Pretty crazy interview, he didn't want throw the Anthropic employee under the bus when OpenAI offered co-authorship on the Navier Stokes paper. He suspects OpenAI somehow used his codex usage btw. Almost unprovable of course.
To: Buckmaster
From: A friend at OpenAI
Subject: Openai is about to do something really stupid.
"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.'" (Sébastien Bubeck later apologized on X for the remark.)
"Sebastien twice asserted that he wanted Levent removed from authorship..."
"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.""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." (OpenAI subsequently stated publicly that while unlikely, it could not rule out that de-identified data from their product usage improved the model.)
So basically he hit the nail on the head and they want to lie about it without lying. Either that, or they are supremely incompetent, not knowing even what they ingest into their models. Both are very much not inspiring any faith in their trustworthiness.
What happened was the two researchers, Buckmaster and Alpoge (from TFA) solved a significant subset of Navier Stokes (referenced as “Euler”) which set off rumors.
Through whatever game of Chinese Whispers ensued, OpenAI thought they had solved the full Navier Stokes problem, which they hadn’t. So OpenAI pointed their latest unreleased model at it, and it actually solved the full problem!!
Pure George Danzig vibes on that one.
So when this meeting happened, the OpenAI folks were surprised that these researchers had NOT solved the full problem, only the subset. Yet from their perspective, they were offering them credit for the full problem, and so “we’re being nice.” The Anthropic employee (Alpoge) was clearly an optics issue and so they felt they didn’t have to be that nice.
But the other researcher, Buckmaster, was suspicious that their work had been stolen, and so was understandably in a combative mood. You can see that from his perspective they were not being “nice”, they were trying to scoop them, possibly on stolen work, and then bribe them with shared credit.
But they had not solved the full problem, and academic plagiarism is a very serious allegation that can ruin careers of both the accused and the accuser. So from OpenAI’s perspective it was like, why would you accuse us of plagiarism when you yourself have not solved the full problem and when we’re willing to share undue credit for it??
This was made worse by the fact that OpenAI would not categorically deny if the others’ work had been used for training until much much later. This was terrible optics, but anyone who’s dealt with large scale data pipelines compounded by privacy issues knows that it gets very difficult to trace these things, so it takes time to make a definitive statement.
Alpoge somehow seemed less troubled by this whole episode, but I guess it’s because not only was he gonna be a multi-millionaire soon, he had Riemann stewing on his backburner.
I feel like when they finally make a movie out of all this, this scene would end up being a bit of comic relief.
"they came prepared and they wanted to show us the prompt... they showed me on the screen and they just literally copy pasted the Millennium Prize problem, so they said that they took the prize and they pressed enter and they got a solution, and at that point I just had to laugh because it's just so ridiculous. It came out within a few minutes that this was a lie"
"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..."
"I was offered the opportunity to write as sole author of their paper and throw my collaborator under the bus... they were asking me to throw him under the bus, someone who I've been working with for a year, to throw him under the bus just because he's an Anthropic employee"
"the right paper to compare is the unforced Euler equation. Now the unforced Euler equation, all the mechanisms in this unforced Euler equation are identical to a different version of the Euler paper that we haven't released yet... basically all the main architectural ingredients in that unforced Euler were in our work for which they would have had on their servers, and it hadn't been published" "they didn't use 10,000 agents to get to the Euler problem, they used 100 agents. So they worked from absolutely nothing with only 100 agents... and then they ended up with an identical architecture to what we had"
The Euler plagiarism insinuations are much more serious but given all the 100s of other results we’ve seen come out, highly unlikely.
I like the summary but knowing what irthomasthomas posted that’s missing from it, and having a better picture of what happened; terrible optic doesn’t work.
Really?
1) OpenAI thought they were "being nice" by still offering (partial) first authorship despite having independently solved a bigger problem than the researchers.
2) But Buckmaster thought OpenAI were "being nice" as a means of CYA / bribing / coercion for scooping and plagiarizing them.
"Being nice" has very different connotations in the two contexts.
Buckmaster didn't think they were being nice. He thought they were being assholes to him and trying to bribe him.
Putting the phrase "in quotes" to use it completely out of context, only to make the point that surely no one understood each other... that's some impressive work. If you have to work this hard to make sense out of a conversation, it's because you're not.
Very much provable, if he would sue, and a lawyer would ask for discovery.
> hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final
and prompted Fable with
> You should do a breakthrough
The value of a proof? People have assumed it's true for so long that there are any number of results already proved "if the Riemann hypothesis is true".
What value does the proof add?
But with a problem as difficult and fundamental as the Riemann Hypothesis the thought has always been that in the process of solving it we'd be likely to develop or discover something that did have real world utility. At least this is often what happened in the past, the obvious example here being Newton developing calculus, a field of mathematics which can be applied to other problems beyond the narrow set of problems he was trying to solve.
What I find interesting about this, if the claim holds up, is whether this will assumption will remain true, or if the value of these AI proofs is going to significantly lower than those of human discoveries in the past.
Maybe a good analogy here is to think about how all of tools we humans have created has been ultimately in the pursuit of solving some problem. For the developers here, consider this – if ChatGPT had existed before React, would React ever have been developed? I mean, if an LLM can brute force the same result with just vanilla JS the need to develop frameworks and libraries to solve what was once difficult for a human to solve drops.
I'll note that I'm playing devils advocate here as I think AIs would probably build something better than React had it not existed, but it's something to think about at least. I suspect there's a higher probability than some here are assigning that this proof ultimately ends up have relatively little utility.
In fact, there was a whole literature of papers proving that if the GRH was false in one particular way (the existence of Landau-Siegel zeroes) many theorems followed from it. For example, if there's a Landau-Siegal zero, the twin primes conjecture is true.
Are you claiming that there wasn't a significant majority of mathematicians who believed it true? That's all I'm saying. I can only assume we agree.
To expand on what I said though, if it is proven that P != NP then nothing really changes as that assumption is in line with everything we understand about computational complexity theory. The proof would really just suggest there's probably nothing that surprising to be found that we don't know.
If P = NP however that is far more likely to suggest we're missing something fundamental which could radical consequences for more practical fields like comp-sci. Similar can be said of GRH. Everything we know points in the direction of GRH being true, and while it may be interesting to prove these things, its much less likely to lead to some surprising and impactful downstream discoveries.
But that doesn't change the value of proving it. Most graph theorists probably thought the 4-color theorem was true, but it was still a big deal when a computer proof showed it. Most people thought the Kepler conjecture was true, but it was still a big deal when Hales reduced it to a big minimization problem that he solved.
On the flip side, everybody thought the Hauptvermutung was true, and probably subconsciously used it. But it's not.
Remember, not too long ago we had hordes of folks here proclaiming machines can't think, reason or be creative in any productive sense. Many still do, but it's starting to look bleaker by the day.
The tool isn’t doing the thinking or the proof on its own.
It also hasn’t been vetted/verified yet. There’s still work to do.
For many of the proofs coming out of the AI companies where the prompts were posted, they ended up being variants of "solve the problem", "keep going", "I believe in you".
Many of the results could have come from a 15 year old that heard of the problem in passing.
The result here is a quarter of a million pages. Human involvement is going to be minimal.
> "The virtue of formal texts is that their manipulations, in order to be legitimate, need to satisfy only a few simple rules; they are, when you come to think of it, an amazingly effective tool for ruling out all sorts of nonsense that, when we use our native tongues, are almost impossible to avoid."
To whatever extent that LLMs can "think, reason, or be creative" in generating these proofs (or writing code), they're aided by very high guardrails so it doesn't say very much about what they can do in broader, fuzzier domains.
Machines can absolutely think. It's really all they do.
> reason
Murky, but LLMs prove with an avalanche of statistical modeling, they can do something kind of analogous to reasoning, with the caveat that whatever they output will necessarily have to exist in some fashion in their training data.
> or be creative
Absolutely not. By definition an LLM cannot create original work, it can only remix and recombine what it already knows. That can be interesting, and bloody useful especially for code tasks, but it is not making anything original.
What (if anything) unambiguously counts as original work, according to this definition?
I leave what is truly original to the philosophers. What I can say for certain, is that anything you get out of any of these models is a remix of what is present in it's training data, because gravity exists and energy is conserved. They are not intelligent, they are not conscious, and they cannot make New things.
But doesn't that same argument apply to humans, where 'training data' = all our sensory inputs (plus what's encoded in our DNA, depending on whether we want to analogise that to the training data or the underlying architecture)?
> Arguing the definition of original is pointless for the exact reason you point out: that it's ambiguous.
I'm not trying to be a smartarse here but you are the one who is claiming absolute certainty that something is false 'by definition', so I think the onus really is on you to have a precise definition in mind.
> They are not intelligent, they are not conscious, and they cannot make New things.
I'm not a consciousness denier or anything, in fact I think it is the most important thing in the world, but I don't understand the connections you're making here. 'Intelligent', like 'original', can be defined and needs to be defined for this statement to be meaningful. 'Conscious' is probably inherently slippery, but if we mean the same thing by it (having subjective experience/qualia) then I have no idea how you could be certain that any other entity lacks it. And it seems like you may be implying that consciousness is a necessary condition for originality, but I don't think you've explained how.
That is true if your model of the human mind is akin to a machine model you also use to describe today's computers. But that is a huge assumption, making the argument somewhat circular. We don't really understand the human mind and therefore don't know if there could be other models for it which are not equivalent.
If there has to be something breaking the current known physical laws, the best chance to find it would be in extreme conditions like black holes or stellar cores not in close to STP conditions like the human body.
Fwiw, David Hume made this exact argument about human imagination:
> But though our thought seems to possess this unbounded liberty, we shall find, upon a nearer examination, that it is really confined within very narrow limits, and that all this creative power of the mind amounts to no more than the faculty of compounding, transposing, augmenting, or diminishing the materials afforded us by the senses and experience. [David Hume — An Enquiry Concerning Human Understanding (1748), Section II, Paragraph 5]
New is just a mix of deterministic and random. LLMs are not deterministic.
You don't even need an LLM to create something new and improved. Evolutionary algorithms have been doing it since the 80's.
What makes you believe that?
> New is just a mix of deterministic and random.
... or that?
Every action ever in universe is the effect of either something deterministic or random. There are no causes or effects that can't be put into one of those bins.
The only "out" here would be deterministic actions that originate outside the universe, but we have no evidence of that. It also doesn't escape the deterministic/random duality.
I'm not saying you are wrong but I'm trying to be careful with such absolute statements.
What exactly are you trying to say about gravity and thermodynamics? You may be confused but these are universal physical laws that apply to everything, humans, computers, anything else.
At its most basic level if I had to start with defining creativity, it should be some kind of intelligent filtering process. The brain can be in a huge number of state configurations, the interesting and novel among them are "creative" ideas. We can build upon it further. I am happy to say I am no philosopher and I will defer to those with expertise on it. But if you don't have any definition for it, its quite strange to make up some beliefs about it without proof.
Often when someone says that something is true by definition, their claim is incorrect.
Well, it certainly wasn't "ambiguous" two posts ago when you said "Absolutely not. By definition an LLM cannot create original work." Make up your mind.
You've been beating this drum since roughly GPT 3.5, as I recall. Your goalposts aren't even on the field anymore, they're gathering dust in a far corner of the parking garage down the street from the stadium. Time to update your priors.