Models & Research

OpenAI launches a science blog, and its first guest is a particle physicist

3 min read AI-generated

Nathaniel Craig had a strategy for a problem that's been open for fifty years, but no proof. He says GPT-6 Astra took the lead in working it out.

Featured image for "OpenAI launches a science blog, and its first guest is a particle physicist"

OpenAI opened a dedicated science blog on Sunday. It’s meant as a place for researchers to show what they’re discovering with AI and how they actually do the work. Three kinds of posts are planned: new results, methods you can adapt (including where they fall short), and firsthand accounts from different fields.

What OpenAI points to so far

The launch post lists what its models have produced in recent months: a proof on the Navier–Stokes problem in September, then a batch of math results in October, among them a tighter error bound for estimating how many primes lie below a given number. From biology, OpenAI cites a project with PerturbAI and Scripps Research, where thousands of GPT-Rosalind agents analyzed data from millions of mouse brain cells in 32 minutes and proposed mechanisms to test in the lab. What the post leaves out: OpenAI had to pull three manuscripts from that October batch over a sign error. It does at least raise the open question of how machine-generated proofs get checked and absorbed into a field.

A lattice for the Standard Model

The first guest post comes from Nathaniel Craig, who leads the particle theory group at UC Santa Barbara and is currently a visiting scientist at OpenAI. His topic: the Standard Model matches experiments across eighteen orders of magnitude, yet it still lacks a full mathematical definition that doesn’t rely on approximation. Physicists usually calculate as if the interactions were switched off and then add them back as small corrections. That doesn’t work for something like the proton’s mass.

One way out is to treat spacetime as a grid of points. For the strong force, that approach has paid off over the past few decades. The full Standard Model is harder, because left-handed particles need the right phase correction on the lattice. Craig had an idea for building that correction up across ever finer grids, but couldn’t prove it worked in every case. In an October preprint he now proposes a definition. According to Craig, GPT-6 Astra worked out how to carry the force fields onto finer grids, showed the comparisons converge to a definite answer, and took the lead on the detailed proof. Three lattice theorists, Maarten Golterman, Andreas Kronfeld and Yigal Shamir, read the proof and pushed back with hard questions.

A candidate, and Craig says so

Craig is careful. He calls the result a candidate that “potentially” answers a question open for more than fifty years, and more an answer in principle than in practice. Experts still need to decide whether the construction has every property it should. He’s currently checking one of them himself, reflection positivity.

I like that tone, and it fits OpenAI’s stated plan to write about where methods fall short. A blog where researchers explain what the model contributed and who checked it is easier to follow than this fall’s big announcements. How much it’s worth will show the first time it reports a result that didn’t hold up.

Sources

OpenAISciencePhysicsGPT-6 AstraResearch