Ex-OpenAI researcher's lab uses 1,300 H200s to build Neon, beating GPT-6 on its benchmark

LiamFedus · x · 2026-09-16

Liam Fedus announced high-throughput materials labs in Menlo Park creating a loop between experiments and models: labs generate fresh data, models learn from it, then help decide what to try next. Using only 1,300 H200s plus months of experimental data, they mid-trained and RL'd an open-source model, Neon, that surpasses GPT-6 Astra on their analysis benchmark. Focus: superconductors, magnets, and semiconductor materials.

Quoted context from xiangfuml: simulated RL environments were a useful start but hit a sim2real gap; coding agents advanced fast because AI researchers are programmers themselves, while in science the task context lives with scientists—useful capabilities require many iterations between the two.

Related event: Periodic Labs Trains Materials Model Beating GPT-6 with Just 1,300 H200s(31 posts)→

Original post →

More from Companies & People

Companies & People channel →