Ex-OpenAI's Periodic Labs beats GPT-6 Astra on materials science with just 1,300 H200s
ying11231 · x · 2026-09-16
Liam Fedus and the Periodic Labs team built high-throughput materials labs in Menlo Park creating a loop between experiments and models. Using only 1,300 H200s plus months of proprietary experimental data, they mid-trained and RL'd an open-source model — Neon — surpassing GPT-6 Astra on their materials science benchmark. Focus areas include superconductors, magnets, and semiconductor materials, with SGLang and Miles powering inference and training. Key takeaway: fresh experimental data plus modest compute can beat frontier general models on specialized benchmarks.
More from AGI Musings
- Alpha School investigation: 17 campuses on probation, Bill Ackman as shadow lead, Holmes involved — benjaminjriley · 2026-11-02
- Google's chief economist: AI will transform more jobs than it kills — gains come from redesigning workflows — danielrock · 2026-11-02
- Daniel Lemire: AI is exploding demand for boring compute, server CPUs to double — lemire · 2026-11-02
- Alignment researcher: the word "aligned" implies consistencies that don't exist — davidmanheim · 2026-11-02
- By 2035, Indian Parents Will Ask How Many Agents You Manage, Not Your Salary — theteknosaur · 2026-11-02
- Under 5% of US consumers pay for AI, yet compute demand is already forcing usage cuts — NinaDSchick · 2026-11-02