OpenAI Gluon Amplitude Result: Formula, Checks and AI Role

A textbook zero turns out not to be zero

The paper, "Single-minus gluon tree amplitudes are nonzero," reconsiders a class of tree-level scattering amplitudes describing one gluon with minus helicity decaying into a group of gluons with plus helicity. These amplitudes had long been presumed to vanish identically. [1]

The authors show the amplitudes are in fact nonzero for certain 'half-collinear' momentum configurations that exist in Klein space or for complexified momenta, and derive a closed-form, piecewise-constant expression for the decay as a function of the gluons' momenta. The formula satisfies established consistency checks, including Weinberg's soft theorem. [1]

Atlas interpretation: Klein space and complexified momenta are not the kinematic regime of an ordinary particle collision. The result corrects a textbook assumption about the mathematical structure of gluon scattering rather than predicting anything a collider would measure directly, which is why the paper reads as a formal amplitudes result rather than a phenomenology paper, despite its cross-listing in hep-ph. [1]

How the model was used

According to Phys.org, the team tried the problem with a chat model in the spring without success, then returned to it with an internal OpenAI system referred to as 'Super Chat,' which proposed a solution and then spent about 12 hours verifying it. The physicists spent roughly a week afterward checking the calculation by hand and preparing it for publication. [2]

Andrew Strominger of Harvard, who led the effort, is quoted saying there was a moment when he felt he was working with a creative person rather than a machine. Alex Lupsasca, one of the paper's authors and, per Phys.org, OpenAI's first hire for its OpenAI for Science effort, called the result the first significant discovery in theoretical physics made by an AI. [2]

Atlas interpretation: "First real discovery" is Lupsasca's characterization, made by a co-author who also works at the company that built the model. Phys.org reports it as a quote, not as an independently adjudicated fact, and this page does the same: the claim is attributed, not endorsed. [2]

Who signed the paper, and where they sit

The arXiv listing names five human authors: Alfredo Guevara, Alexandru Lupsasca, David Skinner, Andrew Strominger, and Kevin Weil. Phys.org's account names and quotes four of the five, Strominger, Lupsasca, Skinner, and Guevara, as the physicists involved and does not mention Weil. [1][2]

Lupsasca is described by Phys.org as a research scientist at OpenAI and a former Harvard graduate student, which is the basis for this event's organization tag. [2]

Atlas interpretation: The summary above describes four physicists and a chat model, which tracks the Phys.org framing rather than the arXiv author list. This page notes the fifth credited human author because a reader checking the primary source would otherwise find a mismatch with no explanation. [1]

Which model, which release

Phys.org's account distinguishes between the public ChatGPT-5.2 proand an internal system called Super Chat that OpenAI made available to the team. The event's title credits GPT-5.2 specifically, but the tool that produced and verified the solution, per the reporting, was the internal variant rather than the public release. [2]