OpenAI's GPT‑5.2 derives physics formulas
- OpenAI said on February 13 that GPT‑5.2 Pro helped conjecture a new particle-physics formula in a preprint with researchers from IAS, Vanderbilt, Cambridge, and Harvard. (openai.com) - The result targets “single-minus” gluon tree amplitudes long treated as zero; the paper says they are nonzero in a special half-collinear regime. (openai.com) - That matters because it pushes frontier AI past summarizing science into proposing formal structures experts can then prove and extend. (openai.com)
The object here is a scattering amplitude — the math physicists use to predict what happens when particles hit each other. That sounds narrow, but it sits right in the middle of quantum field theory, where tiny simplifications can unlock whole new ways of calculating. (openai.com) The gap was a familiar one in theory work: one class of gluon interactions had mostly been written off as zero, so people stopped looking very hard. Then OpenAI published a February 13 preprint saying GPT‑5.2 Pro had guessed the closed-form formula for exactly that “missing” case, and the human team then proved it. ### What did the model actually do? (openai.com) Not “solve physics” in the vague chatbot sense. The paper says GPT‑5.2 Pro conjectured the final formula — specifically Eq. 39 — after the authors had worked out low‑n cases by hand and run into ugly expressions whose complexity grows superexponentially with the number of particles. In plain English, the humans generated hard examples, the model spotted the pattern, and the researchers checked that the pattern was real. ### What are gluon amplitudes? Gluons are the particles that carry the strong force — the force that binds quarks inside protons and neutrons. A scattering amplitude is the compact mathematical object that tells you the probability of a given interaction. (openai.com) Physicists care because these amplitudes often turn out to be much simpler than the brute-force Feynman-diagram calculation suggests, and those simplifications usually point to deeper structure in the theory. ### What was supposed to be impossible? The paper revisits “single-minus” tree amplitudes — cases where one gluon has minus helicity and the other \(n-1\) gluons have plus helicity. (openai.com) Textbook reasoning says these amplitudes vanish at tree level for generic momenta, so they were mostly treated as absent. The new result says that conclusion was too broad: in a precisely defined half-collinear slice of momentum space, the amplitude does not vanish. ### Why does that matter? Because “zero” and “not zero” are different worlds in physics. If something truly vanishes, you can ignore that whole channel. If it survives in a special regime, then there is hidden structure people have been missing. (openai.com) The paper gives a closed-form expression for that regime and says it passes nontrivial checks, including Weinberg’s soft theorem. ### What is this half-collinear regime? Basically, it is a special alignment of particle momenta where the standard vanishing argument breaks. This is not the generic real-world kinematic setup people usually assume. It lives in a mathematically well-defined corner — described in the paper using Klein-space or complexified momenta — where the usual shortcut no longer applies. (openai.com) That is the catch and also the whole point: the old statement was true generically, but not universally. ### Did the AI prove the theorem too? OpenAI’s writeup draws a line here. GPT‑5.2 Pro proposed the formula. The formal proof came afterward, with verification by the authors and what OpenAI describes as an internal OpenAI model helping on the proof side. (arxiv.org) So this is not “the chatbot replaced physicists.” It is closer to a conjecture engine that can notice structure in messy symbolic output faster than a person can. ### Is this already spreading beyond gluons? Yes — at least as a research direction. OpenAI flagged graviton extensions in its writeup, and a follow-on arXiv paper in March pushed the same nonzero single-minus idea into gravity. (openai.com) That does not mean the whole field has been rewritten overnight, but it does mean this was not a one-off curiosity with nowhere to go. ### So what changed here, really? The big shift is not that AI can do algebra. People already knew that. The shift is that a frontier model appears to have proposed a non-obvious, publishable structure in a live area of theoretical physics — one that experts then turned into a proof. (openai.com) That is a different category of use. It moves the model from assistant to hypothesis generator. ### Bottom line This is still a preprint, not a settled rewrite of physics. But the important part already landed: GPT‑5.2 was used to surface a real mathematical conjecture in a hard domain, and humans could not just admire it — they could verify it. (openai.com) That is the threshold people have been waiting for.
