Liam Fedus's Lab Builds Neon With Just 1,300 H200s, Beats GPT-6 Astra on Materials Benchmark
LiamFedus · x · 2026-09-16
Former OpenAI VP of research Liam Fedus detailed his new lab's loop: high-throughput materials labs in Menlo Park generate fresh experimental data, models learn from it, then steer what to try next.
- Using only 1,300 H200s plus months of proprietary experimental data, the team mid-trained and RL'd an open-source model to surpass GPT-6 Astra on their analysis benchmark—dubbed Neon
- First targets: hard problems in materials science including superconductors, magnets, and semiconductor materials
- Framed as evidence AI is collapsing the barrier to hardware creation: 'a hardware engineer in their pocket'
- A landmark AI-for-Science case: vertical data + modest compute beating a general frontier model
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