The headline may make it seem like AI just discovered some new result in physics all on its own, but reading the post, humans started off trying to solve some problem, it got complex, GPT simplified it and found a solution with the simpler representation. It took 12 hours for GPT pro to do this. In my experience LLM’s can make new things when they are some linear combination of existing things but I haven’t been to get them to do something totally out of distribution yet from first principles.
Humans have worked out the amplitudes for integer n up to n = 6 by hand, obtaining very complicated expressions, which correspond to a “Feynman diagram expansion” whose complexity grows superexponentially in n. But no one has been able to greatly reduce the complexity of these expressions, providing much simpler forms. And from these base cases, no one was then able to spot a pattern and posit a formula valid for all n. GPT did that.
Basically, they used GPT to refactor a formula and then generalize it for all n. Then verified it themselves.
> I think this was all already figured out in 1986 though
They cite that paper in the third paragraph...
Naively, the n-gluon scattering amplitude involves order n! terms. Famously, for the special case of MHV (maximally helicity violating) tree amplitudes, Parke and Taylor [11] gave a simple and beautiful, closed-form, single-term expression for all n.
It also seems to be a main talking point.
I think this is a prime example of where it is easy to think something is solved when looking at things from a high level but making an erroneous conclusion due to lack of domain expertise. Classic "Reviewer 2" move. Though I'm not a domain expert and so if there was no novelty over Parke and Taylor I'm pretty sure this will get thrashed in review.
You're right. Parke & Taylor showed the simplest nonzero amplitudes have two minus helicities while one-minus amplitudes vanish (generically). This paper claims that vanishing theorem has a loophole - a new hidden sector exists and one-minus amplitudes are secretly there, but distributional
It bears repeating that modern LLMs are incredibly capable, and relentless, at solving problems that have a verification test suite. It seems like this problem did (at least for some finite subset of n)!
This result, itself does not generalize to open-ended problems, whether in business or in research in general. Discovering the specification to build is often the majority of the battle. LLMs aren't bad at this, per se, but they're nowhere near as reliably groundbreaking as they are on verifiable problems.
I'm not sure if GPTs ability goes beyond a formal math package's in this regard or its just its just way more convienient to ask ChatGPT rather than using these software.
After last month’s Erdos problems handling by LLMs at this point everyone writing papers should be aware that literature checks are approximately free, even physicists.
> but I haven’t been to get them to do something totally out of distribution yet from first principles
Can humans actually do that? Sometimes it appears as if we have made a completely new discovery. However, if you look more closely, you will find that many events and developments led up to this breakthrough, and that it is actually an improvement on something that already existed. We are always building on the shoulders of giants.
You could nitpick a rebuttal, but no matter how many people you give credit, general relativity was a completely novel idea when it was proposed. I'd argue for special relatively as well.
I am not a scientific historian, or even a physicist, but IMO relativity has a weak case for being a completely novel discovery. Critique of absolute time and space of Newtonian physics was already well underway, and much of the methodology for exploring this relativity (by way of gyroscopes, inertial reference frames, and synchronized mechanical clocks) were already in parlance. Many of the phenomena that relativity would later explain under a consistent framework already had independent quasi-explanations hinting at the more universal theory. Poincare probably came the closest to unifying everything before Einstein:
> In 1902, Henri Poincaré published a collection of essays titled Science and Hypothesis, which included: detailed philosophical discussions on the relativity of space and time; the conventionality of distant simultaneity; the conjecture that a violation of the relativity principle can never be detected; the possible non-existence of the aether, together with some arguments supporting the aether; and many remarks on non-Euclidean vs. Euclidean geometry.
Now, if I had to pick a major idea that seemed to drop fully-formed from the mind of a genius with little precedent to have guided him, I might personally point to Galois theory (https://en.wikipedia.org/wiki/Galois_theory). (Ironically, though, I'm not as familiar with the mathematical history of that time and I may be totally wrong!)
Right on with special relativity—Lorentz also was developing the theory and was a bit sour that Einstein got so much credit. Einstein basically said “what if special relativity were true for all of physics”, not just electromagnetism, and out dropped e=mc^2. It was a bold step but not unexplainable.
As for general relativity, he spent several years working to learn differential geometry (which was well developed mathematics at the time, but looked like abstract nonsense to most physicists). I’m not sure how he was turned on to this theory being applicable to gravity, but my guess is that it was motivated by some symmetry ideas. (It always come down to symmetry.)
> The quintic was almost proven to have no general solutions by radicals by Paolo Ruffini in 1799, whose key insight was to use permutation groups, not just a single permutation.
Thing is, I am usually the kind of person who defends the idea of a lone genius. But I also believe there is a continuous spectrum, no gaps, from the village idiot to Einstein and beyond.
Let me introduce, just for fun, not for the sake of any argument, another idea from math which I think it came really out of the blue, to the degree that it's still considered an open problem to write an exposition about it, since you cannot smoothly link it to anything else: forcing.
Even if I grant you that, surely we’ve moved the goal posts a bit if we’re saying the only thing we can think of that AI can’t do is the life’s work of a man who’s last name is literally synonymous with genius.
Not really. Pretty sure I read recently that Newton appreciated that his theory was non-local and didn't like what Einstein later called "spooky action at a distance". The Lorentz transform was also known from 1887. Time dilation was understood from 1900. Poincaré figured out in 1905 that it was a mathematical group. Einstein put a bow on it all by figuring out that you could derive it from the principle of relativity and keeping the speed of light constant in all inertial reference frames.
I'm not sure about GR, but I know that it is built on the foundations of differential geometry, which Einstein definitely didn't invent (I think that's the source of his "I assure you whatever your difficulties in mathematics are, that mine are much greater" quote because he was struggling to understand Hilbert's math).
And really Cauchy, Hilbert, and those kinds of mathematicians I'd put above Einstein in building entirely new worlds of mathematics...
From my reading yes, but I think I am likely reading the statement differently than you are.
> from first principles
Doing things from first principles is a known strategy, so is guess and check, brute force search, and so on.
For an llm to follow a first principles strategy I would expect it to take in a body of research, come up with some first principles or guess at them, then iteratively construct and tower of reasonings/findings/experiments.
Constructing a solid tower is where things are currently improving for existing models in my mind, but when I try openai or anthropic chat interface neither do a good job for long, not independently at least.
Humans also often have a hard time with this in general it is not a skill that everyone has and I think you can be a successful scientist without ever heavily developing first principles problem solving.
The process you’re describing is humans extending our collective distribution through a series of smaller steps. That’s what the “shoulders of giants” means. The result is we are able to do things further and further outside the initial distribution.
So it depends on if you’re comparing individual steps or just the starting/ending distributions.
"GPT did this". Authored by Guevara (Institute for Advanced Study), Lupsasca (Vanderbilt University), Skinner (University of Cambridge), and Strominger (Harvard University).
Probably not something that the average GI Joe would be able to prompt their way to...
I am skeptical until they show the chat log leading up to the conjecture and proof.
I'm a big LLM sceptic but that's… moving the goalposts a little too far. How could an average Joe even understand the conjecture enough to write the initial prompt? Or do you mean that experts would give him the prompt to copy-paste, and hope that the proverbial monkey can come up with a Henry V? At the very least posit someone like a grad student in particle physics as the human user.
I would interpret it as implying that the result was due to a lot more hand-holding that what is let on.
Was the initial conjecture based on leading info from the other authors or was it simply the authors presenting all information and asking for a conjecture?
Did the authors know that there was a simpler means of expressing the conjecture and lead GPT to its conclusion, or did it spontaneously do so on its own after seeing the hand-written expressions.
These aren't my personal views, but there is some handwaving about the process in such a way that reads as if this was all spontaneous involvement on GPTs end.
But regardless, a result is a result so I'm content with it.
In this case there certainly were experts doing hand-holding. But simply being able to ask the right question isn't too much to ask, is it? If it had been merely a grad student or even a PhD student who had asked ChatGPT to figure out the result, and ChatGPT had done that, even interactively with the student, this would be huge news. But an average person? Expecting LLMs to transcend the GIGO principle is a bit too much.
they probably also acknowledge pytorch, numpy, R ... but we don't attribute those tools as the agent who did the work.
I know we've been primed by sci-fi movies and comic books, but like pytorch, gpt-5.2 is just a piece of software running on a computer instrumented by humans.
I don't see the authors of those libraries getting a credit on the paper, do you ?
>I know we've been primed by sci-fi movies and comic books, but like pytorch, gpt-5.2 is just a piece of software running on a computer instrumented by humans.
And we are just a system running on carbon-based biology in our physics computer run by whomever. What makes us special, to say that we are different than GPT-5.2?
> And we are just a system running on carbon-based biology in our physics computer run by whomever. What makes us special, to say that we are different than GPT-5.2?
Do you really want to be treated like an old PC (dismembered, stripped for parts, and discarded) when your boss is done with you (i.e. not treated specially compared to a computer system)?
But I think if you want a fuller answer, you've got a lot of reading to do. It's not like you're the first person in the world to ask that question.
It's always a value decision. You can say shiny rocks are more important than people and worth murdering over.
Not an uncommon belief.
Here you are saying you personally value a computer program more than people
It exposes a value that you personally hold and that's it
That is separate from the material reality that all this AI stuff is ultimately just computer software... It's an epistemological tautology in the same way that say, a plane, car and refrigerator are all just machines - they can break, need maintenance, take expertise, can be dangerous...
LLMs haven't broken the categorical constraints - you've just been primed to think such a thing is supposed to be different through movies and entertainment.
I hate to tell you but most movie AIs are just allegories for institutional power. They're narrative devices about how callous and indifferent power structures are to our underlying shared humanity
Their point is, would you be able to prompt your way to this result? No. Already trained physicists working at world-leading institutions could. So what progress have we really made here?
Hmm feels a bit trivializing, we don't know exactly how difficult it was to come up with the generic set of equations mentioned from the human starting point.
I can claim some knowledge of physics from my degree, typically the easy part is coming up with complex dirty equations that work under special conditions, the hard part is the simplification into something elegant, 'natural' and general.
Also
"LLM’s can make new things when they are some linear combination of existing things"
Doesn't really mean much, what is a linear combination of things you first have to define precisely what a thing is?
When chess engines were first developed, they were strictly worse than the best humans. After many years of development, they became helpful to even the best humans even though they were still beatable (1985–1997). Eventually they caught up and surpassed humans but the combination of human and computer was better than either alone (~1997–2007). Since then, humans have been more or less obsoleted in the game of chess.
Five years ago we were at Stage 1 with LLMs with regard to knowledge work. A few years later we hit Stage 2. We are currently at Stage 3 for an extremely high percentage of knowledge work. Stage 4 will come, and I would wager it's sooner rather than later.
With a chess engine, you could ask any practitioner in the 90's what it would take to achieve "Stage 4" and they could estimate it quite accurately as a function of FLOPs and memory bandwidth. It's worth keeping in mind just how little we understand about LLM capability scaling. Ask 10 different AI researchers when we will get to Stage 4 for something like programming and you'll get wild guesses or an honest "we don't know".
> With a chess engine, you could ask any practitioner in the 90's what it would take to achieve "Stage 4" and they could estimate it quite accurately as a function of FLOPs and memory bandwidth.
And the same practitioners said right after deep blue that go is NEVER gonna happen. Too large. The search space is just not computable. We'll never do it. And yeeeet...
Chess grandmasters are living proof that it’s possible to reach grandmaster level in chess on 20W of compute. We’ve got orders of magnitude of optimizations to discover in LLMs and/or future architectures, both software and hardware and with the amount of progress we’ve got basically every month those ten people will answer ‘we don’t know, but it won’t be too long’. Of course they may be wrong, but the trend line is clear; Moore’s law faced similar issues and they were successively overcome for half a century.
The evolution was also interesting: first the engines were amazing tactically but pretty bad strategically so humans could guide them. With new NN based engines they were amazing strategically but they sucked tactically (first versions of Leela Chess Zero). Today they closed the gap and are amazing at both strategy and tactics and there is nothing humans can contribute anymore - all that is left is to just watch and learn.
Serious questions, I often hear about this "let the LLM cook for hours" but how do you do that in practice and how does it manages its own context? How doesn't it get lost at all after so many tokens?
I’m guessing, would love someone who has first hand knowledge to comment. But my guess is it’s some combination of trying many different approaches in parallel (each in a fresh context), then picking the one that works, and splitting up the task into sequential steps, where the output of one step is condensed and is used as an input to the next step (with possibly human steering between steps)
From what I've seen is a process of compacting the session once it reaches some limit, which basically means summarizing all the previous work and feeding it as the initial prompt for the next session.
In my experience humans can make new things when they are some linear combination of existing things but I haven’t been able to get them to do something totally out of distribution yet from first principles[0].
They want to be seen alongside the big news from their competitors, so it looks less like they are falling behind, let me know when they have a passing pre-train run again (apparently haven't since Ilya left)
I don't want to be rude but like, maybe you should pre-register some statement like "LLMs will not be able to do X" in some concrete domain, because I suspect your goalposts are shifting without you noticing.
We're talking about significant contributions to theoretical physics. You can nitpick but honestly go back to your expectations 4 years ago and think — would I be pretty surprised and impressed if an AI could do this? The answer is obviously yes, I don't really care whether you have a selective memory of that time.
It's a nontrivial calculation valid for a class of forces (e.g. QCD) and apparently a serious simplification to a specific calculation that hadn't been completed before. But for what it's worth, I spent a good part of my physics career working in nucleon structure and have not run across the term "single minus amplitudes" in my memory. That doesn't necessarily mean much as there's a very broad space work like this takes place in and some of it gets extremely arcane and technical.
One way I gauge the significance of a theory paper are the measured quantities and physical processes it would contribute to. I see none discussed here which should tell you how deep into math it is. I personally would not have stopped to read it on my arxiv catch-up
I never said LLMs will not be able to do X. I gave my summary of the article and my anecdotal experiences with LLMs. I have no LLM ideology. We will see what tomorrow brings.
> We're talking about significant contributions to theoretical physics.
Whoever wrote the prompts and guided ChatGPT made significant contributions to theoretical physics. ChatGPT is just a tool they used to get there. I'm sure AI-bloviators and pelican bike-enjoyers are all quite impressed, but the humans should be getting the research credit for using their tools correctly. Let's not pretend the calculator doing its job as a calculator at the behest of the researcher is actually a researcher as well.
If this worked for 12 hours to derive the simplified formula along with its proof then it guided itself and made significant contributions by any useful definition of the word, hence Open AI having an author credit.
How much precedence is there for machines or tools getting an author credit in research? Genuine question, I don't actually know. Would we give an author credit to e.g. a chimpanzee if it happened to circle the right page of a text book while working with researchers, leading them to a eureka moment?
Would it? I think there's a difference between "the researchers used ChatGPT" and "one of the researchers literally is ChatGPT." The former is the truth, and the latter is the misrepresentation in my eyes.
I have no problem with the former and agree that authors/researchers must note when they use AI in their research.
> now you are debating exactly how GPT should be credited. idk, I'm sure the field will make up some guidance
In your eyes maybe there's no difference. In my eyes, big difference. Tools are not people, let's not further the myth of AGI or the silly marketing trend of anthropomorphizing LLMs.
>How much precedence is there for machines or tools getting an author credit in research?
Well what do you think ? Do the authors (or a single symbolic one) of pytorch or numpy or insert <very useful software> typically get credits on papers that utilize them heavily?
Well Clearly these prominent institutions thought GPT's contribution significant enough to warrant an Open AI credit.
>Would we give an author credit to e.g. a chimpanzee if it happened to circle the right page of a text book while working with researchers, leading them to a eureka moment?
Cool Story. Good thing that's not what happened so maybe we can do away with all these pointless non sequiturs yeah ? If you want to have a good faith argument, you're welcome to it, but if you're going to go on these nonsensical tangents, it's best we end this here.
> Well what do you think ? Do the authors (or a single symbolic one) of pytorch or numpy or insert <very useful software> typically get credits on papers that utilize them heavily ?
I don't know! That's why I asked.
> Well Clearly these prominent institutions thought GPT's contribution significant enough to warrant an Open AI credit.
Contribution is a fitting word, I think, and well chosen. I'm sure OpenAI's contribution was quite large, quite green and quite full of Benjamins.
> Cool Story. Good thing that's not what happened so maybe we can do away with all these pointless non sequiturs yeah ? If you want to have a good faith argument, you're welcome to it, but if you're going to go on these nonsensical tangents, it's best we end this here.
It was a genuine question. What's the difference between a chimpanzee and a computer? Neither are humans and neither should be credited as authors on a research paper, unless the institution receives a fat stack of cash I guess. But alas Jane Goodall wasn't exactly flush with money and sycophants in the way OpenAI currently is.
If you don't read enough papers to immediately realize it is an extremely rare occurrence then what are you even doing? Why are you making comments like you have the slightest clue of what you're talking about? including insinuating the credit was what...the result of bribery?
You clearly have no idea what you're talking about. You've decided to accuse prominent researchers of essentially academic fraud with no proof because you got butthurt about a credit. You think your opinion on what should and shouldn't get credited matters ? Okay
Do I need to be credentialed to ask questions or point out the troubling trend of AI grift maxxers like yourself helping Sam Altman and his cronies further the myth of AGI by pretending a machine is a researcher deserving of a research credit? This is marketing, pure and simple. Close the simonw substack for a second and take an objective view of the situation.
If a helicopter drops someone off on the top of Mount Everest, it's reasonable to say that the helicopter did the work and is not just a tool they used to hike up the mountain.
Who piloted the helicopter in this scenario, a human or chatgpt? You'd say the pilot dropped them off in a helicopter. The helicopter didn't fly itself there.
“They have chosen cunning instead of belief. Their prison is only in their minds, yet they are in that prison; and so afraid of being taken in that they cannot be taken out.”
Is every new thing not just combinations of existing things? What does out of distribution even mean? What advancement has ever made that there wasn’t a lead up of prior work to it? Is there some fundamental thing that prevents AI from recombining ideas and testing theories?
> Is every new thing not just combinations of existing things?
If all ideas are recombinations of old ideas, where did the first ideas come from? And wouldn't the complexity of ideas be thus limited to the combined complexity of the "seed" ideas?
I think it's more fair to say that recombining ideas is an efficient way to quickly explore a very complex, hyperdimensional space. In some cases that's enough to land on new, useful ideas, but not always. A) the new, useful idea might be _near_ the area you land on, but not exactly at. B) there are whole classes of new, useful ideas that cannot be reached by any combination of existing "idea vectors".
Therefore there is still the necessity to explore the space manually, even if you're using these idea vectors to give you starting points to explore from.
All this to say: Every new thing is a combination of existing things + sweat and tears.
The question everyone has is, are current LLMs capable of the latter component. Historically the answer is _no_, because they had no real capacity to iterate. Without iteration you cannot explore. But now that they can reliably iterate, and to some extent plan their iterations, we are starting to see their first meaningful, fledgling attempts at the "sweat and tears" part of building new ideas.
For example, ever since the first GPT 4 I’ve tried to get LLM’s to build me a specific type of heart simulation that to my knowledge does not exist anywhere on the public internet (otherwise I wouldn’t try to build it myself) and even up to GPT 5.3 it still cannot do it.
But I’ve successfully made it build me a great Poker training app, a specific form that also didn’t exist, but the ingredients are well represented on the internet.
And I’m not trying to imply AI is inherently incapable, it’s just an empirical (and anecdotal) observation for me. Maybe tomorrow it’ll figure it out. I have no dogmatic ideology on the matter.
"An internal scaffolded version of GPT‑5.2 then spent roughly 12 hours reasoning through the problem, coming up with the same formula and producing a formal proof of its validity."
When I use GPT 5.2 Thinking Extended, it gave me the impression that it's consistent enough/has a low enough rate of errors (or enough error correcting ability) to autonomously do math/physics for many hours if it were allowed to [but I guess the Extended time cuts off around 30 minute mark and Pro maybe 1-2 hours]. It's good to see some confirmation of that impression here. I hope scientists/mathematicians at large will be able to play with tools which think at this time-scale soon and see how much capabilities these machines really have.
Yes and 5.3 and the latest codex cli client is incredibly good across compactions. Anyone know the methodology they're using to maintain state and manage context for a 12 hour run? It could be as simple as a single dense document and its own internal compaction algrorithm, I guess.
It's a bit unclear to me what happens if I do that after it thinks for 30 minutes and ends with no response. Does it start off where it left off? Does it start from scratch again? Like I don't know how the compaction of their prior thinking traces work
AI can be an amazing productivity multiplier for people who know what they're doing.
This result reminded me of the C compiler case that Anthropic posted recently. Sure, agents wrote the code for hours but there was a human there giving them directions, scoping the problem, finding the test suites needed for the agentic loops to actually work etc etc. In general making sure the output actually works and that it's a story worth sharing with others.
The "AI replaces humans in X" narrative is primarily a tool for driving attention and funding. It works great for creating impressions and building brand value but also does a disservice to the actual researchers, engineers and humans in general, who do the hard work of problem formulation, validation and at the end, solving the problem using another tool in their toolbox.
> The "AI replaces humans in X" narrative is primarily a tool for driving attention and funding.
It's also a legitimate concern. We happen to be in a place where humans are needed for that "last critical 10%," or the first critical 10% of problem formulation, and so humans are still crucial to the overall system, at least for most complex tasks.
But there's no logical reason that needs to be the case. Once it's not, humans will be replaced.
The reason there is a marketing opportunity is because, to your point, there is a legitimate concern. Marketing builds and amplifies the concern to create awareness.
When the systems turn into something trivial to manage with the new tooling, humans build more complex or add more layers on the existing systems.
I'm not sure you can call something an optimizing C compiler if it doesn't optimize or enforce C semantics (well, it compiles C but also a lot of things that aren't syntactically valid C). It seemed to generate a lot of code (wow!) that wasn't well-integrated and didn't do what it promised to, and the human didn't have the requisite expertise to understand that. I'm not a theoretical physicist but I will hold to my skepticism here, for similar reasons.
We will be producing them even less. I fear for the future graduates, hell even for school children, who are now uncontrollably using ChatGPT for their homework. Next level brainrot
Right. If it hadn't been Nicholas Carlini driving Claude, with his decades of experience, there wouldn't be a Claude c compiler. It still required his expertise and knowledge for it to get there.
It's interesting to me that whenever a new breakthrough in AI use comes up, there's always a flood of people who come in to handwave away why this isn't actually a win for LLMs. Like with the novel solutions GPT 5.2 has been able to find for erdos problems - many users here (even in this very thread!) think they know more about this than Fields medalist Terence Tao, who maintains this list showing that, yes, LLMs have driven these proofs: https://github.com/teorth/erdosproblems/wiki/AI-contribution...
The reality is: "GPT 5.2 after crunching 12 hours mathematical formulas supervised and prompted by 4 experts in the field" which would be nice and interesting per se.
But the title creates a much bigger expectation.
I wouldn't be surprised if you give an LLM some of the thousands of algos we use there and with proper promoting from experts in the field guiding it through the crunching it found a version that works better for bigger or smaller numbers.
It's easy to fall into a negative mindset when there are legions of pointy haired bosses and bandwagoning CEOs who (wrongly) point at breakthroughs like this as justification for AI mandates or layoffs.
Let’s have some compassion, a lot of people are freaking out about their careers now and defense mechanisms are kicking in. It’s hard for a lot of people to say “actually yeah this thing can do most of my work now, and barrier of entry dropped to the ground”.
For the sake of clarity: Woit's post is not about the same alleged instance of GPT producing new work in theoretical physics, but about an earlier one from November 2025. Different author, different area of theoretical physics.
It would be more accurate to say that humans using GPT-5.2 derived a new result in theoretical physics (or, if you're being generous, humans and GPT-5.2 together derived a new result). The title makes it sound like GPT-5.2 produced a complete or near-complete paper on its own, but what it actually did was take human-derived datapoints, conjecture a generalization, then prove that generalization. Having scanned the paper, this seems to be a significant enough contribution to warrant a legitimate author credit, but I still think the title on its own is an exaggeration.
Many innovations are built off cross pollination of domains and I think we are not too far off from having a loop where multiple agents grounded very well in specific domains can find intersections and optimizations by communicating with each other, especially if they are able to run for 12+ hours. The truth is that 99% of attempts at innovation will fail, but the 1% can yield something fantastic, the more attempts we can take, the faster progress will happen.
They also claimed ChatGPT solved novel erdös problems when that wasn’t the case. Will take with a grain of salt until more external validation happened. But very cool if true!
How was that not the case? As far as I understand it ChatGPT was instrumental to solving a problem. Even if it did not entirely solve it by itself, the combination with other tools such as Lean is still very impressive, no?
My understanding is there's been around 10 erdos problems solved by GPT by now. Most of them have been found to be either in literature or a very similar problem was solved in literature. But one or two solutions are quite novel.
Yeah that was also my take-away when I was following the developments on it. But then again I don't follow it very closely so _maybe_ some novel solutions are discovered. But given how LLMs work, I'm skeptical about that.
I would be less interested in scattering amplitude of all particle physics concepts as a test case because the scattering amplitudes because it is one of the concisest definition and its solution is straightforward (not easy of course). So once you have a good grasp of the QM and the scattering then it is a matter of applying your knowledge of math to solve the problem. Usually the real problem is to actually define your parameters from your model and define the tree level calculations. Then for LLM to solve these it is impressive but the researchers defined everything and came up with the workflow.
So I would read this (with more information available) with less emphasize on LLM discovering new result. The title is a little bit misleading but actually "derives" being the operative word here so it would be technically correct for people in the field.
I' m far from being an LLM enthusiast, but this is probably the right use case for this technology: conjectures which are hard to find, but then the proof can be checked with automated theorem provers. Isn't it what AlphaProof does by the way?
So wait,GPT found a formula that humans couldn't,then the humans proved it was right? That's either terrifying or the model just got lucky. Probably the latter.
I'd say "couldn't in 20 hours" might be more defensible. Depends on how many humans though. "couldn't in 20 GPT watt-hours" would give us like 2,000 humans or so.
I like the use of the word "derives". However, it gets outshined by "new result" in public eyes.
I expect lots of derivations (new discoveries whose pieces were already in place somewhere, but no one has put them together).
In this case, the human authors did the thinking and also used the LLM, but this could happen without the original human author too (some guy posts some partial on the internet, no one realizes is novel knowledge, gets reused by AI later). It would be tremendously nice if credit was kept in such possible scenarios.
Don't lend much credence to a preprint. I'm not insinuating fraud, but plenty of preprints turn out to be "Actually you have a math error here", or are retracted entirely.
Theoretical physics is throwing a lot of stuff at the wall and theory crafting to find anything that might stick a little. Generation might actually be good there, even generation that is "just" recombining existing ideas.
I trust physicists and mathematicians to mostly use tools because they provide benefit, rather than because they are in vogue. I assume they were approached by OpenAI for this, but glad they found a way to benefit from it. Physicists have a lot of experience teasing useful results out of probabilistic and half broken math machines.
If LLMs end up being solely tools for exploring some symbolic math, that's a real benefit. Wish it didn't involve destroying all progress on climate change, platforming truly evil people, destroying our economy, exploiting already disadvantaged artists, destroying OSS communities, enabling yet another order of magnitude increase in spam profitability, destroying the personal computer market, stealing all our data, sucking the oxygen out of investing into real industry, and bold faced lies to all people about how these systems work.
Also, last I checked, MATLAB wasn't a trillion dollar business.
Interestingly, the OpenAI wrangler is last in the list of Authors and acknowledgements. That somewhat implies the physicists don't think it deserves much credit. They could be biased against LLMs like me.
When Victor Ninov (fraudulently) analyzed his team's accelerator data using an existing software suite to find a novel SuperHeavy element, he got first billing on the authors list. Probably he contributed to the theory and some practical work, but he alone was literate in the GOOSY data tool. Author lists are often a political game as well as credit, but Victor got top billing above people like his bosses, who were famous names. The guy who actually came up with the idea of how to create the element, in an innovative recipe that a lot of people doubted, was credited 8th
The headline may make it seem like AI just discovered some new result in physics all on its own, but reading the post, humans started off trying to solve some problem, it got complex, GPT simplified it and found a solution with the simpler representation. It took 12 hours for GPT pro to do this. In my experience LLM’s can make new things when they are some linear combination of existing things but I haven’t been to get them to do something totally out of distribution yet from first principles.
This is the critical bit (paraphrasing):
Humans have worked out the amplitudes for integer n up to n = 6 by hand, obtaining very complicated expressions, which correspond to a “Feynman diagram expansion” whose complexity grows superexponentially in n. But no one has been able to greatly reduce the complexity of these expressions, providing much simpler forms. And from these base cases, no one was then able to spot a pattern and posit a formula valid for all n. GPT did that.
Basically, they used GPT to refactor a formula and then generalize it for all n. Then verified it themselves.
I think this was all already figured out in 1986 though: https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.56... see also https://en.wikipedia.org/wiki/MHV_amplitudes
I think this is a prime example of where it is easy to think something is solved when looking at things from a high level but making an erroneous conclusion due to lack of domain expertise. Classic "Reviewer 2" move. Though I'm not a domain expert and so if there was no novelty over Parke and Taylor I'm pretty sure this will get thrashed in review.
You're right. Parke & Taylor showed the simplest nonzero amplitudes have two minus helicities while one-minus amplitudes vanish (generically). This paper claims that vanishing theorem has a loophole - a new hidden sector exists and one-minus amplitudes are secretly there, but distributional
> simplest nonzero amplitudes have two minus helicities while one-minus amplitudes vanish
Sorry but I just have to point out how this field of maths read like Star Trek technobabble too me.
https://www.youtube.com/watch?v=cn4fW0EInqw
It bears repeating that modern LLMs are incredibly capable, and relentless, at solving problems that have a verification test suite. It seems like this problem did (at least for some finite subset of n)!
This result, itself does not generalize to open-ended problems, whether in business or in research in general. Discovering the specification to build is often the majority of the battle. LLMs aren't bad at this, per se, but they're nowhere near as reliably groundbreaking as they are on verifiable problems.
I'm not sure if GPTs ability goes beyond a formal math package's in this regard or its just its just way more convienient to ask ChatGPT rather than using these software.
You should probably email the authors if you think that's true. I highly doubt they didn't do a literature search first though...
You should be more skeptical of marketing releases like this. This is an advertisement.
They also reference Parke and Taylor. Several times...
Don't underestimate the willingness of physicists to skimp on literature review.
After last month’s Erdos problems handling by LLMs at this point everyone writing papers should be aware that literature checks are approximately free, even physicists.
Still pretty awesome though, if you ask me.
Big difference between “drives new result” and “reproduces something likely in its training dataset”.
I think even “non-intelligent” solver like Mathematica is cool - so hell yes, this is cool.
> but I haven’t been to get them to do something totally out of distribution yet from first principles
Can humans actually do that? Sometimes it appears as if we have made a completely new discovery. However, if you look more closely, you will find that many events and developments led up to this breakthrough, and that it is actually an improvement on something that already existed. We are always building on the shoulders of giants.
Relativity comes to mind.
You could nitpick a rebuttal, but no matter how many people you give credit, general relativity was a completely novel idea when it was proposed. I'd argue for special relatively as well.
I am not a scientific historian, or even a physicist, but IMO relativity has a weak case for being a completely novel discovery. Critique of absolute time and space of Newtonian physics was already well underway, and much of the methodology for exploring this relativity (by way of gyroscopes, inertial reference frames, and synchronized mechanical clocks) were already in parlance. Many of the phenomena that relativity would later explain under a consistent framework already had independent quasi-explanations hinting at the more universal theory. Poincare probably came the closest to unifying everything before Einstein:
> In 1902, Henri Poincaré published a collection of essays titled Science and Hypothesis, which included: detailed philosophical discussions on the relativity of space and time; the conventionality of distant simultaneity; the conjecture that a violation of the relativity principle can never be detected; the possible non-existence of the aether, together with some arguments supporting the aether; and many remarks on non-Euclidean vs. Euclidean geometry.
https://en.wikipedia.org/wiki/History_of_special_relativity
Now, if I had to pick a major idea that seemed to drop fully-formed from the mind of a genius with little precedent to have guided him, I might personally point to Galois theory (https://en.wikipedia.org/wiki/Galois_theory). (Ironically, though, I'm not as familiar with the mathematical history of that time and I may be totally wrong!)
Right on with special relativity—Lorentz also was developing the theory and was a bit sour that Einstein got so much credit. Einstein basically said “what if special relativity were true for all of physics”, not just electromagnetism, and out dropped e=mc^2. It was a bold step but not unexplainable.
As for general relativity, he spent several years working to learn differential geometry (which was well developed mathematics at the time, but looked like abstract nonsense to most physicists). I’m not sure how he was turned on to this theory being applicable to gravity, but my guess is that it was motivated by some symmetry ideas. (It always come down to symmetry.)
From that article:
> The quintic was almost proven to have no general solutions by radicals by Paolo Ruffini in 1799, whose key insight was to use permutation groups, not just a single permutation.
Thing is, I am usually the kind of person who defends the idea of a lone genius. But I also believe there is a continuous spectrum, no gaps, from the village idiot to Einstein and beyond.
Let me introduce, just for fun, not for the sake of any argument, another idea from math which I think it came really out of the blue, to the degree that it's still considered an open problem to write an exposition about it, since you cannot smoothly link it to anything else: forcing.
Even if I grant you that, surely we’ve moved the goal posts a bit if we’re saying the only thing we can think of that AI can’t do is the life’s work of a man who’s last name is literally synonymous with genius.
That's not exactly true. Lorentz contraction is a clear antecedent to special relativity.
Not really. Pretty sure I read recently that Newton appreciated that his theory was non-local and didn't like what Einstein later called "spooky action at a distance". The Lorentz transform was also known from 1887. Time dilation was understood from 1900. Poincaré figured out in 1905 that it was a mathematical group. Einstein put a bow on it all by figuring out that you could derive it from the principle of relativity and keeping the speed of light constant in all inertial reference frames.
I'm not sure about GR, but I know that it is built on the foundations of differential geometry, which Einstein definitely didn't invent (I think that's the source of his "I assure you whatever your difficulties in mathematics are, that mine are much greater" quote because he was struggling to understand Hilbert's math).
And really Cauchy, Hilbert, and those kinds of mathematicians I'd put above Einstein in building entirely new worlds of mathematics...
Agree with you everywhere. Although I prefer the quote:
"Since the mathematicians have invaded the theory of relativity, I do not understand it myself anymore."
:)
> Can humans actually do that?
From my reading yes, but I think I am likely reading the statement differently than you are.
> from first principles
Doing things from first principles is a known strategy, so is guess and check, brute force search, and so on.
For an llm to follow a first principles strategy I would expect it to take in a body of research, come up with some first principles or guess at them, then iteratively construct and tower of reasonings/findings/experiments.
Constructing a solid tower is where things are currently improving for existing models in my mind, but when I try openai or anthropic chat interface neither do a good job for long, not independently at least.
Humans also often have a hard time with this in general it is not a skill that everyone has and I think you can be a successful scientist without ever heavily developing first principles problem solving.
Go enough shoulders down, and someone had to have been the first giant.
Probably not homo sapiens.. other hominids older than us developed a lot of technology
Pythagoras is the turtle.
Depends on what you think is valid.
The process you’re describing is humans extending our collective distribution through a series of smaller steps. That’s what the “shoulders of giants” means. The result is we are able to do things further and further outside the initial distribution.
So it depends on if you’re comparing individual steps or just the starting/ending distributions.
"GPT did this". Authored by Guevara (Institute for Advanced Study), Lupsasca (Vanderbilt University), Skinner (University of Cambridge), and Strominger (Harvard University).
Probably not something that the average GI Joe would be able to prompt their way to...
I am skeptical until they show the chat log leading up to the conjecture and proof.
I'm a big LLM sceptic but that's… moving the goalposts a little too far. How could an average Joe even understand the conjecture enough to write the initial prompt? Or do you mean that experts would give him the prompt to copy-paste, and hope that the proverbial monkey can come up with a Henry V? At the very least posit someone like a grad student in particle physics as the human user.
I would interpret it as implying that the result was due to a lot more hand-holding that what is let on.
Was the initial conjecture based on leading info from the other authors or was it simply the authors presenting all information and asking for a conjecture?
Did the authors know that there was a simpler means of expressing the conjecture and lead GPT to its conclusion, or did it spontaneously do so on its own after seeing the hand-written expressions.
These aren't my personal views, but there is some handwaving about the process in such a way that reads as if this was all spontaneous involvement on GPTs end.
But regardless, a result is a result so I'm content with it.
That's kinda the whole point.
SpaceX can use an optimization algorithm to hoverslam a rocket booster, but the optimization algorithm didn't really figure it out on its own.
The optimization algorithm was used by human experts to solve the problem.
In this case there certainly were experts doing hand-holding. But simply being able to ask the right question isn't too much to ask, is it? If it had been merely a grad student or even a PhD student who had asked ChatGPT to figure out the result, and ChatGPT had done that, even interactively with the student, this would be huge news. But an average person? Expecting LLMs to transcend the GIGO principle is a bit too much.
hey, GPT, solve this tough conjecture I've read about on Quanta. make no mistakes
make no mistakes *please*
"Hey GPT thanks for the result. But is it actually true?"
Lupsasca is at OpenAI:
https://lupsasca.com/
Certainly the result looks very much desired by an OpenAI researcher.
The paper has all those prominent institutions who acknowledge the contribution so realistically, why would you be skeptical ?
they probably also acknowledge pytorch, numpy, R ... but we don't attribute those tools as the agent who did the work.
I know we've been primed by sci-fi movies and comic books, but like pytorch, gpt-5.2 is just a piece of software running on a computer instrumented by humans.
I don't see the authors of those libraries getting a credit on the paper, do you ?
>I know we've been primed by sci-fi movies and comic books, but like pytorch, gpt-5.2 is just a piece of software running on a computer instrumented by humans.
Sure
And we are just a system running on carbon-based biology in our physics computer run by whomever. What makes us special, to say that we are different than GPT-5.2?
> And we are just a system running on carbon-based biology in our physics computer run by whomever. What makes us special, to say that we are different than GPT-5.2?
Do you really want to be treated like an old PC (dismembered, stripped for parts, and discarded) when your boss is done with you (i.e. not treated specially compared to a computer system)?
But I think if you want a fuller answer, you've got a lot of reading to do. It's not like you're the first person in the world to ask that question.
It's always a value decision. You can say shiny rocks are more important than people and worth murdering over.
Not an uncommon belief.
Here you are saying you personally value a computer program more than people
It exposes a value that you personally hold and that's it
That is separate from the material reality that all this AI stuff is ultimately just computer software... It's an epistemological tautology in the same way that say, a plane, car and refrigerator are all just machines - they can break, need maintenance, take expertise, can be dangerous...
LLMs haven't broken the categorical constraints - you've just been primed to think such a thing is supposed to be different through movies and entertainment.
I hate to tell you but most movie AIs are just allegories for institutional power. They're narrative devices about how callous and indifferent power structures are to our underlying shared humanity
Their point is, would you be able to prompt your way to this result? No. Already trained physicists working at world-leading institutions could. So what progress have we really made here?
It's a stupid point then. Are you able to work with a world leading physicist to any significant degree? No
It's like saying: calculator drives new result in theoretical physics
(In the hands of leading experts.)
No it's not like saying that at all, which is why Open AI have a credit on the paper.
Open AI have a credit on the paper because it is marketing.
Hmm feels a bit trivializing, we don't know exactly how difficult it was to come up with the generic set of equations mentioned from the human starting point.
I can claim some knowledge of physics from my degree, typically the easy part is coming up with complex dirty equations that work under special conditions, the hard part is the simplification into something elegant, 'natural' and general.
Also "LLM’s can make new things when they are some linear combination of existing things"
Doesn't really mean much, what is a linear combination of things you first have to define precisely what a thing is?
When chess engines were first developed, they were strictly worse than the best humans. After many years of development, they became helpful to even the best humans even though they were still beatable (1985–1997). Eventually they caught up and surpassed humans but the combination of human and computer was better than either alone (~1997–2007). Since then, humans have been more or less obsoleted in the game of chess.
Five years ago we were at Stage 1 with LLMs with regard to knowledge work. A few years later we hit Stage 2. We are currently at Stage 3 for an extremely high percentage of knowledge work. Stage 4 will come, and I would wager it's sooner rather than later.
With a chess engine, you could ask any practitioner in the 90's what it would take to achieve "Stage 4" and they could estimate it quite accurately as a function of FLOPs and memory bandwidth. It's worth keeping in mind just how little we understand about LLM capability scaling. Ask 10 different AI researchers when we will get to Stage 4 for something like programming and you'll get wild guesses or an honest "we don't know".
> With a chess engine, you could ask any practitioner in the 90's what it would take to achieve "Stage 4" and they could estimate it quite accurately as a function of FLOPs and memory bandwidth.
And the same practitioners said right after deep blue that go is NEVER gonna happen. Too large. The search space is just not computable. We'll never do it. And yeeeet...
Chess grandmasters are living proof that it’s possible to reach grandmaster level in chess on 20W of compute. We’ve got orders of magnitude of optimizations to discover in LLMs and/or future architectures, both software and hardware and with the amount of progress we’ve got basically every month those ten people will answer ‘we don’t know, but it won’t be too long’. Of course they may be wrong, but the trend line is clear; Moore’s law faced similar issues and they were successively overcome for half a century.
IOW respect the trend line.
The evolution was also interesting: first the engines were amazing tactically but pretty bad strategically so humans could guide them. With new NN based engines they were amazing strategically but they sucked tactically (first versions of Leela Chess Zero). Today they closed the gap and are amazing at both strategy and tactics and there is nothing humans can contribute anymore - all that is left is to just watch and learn.
Serious questions, I often hear about this "let the LLM cook for hours" but how do you do that in practice and how does it manages its own context? How doesn't it get lost at all after so many tokens?
I’m guessing, would love someone who has first hand knowledge to comment. But my guess is it’s some combination of trying many different approaches in parallel (each in a fresh context), then picking the one that works, and splitting up the task into sequential steps, where the output of one step is condensed and is used as an input to the next step (with possibly human steering between steps)
From what I've seen is a process of compacting the session once it reaches some limit, which basically means summarizing all the previous work and feeding it as the initial prompt for the next session.
In my experience humans can make new things when they are some linear combination of existing things but I haven’t been able to get them to do something totally out of distribution yet from first principles[0].
[0]: https://slatestarcodex.com/2019/02/19/gpt-2-as-step-toward-g...
They want to be seen alongside the big news from their competitors, so it looks less like they are falling behind, let me know when they have a passing pre-train run again (apparently haven't since Ilya left)
Absolutely no way this is true right? Ilya left around the time 4o was released. I can't imagine they haven't had a single successful run since then.
When's the last time they talked about it?
I heard this from people who know more than me
Can't say, just seems implausible, but I am a nobody anyways ¯\_(ツ)_/¯
I don't want to be rude but like, maybe you should pre-register some statement like "LLMs will not be able to do X" in some concrete domain, because I suspect your goalposts are shifting without you noticing.
We're talking about significant contributions to theoretical physics. You can nitpick but honestly go back to your expectations 4 years ago and think — would I be pretty surprised and impressed if an AI could do this? The answer is obviously yes, I don't really care whether you have a selective memory of that time.
I don't know enought about theoretical physics: what makes it a significant contribution there?
It's a nontrivial calculation valid for a class of forces (e.g. QCD) and apparently a serious simplification to a specific calculation that hadn't been completed before. But for what it's worth, I spent a good part of my physics career working in nucleon structure and have not run across the term "single minus amplitudes" in my memory. That doesn't necessarily mean much as there's a very broad space work like this takes place in and some of it gets extremely arcane and technical.
One way I gauge the significance of a theory paper are the measured quantities and physical processes it would contribute to. I see none discussed here which should tell you how deep into math it is. I personally would not have stopped to read it on my arxiv catch-up
https://arxiv.org/list/hep-th/new
Maybe to characterize it better, physicists were not holding their breath waiting for this to get done.
Thank you!
Not every contribution has immediate impact.
That doesn't answer the question. That statement just admits "maybe" which isn't helpful or insightful to answering it.
I never said LLMs will not be able to do X. I gave my summary of the article and my anecdotal experiences with LLMs. I have no LLM ideology. We will see what tomorrow brings.
> We're talking about significant contributions to theoretical physics.
Whoever wrote the prompts and guided ChatGPT made significant contributions to theoretical physics. ChatGPT is just a tool they used to get there. I'm sure AI-bloviators and pelican bike-enjoyers are all quite impressed, but the humans should be getting the research credit for using their tools correctly. Let's not pretend the calculator doing its job as a calculator at the behest of the researcher is actually a researcher as well.
If this worked for 12 hours to derive the simplified formula along with its proof then it guided itself and made significant contributions by any useful definition of the word, hence Open AI having an author credit.
> hence Open AI having an author credit.
How much precedence is there for machines or tools getting an author credit in research? Genuine question, I don't actually know. Would we give an author credit to e.g. a chimpanzee if it happened to circle the right page of a text book while working with researchers, leading them to a eureka moment?
>How much precedence is there for machines or tools getting an author credit in research?
For a datum of one, the mathematician Doron Zeilberger give credit to his computer Shalosh B. Ekhad on select papers.
https://medium.com/@miodragpetkovic_24196/the-computer-a-mys...
https://sites.math.rutgers.edu/~zeilberg/akherim/EkhadCredit...
https://sites.math.rutgers.edu/~zeilberg/pj.html
Interesting (and an interesting name for the computer too), thanks!
Not exactly the same thing, but I know of at least two professors that would try to list their cats as co-authors:
https://en.wikipedia.org/wiki/F._D._C._Willard
https://en.wikipedia.org/wiki/Yuri_Knorozov
That is great, thank you!
I have seem stuff like "you can use my program if you will make me a co-author".
That usually comes up with some support usually.
it's called ethics and research integrity. not crediting GPT would be a form of misrepresentation
Would it? I think there's a difference between "the researchers used ChatGPT" and "one of the researchers literally is ChatGPT." The former is the truth, and the latter is the misrepresentation in my eyes.
I have no problem with the former and agree that authors/researchers must note when they use AI in their research.
now you are debating exactly how GPT should be credited. idk, I'm sure the field will make up some guidance
for this particular paper it seems the humans were stuck, and only AI thinking unblocked them
> now you are debating exactly how GPT should be credited. idk, I'm sure the field will make up some guidance
In your eyes maybe there's no difference. In my eyes, big difference. Tools are not people, let's not further the myth of AGI or the silly marketing trend of anthropomorphizing LLMs.
>How much precedence is there for machines or tools getting an author credit in research?
Well what do you think ? Do the authors (or a single symbolic one) of pytorch or numpy or insert <very useful software> typically get credits on papers that utilize them heavily? Well Clearly these prominent institutions thought GPT's contribution significant enough to warrant an Open AI credit.
>Would we give an author credit to e.g. a chimpanzee if it happened to circle the right page of a text book while working with researchers, leading them to a eureka moment?
Cool Story. Good thing that's not what happened so maybe we can do away with all these pointless non sequiturs yeah ? If you want to have a good faith argument, you're welcome to it, but if you're going to go on these nonsensical tangents, it's best we end this here.
> Well what do you think ? Do the authors (or a single symbolic one) of pytorch or numpy or insert <very useful software> typically get credits on papers that utilize them heavily ?
I don't know! That's why I asked.
> Well Clearly these prominent institutions thought GPT's contribution significant enough to warrant an Open AI credit.
Contribution is a fitting word, I think, and well chosen. I'm sure OpenAI's contribution was quite large, quite green and quite full of Benjamins.
> Cool Story. Good thing that's not what happened so maybe we can do away with all these pointless non sequiturs yeah ? If you want to have a good faith argument, you're welcome to it, but if you're going to go on these nonsensical tangents, it's best we end this here.
It was a genuine question. What's the difference between a chimpanzee and a computer? Neither are humans and neither should be credited as authors on a research paper, unless the institution receives a fat stack of cash I guess. But alas Jane Goodall wasn't exactly flush with money and sycophants in the way OpenAI currently is.
>I don't know! That's why I asked.
If you don't read enough papers to immediately realize it is an extremely rare occurrence then what are you even doing? Why are you making comments like you have the slightest clue of what you're talking about? including insinuating the credit was what...the result of bribery?
You clearly have no idea what you're talking about. You've decided to accuse prominent researchers of essentially academic fraud with no proof because you got butthurt about a credit. You think your opinion on what should and shouldn't get credited matters ? Okay
I've wasted enough time talking to you. Good Day.
Do I need to be credentialed to ask questions or point out the troubling trend of AI grift maxxers like yourself helping Sam Altman and his cronies further the myth of AGI by pretending a machine is a researcher deserving of a research credit? This is marketing, pure and simple. Close the simonw substack for a second and take an objective view of the situation.
If a helicopter drops someone off on the top of Mount Everest, it's reasonable to say that the helicopter did the work and is not just a tool they used to hike up the mountain.
Who piloted the helicopter in this scenario, a human or chatgpt? You'd say the pilot dropped them off in a helicopter. The helicopter didn't fly itself there.
“They have chosen cunning instead of belief. Their prison is only in their minds, yet they are in that prison; and so afraid of being taken in that they cannot be taken out.”
― C.S. Lewis, The Last Battle
"For me, it is far better to grasp the universe as it really is than to persist in delusion, however satisfying and reassuring."
— Carl Sagan
Shall we snipe each other with quotes neither of us came up with all afternoon? Bonus points for whomever's LLM researches the most pithy one!
I read the narnia series many times as a kid and this one stuck with me, I didn't prompt for it.
I have no real way to demonstrate that I'm telling the truth, but I am ¯\_(ツ)_/¯
Is every new thing not just combinations of existing things? What does out of distribution even mean? What advancement has ever made that there wasn’t a lead up of prior work to it? Is there some fundamental thing that prevents AI from recombining ideas and testing theories?
> Is every new thing not just combinations of existing things?
If all ideas are recombinations of old ideas, where did the first ideas come from? And wouldn't the complexity of ideas be thus limited to the combined complexity of the "seed" ideas?
I think it's more fair to say that recombining ideas is an efficient way to quickly explore a very complex, hyperdimensional space. In some cases that's enough to land on new, useful ideas, but not always. A) the new, useful idea might be _near_ the area you land on, but not exactly at. B) there are whole classes of new, useful ideas that cannot be reached by any combination of existing "idea vectors".
Therefore there is still the necessity to explore the space manually, even if you're using these idea vectors to give you starting points to explore from.
All this to say: Every new thing is a combination of existing things + sweat and tears.
The question everyone has is, are current LLMs capable of the latter component. Historically the answer is _no_, because they had no real capacity to iterate. Without iteration you cannot explore. But now that they can reliably iterate, and to some extent plan their iterations, we are starting to see their first meaningful, fledgling attempts at the "sweat and tears" part of building new ideas.
"Sweat and tears" -> exploration and the training signal for reinforcement learning.
For example, ever since the first GPT 4 I’ve tried to get LLM’s to build me a specific type of heart simulation that to my knowledge does not exist anywhere on the public internet (otherwise I wouldn’t try to build it myself) and even up to GPT 5.3 it still cannot do it.
But I’ve successfully made it build me a great Poker training app, a specific form that also didn’t exist, but the ingredients are well represented on the internet.
And I’m not trying to imply AI is inherently incapable, it’s just an empirical (and anecdotal) observation for me. Maybe tomorrow it’ll figure it out. I have no dogmatic ideology on the matter.
Just wait until LLMs are fast and cheap enough to be run in a breadth first search kind of way, with "fuzzy" pruning.
"An internal scaffolded version of GPT‑5.2 then spent roughly 12 hours reasoning through the problem, coming up with the same formula and producing a formal proof of its validity."
When I use GPT 5.2 Thinking Extended, it gave me the impression that it's consistent enough/has a low enough rate of errors (or enough error correcting ability) to autonomously do math/physics for many hours if it were allowed to [but I guess the Extended time cuts off around 30 minute mark and Pro maybe 1-2 hours]. It's good to see some confirmation of that impression here. I hope scientists/mathematicians at large will be able to play with tools which think at this time-scale soon and see how much capabilities these machines really have.
Yes and 5.3 and the latest codex cli client is incredibly good across compactions. Anyone know the methodology they're using to maintain state and manage context for a 12 hour run? It could be as simple as a single dense document and its own internal compaction algrorithm, I guess.
https://developers.openai.com/cookbook/articles/codex_exec_p... might be what you're looking for
after those 30 min you can manually ask it again to continue working on the problem
It's a bit unclear to me what happens if I do that after it thinks for 30 minutes and ends with no response. Does it start off where it left off? Does it start from scratch again? Like I don't know how the compaction of their prior thinking traces work
AI can be an amazing productivity multiplier for people who know what they're doing.
This result reminded me of the C compiler case that Anthropic posted recently. Sure, agents wrote the code for hours but there was a human there giving them directions, scoping the problem, finding the test suites needed for the agentic loops to actually work etc etc. In general making sure the output actually works and that it's a story worth sharing with others.
The "AI replaces humans in X" narrative is primarily a tool for driving attention and funding. It works great for creating impressions and building brand value but also does a disservice to the actual researchers, engineers and humans in general, who do the hard work of problem formulation, validation and at the end, solving the problem using another tool in their toolbox.
> The "AI replaces humans in X" narrative is primarily a tool for driving attention and funding.
It's also a legitimate concern. We happen to be in a place where humans are needed for that "last critical 10%," or the first critical 10% of problem formulation, and so humans are still crucial to the overall system, at least for most complex tasks.
But there's no logical reason that needs to be the case. Once it's not, humans will be replaced.
The reason there is a marketing opportunity is because, to your point, there is a legitimate concern. Marketing builds and amplifies the concern to create awareness.
When the systems turn into something trivial to manage with the new tooling, humans build more complex or add more layers on the existing systems.
Actually, the results were far worse and way less impressive than what the media said.
the c compiler results or the physics results this post is about?
I'm not sure you can call something an optimizing C compiler if it doesn't optimize or enforce C semantics (well, it compiles C but also a lot of things that aren't syntactically valid C). It seemed to generate a lot of code (wow!) that wasn't well-integrated and didn't do what it promised to, and the human didn't have the requisite expertise to understand that. I'm not a theoretical physicist but I will hold to my skepticism here, for similar reasons.
> for people who know what they're doing.
I worry we're not producing as many of those as we used to
We will be producing them even less. I fear for the future graduates, hell even for school children, who are now uncontrollably using ChatGPT for their homework. Next level brainrot
Right. If it hadn't been Nicholas Carlini driving Claude, with his decades of experience, there wouldn't be a Claude c compiler. It still required his expertise and knowledge for it to get there.
It's interesting to me that whenever a new breakthrough in AI use comes up, there's always a flood of people who come in to handwave away why this isn't actually a win for LLMs. Like with the novel solutions GPT 5.2 has been able to find for erdos problems - many users here (even in this very thread!) think they know more about this than Fields medalist Terence Tao, who maintains this list showing that, yes, LLMs have driven these proofs: https://github.com/teorth/erdosproblems/wiki/AI-contribution...
It's an obvious tension created by the title.
The reality is: "GPT 5.2 after crunching 12 hours mathematical formulas supervised and prompted by 4 experts in the field" which would be nice and interesting per se.
But the title creates a much bigger expectation.
I wouldn't be surprised if you give an LLM some of the thousands of algos we use there and with proper promoting from experts in the field guiding it through the crunching it found a version that works better for bigger or smaller numbers.
It's easy to fall into a negative mindset when there are legions of pointy haired bosses and bandwagoning CEOs who (wrongly) point at breakthroughs like this as justification for AI mandates or layoffs.
Let’s have some compassion, a lot of people are freaking out about their careers now and defense mechanisms are kicking in. It’s hard for a lot of people to say “actually yeah this thing can do most of my work now, and barrier of entry dropped to the ground”.
It is not only the the peanut gallery that is skeptical:
https://www.math.columbia.edu/~woit/wordpress/?p=15362
Let's wait a couple of days whether there has been a similar result in the literature.
For the sake of clarity: Woit's post is not about the same alleged instance of GPT producing new work in theoretical physics, but about an earlier one from November 2025. Different author, different area of theoretical physics.
It would be more accurate to say that humans using GPT-5.2 derived a new result in theoretical physics (or, if you're being generous, humans and GPT-5.2 together derived a new result). The title makes it sound like GPT-5.2 produced a complete or near-complete paper on its own, but what it actually did was take human-derived datapoints, conjecture a generalization, then prove that generalization. Having scanned the paper, this seems to be a significant enough contribution to warrant a legitimate author credit, but I still think the title on its own is an exaggeration.
Many innovations are built off cross pollination of domains and I think we are not too far off from having a loop where multiple agents grounded very well in specific domains can find intersections and optimizations by communicating with each other, especially if they are able to run for 12+ hours. The truth is that 99% of attempts at innovation will fail, but the 1% can yield something fantastic, the more attempts we can take, the faster progress will happen.
They also claimed ChatGPT solved novel erdös problems when that wasn’t the case. Will take with a grain of salt until more external validation happened. But very cool if true!
Well they (OpenAI) never made such a claim. And yes, LLMs have made unique solutions/contributions to a few erdos problems.
How was that not the case? As far as I understand it ChatGPT was instrumental to solving a problem. Even if it did not entirely solve it by itself, the combination with other tools such as Lean is still very impressive, no?
It didn't solve it, it simply found that it had been solved in a publication and that the list of open problems wasn't updated.
My understanding is there's been around 10 erdos problems solved by GPT by now. Most of them have been found to be either in literature or a very similar problem was solved in literature. But one or two solutions are quite novel.
https://github.com/teorth/erdosproblems/wiki/AI-contribution... may be useful
Some of these were initially hyped as novel solutions, and then were quietly downgraded after it was discovered the solutions weren’t actually novel.
Yeah that was also my take-away when I was following the developments on it. But then again I don't follow it very closely so _maybe_ some novel solutions are discovered. But given how LLMs work, I'm skeptical about that.
Wasnt that like some marketing bro? This is coming out the front door with serious physicists attached.
I would be less interested in scattering amplitude of all particle physics concepts as a test case because the scattering amplitudes because it is one of the concisest definition and its solution is straightforward (not easy of course). So once you have a good grasp of the QM and the scattering then it is a matter of applying your knowledge of math to solve the problem. Usually the real problem is to actually define your parameters from your model and define the tree level calculations. Then for LLM to solve these it is impressive but the researchers defined everything and came up with the workflow.
So I would read this (with more information available) with less emphasize on LLM discovering new result. The title is a little bit misleading but actually "derives" being the operative word here so it would be technically correct for people in the field.
The preprint: https://arxiv.org/abs/2602.12176
I' m far from being an LLM enthusiast, but this is probably the right use case for this technology: conjectures which are hard to find, but then the proof can be checked with automated theorem provers. Isn't it what AlphaProof does by the way?
Cynically, I wonder if this was released at this time to ward off any criticism from the failure of LLMs to solve the 1stproof problems.
All I saw was gravitons and thought we’re finally here the singularity has begun
So wait,GPT found a formula that humans couldn't,then the humans proved it was right? That's either terrifying or the model just got lucky. Probably the latter.
> found a formula that humans couldn't
Couldn't is an immensely high bar in this context, didn't seems more appropriate and renders this whole thing slightly less exciting.
I'd say "couldn't in 20 hours" might be more defensible. Depends on how many humans though. "couldn't in 20 GPT watt-hours" would give us like 2,000 humans or so.
5.2 is the best model on the market.
Well, anyone can derive a new result in anything. Question is most often if the result makes any sense
I'll read the article in a second, but let me guess ahead of time: Induction.
Okay read it: Yep Induction. It already had the answer.
Don't get me wrong, I love Induction... but we aren't having any revolutions in understanding with Induction.
I guess the important question is, is this enough news to sustain OpenAI long enough for their IPO?
Well it’ll be at least a whole month before some other company announces similar capability. The moat will hold!
I believe Gemini holds the moat now.
I like the use of the word "derives". However, it gets outshined by "new result" in public eyes.
I expect lots of derivations (new discoveries whose pieces were already in place somewhere, but no one has put them together).
In this case, the human authors did the thinking and also used the LLM, but this could happen without the original human author too (some guy posts some partial on the internet, no one realizes is novel knowledge, gets reused by AI later). It would be tremendously nice if credit was kept in such possible scenarios.
Interesting considering the Twitter froth recently about AI being incapable in principle of discovering anything.
Anything but recent.
Don't lend much credence to a preprint. I'm not insinuating fraud, but plenty of preprints turn out to be "Actually you have a math error here", or are retracted entirely.
Theoretical physics is throwing a lot of stuff at the wall and theory crafting to find anything that might stick a little. Generation might actually be good there, even generation that is "just" recombining existing ideas.
I trust physicists and mathematicians to mostly use tools because they provide benefit, rather than because they are in vogue. I assume they were approached by OpenAI for this, but glad they found a way to benefit from it. Physicists have a lot of experience teasing useful results out of probabilistic and half broken math machines.
If LLMs end up being solely tools for exploring some symbolic math, that's a real benefit. Wish it didn't involve destroying all progress on climate change, platforming truly evil people, destroying our economy, exploiting already disadvantaged artists, destroying OSS communities, enabling yet another order of magnitude increase in spam profitability, destroying the personal computer market, stealing all our data, sucking the oxygen out of investing into real industry, and bold faced lies to all people about how these systems work.
Also, last I checked, MATLAB wasn't a trillion dollar business.
Interestingly, the OpenAI wrangler is last in the list of Authors and acknowledgements. That somewhat implies the physicists don't think it deserves much credit. They could be biased against LLMs like me.
When Victor Ninov (fraudulently) analyzed his team's accelerator data using an existing software suite to find a novel SuperHeavy element, he got first billing on the authors list. Probably he contributed to the theory and some practical work, but he alone was literate in the GOOSY data tool. Author lists are often a political game as well as credit, but Victor got top billing above people like his bosses, who were famous names. The guy who actually came up with the idea of how to create the element, in an innovative recipe that a lot of people doubted, was credited 8th
https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.83...
End times approach..
Car manufacturers need to step up their hype game...
New Honda Civic discovered Pacific Ocean!
New F150 discovers Utah Salt Flats!
Sure it took humans engineering and operating our machines, but the car is the real contributor here!