Did GPT 5.2 make a breakthrough discovery in theoretical physics?

A few days ago, OpenAI published a blog post called GPT-5.2 derives a new result in theoretical physics, accompanying the release of preprint with a more opaque title Single-minus gluon tree amplitudes are nonzero.

This announcement sparked many debates online, with reactions going from "physics will never be the same anymore" to "it's just a fancy calculator."

It is hard to tell from the actual paper what was really the contribution of OpenAI's models, and almost no details have been given regarding the prompts, the scaffolding, the back-and-forth between GPT 5.2 and the human researchers.

But at least, let's try to understand the physics part of this !

As a theoretical physicist by training, I would like to walk you through the context and the significance of the results, and explain how they relate to the broader goal of better understanding the laws of the universe.

Gluons talk to each other

You may have already heard that physicists distinguish four fundamental interactions in nature: gravity, electromagnetism, the weak nuclear force, and the strong nuclear force. Those forces act on matter particles : electrons, quarks, muons... more generally called fermions.

One of the greatest achievements of twentieth-century physics has been the realization that three of those interactions (all except gravity) can be described using the same formalism: Yang-Mills gauge theory. According to this framework, a specific nonlinear differential equation, the Yang-Mills equation, is enough to explain the internal dynamics of electromagnetism and both nuclear forces.

The details are not important, but let me show you the key equation because this is an important one !

$${\mathcal{L}}_{YM}[A]=-\frac{1}{4}({\mathrm{\partial }}_{\mu }{A}_{\nu }^a-{\mathrm{\partial }}_{\nu }{A}_{\mu }^a)({\mathrm{\partial }}^{\mu }{A}^{a\nu }-{\mathrm{\partial }}^{\nu }{A}^{a\mu })-g{f}^{abc}({\mathrm{\partial }}_{\mu }{A}_{\nu }^a)A^{b\mu }{A}^{c\nu }-\frac{g^2}{4}{f}^{abc}{f}^{ade}{A}_{\mu }^b{A}_{\nu }^c{A}^{d\mu }{A}^{e\nu }\backslash mathcal\{L\}\_\{YM\}[A]\; =\; -\backslash frac\{1\}\{4\}(\backslash partial\_\backslash mu\; A^a\_\backslash nu\; -\; \backslash partial\_\backslash nu\; A^a\_\backslash mu)\; (\backslash partial^\backslash mu\; A^\{a\backslash nu\}\; -\; \backslash partial^\backslash nu\; A^\{a\backslash mu\})\; -\backslash ,\; g\backslash ,\; f^\{abc\}\backslash ,\; (\backslash partial\_\backslash mu\; A^a\_\backslash nu)\backslash ,\; A^\{b\backslash mu\}\; A^\{c\backslash nu\}\; -\backslash ,\; \backslash frac\{g^2\}\{4\}\backslash ,\; f^\{abc\}\; f^\{ade\}\backslash ,\; A^b\_\backslash mu\; A^c\_\backslash nu\; A^\{d\backslash mu\}\; A^\{e\backslash nu\}$$

In this equation $AA$ is a field called the gauge field, the quantity defined here ${\mathcal{L}}_{YM}\backslash mathcal\{L\}\_\{YM\}$ is called the Lagrangian of the theory, and it defines all its dynamics. An important parameter of the formula is $gg$, the so-called coupling constant, which essentially governs how strong the interaction will be. This is a classical field theory, here "classical" essentially means "non-quantum".

Now to go from the classical to the quantum Yang-Mills field theory, you have to do what we call a path integral. Roughly speaking, you have to exponentiate the previous formula, and integrate it over all possible trajectories that the field A can take. We can write this idea in a naive way like that:

$$Z=\int \mathcal{D}A\exp \text{\hspace{0.17em}⁣}(\frac{i}{\mathrm{\hslash }}\int {\mathcal{L}}_{YM}[A])Z\; =\; \backslash int\; \backslash mathcal\{D\}A\; \backslash exp\backslash !\backslash left(\; \backslash frac\{i\}\{\backslash hbar\}\; \backslash int\; \backslash mathcal\{L\}\_\{YM\}[A]\backslash right)$$

Again, no need to understand the details of the formula, the important message is that to compute anything useful in quantum field theory, you have to integrate over a huge amount of paths, and here $\mathcal{D}A\backslash mathcal\{D\}A$ is the "path" equivalent of the $dxdx$ of simple integrals.

An important consequence of this formalism is that at the quantum level, the fundamental interactions are all mediated by a specific class of particles called gauge bosons. For electromagnetism, the gauge bosons are the photons; for the weak nuclear force, the W and Z bosons; and for the strong nuclear force, the gluons.

But if we use the same Yang-Mills equation to describe the three fundamental forces, how are they different? What distinguishes them? Actually this is thanks to a special ingredient that is required by the Yang-Mills equation: an algebraic structure called a group; think about a group of matrices for instance. If as a group in the Yang-Mills equation you use U(1) (the group of complex numbers of unit modulus), you get electromagnetism. If instead you use SU(2), you get the weak nuclear force. And if you use SU(3), you get the strong nuclear force. Change the group, change the type of interaction described.

And more than that : if you use the combined group $U(1)\times SU(2)\times SU(3)U(1)\; \backslash times\; SU(2)\; \backslash times\; SU(3)$, then you get a unified model describing the three interactions at once! This is the key ingredient for what is called today the Standard Model of particle physics. Well, actually I'm oversimplifying, but this is good enough for now; the important thing to realize is that Yang-Mills theory is the key to the whole world of particle physics as we know it today. When we couple it to the equations describing matter and the Higgs boson, we essentially have everything there is to know about the fundamental laws of the universe, except for gravity. And the agreement with experiment is extraordinary: some predictions of the Standard Model have been verified to more than ten decimal places, and the LHC has confirmed every particle the theory predicted, including the Higgs boson.

But unfortunately, the Yang-Mills equations are very hard to work with in practice. One illustration of that is that even if we remove matter entirely — keeping just the pure Yang-Mills equations — there are many things we don't know how to solve. For electromagnetism, things are easy : the pure Yang-Mills equation in that case is trivial because the U(1) group is commutative, it's called an abelian group. The main consequence is that the gauge bosons of electromagnetism — the photons — can't interact, the equations become linear and can be reduced to Maxwell's equations of classical electromagnetism.

But electromagnetism is really just a special case : for any other non-abelian group — SU(2), SU(3), more generally SU(N) and even the exceptional Lie groups — things become tricky. Concretely that means that the classical equation is non-linear, and at the quantum level the gauge bosons associated can interact with themselves : one can decay into two, or two can collide and become one; and understanding this self-interaction is crucial to the theory.

But there is one good news: the tricky part does not depend on the choice of the group, whether its SU(3) like for the strong nuclear force, or any other non-abelian gauge group. There is a kind of factorization that happens so that the "group specific" parts separates from the rest, so we can just study an abstract Yang-Mills theory, and the results will be valid for all of them. (For expert readers, this is why in the paper, they talk about the color-ordered scattering of gluons : color-ordered means we just consider the kinematic part of the amplitude and we don't really care about the underlying non-abelian group; we can then use gluons as a generic name for any non-abelian gauge boson.)

So here we are: we want to understand the self-interaction of gluons in non-abelian Yang-Mills theory.

What physicists actually compute

Now, what do we mean by "understanding self-interactions of gluons"? In general, when physicists want to understand how a bunch of particles might interact, they consider a particular quantity called the scattering amplitude.

Imagine two particles being originally far from each other, moving in the same region, going past each other, interacting and parting ways. To know the probability of the final state (their direction and momentum) as a function of the initial state, physicists compute a complex number called the amplitude, whose squared magnitude gives the probability. There is an amplitude when particles are going past each other, but also when they collide or decay. In that case, the number of incoming and outgoing particles is not the same : we can have for instance one incoming particle decaying into two, or two particles colliding to form only one. Or more complicated situations.

In principle, this amplitude is directly dictated by the path integral formulation Yang-Mills theory, the big $ZZ$ that I wrote above. But in practice, because this is a monstrous mathematical object, it is almost impossible to compute directly.

The reason is that to express this amplitude, we somehow have to sum over all possible paths and intermediate processes the particles can undergo. That looks impossible, but the great physicist Richard Feynman found a way...

Feynman's insight was to turn the big integral into a power series. I already mentioned that in the Yang-Mills equation, the coupling constant, generally denoted $gg$, governs how strong the interaction is. And fortunately, it is generally small and we can use that to expand the big path integral into powers of $gg$. You can understand the logic of this expansion by considering a much simpler numerical case. Look at the following quantity $$Z=\int dx{e}^{-x^2-g{x}^4}\mathrm{.}Z\; =\; \backslash int\; dx\backslash ;\; e^\{-x^2\; -\; gx^4\}.$$ You can always expand the $e^{-g{x}^4}e^\{-gx^4\}$ part and get $$Z=\int dx{e}^{-x^2}(1-g{x}^4+\frac{1}{2}(g{x}^4{)}^2-\frac{1}{6}(g{x}^4{)}^3+\mathrm{.}\mathrm{.}\mathrm{.})Z\; =\; \backslash int\; dx\backslash ;\; e^\{-x^2\}\; \backslash left(1\; -\; gx^4\; +\; \backslash frac\{1\}\{2\}\; (gx^4)^2\; -\; \backslash frac\{1\}\{6\}\; (gx^4)^3\; +\; ...\backslash right)$$ ...and from this, you are free to keep only enough terms so that the approximation is good enough. And the nice thing is that now, you get a bunch of integrals that are much more tractable.

Again, I'm simplifying a lot ! But essentially it works the same way with path integrals. Except that each term of the expansion is very complicated, and is itself made of a lot of sub-terms.

To classify those sub-terms, Feynman realized you can use diagrams, now called Feynman diagrams: pictures showing one possible "story" of how particles might interact on their way from in to out. And the diagrams can be organized by complexity. Here you see two such diagrams, representing two possible ways a unique incoming gluon can decay into two outgoing gluons (gluons are conventionally represented by small spring-like lines)

In these diagrams, what matters is the overall topology of the graph, not the actual shape, size or orientation. For each, there is a corresponding integral that has to be computed, but a much simpler one than the big path integral we started from.

Importantly, Feynman diagrams can be classified by the number of closed loops they contain. What we call tree-level diagrams have no loops (like the left one above); they look like branching trees and represent the classical (non-quantum) limit of the theory. For the other diagrams, each loop adds a quantum correction by introducing a power of $\mathrm{\hslash }\backslash hbar$, the Planck's constant. Taking into account more loops leads to a better approximation in the expansion.

The good news is that in many situations, we don't have to go too far in the loop expansion. One or two loops are enough for most phenomena that we actually can measure in particle accelerators. However, the bad news is that even at tree-level, the number of diagrams to consider grows ferociously, essentially as the factorial of the number of particles involved. In the case of pure gluon interactions, with 6 particles you need to consider 220 diagrams, each diagram being a complicated algebraic expression you have to sum. As soon as you go past a handful of particles, it becomes really intractable.

But sometimes, magic happens...

The observation that launched a field

The first breakthrough came from Parke and Taylor in 1986. To state their result I need to introduce one more ingredient: helicity. You might know that light can have polarization. At the fundamental level, there are two circular polarizations, and a photon can be found either in the "clockwise" or "anticlockwise" polarization. The same logic applies to all massless gauge bosons, like gluons. Instead of polarization, we talk about helicity, and we count it either as positive (+), or negative (-).

So when we compute scattering amplitudes, we have to specify the helicity of the incoming and outgoing particles: how many are $++$ and how many are $--$. And because there are certain symmetries and conservation laws that must apply, some patterns always lead to a zero amplitude.

If all particles have the same helicity (all plus or all minus), the amplitude is automatically zero. If there is a single minus (and all the other ones plus), also zero...(or so everyone assumed for 40 years !)

The first clear nonzero pattern has exactly two $--$ and the rest $++$. As this is the most helicity "imbalanced" pattern you could imagine, this is called the MHV amplitude ("maximally helicity violating"). And for that specific configuration, Parke and Taylor showed that the very complicated sum of factorially-many tree-level diagrams dramatically collapses. No matter how many particles there are, it reduces to a one-line expression, instead of a sum of $n!n!$ complicated terms.

This almost miraculous simplification was considered as a hint that something deep might have been missed about those quantum field theories. Maybe the Feynman diagram expansion is hiding something deeper, that you can only realize by putting back together some of those terms for the quantum amplitudes.

That result ended up launching an entire subfield, as the surprises kept coming. Witten (2003) reformulated things in twistor space and explained why Parke-Taylor is so simple: the amplitude localizes on a curve in this abstract space. Britto, Cachazo, Feng, and Witten (2005) found a recursion formula that builds tree amplitudes from smaller pieces, no Feynman diagrams at all. And then there's the amplituhedron (Arkani-Hamed and Trnka, 2013), which is the most striking of the lot: it's a geometric object whose volume is the amplitude.

The pattern is always the same: Feynman diagrams work, but apparently they're hiding something simpler. Below you can see the citation count of the Parke-Taylor paper, clearly showing that something happened after Witten's paper in 2003!

The loophole nobody noticed

So I told you that the first case of dramatic simplification, the Parke-Taylor formula, was established for the "double minus" (or MHV) configuration, when only two particles have a minus helicity, and all the others have plus. I also mentioned that when all helicities are identical, the amplitude is automatically zero. And the same thing happens when only a single one is minus...except that for this "single minus" case, there was a flaw in the argument leading to this conclusion!

I won't detail the proof, but essentially it amounts to choosing a clever reference object in the calculation and showing it leads to vanishing amplitude. Except there are some very special conditions in which picking such a reference is not possible. But for that to happen, you have to go into a different spacetime!

You know we live in a world where there are 3 dimensions of space, and 1 dimension of time. This is called the (1,3) Minkowski spacetime. But there's a mathematically valid alternative called the (2,2) Klein spacetime: two time dimensions, two space dimensions. Of course, we don't live there, but you can study field theories in it just fine. And as we will see, it might have consequences for our world.

The key is that the (2,2) Klein spacetime opens up a kinematic regime that cannot exist in our world: "half-collinear" particles. If we impose this as the initial condition for our interacting gluons, the proof that the amplitude vanishes breaks down. It looks like a minor detail. Actually Edward Witten may have noticed this possibility in passing in 2003, but nobody followed up and tried to compute what this nonzero amplitude for the "single-minus" case might be. Until now.

What the paper computes

This is where the new paper enters the picture.

Long story short, it proves that in a particular regime, the amplitudes are indeed nonzero. But it does more than that: it finds a very simple formula that collapses all the tree-level Feynman diagram contributions. Something very similar to what Parke and Taylor did for the "double minus" case called MHV.

To arrive at these results, the researchers first use their brain and their pen to compute amplitudes for a small number of particles. To do this, they use a recursion relation from Berends and Giele (1988), which lets you build bigger amplitudes from smaller ones. The authors specialized the recursion to the half-collinear regime and solved it one by one for a small number $nn$ of particles. You can have a look at equations 29 to 32 in the paper: when $n=3n=3$, there is only 1 term; $n=4n=4$, 2 terms; $n=5n=5$, 8 terms; and finally $n=6n=6$: 32 terms. Here is the amplitude expression in that case

As expected, the growth is super-exponential. You can see that at $n=6n=6$ the expression fills a quarter page of dense algebra. We can understand why they didn't want to go to $n=7n=7$! And then, supposedly, entered GPT 5.2 Pro.

The first AI contribution

According to researcher Alex Lupsasca's thread on X, they engaged in back-and-forth work with GPT-5.2 Pro, which apparently helped identify a restricted kinematic region (called ${\mathcal{R}}_1\backslash mathcal\{R\}\_1$ in the paper), where things got simpler.

Remember we were already in a very special regime (the "half-collinear" one) in a weird spacetime. If on top of that, you assume that among your $nn$ particles contributing to your diagram, this is actually one decaying into $n-1n-1$ others, the expressions simplify drastically. You can see it by comparing equations (29)-(32) to their simplified versions (35)-(38). The change is drastic! Those 32 terms at $n=6n=6$ collapse to a product of 4 factors.

Note that this simplification is more involved than just "expand and reduce." All the quantities that are the building blocks of the formulas are interdependent. They are related by relations like momentum conservation in a non-obvious way. Think of it as recognizing that a messy polynomial factors, but only after the right substitutions.

Then comes what is to me the most interesting step. From the simplified expressions for n=3, 4, 5, and 6, GPT 5.2 Pro conjectured a general formula: the amplitude for $nn$ particles is a simple product of $n-2n-2$ factors:

$$A_{1\cdots n}{\mid }_{{\mathcal{R}}_1}=\frac{1}{2^{n-2}}\prod _{m=2}^{n-1}({\text{sg}}_{m,m+1}+{\text{sg}}_{1,2\cdots m})A\_\{1\backslash cdots\; n\}\backslash big|\_\{\backslash mathcal\{R\}\_1\}\; =\; \backslash frac\{1\}\{2^\{n-2\}\}\backslash prod\_\{m=2\}^\{n-1\}\backslash left(\backslash text\{sg\}\_\{m,m+1\}\; +\; \backslash text\{sg\}\_\{1,2\backslash cdots\; m\}\backslash right)$$

At first it looks simple, but actually it is not. And it's not clear how long an experienced physicist would have taken to come up with the same conjecture. Some people have said that AI will be truly considered intelligent not when it is able to prove new theorems, but when it can come up with interesting new conjectures...if the story is true, there we are!

First of all, the human researchers applied some sanity checks to this conjecture. It happens that amplitudes have to satisfy certain rules like the so-called Kleiss-Kuijf relations, or Weinberg's soft theorem. The conjecture proposed by GPT 5.2 Pro passed all those tests... so maybe it is true?

The proof

This is where apparently an internally scaffolded version of GPT-5.2 took off, and found the proof after about 12 hours of autonomous work (according to OpenAI press release, although per Alex Lupsasca's thread, it might have been a back and forth?)

What is beautiful is that the proof is fairly short, less than a single page in the paper, and goes in three steps. My knowledge of the field is not strong enough to appreciate how hard the proof might have been to come up with. It seems that the difficulty in it is not algebraic manipulations, but really coming up with the winning strategy, like those specific three steps. While steps 2 and 3 look like mostly applying relations and combining with previous equations and identities, the first step seems to be the most creative. The argument is a pigeonhole-type observation: among a sequence of "weighted averages" of the particle positions, at least one must land between consecutive values, and a step function kills that factor.

What to make of this

So here is the final result: we have a computation that in Feynman-diagram form involves super-factorially many terms, and in a particular setting it collapses into a closed-form product of $n-2n-2$ terms. Something like the Parke-Taylor formula for MHV, but for the "single minus" case this time.

Remember that the setting is narrow, though: (2,2) signature instead of physical spacetime; the half-collinear regime, a measure-zero slice of the full configuration space; and one specific decay kinematics (one particle decaying into n-1). So why should anyone care? Well, it might be the sign of something deeper for our physical world.

First of all, we are talking about tree-level amplitudes. But those amplitudes are the classical limit of the theory. So new nonzero tree amplitudes mean that there might be some previously overlooked classical solutions of the Yang-Mills equation.

Then from a mathematical perspective, it seems to mean something for the Self-dual Yang-Mills (SDYM), a simplified version of Yang-Mills where the field equals its own dual. SDYM has rich classical dynamics but its perturbative scattering amplitudes were thought to vanish beyond three particles. This is surprising, a theory with nontrivial classical solutions should scatter! So a nonzero single-minus amplitude might resolve that tension.

Finally there might be a practical angle too. A few decades ago, people realized that the (classical) theory of general relativity (which describes gravity) can be built as a "double copy" of Yang-Mills. It is not just a theoretical realization, now this link is a major tool for predicting gravitational waves from merging black holes, the kind of signals LIGO and Virgo detect. So if we have new Yang-Mills amplitudes, that might eventually mean something for calculations in general relativity?

An "unphysical" spacetime?

Now let's talk about what seems like the big limitation of this work. Why care about a spacetime we don't live in?

The reason is that tree-level amplitudes are rational functions of the kinematic variables (momentum, etc.). And different signatures for spacetime (like (2,2) in the paper or (1,3) in our real world) are different real slices of the same complex-analytic object. So even in this weird spacetime, we are deriving results that might have an impact on the real physical spacetime. (Though I should flag: this amplitude lives on a measure-zero locus, so the analytic continuation story is not straightforward and the implications for Minkowski physics are unclear.)

More to the point, the amplitudes community keeps finding that "unphysical" settings produce important results. Twistor theory is most natural in (2,2) signature, and that's where Parke-Taylor was first explained. BCFW recursion uses complex-shifted momenta. The amplituhedron lives in a space with no physical meaning. Instantons are solutions in Euclidean signature (four space dimensions, zero time). In each case, the apparently "unphysical" result turned out to matter.

Also, I have to mention I spent my PhD studying (1,2) and (0,3) quantum gravity, so who am I to judge!

The AI part, honestly

Since some readers are here for the AI angle, after all this, let's address this as honestly as possible.

First of all, the physics (going to the (2,2) Klein signature, the half-collinear regime, the loophole in the vanishing proof, the recursion, the connection to SDYM) is apparently all human work. That's probably the hardest part, and it comes from decades of expertise!

The conjecture, recognizing a pattern in the small $nn$ data, may not be the hardest step, but it is one that brings me joy. This is a beautiful use of AI, that goes beyond brute force symbolic manipulation, and shows the kind of creative breakthrough that comes out of it.

Once expressions are simplified in the right region, the product structure starts to show. The proof uses standard tools and a good amplitudes physicist could probably have found it in a few weeks. But the specific idea to show $V=0V=0$ first, the creative entry point it seems, was coming from the model.

But I have to say I would have appreciated more details on how AI was used: which scaffolding, the back and forth, etc.

As an optimistic note, let's end on the paper's last line: "We suspect that there are more interesting insights to come with our methodology and hope that this paper is a step on the road to a more complete understanding of the inner structure of scattering amplitudes."