Our recent preprint on gluon amplitudes has sparked a lot of discussion, so I want to share the backstory — including how AI helped crack a problem that had stumped us for a year. I'll also be giving a public lecture at Harvard this week. Details at the end.

About a year ago, world experts on this problem, Alfredo Guevara (IAS), David Skinner (Cambridge) and Andy Strominger (Harvard) realized something surprising: single-minus amplitudes should not be identically zero, despite the arguments in some textbooks to the contrary (which have a loophole when the interacting particles are collinear). The question became: what should these amplitudes actually be?

Last October, I joined the newly formed OpenAI for Science division, led by @kevinweil & @markchen90, with the goal of improving the scientific capabilities of our frontier models. I was so excited about what the latest internal models could do for physics that I invited my coauthor and PhD advisor Andy to work together with them on a problem so that he could see for himself.

Alfredo, David, and Andy had spent the past year trying to find a simple formula for the nonvanishing single-minus amplitudes, analogous to the Parke-Taylor formula obtained in the 80s for the double-minus ("MHV") amplitudes. Alfredo had obtained a complicated expression, Eq. 21 in the preprint, but it was unwieldy: a sum over Feynman diagrams whose complexity grows superexponentially in the number of interacting particles.

Using this formula, it was hard to compute the amplitudes up to n=6 (as shown in Eqs. 29--32), but we thought that a much simpler expression should exist, just like the Parke-Taylor formula had simplified the horrendous Feynman-diagram computations for MHV amplitudes (this is well explained in Part N.2 of Zee's QFT book). But such a formula was still proving elusive.

We engaged in a lot of back-and-forth with GPT-5.2 Pro, which helped us identify the kinematic region R_1 in which we should search for a simplified formula. It also found surprising nontrivial simplifications for the amplitudes that we had, shown in Eqs. 35--38, which led it to guess Eq. 39 for the general pattern. With this target in sight, we formulated a sharp question for an internal model to tackle, and it independently came up with the simple Eq. 39 and then proved it.

The part of the preprint following Eq. 39 is essentially the proof provided by the model and verified by us. This was truly a collaborative effort between humans and AI, with GPT-5.2 and its scaffolding contributing at the level of a very talented contributor. As Andy said, the final result it provided had eluded the team for a year and may not have been found for a while longer. To my mind, we have crossed a threshold for AI in physics.

I think that this year is going to be an inflection point for science, and that AI will do to physics in 2026 what it did to coding in 2025. I will be talking about these recent results and future directions at Harvard on Tuesday at 2pm in Science Center hall A.