OpenAI used GPT-5.2 Pro to extend a gluon amplitude result into quantum gravity

What the graviton result actually claims

The preprint, "Single-minus graviton tree amplitudes are nonzero," targets a specific configuration: one particle with negative helicity and the rest positive. Standard textbook arguments say such amplitudes vanish at tree level, where only the most direct interaction diagrams matter and quantum loop effects are ignored.

The authors' claim is narrow and precise. That vanishing conclusion depends on assuming generic particle motion. When momenta satisfy a special alignment — the half-collinear regime — the usual argument breaks. In that regime the amplitudes do not vanish; they exist as well-defined mathematical distributions supported on a restricted region of momentum space. The paper derives explicit formulas and ties them to an infinite-dimensional "w-(1+∞)" symmetry that Penrose identified in classical gravity a half century ago.

This is not a claim to have quantized gravity. The announcement frames it as "a small step" toward reconciling quantum mechanics with general relativity — a concrete extension of a recent gluon result into the gravitational setting, where the two theories share structural features even though the underlying forces differ.

The gluon paper as anchor, and the model as author

The methodology is the part worth reading closely. A prior gluon result had already shown that a neglected helicity configuration could produce nonzero amplitudes under special conditions. That completed gluon paper was handed to GPT-5.2 Pro as context, and the model was asked to construct the corresponding gravity amplitudes — an extension the announcement says would have taken human authors considerable time.

GPT‑5.2 Pro not only solved this problem using a beautiful and surprising technique (the directed matrix-tree theorem), it also produced an excellent preliminary draft of the paper.Montana Labs

OpenAI published a transcript of that initial exchange. The specificity here matters: the model didn't just gesture at an approach, it reached for the directed matrix-tree theorem, and a further round of interaction connected the amplitudes to the Penrose symmetry. The author list mixes OpenAI staff (Lupsasca, Weil) with physicists at the Institute for Advanced Study, Vanderbilt, Cambridge, and Harvard.

Verification became the dominant cost

The most transferable observation in the announcement is about where the labor went. The final formulas were verified analytically and checked for consistency with known physical limits using standard methods. Between the earlier gluon result and this one, the authors report that most elapsed time was spent confirming derivations, checking consistency, and preparing formal write-ups — not generating the initial conjectures.

That inverts the usual research economics. When the conjecture-generation step compresses, the bottleneck shifts to proof, cross-checking, and exposition. The announcement calls this "a significant shift, with verification and exposition representing the dominant share of effort." It's a claim about workflow, and it's testable against the published transcript and preprint.

The implication: transfer between neighboring theories, checked by hand

The specific lesson of this preprint is that supplying a completed result in one theory — gluons — as an anchor let the model explore a structurally related theory — gravity — and land on a construction that was then proven by conventional analytic methods. The transfer worked precisely because the two settings share features, and because humans retained the verification loop.

For teams building AI-assisted reasoning tools, the pattern to note is not "AI did physics" but the division of labor: a strong prior result as context, a model that proposes a nonobvious technique and a draft, and a human verification standard that remains unchanged. OpenAI frames the ongoing goal as understanding how AI can participate in theoretical research "while maintaining conventional standards of mathematical verification and scientific rigor" — an admission that the value depends entirely on that second half holding.