GPT-6 Astra Took the Lead on a 50-Year Physics Proof. Three Humans Checked It
The Standard Model has matched experiments across eighteen orders of magnitude for more than fifty years, and nobody can give the full theory a rigorous nonperturbative definition. OpenAI's new science blog says GPT-6 Astra helped produce a candidate one. Post: https://openai.com/index/step-toward-defining-the-standard-model/ , preprint from October: https://arxiv.org/abs/2610.09093
The problem in plain words: put the weak force on a lattice, a grid standing in for spacetime, and the natural rules hand you a right-handed partner particle you never asked for. Physicist Nathaniel Craig had a strategy, representing the fields on finer and finer grids and summing the phase corrections, but could not prove it worked in every case it needed to cover. Per the post, Astra suggested how to carry the force fields onto finer grids, showed the comparisons would converge, and took the lead in writing the detailed proof. Lattice theorists Maarten Golterman, Andreas Kronfeld and Yigal Shamir went over the proof and asked hard questions.
Read the hedges, because Craig wrote them in. It is a candidate. Scientific dialogue is still required. It is "far more an answer in principle than an answer in practice," with open questions like reflection positivity at finite lattice spacing. The post does not say whether any of it was formalized. So the checker here is three named experts, which is how physics has always worked, and also a much slower verifier than a compiler.
That is why Kevin Buzzard's post from Nov 12, which reached the HN front page in this window, is the useful companion read: https://xenaproject.wordpress.com/2026/11/12/and-the-proof-is-checked-in-lean/ . Buzzard notes OpenAI's 722-paper theorem dump came about 45% with Lean code, and three of the papers without Lean have since been retracted over a sign error. The same day OpenAI also published a study of 1,819 scientists in which users estimated work took 1.9 times longer without AI. The speed is real. The bottleneck is moving from finding the proof to deciding whether to believe it.
← Back to all articles
The problem in plain words: put the weak force on a lattice, a grid standing in for spacetime, and the natural rules hand you a right-handed partner particle you never asked for. Physicist Nathaniel Craig had a strategy, representing the fields on finer and finer grids and summing the phase corrections, but could not prove it worked in every case it needed to cover. Per the post, Astra suggested how to carry the force fields onto finer grids, showed the comparisons would converge, and took the lead in writing the detailed proof. Lattice theorists Maarten Golterman, Andreas Kronfeld and Yigal Shamir went over the proof and asked hard questions.
Read the hedges, because Craig wrote them in. It is a candidate. Scientific dialogue is still required. It is "far more an answer in principle than an answer in practice," with open questions like reflection positivity at finite lattice spacing. The post does not say whether any of it was formalized. So the checker here is three named experts, which is how physics has always worked, and also a much slower verifier than a compiler.
That is why Kevin Buzzard's post from Nov 12, which reached the HN front page in this window, is the useful companion read: https://xenaproject.wordpress.com/2026/11/12/and-the-proof-is-checked-in-lean/ . Buzzard notes OpenAI's 722-paper theorem dump came about 45% with Lean code, and three of the papers without Lean have since been retracted over a sign error. The same day OpenAI also published a study of 1,819 scientists in which users estimated work took 1.9 times longer without AI. The speed is real. The bottleneck is moving from finding the proof to deciding whether to believe it.
Comments