OpenAI has published a candidate answer to a question that has stood open in physics for more than half a century: how to define the Standard Model of particle physics on a lattice. The work, written by physicist Nathaniel Craig with the help of OpenAI's newest model, GPT-6 Astra, appeared on November 15 as the first guest post on OpenAI's new science blog.

The Standard Model is one of science's great successes, describing the known elementary particles and three of the four fundamental forces. Yet it still lacks a rigorous definition that works without treating interactions as small corrections — the missing nonperturbative definition.

A lattice — spacetime treated as a grid of points — is the most promising way to close that gap, because a finite grid can be put on a computer and studied exactly. The obstacle is the weak force: it treats left-handed and right-handed particles differently, and that asymmetry has defeated attempts to place the full Standard Model on a lattice for decades.

A grid, finer grids and a phase rule

Craig, who leads the particle theory group at the University of California, Santa Barbara and is a visiting scientist at OpenAI, had a strategy but could not prove it worked in every case. His idea: take the force fields on the grid and represent them on a succession of ever finer grids, comparing the allowed fermion states on each pair of neighbouring grids. Adding up those comparisons gives the phase rule — the correction that keeps left-handed fermions consistent as the fields change.

That is where Astra came in. According to Craig, the model suggested the right way to carry force fields onto finer grids, showed that the comparisons would converge to a definite answer, and took the lead in developing the detailed proof. The proof shows the correction is right around every allowed loop, supplying a measure for the Standard Model at a finite lattice spacing, so long as field strengths stay bounded.

Human experts played a part too: lattice field theorists Maarten Golterman, Andreas Kronfeld and Yigal Shamir reviewed Astra's proof and asked probing questions. The underlying preprint, with Craig as sole author, was submitted in October and runs to 123 pages; Astra also carried out tests of the method, including formal verification in Lean.

A beginning, not an end

Craig is careful about what the result is — and what it is not. He calls it a candidate definition and "much more a beginning than an end": other physicists still have to decide whether the construction has all the properties they want, and he is still checking whether it satisfies a consistency condition known as reflection positivity. It is, in his words, far more an answer in principle than an answer in practice, and it cannot yet be used for realistic simulations.

Some independent physicists are more cautious still. Columbia mathematician Peter Woit described it as a modest advance over earlier work by Martin Lüscher, and a long way from a full definition in the continuous spacetime limit. OpenAI's recent proof that Yang-Mills theory exists may help, Craig notes, but the continuum limit remains unfinished business.

AI as a research collaborator

Whatever the verdict, the post is the latest sign of AI models moving from assistants to collaborators in fundamental research — just last week, physicist Edward Witten credited Anthropic's Claude Opus 5.5 in a paper on lattice gauge theory. The test now is whether Craig's construction survives scrutiny from the wider lattice community.