Key takeaways
- Physicist Nathaniel Craig described a candidate lattice definition of the Standard Model on OpenAI's science blog on November 15, 2026, based on his October arXiv preprint.
- Craig devised the strategy of comparing fermion states across finer and finer grids, and he says GPT-6 Astra turned it into a construction and led the detailed proof.
- Lattice theorists Maarten Golterman, Andreas Kronfeld and Yigal Shamir reviewed the proof, but practical simulations, reflection positivity and the continuum theory remain open.
Physicist Nathaniel Craig says he and OpenAI's GPT-6 Astra have found a candidate lattice Standard Model: a way to define the full theory of particle physics on a grid of points that stands in for spacetime. He described the work on November 15, 2026, in the first guest post on OpenAI's new science blog. The math is in his October preprint, "A gauge-invariant measure for lattice chiral fermions". It targets a problem lattice theorists have worked on for more than 25 years.
This is a careful claim. Craig calls the result a candidate, "far more an answer in principle than an answer in practice." What makes it news outside physics is the split of labor. By Craig's account, the AI did most of the proof.
Why the Standard Model has no lattice definition yet
The Standard Model is the theory of every known particle and of three of the four forces. Physicists mostly use it through perturbation theory: start from free particles and add the interactions as small corrections. That works for a lot of quantities and fails for others. Craig's example is the proton's mass, which comes mostly from the strong force's own dynamics, where the corrections are not small.
The usual fix is lattice field theory. You chop spacetime into a grid, put the fields on the points and links, and compute. For the strong force this has worked for decades. For the whole Standard Model it hasn't, so the theory still has no complete nonperturbative definition, meaning one that doesn't lean on the small-corrections trick.
The problem: chirality and fermion doubling
The obstacle is chirality. The weak force treats left-handed and right-handed fermions, the particles that make up matter, differently. Only the left-handed ones feel it.
Put fermions on a grid the obvious way and you get doublers: extra particles with the wrong handedness. In Craig's words, "the most natural rules give you a right-handed partner you never asked for." His analogy is a box with ordinary mirrors. Shine left-circular light in and it bounces back right-circular, so both kinds end up inside.
Earlier work solved most of this. Ginsparg and Wilson described a lattice version of chiral symmetry in 1982. Narayanan and Neuberger introduced overlap fermions, which keep one handedness without doublers. Lüscher showed how to tell the handednesses apart and solved the full problem for simpler, abelian forces like electromagnetism. Kadoh and Kikukawa applied these ideas to the Standard Model.
One piece was still missing: the fermion measure. When the theory sums over all possible fermion states, each lattice field setup needs a phase, a kind of reference angle. For the Standard Model's forces, nobody had shown how to choose those phases so they stay consistent as the fields change. Get it wrong and the theory loses gauge invariance, the symmetry the forces are built on.
What Craig proposed
Craig's idea was to put the same force fields on a sequence of finer and finer grids and compare the allowed fermion states on each pair of neighboring grids. Each comparison yields a small phase correction. Add them all up and you get a correction on the original, coarse grid.
The paper proves the sum converges, and that the resulting phase comes out consistent around every closed loop of field changes, which is what gauge invariance requires. The construction starts with SU(5) and the fermions of one Standard Model generation, then extends to the Standard Model's own gauge groups.
It comes with conditions. The proof holds only on "admissible" fields, where the field strength on each smallest square of the grid stays under a fixed bound. The theory has to be free of gauge anomalies, a standard requirement, plus an extra condition on certain fermion representations. The paper also leaves one constant phase per disconnected piece of field space unfixed.
What the AI did and what Craig did
Craig's post gives a clear split:
- Craig came up with the strategy of comparing grids at successive spacings.
- GPT-6 Astra turned it into a working construction, suggested how to carry force fields onto finer grids, showed the comparisons converge, and led the detailed proof. Craig writes that Astra "took the lead in developing the detailed proof."
- Tests that Craig and colleagues proposed, including a formal check in the Lean proof language, were carried out by Astra, according to the paper's acknowledgments.
The acknowledgments add that the manuscript mixes sections Craig wrote with heavy rewrites of the proof Astra composed. The arXiv listing names Craig as the only author.
GPT-6 Astra is the OpenAI model covered here when it launched. OpenAI has not said how the model was set up for this project.
Who checked it
Three lattice theorists, Maarten Golterman, Andreas Kronfeld and Yigal Shamir, reviewed Astra's proof and asked probing questions, Craig writes. The paper thanks all three. The preprint itself has one version on arXiv, posted October 6, and lists no journal publication.
The paper reports three kinds of checks. The SU(5) construction and its Standard Model restriction are formalized in sixteen Lean 4 modules, which the paper says compile without errors and use only standard axioms. A two-dimensional version of the construction is worked through on explicit examples. And a second, independent construction agrees with the main one up to the small local differences expected between equivalent definitions. The paper is explicit that Lean checks the formal statements under their stated assumptions, not the physics.
What is still open
Plenty, by Craig's own list:
- Whether it has every property physicists want. "Scientific dialogue among experts is required," he writes.
- Practice. Running realistic Standard Model simulations this way would take substantial further work.
- Reflection positivity, the property that ties a lattice model to a quantum theory whose probabilities add up. Craig says he is studying it now.
- The continuum. The construction works at a finite grid spacing. Defining the interacting theory in continuous spacetime is still open. The post suggests OpenAI's recent work on Yang-Mills theory might supply that step.
The next thing to watch is the lattice community's response: follow-up papers, independent checks of the Lean files and the companion proofs on Zenodo, and whether the preprint passes peer review at a journal.
Frequently asked questions
What did OpenAI do with the Standard Model?
OpenAI's GPT-6 Astra worked with physicist Nathaniel Craig on a candidate lattice definition of the Standard Model. Craig devised the strategy and the model led the detailed proof, which is in his October 2026 arXiv preprint.
What is the fermion doubling problem?
When fermions are put on a lattice in the obvious way, extra particles with the wrong handedness appear. That breaks the Standard Model's weak force, which acts only on left-handed fermions.
Who is Nathaniel Craig?
Nathaniel Craig leads the particle theory group at UC Santa Barbara, is a permanent member of its Kavli Institute for Theoretical Physics, and was a visiting scientist at OpenAI.
Has the lattice Standard Model result been peer reviewed?
Three lattice theorists reviewed the proof before release, according to Craig. The preprint has one version on arXiv, posted October 6, 2026, and lists no journal publication.
Sources
- A step toward defining the Standard Model on a latticeOpenAI
- A gauge-invariant measure for lattice chiral fermions (arXiv:2610.09093)arXiv
- Our new blog for scienceOpenAI
- P. H. Ginsparg and K. G. Wilson, Phys. Rev. D 25, 2649 (1982)Physical Review D
- R. Narayanan and H. Neuberger, Nucl. Phys. B 412, 574 (1994)arXiv
- M. Lüscher, Nucl. Phys. B 549, 295 (1999)arXiv
- D. Kadoh and Y. Kikukawa, JHEP 05 (2008) 095arXiv
