Periodic Labs beats GPT-6 Astra on X-ray analysis with Neon, trained on just 1,300 H200s
hsu_byron · x · 2026-09-16
Periodic Labs (founded by ex-OpenAI's Liam Fedus) unveiled Neon, its first model: it built high-throughput materials labs in Menlo Park, mid-trained and RL'd an open-source model using only 1,300 H200s plus months of proprietary experimental data, surpassing GPT-6 Astra on X-ray diffraction analysis and boosting Kimi K2.6 by 20x. The key twist is RL directly on real-world experimental data, in a closed loop where lab results inform the next experiments. Initial focus: superconductors, magnets, and semiconductor materials.
More from AGI Musings
- Why recursive self-improvement to superintelligence likely won't happen — tawnniee · 2026-09-16
- Gradio founder: let users read full unencrypted reasoning traces to align LLMs — mmitchell_ai · 2026-09-16
- A 1797 Goethe poem explains why AI researchers fear recursive self-improving systems — OmarUFlorez · 2026-09-16
- Everyone using AI at work risks becoming a useless middle manager — jamesluke585 · 2026-09-16
- Researcher pushes back on AI biorisk alarmism: threat vectors haven't changed since o1 — anshulkundaje · 2026-09-16
- Maybe human intelligence is near the ceiling of stability evolution allows — games-and-games · 2026-09-16