On November 15, OpenAI published a guest post on its newly launched science blog: Nathaniel Craig, a physicist at the University of California, Santa Barbara and a visiting scientist at OpenAI, announced that he and GPT-6 Astra have produced a candidate lattice definition of the Standard Model. In Craig's words, Astra did more than help with the arithmetic — it took the lead in developing the detailed proof. An AI model has, for the first time, led the proof work on a foundational problem in theoretical physics.
The Standard Model describes the known elementary particles and three of the four fundamental forces — the strong force, the weak force and electromagnetism — and its predictions match experiments across eighteen orders of magnitude. But for more than fifty years it has lacked a nonperturbative definition, one that does not treat interactions as small corrections. Perturbation theory can predict the electron's magnetism to a few parts in a trillion, yet it cannot compute problems dominated by the strong force, such as the mass of the proton. Cutting spacetime into a lattice and computing directly is the accepted way out, and lattice field theory has already derived the proton's mass from first principles.
The hard part: the weak force only likes the left hand
The difficulty is the weak force. The strong force and electromagnetism treat left- and right-handed fermions alike; the weak force acts only on left-handed ones. Put a left-handed fermion on the lattice in the most natural way and the rules conjure up an unwanted right-handed "mirror" particle, giving the wrong particle content. Overlap fermions solved the mirror problem, but the phase rule stayed open: Martin Lüscher worked out the phase prescription for a simpler class of chiral theories, while the full Standard Model's phase rule had been open for more than 25 years.
Craig's October preprint offers a new approach: carry the force fields from the target lattice onto a series of ever finer lattices, compare the allowed fermion states on neighbouring lattices, and add up the phase corrections piece by piece until the phase rule on the original lattice emerges. Astra suggested how to carry the fields across, showed that the comparisons converge to a definite answer, and led the detailed proof. Three lattice field theorists — Maarten Golterman, Andreas Kronfeld and Yigal Shamir — reviewed the proof and put probing questions to it.
A candidate answer, not a verdict
Craig himself calls the result more an answer in principle than an answer in practice: it is only a candidate definition, and whether it can be used in realistic simulations or satisfies properties such as reflection positivity is still being checked; the definition of the Standard Model in continuous spacetime remains unsolved. For the AI industry, the real news is not the physics result itself but Astra's role — from writing code and solving maths problems to leading a full theoretical proof, general-purpose models are moving into territory that takes human experts years to advance. Astra has already made headlines before, including stopping a suspicious task during a safety evaluation and sharply improving safe-response rates in cyberattack scenarios. The blog post is a signal in itself: OpenAI has opened a dedicated science blog, and scientific discovery is becoming a stage for showing off model capability.