Latent Space: The AI Engineer Podcast - 🔬Doing Vibe Physics — Alex Lupsasca, OpenAI

Episode Date: May 5, 2026

Some people are going crazy over GPT 5.5. Some people. This is the story of the Jagged Frontier. People who use AI to write emails or even code implementation work find the lift moderate whereas people pushing the limits of the model are figuring out that the limits just moved outwards.Alex Lupsaska has been tracking this limit for a year and a half now. “When GPT5 came out, it was able to reproduce one of my best papers (that took a very long time to come up with) in 30 minutes.”But Alex also notes that this shift was mostly invisible.I remember when GPT-5 came out… on Twitter, the reception was lukewarm. A lot of people were like, well, we expected a lot more, and it’s not better at writing email. And I remember thinking, well, okay, GPT-3 could write email. How much better can it get at writing email? That’s not the point. But at the science frontier, the capabilities were really taking off.We walk through his paper and more with him in today’s Science pod! Watch here.The “Oscar for physics”Alex made an early splash in his career with breakthroughs in our understanding of black holes. He’s also known for Black Hole Explorer and an iPhone app that makes visualizing black holes fun and interactive to regular audiences. Alex won the 2024 New Horizons in Fundamental Physics Breakthrough Prize. Known as the “Oscar for physics” this is arguably the most prestigious prize an early stage theoretical physicist can win.Alex first saw promise for AI in theoretical physics after he asked o3 for help on his research. In the podcast, Alex recalls asking GPT for help with a calculation that would have taken days, and getting a result in eleven minutes. He immediately recognized how impactful AI would be for his work even as though his physicist colleagues and the larger community gave it a lukewarm or skeptical reception.The Move 37 Moment for AI x PhysicsGPT-5 had just been released, and Alex tried asking it to solve a problem in a just published paper. GPT-5 said no answer. But Mark Chen, CRO of OpenAI, pushed a bit harder, and had Alex prime the model with a textbook warmup problem, which it easily solved. After using this “priming” trick, GPT-5 was able to reproduce his full result in eleven minutes (yes, the paper was released after the model’s training cutoff).“This changes everything.” Alex notes that we seem to be on the edge of a massive change in theoretical physics reasoning. A year prior LLMs were just starting do correct math. Now ChatGPT could reproduce his hardest paper in the time it takes to get a coffee.Alex was on sabbatical at Vanderbilt, and he joined OpenAI to start pushing the boundary of AI’s ability to accelerate physics.“AI solved the problem before the plane landed”Alex began to put GPT through it’s paces, reaching out to colleagues for problems they were stuck on. His old PhD advisor (Prof. Andrew Storminger at Harvard) had an insidght about certain physical quantities known as “single-minus gluon tree amplitudes”. In certain cases, these amplitudes may be non-zero when previously shown to always vanish. The team pushed this intuition forward, and came up with a formula for these quantities that appeared nonzero, but which was otherwise completely intractable. Spending over a year on this problem, no real progress was made.Prof. Storminger planned to visit OpenAI to work on the problem the week after the initial conversation started. In that one week ChatGPT fully solved the problem, as Alex recalled, before Prof. Storminger’s plane even landed.What was interesting is not only that ChatGPT solved this problem, but how it solved it. The model quickly realized found a limiting case (known as the “half-collinear regime”), that in hindsight has a nice intuitive explanation. Taking this limit, the gnarly results collapsed down to a simple and intuitive formula!The last step was to prove this intuitive formula. The team started with a fresh session, gave a prompt with the context of what they previously learned, and let the model loose. Not only was ChatGPT able to reproduce the previous result, it was able to prove it using a technique unknown to the authors!The Vibe Physics momentWith a concrete success in the bag, the team asked if they could generate new physics from scratch using ChatGPT. They took on what they felt to be a harder problem, looking at the graviton, a proposed particle that should appear when one combines gravity and quantum mechanics. They wrote up a simple prompt asking ChatGPT to perform the same research as the gluon paper but instead for gravitons. And then hit go!What came next was truly “vibe physics”, with ChatGPT pushing out 110 pages of novel physics, new calculations, and novel techniques. This was over the course of a day, with most interactions the familiar following the now familiar pattern for anyone who uses a coding agent:GPT: Here's your . Would you like me to do ? Alex: Yes, please do! GPT: And for those who look deeply, this really was not just a direct 1-1 mapping between gluons and gravitons. ChatGPT imported new techniques that were necessary due to the nature of gravitons, and used them flawlessly.They spent the next three weeks verifying all the results. And voila! A new paper featuring novel results in quantum gravity, generated in less than three days total. Truly a “Feel the AGI moment”.For those interested, there’s a blog post with the full transcript from initial prompt to final paper. Even if you know no physics, it’s crazy seeing pages of correct calculations fall out of simple prompts such as “Yes calculate outside of SD first. This is the first step.”Out-of-domain = new knowledgeThe thing that is qualitatively different between Vibe Physics and Vibe Coding is that Vibe Physics means actually extending the frontier of human knowledge. Looking at the Gluon and Graviton results, they seem in retrospect, like many results in physics and math, like natural extensions of what we already know. This is in fact part of what makes them beautiful. But this was a problem that stumped experts in the domain for a year. Although it does still have a bit of a recombinant flavor, this thing has never been done before.It may be that there are still large classes of problems that AI won’t do well on, and approaches that an AI might not think to take. This is the “taste” that everyone has been talking about. Alex told us that these capabilities, however, allow him to explore many possible avenues in order to map out much more ambitious problems to tackle. With AI able to output results basically as fast as we can conceive and validate them, the scope of what one theorist can hope to achieve has just gotten a lot, lot bigger. This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe

Okay, so I think we're at this special time now where, at least in some directions, AI has become superhuman, at least on certain tasks. And that's what led to these recent papers that resolve a problem that was puzzling physicists, experts in the field for over a year, and they weren't able to resolve it. And AI was able to do it very quickly. So I think that's a certain milestone that we've passed. by that you guys are bringing attention to this because I think maybe for the average person

on the street who doesn't care about theoretical physics, this is not very noticeable. But I think it's a very profound change, and we've really passed some kind of a threshold. Welcome to the A for Science podcast, part of Lean Space Network. I'm Brandon. I develop RNA therapeutics using AI at At Atomic AI.

I'm joined by my co-host, R.J. Honakey, a CTO and founder of Mirroromics. Yeah, it's a pleasure to introduce Alex Ljoshaska, professor at Vanderbilt University and fellow at OpenAI. For a young researcher, he has quite a storied background. Amongst other things, he's the winner of the 2024 New Verizon's Breakthrough Prize. It's the call it the Oscars for Science. I asked Chattebti. Is this the most prestigious award?

Someone of his career could win, and it recommended a second one called the IUPAP award, which turns out to you had also won. Anyway, right now he's having fun at OpenAI doing some really cool research of pushing the foundation of theoretical physics using GPT models. A pleasure to be here. The one message I wanted to convey is that I think we're on this trajectory, which I personally find very surprising and yeah, kind of surreal, but also amazing, where I would say a little over a year ago, AI was very useful for email, but not the kind of work that I do that I consider important

theoretical physics calculations. I thought, oh, that's special, much harder than email, and AI is not going to be able to do that. And then there were a series of developments that came in rapid succession that completely changed my mind. And I can walk you through some of these examples

specifically. In particular, Chad GPT-O-3 was the first really strong reasoning model that could do actual math that was used one for my research and could save me a lot of time. That's when I started to really, attention and use it a lot more.

And I thought, wow, this is a great tool. I get ahead of this and learn how to integrate into my workflow. Then when GPT5 came out, it was able to reproduce one of my best papers that took be a very long time to come up with in like 30 minutes. And that's when I really became AI pill. I thought, oh my God, this changes everything. It's the most important discovery in my lifetime. It's going to affect everything about how we do research.

And frankly, a lot of my colleagues. I would go around telling them this crying, you think happy attention. Yeah, I was getting lots of different reactions, but I think people weren't quite getting it. But I talked to Open AI.

They were also really excited and I thought, I don't know that much about AI, but I have to get in on this. To understand that this is happening and not be a part of it is a huge mistake, so I have to go to Open AI. So I was on sabbatical.

It was very easy to come here and join the company. and then it just kept ramping up even beyond that. And to the point where now I think most of my senior colleagues in physics are aware of where things are ahead of and they're all getting on board. So, yeah, I think that's an awesome story. Sorry, I was just to say, like, I find it really funny that you, that story because it almost reminds me of a lot of different people who had the same realization with Codex starting sometime last. fall especially, it just took off and a bunch of people are like, even like Andre Carfothy went from, oh man, this is, you know, 21st of my work, it's kind of a nice,

you know, assistant to, oh crap, what just happened? Well, yeah, in August, actually I remember when GPT5 came out, at that point I was really following AI pretty closely and I think on Twitter the reception was lukewarm. A lot of people like, well, we expected a lot more and it's not better to write an email. And I remember thinking, well, okay, GPT3. could write email. How much better can it get writing email? That's not the point. But the science frontier, the capabilities were really taking off. Yeah, there was a lot of attention, I think, paid even to 03. Yeah, but then, but presumably was a huge jump. And I think 5.4 is also a huge jump.

I don't know how noticeable it is on the outside, although I did hear some, I saw some chatter online people are running these independent benchmarks, which do show this. So I think it's, people are realizing. And also anyway, in practice, researchers are now all over AI using it. And yeah, I'm getting inbound all the time because I'm the resident scientist doing physics at Open AI. And so everybody is sending me papers, chats. Like, oh my God, this happened. I got one just this week. Somebody said, Codex just wrote up a simulation of the S.YK models. This is like a very technical thing in quantum mechanics and gravity. And like, yeah, a lot of research groups have been trying to run this simulation.

and it couldn't do it, and Codex did it in 10 minutes. Just because setting it up was so hard. Well, I think partly it's because of the Venn diagram where you look at the people who have the physics knowledge and the people who have the top coding skills and maybe the overlap is not that large, although I think it's been growing. But I think in this example,

there are a lot of really good people in physics with coding skills who have been trying to simulate these things. So in codex, it's just really good now. Okay. Yeah. Nice. Okay, so I think we're at this special time now where at least in some directions, AI has become superhuman, at least on certain tasks.

And that's what led to these recent papers, which maybe we should talk about, that resolve a problem that was puzzling physicists for experts in the field for over a year, and they weren't able to resolve it, and AI was able to do it very quickly. So I think that's a certain milestone that we've passed. And I'm glad that you guys are bringing attention to this because I think maybe for the average person on the street who doesn't care about theoretical physics, this is not very noticeable.

But I think it's a very profound change, and we've really passed some kind of a threshold. Specifically focus on the glue-on paper and the physics part, and we can get to the AI part later. Okay, so in physics, there are two basic principles. principles of nature that we think every law should respect or every theory should respect. On the one hand, there's the principle of relativity, which at some very high level declares

there's an absolute law that cannot be broken, which is that you cannot transmit information faster than the speed of light. But then there's another principle, which is the uncertainty principle that underlies quantum mechanics, which says that everything is a little fuzzy, you know, or position of a lot there's a little fuzziness to that. And so you can see immediately at this level of description already there's a tension between these two

principles because one is an absolute law declaring you cannot confess and speak light and the other one is saying it's a little bit fuzzy. And this is just to give a sense of how when you try to write down these principles in mathematics the equations don't really play it nicely with each other. And so it's been a real struggle

to come up with physical theories that can reconcile simultaneously both principles to describe the physical world around this. And I would say that the great achievement of 20th century physics, which is really one of the greatest triumphs in human

thought as far as I'm concerned, is the elaboration of this framework called quantum field theory, which is a general framework that can describe the physical forces of nature in a way that accommodates both of these principles.

And in quantum field theory, which is our best theory to date. Obviously, it gets a little bit technical, but again, try to keep it pretty high level. What you're trying to compute or describe are the probabilities for certain events to occur. Because you're in this quantum mechanical setting, you can't say with certainty what's going to happen

when you have a certain experiment, but you want to predict probability distributions. And in quantum mechanics, probably distributions are obtained by squaring some complex quantity. and by complex, I don't mean, complicated, I mean they're not real numbers. They're real plus symmetry numbers, which we call quantum amplitudes. So the goal of a theory is to predict quantum amplitudes,

which are these complex objects that square the quantum probabilities, and that's the most you can say about the outcome of an experiment. And these quantum amplitudes, in particular, there's a variety of them called scattering amplitudes, which describe the following scenario. Suppose you have a bunch of particles that you throw at one another. This is what happens in particle colliders like the LHC at CERN in Geneva. You take a bunch of particles, you smash them together.

Stuff happens. They interact via the physical laws of nature. Various processes occur, and then other particles come out as a result at the end of the interaction. And so scattering amplitude is the object that describes the, the probability for a particular type of interaction. We have some particles coming in with some energies and momentum and some other particles coming out with other energies and momentum. And so these scattering amplitudes,

they're functions of all the data describing the particles coming in and the particles coming out. So in general, you can have arbitrarily many particles involved in an interaction, and this is one of the hallmarks of quantum field theory that particles can be destroyed, So you don't have the same number of particles at the end necessarily as you had in the beginning. Particles can be created. Lots of things can happen.

And in general, you want to describe all the possibilities. And so you want to have an amplitude for an arbitrary number N of particles. So that's called an endpoint amplitude because there's N particles coming in and out. And it turns out in quantum field theory that if you have a particular force and you're able to compute the endpoint amplitudes, these functions of the end parameters of the functions that squirt the probabilities, then you know everything about the theory, more or less. There's always an asterisk, but it's basically the entire content of the theory. So if you have a theory that tells you any number of particles come in and go out, then

I can say, I can declare anything about that system. Exactly. Then you know everything. And importantly, these amplitudes, they're not just number. numbers, their functions, because the probabilities that they compute depend on how much energy do the particles have, what are their momentum? And also, a particle has something called a lot of particles like the photon, which is the particle of light, has a polarization. So when you look at the surface of a lake and you have polarized sunglasses and you turn your head, you can see more or less sunlight reflected off of the lake. and that's because a photon, which you can think of as a little particle of light, as it propagates,

it carries a little arrow perpendicular to the direction of propagation, which is called the polarization. And this polarization has a direction, and sunglasses can selectively let in light with one polarization, and not the other. And this polarization actually, as light travels, it can rotate, it can wind, it can do its own thing. and in general, if it winds in a right-handed way, so as the particle travels, if the polarization winds to the right, we call that a positive helicity

or a right-handed polarization, and if it winds in the other direction, we call that a left-handed holicity or negative helicity. So in general, these amplitudes, which are the fundamental object in quantum field theory, that you want to contain all the information there is to know about physical forces, these amplitudes depend on not just the energies and momentum but also the polarizations.

Now I've told you about how there's two basic principles of nature, relativity and quantum mechanics. They come together in this framework quantum field theory, and I keep talking about forces. So there's four fundamental forces of nature. There's electromagnetism, which is responsible for basically the properties of atomic elements and the periodic table and therefore chemistry and biology and everything that you see touch, feel pretty much is all due to electromagnetism, textures, colors. And this force is mediated by the photon, which is the particle of light.

That one is the most familiar to us. Then there's gravity, which is another force that we feel very much, because it keeps us to the ground. And then there's two nuclear forces, the weak and the strong nuclear force, which we don't really notice directly in our daily lives, but the weak nuclear force is responsible for radioactive decay and other such process. and the strong force, which is the strongest of them all, is what binds the nucleus together. So you're learning high school that light charges repel, but if so, then why do protons

stick together inside the nucleus of the atom? They should repeal one another. And indeed, that's the case, but if you bring in really close, then the strong force kicks in and overwhelms the relatively weaker electromagnetic force. So the strong force is mediated by the exchange of the particles of the strong force, which are called gluons, because they're what glues together the nucleus of the atom. So gluons are the particles of the strong force,

and gravity is mediated by gravitons. I think, like, the glue on paper, I think, was sort of the, maybe the starting point for this, maybe not, but, like, the glue on paper had, like, a really, like, specific result, right? Yeah, absolutely. So,

maybe let me just flash the paper itself. So we put this on the archive, a little over a month ago now. And here's the paper. Let me explain in a few sentences now that I've given a lot of background what the title means. Yeah.

So the title says single minus gluon tree amplitudes are non-zero. This might sound forbidding, but I think we can unpack this for the audience. So gluons are the particles that carry the strong force. And gluon amplitudes

are functions that describe the quantum probabilities for gluons to interact via the strong force. Now, the word tree here is a little bit of a technicality. It means we're only considering processes where no gluons are created or destroyed. If gluons are created or destroyed, then you get loops, which we can explain later. But this is just a technicality. So we're considering special interactions where the same luans that come in also come out. So for anyone who's ever fit a polynomial, you can think of trees being like a linear term,

and then loops can be higher order terms. Exactly. Way more complicated than that. But conceptually, it's like kind of the lowest order in a series. And so single minus, now I have to explain that. So remember I told you earlier how particles have polarization. So when you try to study glue on amplitudes, this is like a whole industry of physics. It's a very complicated field.

People have written thousands of papers over the decades. So you always want to try to understand the simplest examples first. That's why you start with a triumplet, you store the leading effects, and then you worry about the loop corrections. So you might think that the simplest example to start with is one in which all the particles have the same holisticity. So say they're all right-handed, or that is to say,

they're all plus-holicity particles. It's been known for a long time that actually in that case the amplitude is just zero which means the interaction is forbidden and cannot happen. That's one way time. It's just a symmetry just explicitly

forbids this. You don't even have to calculate anything you just know. Yeah just dimensional analysis. Yes. It's a very general argument. Yeah, you don't need to do very much work. And so yeah, it's true that it's the simplest example but it's so simple that nothing happens. So, okay,

be it's true. You might ask, what about the next level up? What about it? Oh, no, I want to understand this. You have, like, a bunch of gluons. They're coming into an interaction. Yeah. They're all in the same helicity. Yeah. And then you're just saying, that just can't happen. Yeah. Okay. Like, because, like, I can't, I take my glen gun and shoot, and he takes his glen gun and shoot, and they go there, and then that just can't happen. We'll just go right through each other. Oh, so they just won't interact. They won't interact. Oh, okay. Yeah, yeah, yeah, yeah.

That's a good clara for Kijit. Yeah. And now you might ask, what if one of them has the opposite holisticity, but all the others have plus solicity, but one of them has a minus helicity. So that's what we would call a single minus amplitude. And if you look at the lecture notes and textbooks that have been written on this, the same argument that rules out the all plus amplitudes also appears to rule out

the single minus amplitude. They're too simple. They can't really interact. Nothing to see here. Move on. So then you might ask, okay, well, what about the next thing where there's two particles

that are minus, helicity and all the others. So if there's n of them, there's n minus two others that have positive felicity. So these would be double minus amplitudes. And people in the 80s studied and computed these amplitudes,

they're not zero. And in particular, there were two physicists, Park and Taylor, who found this beautiful result. They did a lot of really hard work and computed these amplitudes, very technical, difficult calculation. But at the end, you get all these terms and you have to sum them all up and almost all of them cancel. And at the end, you're left with this very simple formula that fits in half a line, which is now known as the Park Taylor formula. for these amplitudes. And these amplitudes are now called MHV amplitudes, which stands for maximally helicity violating.

Because they have the largest, or so we thought, possible asymmetry between the plus and the minus holistic particle, the most asymmetry. Now let's get to this paper, which came out last month. So this is a paper written with Alfredo Guevara, who's a post. postdoc at the Institute for Advanced Study, David Skinner, a professor at Cambridge University,

Andrew Strominger, a professor at Harvard, used to be my advisor, and also Kevin Wheel, who studied as a particle physicist in a previous life. So how did this happen? Well, maybe we'll get into how I ended up at Open AI a little bit later, but I ended up at Open AI, started to improve the model's abilities to do physics. The models got really, really good at physics, and I thought, okay, it's so good now we should try to solve some actual research problems at the frontier. And I called up Andy, who used to be my advisor, and I said, hey, Andy, do you want to come here to the SF, visit Open AI? And we can try to solve one of your problems in physics. And I thought, you know, it's probably not going to work, but if it doesn't work, at least we'll figure out why it doesn't work.

And, you know, I can do this with a different physicist every month, and eventually something will work. And in the meantime, we'll learn how to improve the models. It's all fun and useful. And so Andy was the first one that I invited to do this. And he said, well, I have this perfect problem that I've been thinking about with Alfredo and David for the past year. I'll explain out the problem. But the amazing thing is that we decided to start working on it using AI a little bit before Andy was scheduled to come, like the week before. And in fact, using Chad GPT, we solved the problem before he even got off.

the plane, which was a huge surprise to him. To him, to me, to be honest, I had not expected that. And it was a really cool story. So Andy, David, and Alfredo understood a year ago that this statement that the single minus amplitudes, the statement that there is zero is not exactly correct. because the usual argument in the electron nodes and textbooks has a loophole. And the loophole is that it assumes that the particles are coming from generic directions. But in a certain regime where the particles are exactly aligned with one another,

we say they're collinear, then the usual argument has a loophole, and it's possible for the amplitudes to not be zero. But then if they're not zero, what are they? So suddenly these really simple amplitudes previously thought to be zero, if they're not zero, we should compute them, and they should be something really nice and simple and special. Now, I'm burying a lot of, I'm sweeping a lot of details under the rug here. This has to work in some different signature space time. It connects lots of other things they've been worrying about. We're not going to worry about this.

I mean, I was actually hoping at the end maybe we could talk about what it means to be two dimensions in space and two dimensions in time. But, yeah, I mean, I think part of this is doable. The loophole is one about, you know, the alignment of the particles, but it's also a loophole about the space time of physics that the universe we're living in, and this is not so. Yeah, this is really mind-bending stuff. So they understood that they're not zero, and they started to compute them.

And Alfredo is really, I think, the unsulling hero of this story because he did a lot of really hard work to compute these things by hand. And I'll just show you an example. So in the paper, there's a lot of formalism. So here is the beginning of the definition of the general answer. Yeah, it's very hard to unpack. But it starts here. Then you have to define these vertices, objects, V.

They're complicated. They involve sign and theta functions of spinners. And then you have this recursive formula. Okay, it's a whole mess. And concretely, if you try to unpack this definition, remember these amplitudes are a function of the number of particles involved. So there's a three-point amplitude where there's only three gluons in the interaction. And, you know, the answer is pretty simple.

This is some function that we've defined here, not that complicated. Then this is the four-point amplitude. But now there's four particles. And you can see that we go from one term to a sum of two terms here. But then once you get to five particles, you start to get a lot. more terms. There's eight of them being summed here. And by the time you get to six terms, it's in your face. For those people not watching this on YouTube and listening, this equation takes up a quarter of the page is 32 terms, each of which is a product of four terms, each of which

is itself encapsulating a rather complicated formula. Yeah. So this is super nasty, and that's as far as Alfredo got or anyone else. So Alfredo, is this just a expansion of some sort of how hard is it to do this expansion very hard okay yeah and there's a nice graphical way to understand this in terms of Feynman diagrams i hadn't planned to explain this but uh there's a visual this is kind of a visual subject so the math is very complicated and already back in the 40s richard feyman who's one of the pioneers of quantum peel theory came up with this very visual way to organize our understanding of the subjects,

you can doodle these little cartoons that represent possible interactions. And the rules of quantum mechanics actually say that in these amplitudes where you scatter a bunch of particles, you get to fix what comes in and what comes out because that's the question you're asking, what's the probability for a certain traction.

But then everything that happens in between, you don't get to choose that because the physical laws determine what happens. And actually in quantum mechanics, you're supposed to consider all the possibilities, all the ways in which the incoming particles can interact and transform into the outgoing particles, and you're supposed to average or sum over all the possibilities to get the final amplitude for the process as a sum over the amplitudes for each individual possibility. First, I'll be you could get there.

So just to be clear, there's incoming particles. Yeah. They interact. There's all these different, they each have their own amplitudes. And then it's sort of like, I select for this one, one possibility, and this one, one possibility. And then I get like one possible interaction. And then there's an infinite number of those for each.

And then I sum those infinite blasphos. And I get that. Yeah. So in principle, there are infinitely many pictures to sum over. But that's why we organize them by how complex they are. And it turns out that every time you get an interaction, every time there's a, the vertex where it lines meet. That point interaction

comes with a power of the coupling constant which controls the strength of the interaction and it turns out that every additional interaction makes the amplitude more suppressed. So it contributes less to the final answer. And so you want to first consider the diagrams with the fewest possible number of attractions

because they will give you most of the total final amplitude. And then if you're trying to get a more and more refined answer, you then consider the more and more complicated cartoons with more and more interactions. And in fact, this is one of the ways in which the diagrams can get complicated is that they can have loops. So for instance, here you have a particle that decays into two particles, creating this loop because then they meet it up again and disappear. So in this interaction, you have intermediate particles being created and destroyed. But whenever that happens, you get two extra vertices in your graph.

So these diagrams are suppressed because it's less likely to happen that you get these extra felicitous interactions. And so you don't need to worry about this as much. It's like a small correction. And of course, in principle, you can keep going, but you're never done excess in very special circumstances. The Hieroper Pallor is in a polynomial or something. Right.

Or a Taylor series. And so to go back to the story back in the 80s with an MHV amplitudes, which I think now is a bit a misnomer, I would call them double minus amplitude, because that's where we're going to get to in a second, right? There was this heroic calculation where a lot of Feynman diagrams were summed, and they were considering more and more interactions with more and more particles, and every time there were more and more terms, but they all canceled, to the end, always give a simple answer. And in fact, that's what this PT term, PT stands for Park Taylor. These formulas are, you know, they fit in the line, so it's not that complicated, but it's very surprising that such a messy calculation at the end would clean up

into such a simple result. And so what Alfredo Andy and David did was to understand that these single minus amplitudes in this special case where some of the particles are aligned, they don't have to be zero. And then you can do this very complicated Feynman diagram expansion to get the answer, which is not zero. But the problem is if you do it this way, well, you can represent the answer in some horrendously messy, complicated way. But if you unpack it, it's extremely complicated. It's complex in the following sense. When you consider the endpoint amplitude,

so the probability of N particles interacting, the number of terms in your answer, which correspond to the number of diagrams roughly that you have to add up, it grows factorially in N, the number of particles, and factorial growth is really bad. It's super exponential. It grows faster than an exponential.

So it blows up in your face. This is what you're seeing here. And that's because roughly you have to draw all the possible cartoons, and the possible combinations is a combinatorial problem, and that's where the factorial behavior comes from. But we know from the 80s that in the actually more complicated double-minus case, Park and Taylor found this miraculous simplification. And so Andy, Alfredo, and David spent the last year chasing the analog of the Park-Taylor formula,

the very simple answer that was obtained in the 80s for the double minus amplitudes, but now for these single minus amplitudes, which they understood are not zero, but then what are they? And they were getting this really complicated answer. And, okay, you never know in physics ahead of time if something will simplify. You have to believe in it to find this simplification. But because the double minus one simplified, it felt like these should simplify too. And we think they're important for lots of things,

and that these are somehow really important objects that are very fond of. the metal, and so they should have a nice description. And so they spent a year looking for that. There's a funny, the next line, if you scroll down, it's something like, we need a simpler formula. Right. So when we were on the paper, we need a market size formula is needed. Yeah, a market size formula is needed.

And this is where AI comes in. Because when I asked Andy, hey, do you have a problem in your pocket that we should use AI to target? He said, well, I have just a perfect thing for you. We've been puzzled about this. It's really important. It's really interesting. It connects to all these things.

And we don't know the answer. Yeah, I mean, like, when I was a grad student, if I had approached something like this, I would have probably plugged it into a computer algebra system, chugged along, tried a few limiting cases, see if there's any, like, magical, you know, simplifications which happen. This type of thing is, you know, something that oftentimes you see this thing.

You're like, we need a different approach. Exactly. Then, before Eddie even comes, here we started to play with chat gpt and alfredo andy and i were trying different things lots of different chats happening going back and forth david as well and the first thing that happened is that we fed the five point amplitude into chat gpt and we're like can you simplify this and it's like you know there's a special region so there's an extra assumption that you can make in which this answer simplifies to this one.

So this assumption is equivalent to, you have one particle coming in and it decays into in minus one. That's one way to think about it roughly. Okay. But we're in two time dimensions. Yeah, it's complicated. Basically, you can look at what we call face space.

It's the entire space of possibilities for all the energies of incoming particles in the mementa. And there's a special region in that phase base, where one particle has one different sign of its frequency compared to the other. and in that region, there's a big simplification that happens that Chad GPT found, and I should say this was the public model, but the pro version that thinks really hard.

So was that an unknown fact that it just was able to relate to the problem, or was that something that it put together? As far as I know, it put that together. It said, you know, this five-point function, which is a sum of eight terms, each one of which is a product of three terms, they're all pretty complicated. It said, hey, actually, this simplifies to this product of only three terms. And we stared at this and thought, wow, that's really nice.

We didn't know this. It's actually, in hindsight, once you know, you can redirect this, but it takes a while to understand where this comes from. So I think that was a leap of insight that the AI had. And I think what it did, I mean, at some point said, I wrote a Python code and I ran through all 5,000 possibilities. And I, okay, I did you. So it's the equivalent of running.

his computer algebra system, but it just decided to do it on its own and came up with a huge simplification. So great. This was after making the assumption. This is after the K, one particle decay assumption. Yeah, so it made, it figured out, there was a lot of exchange. This is a very experimental, but we're talking about it a lot. To figure out there's some form region in which things simplify it in that region. It said, okay, this thing simplifies it. CBT came up with that simplification as well. of my album? Yeah, yeah.

And then we were like, okay, well, let's give it the six-point function, which Alfredo heroically computed, and by we didn't have the center point function. I don't think anybody could use the identity to expand it. It would be disgusting. And then Chad GPT does its little thing, and then it's like, hope, yep, simplifies to this. And we thought, whoa, okay, that is really nice. Because now instead of 32 terms, it reduces to just four terms. And it's not a sum of 32 terms.

it's a product of only four terms. And then we asked Chad GPT, okay, well, can you guess the general formula for all N? And that step, by the way, I mean, you could imagine using some programming language or symbolic manipulation software to do these reductions in certain examples. But to tackle the general case, I don't know how to use a computer to do that. But Chad GPT said, yeah, this is the answer in the general case. Boom. How long does that take? You know, it's like using pro, it thinks for 20 minutes. at a time, you go back.

But it wasn't like six days or something. No, no, no, it's just like over the course of several interactions. And the amazing thing is that the formula that it proposed, instead of having this factorial growth, which is super exponential, where the number of terms, as you consider an increasing number, N of particles, the number of terms blows up, here it's actually linear. So if you double the number of particles, you only doubled the number of terms. It's the nicest possible behavior you could imagine.

the equivalent, I think, of the Park-Taylor formula for the double-minus amplitudes that was known back in the 80s, but now for the single minus amplitudes. And this was guessed by GPT, I think it was 5.2 at the time, GPT 5.2 Pro.

But it couldn't quite derive it. So I said, hmm, looks like this, but I don't know how to prove that. Yeah, I think the model was not quite strong enough. Okay, prove it. But part of my work at Open AI has been to develop stronger physics capabilities in the models, and a lot of people have been adding lots of,

you know, it's not just my singular contributions, there's a lot of great research happening, and it all comes together, you know, takes a village. But we had this internal model that could think for a very long time and was extra strong at physics. So we gave it the whole problem from scratch without actually giving it this. We just formulated the problem in a very sharp way. asked the model to solve, to find the answer for the amplitude in the general case in this region, because now we'd identified that this was the special place to look. And it took 12 hours, which is a long time, but it came back with the same formula, which we had not given it. So it rediscovered the correct formula, but this time it also found the proof that the formula is correct,

that derived it. And in fact, the remainder of the paper after we state the equation is devoted to the proof that is basically what came out of the AI. So we say the rest of this work is devoted to proving that the conjecture is correct. There's three steps. First you show this. Second, you show blah, and third you show blah, and then this is basically what the AI came up with. So now I can finally summarize the paper. The title is single minus gluon tree amplitudes are non-zero. So these special interactions between gluons where only one of them has a different holisticity from the others, which were previously thought to never occur.

Actually, these interactions can happen, so the amplitudes are non-zero. That's the main claim of the paper. I think it's quite surprising. I think it's like a really nice paper. And the final result, I guess there's two results. One is understanding that it's not zero.

That came from the humans like a year ago, but they were trying really, really hard to find this simple answer for what the amplitude is. And they were kind of stumped for a year. They were able to get this indirect representation that's extremely complicated in terms of Feynman diagrams, but they were looking for the simple formula

that is the analog of this Park Taylor work from the 80s for the more complicated aptitudes, and that was done with the AI. And so I think that's a really interesting result. Yeah, it's amazing. It totally changes the way you should think about where we are in physics and how AI is going to change that. You know, it's not just hype.

I mean, this is like a real thing to happen. It's a result that top researchers in this field were thinking about for a year, and then the AI solved it. So I find that's interesting. There's several things about the story, which I think people didn't understand on Twitter. If maybe scroll down to like equation 38 to, what's it, 35 to 38. Yeah.

So I would say most even intro grad students would look at 35 to 38 and say 39 is actually a very natural extension of this. Like that is, I don't think, you know, that surprising. I think it's interesting. I didn't know until just now that you can, that when you prove, 39, that was a fresh session. That was without the limiting cases, you started from scratch. Yes. Why did you do it that way?

Because I guess it's an extra way to be confident in the answer. If a different model independently comes up with it from scratch, then you're not just spoon-feeding the answer that you think is correct. That's an extra confirmation. But yeah, I think we thought a lot about how to put this out into the world. and there's no perfect way to do this. Clearly we could have done a better job of communicating it. One thing that was important to us is to not make this paper about AI

because I think this is a really interesting physics result. You know, people will keep reading this paper, I hope, for a long time. We didn't put AI in the abstract because this is a physics result that stands on its own. There's one paragraph really about AI where we just say the final formula was first conjecture to by GPD 5.2 Pro and then proved by an internal open A model because that's what happened. It's true. But we didn't really want to get into it because I think that's not the point of the paper. I mean, it's really interesting how it happened, but the result stands on its own. And I think if you read a paper today that was written 20 years ago that used the computer

to do some critical step in the argument, and it had this whole discussion of how, well, I loaded MS-DOS 3.1 and it had five floppy disks. I had to swap my floppy disk after. You wouldn't care. That's not why you're reading the physics paper today. So we didn't really want to go into that in the paper. And we talked a little bit about it in the blog post that we released with the OpenEI, which is this one. And then I guess on Twitter there were a lot of questions.

And I wrote some tweets that I think clarified it. And there was somebody who is a physicist who wrote a great blog post, like actually understanding the story. And the economist also put out a great article about it, which, they really understood what happened, and I thought it was a great, great coverage. Science magazine also read about it, Harvard and the Institute for Event Study put out of press releases. So I think it got a lot of attention,

but it's kind of a subtle thing to explain. It took us an hour to go through what happened and what was done. So, you know, it's hard to explain. I think it would have been kind of a distraction from the physics point of the paper to go into that. Okay, let's talk about the physics, and give us a sense because, you know, my theoretical physics on the frontier,

it comes from PBS space time right like I'm you know it's a great channel yeah and fantastic and I place and gives you a great high level picture but hard to know how this sits in the pantheon of papers that represent the cutting edge of theoretical you're asking me how good is the paper not exactly that I want to just understand like it seems like you're comparing it to this previous result that is pretty significant and like you know like highly cited and very important. How does this compare to that? Okay, you're putting me in a bit of a tough spot. I will say

I think the result is surprising. That's why the title is what it is. You know, single minus empluses are a non-zero and if you're somebody who works in this field, that should catch your attention. Ultimately, it's very hard to know in science when you release something

into the world, how it's going to be received and how impactful it will be. I think the true value of a paper can only be assessed this disappears into the future based on how much future work it leads to and what developments it opens up. Maybe a better way of asking this. So my understanding is that that previous paper kind of opened up a whole line of thinking

about. I think this is a great segue to the second paper that came out just three breaks later. Perfect. So it got its own blog post. This was March 4th, I guess, two weeks ago now. So we're talking earlier about how there's four forces, strong force mediated by gluons, and then gravity that's mediated by gravitons, except gluons we can produce at the LACC, and we can measure their effects fairly directly.

Gravitons, we think, are also around us being produced all the time, even as I move my hands, but we've never done an experiment that directly measures gravitons, but they're supposed to be the quantum of gravity. they're really interesting from a theoretical standpoint. So going back to RJ's question earlier, what is it gravitation? These different answers we could give, ultimately the correct answer depends on what the theory of quantum gravity, which we don't know yet. If you just naively try to take all of your tricks from field theory that we know from the standard model,

apply to gravity, things just break down. The theory is not self-consistent, is some definition. various problems. Just like if you took in this room there's light flowing around there's some indivisible bit of light that you at some point can't break up into

smaller bits. That's the quantum of light. We call that the photon. And the gravitational force is mediated by the exchange of gravitational force or gravitational waves. If you try to take a gravitational wave

and break it up into smaller and small and small pieces at some point you get a quantum that you can't break up anymore, and that would be the graviton. That's how we understand them. So there's the idea being that, like, you can't, you get to a certain point and you can't have less gravity than that. You either have some or none. Right. That's one way to think.

Yeah. And so we wrote this paper, which is called single minus graviton tree amplitudes or not zero. So it's the same title almost, except with graviton instead of gliton. instead of gluon. And that's our purpose, because we wanted to extend the result. And it's the same story in the sense that it was thought that all symbol minus amplitudes are zero,

but actually it's not true and also for gravity. But gravity is a lot more complicated. So now if you want to compute the graviton amplitudes, it's potentially a lot harder. Do avatons have phase the same way that glons do? So they actually have spin two rather than spin one is getting into the weeds. So the amount of the numbers you have to use to describe them are a little bit different. They're doubled in some sense.

So their polarization is more complicated. I see. But this is really getting into the weeds. But the special region in which the final answer simplifies has two labels because it's a spin two particle. Whereas in the glue one case, there was only one label because it was a spin one particle. So it's not the same math. gluons and gravicons do have some spiritual similarities compared to other types of particles. Yes.

In the sense of... They're particles of force. Yeah, yeah. But they're like sort of doubled. Yeah. They're sort of doubled now. Yeah, I mean, okay, I guess the people watch this podcast probably like to geek out all this. So the modern definition of a particle in quantum field theory, which is our best verified framework for nature, is that particles are irreducible representations of the Pocaray group.

We just lost 90% of our audience. Yeah, okay, maybe we cut this. So there's mathematical representations, and they've all been classified. All the possibilities are known by Vigner, actually, a brain physicist, and it turns out that the representations or possible particles are completely labeled by the mass

and the spin and the charge of the particles. So these are the three quantum numbers, and particles of long-range forces, like gravity and electromagnetism have zero mass. And they have to have integer spin. And spin one is three of the four forces, and spin two is gravity. And then that's it.

But let's set that aside. The really cool thing about this paper is that, well, first of all, it came out three weeks after the first one, which is really fast. And I think this is a great example of AI accelerating science. And in fact, we could have put this paper out three days

after the first one, because that's how fast we got the answer out of Chad GPT. But it took us three weeks because we wanted to check very carefully if that it's correct. But most of the time we spent verifying the answer, not directing, which is insane. Actually, if you take a step back, if you told me a year ago, yeah, like, you're going to have this AI that just does really hard calculations for you. And then most of the human effort goes to verifying the answer. I've thought that, you know, you're crazy. So it's very surreal. And then we also had to write it up as a nice paper,

which they put in the citations and references that takes time. And also had a baby in the meantime. So it was a lost time there. But we did this really fast. So I think it's an example of xerate science. Another really cool thing is that for this paper, we didn't have to use an internal open AI model that had to think for hours. This was all done using the publicly available GPT Pro.

In fact, we shared one of the main prompts that we used. If you go to the blockpost, extending single minus amplitudes to gravitons, and you scroll down to the text, there's a link to one of the chats that we used. So you can see we use chat GPT 5.2 Pro. And the amazing thing about this is that we gave it the glue on paper as a seed.

And we said, we understand the paper. Make sure you understand the manipulations and the appendices because that's where most of the hardware goes. and but it comes back and it says yep um i'm justed the paper let me focus on the appendices here's what happened and basically the punchline is that gpt pro with the glue on paper as an anchor was able to do the graviton calculation which is really different mathematically completely on its own from well i guess not from scratch from the glue on paper but it's it's just a different thing and it was strong enough to do it completely so it took the conceptual leap from the paper

the previous paper and just said, okay, what math do I need to make that same consensual? And it's different math. That's an important thing to emphasize. So in particular, there's a crucial application of something called the directed matrix tree theorem. And Alfredo and David, we've been thinking about these things for a very long time. We're like, oh, that's really cool. That's surprising. We hadn't thought of that or seen that before. That was like known math, but it was because maybe it has such broad understanding of math and physics that is able to say, this is a good thing to apply in this case. Yeah, exactly.

And so here it understood the paper, the glue-on-one, and then we said, okay, well, the task is to generalize this paper to the gravity case. Here are two key changes, but otherwise manipulations should be similar. So we tweak some things at the get-go, and then we said, good luck, you're a brilliant theoretical physicist. So it's like a, you know, we give it two paragraphs. So we gave it a glue-on paper, a couple paragraphs. And it said, good luck.

Thought for 20 minutes. And boom, it starts the thing. So it starts at the beginning. It works through the implications. It's all like really interesting stuff. And then it says, here's what I would do next to turn this into the gravity paper. If you want, I can do blah. And so we say, yeah, go ahead.

So. No, no, thought for 31 minutes. Thoughts, 31 minutes. Yeah. So this exchange is 110 pages. But I think it's hilarious. I would describe this as vibe physics. because you can see

so now it goes away there's a lot of hard work goes lots of equations it's starting to do the okay so now you have to use this different math you have to use these tree calculations LAC reduction formula is a lot of stuff happening

simple for trees concrete checks it's starting to yeah well this is one of the things I love is that it's able to do the same things that a human would do which is check some basic cases of sannie check and to get intuition and so it comes back every 30 minutes

said, well, here's what remains to finish the full gravity paper. And there's a list. If you want, I can write the gravity analog. Yes, do that. This is the first step. Okay, it goes back. Thanks for 34 minutes. Half-colline your support.

This starts the best... Okay, these formulas actually made in the paper in some form. This is all correct. There's a bunch of stuff. At the end, it says, if you want, the next most useful thing I can do is do this. And we're like, yeah, verify this by performing the explicit. check and it goes on just to cut to the end. Finally, we say, okay, write up the paper. And you can see

the paper that it writes. And it's very close to the final thing that we actually put on the archive. So did it make suggestions that were not what you would have suggested as the next steps? It's very smart. It knows kind of where to go. It's useful to steer it. If you can Compared what it came up with, with the actual paper that we put in, the intro, the abstract and introduction were written by Andy, who's an amazing writer. And I think he gave this wider perspective on the problem and how it fits into physics and how it connects to other things that, you know, the AI didn't do it. Just the introed road was more generic. But, okay, AI could write really well. We didn't really try to make it.

Yeah. And the other thing is we added the section, this section two, which was not part of that initial exchange, is about how these graviton amplitudes transform under certain symmetries of physics. And that's something that we're really, really interested in because we eventually want to understand quantum gravity, as I mentioned earlier. And typically the first step to uncovering a new theory is to understand what are its symmetries. That's something that gives you some kind of ground to stand on. Andy in particular has been pushing this program of celestial holography, which is like a whole thing we could get into, but it's an exploration of the symmetries of quantum gravity,

and he really wanted to understand this. And there's a separate chat. We didn't share that one where we led the AI to explaining how these answers fit into the symmetries that we know the theory should have, and that's something they went in there. but actually I think from Section 3 onwards, it's pretty much very close to what the AI wrote. So I would say this is really remarkable.

It's a real solid result in quantum gravity that was done pretty much completely by an AI with humans steering it and asking kind of the right questions, but all the math was derived by Chad GPT Pro, the public model you can access. And most of the time spent was, by vast, the humans was like checking everything and writing it up. And that's really wild. I mean, as a physicist, you find yourself where a lot of coders have found themselves,

where there's a kind of a fundamental, maybe epistemological question here that if now as a physicist, like I could have done that, right? Like maybe like I needed a little more background. But a lot of it was like, yeah, go ahead, right? Like take this paper, give it some prompt. You guys obviously prompted it very well, but there wasn't like maybe an undergraduate in physics could have come up with a lot of that. And so the question is, how does the undergraduate in physics now learn when they don't have to do the hard calculation themselves? Similar to the how does the undergraduate coder? Actually, you're opening up many different strands of conversation.

from Charles super interesting. So let's try to unpack that a little bit. So the most direct thing you asked is how does the next generation learn? Yeah. That's a really good question. I think about this a lot. And now that a lot of senior physicists in the field are coming to grips with these new capabilities, one of the questions that comes up very quickly is how do we train the next

generation? Because the way we were trained is by going through these, you know, these, these difficult rights of passage where you have to do these really arduous calculations and this is how you build confidence in your own abilities and check test your knowledge and it's not just about what you're capable of doing it's about knowing that you're capable of doing it and proving it to yourself and building that self-confidence is important and we don't have a good answer this is something that academia is going to have to grapple with so one thing that is especially difficult is that as a professor I have graduate students.

And the gap between where classes take you, even graduate courses, they only go so far. They go very far, but only so far. And the gap between where that ends and research begins is actually huge. And it's growing wider. Usually as a professor, what you do is when you take on new students, you keep in your pocket a few easy problems in the sense that you know they're going to work. some questions that you know in principle you could work out

not that difficult, but you give them to a student so that they go through the exercise of learning everything around the question, developing the technology, and then you know enough about the problem that you're sure there's an answer that you can get there and you can advise the student

in the process of discovering it. And I think the issue is that many such problems now, I would say these models can probably crush. Yeah. These are problems that we, usually take, again, you know, time scale for theoretical physics paper is six months to a year. That's pretty typical. So if you tell a student, go away and think for six months about this one question, and you have to work really hard, learn a lot of stuff around it, and do lots of calculations,

even the most determined students, would they not, within the course of six months, ever asked? Yeah, that's a little bit weird. Now, it's also an opportunity because I remember that time in my graduate school career in my second year of grad school. I took all my graduate courses my first year and then my second year was my first project. And it was actually the hardest time for me in graduate school to traverse the desert for more classes to take you to the research frontier. It's very hard. And there's a lot of time spent banging your head against the wall. Like all the time you're confused. You don't understand things just because you need

to absorb so much knowledge. And AI can totally help. you with that. It's the best teacher. It knows everything. It can unpack any complicated fact to any desired level of detail. Actually, my experience as a trained professional physicist working on my own research using GPT now is that I would say there's two key ways in which my research has completely changed. One is that I spend much less time being confused. So I'll do a calculation, get an answer and I think, how does this fit in with this other fact that I know. Like, how do I reconcile these things in my mind? I'm confused. Yeah, I do that all that time. Yeah. In research, usually you take a step, then you're like, you hit a roadblock,

an obstacle, you're confused, then you have to think for a few days, maybe you go for a walk, or you work on another project, come back, get a new idea, but you spend a lot of time confused. That's nature research. With GPT, I'm like, hey, I just did this, I found this, how does this match with this other thing? And then it's like, oh, well, you forgot this thing, or, oh, you didn't quite think about it correctly or does the standard fact, you know, and so the amount of time you spend confused just dramatically shrinks and you move so much faster, that's one of the accelerating effects. The other accelerating effect is that, you know, I only have so much free time and energy, especially when you become a professor, you have to teach, you have students, you have a grant

administrator, there's a lot of things you have to do. So you're free time to think about research without distractions, shrinks. And also, you know, you only have so much energy to do hard calculations. And so what you would usually do is if you have a problem, you know, you're at point A and you want to get to point C, you think about the route, oh, I have to go through point B first, and actually maybe there are multiple points, and you try to plot in your mind the course that you're going to take before you actually go start through the hard work. You try to think really hard about where you're going into the charter course. With AI, actually, you can launch 10 instances of chat and have each one try a different route and send it as a scout that moves very

fast into the unknown, pushing outwards. And you can just very quickly get some feedback to see, okay, these approaches are not promising. These are much more promising. And then if you follow them, there's a huge difference between being the first to push into the unknown versus following someone ahead of you. And even if Chad GPT doesn't always get everything right, just kind of having a scout that signposts some key steps along the way that you can use to anchor your own movement is extremely helpful. So this is, there are like two concrete ways that AI has changed the way I work. And I think

if you're entering research, having an assistant that can help you find your way to where you're trying to go can be very good. So I think it's, it's inevitably going to change how we work, how we operate and how we train students. And, you know, part of what's exciting about my job is trying to figure how all of this works, but it's not just a job for open AI. It's actually a job for every researcher and professor more generally to think about this. I think the future is very bright because we have some challenges to overcome, but on balance, this is such an amazing tool. I think it's going to give human physicists AI superpowers because of what I just described,

you can do so much more. And I think actually the kind of skill that is really useful to get great results out of AI is very similar to the kind of skill that you develop as an academic collaborating with other humans. This is like a collaborator. And if you're a professor who's been advising students and postdocs,

a lot of what being a professor involves is knowing for each student postdoc that you're working with exactly what question to give them. So matching the problem to the person and knowing how to give them the question in what way with how much deep, what level of detail,

not too much, not too little. And that's actually kind of what you have to think about when you interact with Chad GPT. So I think that it's a transferable skill, and people who are good at this are about to get AI student powers. What you just described there

reminds me of several conversations we've had on the podcast as far, which keep coming up to this concept of paste. One of the things that, especially you say theoretical physics, high energy physics has maybe had a problem with, I'm not sure if you want to describe it that way, but it can be very trending that there are certain things which become in fashion

because maybe right now we're in a world where we don't have the data to define new directions to really guide or constrain where we're going. I'm curious, like, how does essentially something which is superhuman in that it has basically all known physics and interact with a field where at its core what oftentimes can be. be popular or people start working on is more based upon general aesthetics or what, you know, the community collectively thinks is cool at the time. Because I can imagine Iku could vibe so many different worlds. You know, like, for example, just using Klein Space, using this sort of two-time, two-spatial dimensions for this, was already sort of an assumption that I think is actually kind of important in some ways and does provide feedback to our

world. But in the concept of, you know, you could have asked ChachyBTBT to solve this problem in all sorts of a number of ways. And, you know, maybe it could come up with all sorts of things which don't really align with maybe the useful paste. As a community, how do you actually deal with that? Like a proliferation of really interesting results, but it's actually not clear what is where the field should go. You're getting at the heart of what does it mean to do progress in theoretical physics and research. This is a hard question, and there is a simple answer. If there were, it would be research.

Let me say a couple of things. The first one is, when you go to graduate school in physics, it's usually because you're really interested in the big questions. Why are there three dimensions of space? What happened to the big bang? What's inside a black hole? These are the things that I was thinking about because of sci-fi movies and books that article. What you realize is that actually these questions, even though they're really cool and exciting,

they're not really the most fruitful scientific questions because at any given time there's an edge of knowledge. And the role of scientists is to extend the edge of knowledge, is to push into the unknown. And to do that, you want to find the questions that are right at the edge or just beyond the edge, but not so far that you can't grapple with them. So the question of why there are three dimensions of space, that's a really cool question, but I don't know of anyone who said anything really compelling about that. So it's just a question that's beyond the edge. So it's not, as a professional physicist, I don't spend my time thinking about this because I just don't know any pathway to solving this question.

It's not useful to think about. So really, the process of training as a physicist involves coming to grips with what the edge of knowledge is, because that's where the interesting, fruitful questions to make progress on as a scientist. That's where they live. oftentimes when you go through graduate studies, you worry, oh my God, I have to learn about Feynman diagrams and all this math and all these calculational methods. And it's true that's a really hard thing to learn.

It takes a lot of work. But in some sense, once you become a professional physicist, you should feel like you can learn any tool, you can pick up any tool that is needed for the task at hand. And you should develop that confidence. And that's what makes a competent physicist. The competent physicist is one that can learn any new mathematical tool or, you know, piece of code or whatever that is needed to solve the problem at hand. And that makes you a good physicist or a competent one if you pick up this skill.

And in graduate school, you know, it's daunting you have to learn a lot, but by the end, you should have a lot of skills in the toolkit and the confidence to pick up any new one is needed. The difference between a good physicist and a great physicist is knowing what is the right question to ask. that is actually the hardest part of being a scientist. It's knowing what is the next fruitful question to tackle. And I think AI right now is a very good physicist. In fact, maybe superhuman when it comes to certain computations. But it's like this extremely technically skilled graduate students that you can give a sharp, well-posed question to.

It will do incredibly hard calculations. correctly now and come back to you with the answer. So it's super competent. But one of the things that it doesn't quite have yet is knowing what is the right question to ask. And I think just like with humans, that is actually the hardest skill to pick up. That's the one that comes last. I know you're not working directly on the AI so much. I mean, I don't know exactly how much you. But do you get a sense for, you know, you can imagine a future where you just do better reinforcement learning, maybe you change the architecture of the model completely so it is some other, you know, like whatever, not transformer. And the trajectory just keeps going like this because it's

been very, very rapid increase since 01 of the recent capabilities. Or do you get a sense that we're getting near the edge of the frontier of knowledge now so that the sort of the ability, of the model to recombine knowledge in somewhat novel ways is, you know, like, that's kind of, in the, it seems like not to just, or like not to play down any of these results, but that it seems like there was a lot of what it did, and maybe there's some not like this, but a lot of what it did was recombination of known facts. Okay. Yeah. But do you get a sense that that's, do you have any reason to believe that will continue or if we're going to just like sort of, okay, we know how to recombine stuff really well,

and we can't push beyond that. Without getting too philosophical, I'm not sure that any of us are anything more than recombination of the machine. Fair enough. Working with GPD Pro on this problem, to me, feels like working with a creative collaborator.

It did stuff that I didn't know that I found surprising. And so I think, I'm not sure there's a qualitative, I think it's just a matter of degree. Yeah. Okay, that as we continue scaling the capabilities, which is certainly happening, I don't see why it's going to stop. Like, we definitely have a bunch of things in the pipeline that are going to keep coming this year.

And, you know, my horizon for seeing it to the future is like not that good, but beyond the year. But like, definitely we're going to keep scaling up this year. And I don't say any reason why it's going to stop. And I think that's going to make these models display feats of insight. that looked to us like real creativity. I would say this already happened in this project. At least, you know, what is creative insight?

Is it in the eye of the beholder? I mean, AlphaGo, right? Yeah. I would to come up with moves. It were very... I talked to Terry Tao a couple of weeks ago at UCLA. We had an open AI event with IPAM, which is this Institute of Mathematics there.

And I talked to Terry Tao, and he said that in his view, all of the proofs that he's seen AI come up with in math, even the ones that had first seen creative and surprising later on were tracked down and found to have really pulled facts out of some obscure reference so I don't want to put words in this valve but my understanding was that Terry Tape has not yet been impressed

by a creative move in math but you know Terry Tao is a unique individual I have been impressed I consider myself my bar is lower and I think as we keep scaling this up I can't go into the details but there's a lot of effort at Open AI there's a lot of work really smart, hardworking people that are pushing very hard to take this next step. And I think it's going to come eventually. I mean, just look at the trajectory route that we're on.

So a year ago, I was black hole physicist in academia, not really paying too much attention to AI. I thought AI is cool for emails, but it's not going to do what I do, which is special. O3, which was the first, really strong reasoning model, came out and was able to do. do a calculation for me that would have taken me days. You did it 11 minutes. And I thought, wow, that was shocking to me. And we could go into the details of free of time. I could show you the example because I had a safe and it was really surprising to me. And then I thought, okay, I got to really start using this tool. There's no other software that could do this kind of calculation

as far as I know. It was really surprising and really cool. And then GPT5 came out six months later. and that was able to reproduce one of my hardest calculations, which I think the number of people in the world that could do that, you could count on your hands. When you say reproducing, this has been published or not published? It was a secret or internal...

So last summer in June, I put out this paper, which I really like, in which I found... It's called... Why is there no love in black holes? Yeah, and love is actually a technical term. It refers to Augustus Love, a British mathematician, who studied the tides.

So when you have an object like the moon going around the earth, it exerts tidal forces on the oceans. And so you can measure the tidal response, say, of the earth and its oceans to the moon, via some coefficients that encode the strength of the tidal response. And these are called love numbers in reference to Augustus Love. But famously, black holes do not experience tithes, so they have no love. and there's been a resurgence of interest in this fact in the last five years

because people understood that this can be connected to symmetry principle. So in physics, whenever something is zero, like, why should black holes never experience tides? That's surprising? Well, oftentimes, the answer is because there's a symmetry principle at work that forbids the existence of tides

that protects the structure of the black hole. And so I found these new symmetries, so these are differential operators that act on solutions to this equation that describe perturbations of a black hole. And these generators are symmetries because if you act on a solution to this equation, you get a new solution. You know, I thought this was very beautiful. I liked it very much.

And it came out in June on the archive. And in August, GPT 5 came out. And the cutoff date for its training set precedes the release of this paper. So, GBT did not see this paper in training. And when it came out, I was like, okay, I'm going to got to meet Mark Chen, who's chief research officer at OpenAI. And he said, give GPT a really hard problem. Let's see how good it is. And I was like, you want a hard problem?

I'll get you. I'll get a hard problem. I just solved this problem. And I wrote a paper. I was very excited about it. I thought this is really deep, but cool. And I gave GPT the equation here. And I said, what are the symmetries?

I didn't tell it that there are symmetries because by the full of the answer should be that there aren't any. and thought for five minutes and said, yeah, there are no symmetries, which is what usually happens. And that was wrong. And March, that was vis-a-crest-fault. It's like, oh, well, okay, why don't you give it an easier question?

And so then I gave it the same question, but not for a black hole space time, but for an empty flat space time, which is a simpler problem, but that's actually how I approach this problem myself. You know, you warm up on the easier question first. So I gave it the flat space question, which is,

in this paper. Also, so it's this equation, which looks much simpler. Then this also has three symmetry generators, which are shown here. This is not new.

These equations have been studied for 200 years. Everything in flat space has been known forever. And GPT5-4 thought for, it was like nine minutes, and it came up with the answer.

Very beautiful answer. Perfectly structured, perfectly correct. Actually, at the time, I also tried the other models from our competitors, and none of them could get this at the time. So GPT pros really ahead, and I think it continues to be the best model

for this kind of mathematical and physics work. And Martian was like, okay, well, this is great, but now that it's done, the warmer problem, in the same instance of chat, tried the full problem again. Now that it's been prime, I thought, okay, why not? And so I gave it the same question as before, What are the symmetries of this equation?

Now the full black hole problem. And this time I thought for 18 minutes, which I'd never seen before. And it came up with the answer. So basically in under 30 minutes, with one hint, which is the obvious warmer problem to prime the model on first, it completely solved this problem, which, you know, was one of the nicest calculations that I've ever done. And that really blew my mind.

That was my move 37th. Seven moment. Yeah, that's how we call it in the AI world. And once I saw that, I thought, okay, we're on this crazy trajectory where, you know, 18 months ago was not useful. A year ago, it could do really hard calculations that would take me days. Eight months ago, it could reproduce some of my best work in like under 30 minutes. And then now in the last month, it solved these questions that we've discussed at length, which the world experts had. spent a year thinking about without being able to get to the answer.

So, you know, I think it's just going to keep getting better. Where are we going to be in six months or a year? I don't see any reason why it would stop. And I think we're going to be having a very exciting year. Going back to this like these thoughts about scientific discovery and what can these models do versus just being very good at superhuman at solving physics. The people keep asking this question, hypothetically, could we train a version of chat GPT where it's never seen anything post, you know, 1904 and could it rediscover relativity? I think that there's a very analogous like question we could ask right here, which is new conceptual result about a single minus glue on amplitudes, which was sparked by a human insight.

and there were some certain very specific assumptions which went in here like understanding working in Kerr space time is something that people have been thinking about and has some useful like transferable insight people have been thinking about you know maximally illicitive violating amplitudes for quite some time have you ever tried kind of using a model right before the cutoff date of this and ask given a Kerr metric is there anything interesting with regards to helicity violation or maybe turning around saying, you know, it's long been thought, or it's long been known that with exception of some set of measure zero due to what,

and there's no, you know, single minus non-zero amplitudes. Have you tried either of these directions and asked it to like discover a new insight, push the boundary as you were just talking about, and make a leap in addition to not just solve a problem like you can give it, but actually could you get the intuition? Yes. Yeah, you will try this. Not exactly the counterfactual version that you're describing. I personally haven't done that,

but pushing the models at the frontier to try to make this type of leap is something that we're very focused on. Yeah, and I think, well, okay, I don't want to talk about the eternal research we're doing, but I can say something publicly, I think, which is you can take this page of this paper, and you can feed it to chat GPT Pro, say, I like the best model we head out right now, and you can ask it, what should I do next? Give me the top three follow-up questions to ask based on this paper. I've done this experiment,

and the top three questions it comes up with are like my top three questions for what I should do next. So I think the models are smart enough now and have enough background knowledge that, you know, for this paper, I'd say GPT is about as good as me at finding the next day to ask. And so that's really interesting and it opens up a lot of... So can you just, you know, what is the name of the loop that the agent loop that everyone is talking about where you just say like, okay, what's the next question? Go ahead and solve that. What's the

next question going on? So I guess, and this goes back to the question I was asking before, which is if you do that and probably has been, you've tried it or someone that opening it I presume you get to some plateau, right, where you're not pushing the boundary of knowledge anymore, or is it just like the plateau is money and if you had more money, you could go further? Yeah, just to be very explicit, because I haven't said this quite out loud, I think we now have bottles that can really churn out papers that are as good as humid-ridden papers. In fact, this is a bit of a problem because when a professional physicist uses this tool and they steer the model and they check the answer, they can get amazing results.

But there are also people that feed it kind of wrong questions that go off the deep end. They submit that to archive. And this is a problem that the academic community is trying to come to grips with now, which is this problem of AI slop, but for some. science. This is something we have to figure out. But, you know, I would say that with proper steering, you could probably churn out a paper a day now. You know, just, I don't know, it's like give the question to chat GPT. It'll solve it. If it's not that hard of a question or it's a similar calculation to stuff that's already been done, it can totally do it in 30 minutes. And

then you can say write it up as a paper and you can send it to archive. Okay. So I think that we're already in this moment. We've passed that threshold. This is the new reality. And more and more people are catching onto this all the time. And so some of them are doing this. And this is why the archive is now inundated with submissions. So what's the correct response to this?

I think we put out these two papers in very fast succession. We could spend the rest of the year writing 30 more papers like this. I don't think that's what we should be doing. Instead, I think now that we have this new tool that gives us AI superpowers, I think we should just raise the bar for what it means to write a good paper. Like, we should aim higher, basically. One thing that I'm excited about is that I think these single minus amplitudes papers, they open the way now to a whole direction of research, which I think is a line of attack

on really interesting questions in quantum gravity. To go back to the start of the session, this is the missing piece of the puzzle, fundamental theoretical physics. And I think we have a pretty clear line of attack through a series of questions, all of which I think will be amenable to solution with AI. And so I think, you know,

I'm excited to spend a good part of this year trying to follow this path, but really solve harder and harder problems. And, you know, this is a, this paper gave an answer to a question that had stomped Andy Alfredo and David, who were experts in,

this for a year, but we haven't seen an AI yet solve a question that stumps an entire community of physicists for decades. That hasn't happened yet. But I think given the trajectory that we're on, at some point, hopefully not too far into the future, we should see that. And so I think that's the exciting thing to try to go towards, which is pushing the envelope of what can be done. We wanted to start asking Oligas question, which is if you could remove one biol-knack for your domain. So in this case, maybe it's AI for physics or maybe it's physics or maybe it's mostly AI. But if you could remove one bottleneck for your domain, what would that be and why? Well, off the top of my head, you know, I spent so much of my time writing papers. And the way I think

now is so far from papers, it just feels like not the right way somehow to store and communicate knowledge. I think an extreme version of this, which makes the problem more apparent, is math. Especially certain parts of math where papers are very terse and they take four pages. I had this experience when I was learning algebraic geometry in graduate school,

going to a mathematician and say, what's going on in this like four-page paper? It's just like very tourist notation. And he said, oh, forget what's in the paper. And he took me to the blackboard and he started to draw pictures. You know, he's like, this is how you should think about it. And I was like, oh, wow, this is amazing. But like, one of that is in the paper. And, you know, mathematicians, I think, have this cultural norm that they kind of hide the messy

work and they were writing these beautiful, short, pristine papers. It depends on the subfield. But oftentimes that's the case. And the way they actually think about the subject is a living, breathing entity is very different from the way in which it's recorded in papers. Some of that is also true for physics. You know, I love doing calculations, coming up, with questions finding the answer. And then I would say a huge bottleneck is writing it up. So somehow it feels like papers are not quite the way of the future, or at least the way that we currently operate with,

I write it up, send it to a journal, it takes six months. I don't know, it was just like, why are we doing all of this? It feels like maybe there should be something better. I mean, you could, you know, if you want to understand this paper, thing you can do is upload it into chat GPT and ask it to explain it to you. And you can keep unfolding the complexity into more and more detailed explanations. And so if we move into a world where we use AI to do the calculation, get the result, then we have the step condensing it into a paper and then, you know, I send the paper to Brandon and he puts it back into an AI that we'll have

him, let's get. Like, why are we doing this? Yes. Right? That's a little bit funny. So I feel like, you know, if you ask me, would I be confident that in 20 years we'll have these sort of like static documents in which we publish our results as papers? I would think not. Like that doesn't seem like the best thing we could be doing. Maybe some kind of interactive paper, which lives in some LLM. Maybe your whole paper is some chat GPT page. And, you know, there's a chat bot attached to the paper and you can say, explain the big picture and like zoom into this fact, I think we're going to head in that direction.

That would be a cool thing to see. Writing a paper, though, is a useful exercise because it forces you to condense your thoughts and make them really clear. So I'm not saying it's a bad thing to do in general, but just the way we do it is very slow. But that's the first thing they came to my. Maybe another answer is, in this project, the Graviton paper, we got to a draft of a paper extremely fast. And then we spent most of our time checking the answer. So I think that will effectively be a big, like maybe the next big bottleneck. And that is one of the things that the models, I would say,

if you ask me what are, what is missing in the models? Like, what can we really improve for scientific research? I think we've, we've kind of touched on the two big things already, but just to spell them out. One is creativity. and the spark of invention and really taking the next step. I think that will come as we scale up the intelligence. Well, we'll see, but I don't know that there's something missing inherently. I think it's just it's starting to make these leaps for me, but maybe we should encourage the models to try to make bigger leaps

because large language models, after all, they're trained to give you the middle of the road answer. Like if you ask an AI, like Chad GPT, write me an email about blah, you wanted to give you kind of the expected answer, not to sample from the tails. like wacky email you kind of wanted to give you a reasonable thing so for most tasks you want that

but for science research sometimes you want the idea that comes out of left field the thinking outside the box or really sampling far out of the distribution and that's something we could do in principle but that's not how the models are you know we're not really favoring that

so we might have to do tweaks of this kind to make the models be able to take bigger leaps and then the second thing is verification because we're now in this new regime where the models are so capable that for very hard computations of the frontier of knowledge, they can just do the whole thing. But, you know, is it correct? In this case, it was correct.

You know, sometimes I get emails from people saying, oh, this is really long calculation, but there was a mistake. So I'm more, disappointing. Okay. I mean, the calculations are getting more and more complicated and longer and longer. But, yeah, sometimes they mess up. And so I think improving verification or even just having the model indicate

more directly how confident it is in its answer because I think they're smart enough to know what they're very confident in the answer versus when they're just kind of guessing in some step and getting the AI to be more explicit about that is I think a way to improve them for research. And that verification step, I think, is going to become maybe a bigger bottleneck this year.

Yeah, Karina Hang from Axiom would agree with you emphatically. The formal verification is their thing, right? Yeah, I think, you know, it's interesting. A year ago, I would have said it's super important to have formal verification. Then the models got so smart that I thought, well, you know, if Brandon and I talk about a mathematical proof and we go over it, we're not going to formalize it in set theoretic notation or, you know, we don't think the way

lean, which is this language for formal verification reasons, we reason through the proof in natural language. We use words. So if a model is really smart enough, then it should be able to do the same thing. And we've been saying this huge increase in capability

for mathematical reasoning and development of using natural language. So then maybe for a while it's looked like that wasn't the thing to really focus on. But now that we're at this regime where you can just get Chad GPT to tackle thousands of questions at the same time, and it will return proofs for a significant fraction of them. Now, actually, the onus is back on the

humans to verify all the outputs. And so, yeah, as that becomes a bottleneck, I think formalizing math and automating verification will become more valuable it looks like to me, and that's something we're thinking a lot about as well. Thanks. What do you want the audience to take away from today? Is there one message that you want them to leave with? Yeah, I think it's important to get the word out that the models that we're developing at OpenAI are becoming really capable of scientific research.

I myself was a bit of an AI skeptic a year plus ago because I thought the models are very good at writing tasks but not mathematical tasks. That changed with 03, the first strong reasoning, models, and then GPT-5 was able to do some of the hardest calculations that I can do and reproduce them correctly. And now recently, in the past month, we've seen models solve open questions in theoretical physics, and now they're solving problems in quantum gravity and quantum field theory.

So if you just extrapolate added into the future, imagine where we're going to be in six months or a year. I think it's kind of surreal to live through this time, but it's sort of. really happening is really amazing and I think we're going to see a lot of big changes happening in research so I that's yeah pay attention to the space that's day too awesome so cool thank you so much for taking the time this is like I learned a lot from our discussion and and I'm going to definitely keep up with what you're what you're doing thank you it's been great to be here thank you