OpenAI says GPT‑5.2 has independently discovered and then formally proved a new formula in theoretical physics, overturning what many particle physicists assumed was a settled result. The work is now on arXiv as a co‑authored preprint, with human physicists from Harvard, Cambridge, Vanderbilt, and the Institute for Advanced Study in Princeton validating the math.
What did GPT‑5.2 actually discover?
The paper looks at “single‑minus” gluon scattering amplitudes, a type of particle interaction most experts believed had to vanish at tree level under normal assumptions. GPT‑5.2 helped show that in a specific slice of momentum space called the half‑collinear regime, that amplitude is not zero after all, meaning a supposedly forbidden interaction can occur under the right conditions.
Human researchers first cranked through the hard part by hand, deriving extremely messy expressions for these amplitudes for small numbers of gluons, up to six particles. GPT‑5.2 Pro then spotted a pattern in those results, conjectured a simple general formula (their Equation 39), and an internal scaffolded version of GPT‑5.2 spent about 12 hours producing a full formal proof.
According to OpenAI’s write‑up, the proof was checked using standard tools in quantum field theory, including the Berends-Giele recursion and soft‑limit tests, and then independently verified by the human co‑authors. OpenAI product lead Kevin Weil is listed as a co‑author “on behalf of OpenAI,” while Harvard’s Andrew Strominger has praised the result and noted that the AI followed a line of attack no human in the group would likely have tried.
So what’s the big deal with all these?
For me, the bigger story isn’t just that an LLM helped clean up some equations; it’s that it generated a fresh conjecture and then proved it in a frontier area of physics, not a toy benchmark. We’ve already seen AI move into safety and moderation roles, like Meta’s planned “adult classifier” for spotting underage Instagram users that I covered earlier on this site, but this is AI stepping directly into the role of scientific collaborator.
Skeptics will (and should) keep asking whether these are truly new ideas or just patternmatching on steroids, especially when the training data almost certainly included decades of physics literature. But when a model helps overturn a textbook assumption, forces experts to rewrite parts of their theory, and earns its place in the author list on a serious preprint, dismissing it as mere autocomplete starts to feel intellectually lazy.
If you want to dive into the technical details, OpenAI’s announcement post is a good high‑level starting point, and the full paper titled, “Single‑minus gluon tree amplitudes are nonzero,” is available on arXiv for anyone willing to wrestle with the math. For me, the future AI looks to be getting interesting.
