At a recent OpenAI Forum, Kevin Weil (OpenAI for Science) spoke with Brian Spears, who leads AI initiatives at Lawrence Livermore National Laboratory and helps direct the Department of Energy’s Genesis Mission. Their focus was how frontier reasoning models took LLMs beyond chart partners to become usable instruments for working scientists.
Spears described his “o1 moment” from a year ago. He prompted a model with a simple but extreme setup: a one‑megabar shock (about a million times atmospheric pressure) slamming into stainless steel. The model moved through a textbook explanation toward the working language of simulation: writing down the conservation laws—mass, momentum, energy—in the partial differential equations engineers use, in a form designed to translate those quantities accurately to computer code. Then it unexpectedly referenced LLNL’s export‑controlled Hydra radiation‑hydrodynamics system, a system that normally takes years to learn. The model still couldn’t run Hydra in that setting, but the leap showed in the reasoning: it could navigate from physical scenario to the equations and tools that real labs use.
Spears then jumped to the current state of play, where the AI loop can include execution. In one project, a long back‑and‑forth thread about noisy plots, repeated adjustments, and “blowups” where the math became numerically unstable finally converged on an “optimal ridge” in a simplified “burn‑wave model.” In plainer terms: the OpenAI model found the shape of a density transition to move heat from already‑hot fuel into colder fuel, the sort of starting point that can make the lab’s expensive high‑performance computing runs far more targeted, saving money and time. What once took months of trial and error can now begin with a better first guess.
That matters at national labs, where OpenAI has deployed a frontier model on Venado, a classified supercomputer. Spears described agent‑style workflows that can set up simulations, plan parameter sweeps across thousands of runs, write post‑processing scripts, and return results in a form that a scientist can interrogate, all while keeping humans responsible for high‑consequence judgment and verification. Models can speak the language of physics smoothly, he noted, but they often can’t tell when they’ve produced something meaningful until a person course-corrects them.
In a wider discussion of fusion and policy, Spears reflected on the National Ignition Facility’s ignition milestone on December 4, 2022, and framed Genesis Mission as a bet on an AI “platform” that wraps decades of capability—supercomputing, experiments, manufacturing—into tighter cycles of hypothesis and test, with explicit national‑security stakes. Weil tied these near‑term gains to OpenAI for Science’s broader aim, which is to give researchers tools that compress the time between question, evidence, and insight.
We are grateful to Kevin, Brian, and the OpenAI Forum community for showcasing what human-led AI acceleration can look like in science!
