The most interesting detail here isn’t the physics claim. It’s who wrote the proof. Physicist Nathaniel Craig says GPT-6 Astra didn’t just assist with a strategy he already had, it took the lead on the detailed proof. That is a much stronger claim than the usual “AI helped with my literature search.”
In the first guest post on its science blog, OpenAI published Craig’s account. He leads the particle theory group at UC Santa Barbara and is a visiting scientist at OpenAI. His October preprint proposes a candidate lattice definition of the full Standard Model.
The problem being attacked
The Standard Model matches experiment across eighteen orders of magnitude in energy. Yet after more than fifty years, it still has no nonperturbative definition. Perturbation theory works by switching interactions off, solving the simple case, and adding the interactions back as small corrections. That gives astonishing precision for the electron’s magnetism, accurate to a few parts in a trillion. It fails for something like the proton’s mass, which comes mostly from strongly interacting quarks and gluons. Turn off the strong force and the proton falls apart.
Lattice field theory, which puts spacetime on a grid, already handles the strong force well and has produced the proton mass from first principles. So why not the whole model?
Why chirality blocks the lattice
The weak force is chiral. It treats left-handed and right-handed fermions differently, like owning only a left-handed glove. Put a left-handed fermion on a lattice using the natural rules and you get an unwanted right-handed partner, a problem called doubling. That gives the wrong particle content for the weak force, and it is a main reason the definition has been missing.
Craig had a strategy to get around this but couldn’t prove it worked in every case required. Astra turned the strategy into a construction and developed the proof.
Does it matter?
Two caveats. This is a candidate definition in a preprint, and the physics community still has to check it. And the post is published by the vendor, so treat the framing with some care.
Still, the signal is worth tracking. Google DeepMind has pushed AI into mathematics and science, and Anthropic and others are courting researchers too. Most of those claims involve search, pattern finding, or known-problem benchmarks. A theorist crediting a model with the proof in an open foundational problem is a different kind of evidence. If the proof survives scrutiny, expect more physicists to work this way. If it doesn’t, we learn where the limits are. Either result is useful, and the verdict will come from referees, not blog posts.



