the model for everyone" setting enabled, which I find unlikely. But that doesn't exclude option three from this list.
>our proofs differ significantly and even the precise results proved are different in the Euler case (forced vs unforced)