Periodic's Neon beats frontier models on materials science with just 1,300 H200s and a self-built lab loop
vwxyzjn · x · 2026-09-16
Periodic unveiled Neon, an open-source model mid-trained and RL'd on data from its high-throughput Menlo Park materials labs, forming a closed loop: labs generate fresh data, models learn from it and decide what to try next. Using only 1,300 H200s — orders of magnitude less than rivals — Neon surpassed GPT-6 Astra and Fable 5.1 on Periodic's analysis benchmark. The team targets superconductors, magnets, and semiconductors next, arguing the experimental data loop, not compute scale, is the key.
More from Research
- Single-axis random features beat full weight sets; a Markov-chain POS tagger trains in an hour — cephaloform · 2026-09-16
- Odyssey unveils Odyssey-3 foundation world model for robots, humanoids and driving — chris_j_paxton · 2026-09-16
- Jason Wei: Wet-Lab Data Lets Specialized Models Beat GPT-6 Astra at Frontier Science — vwxyzjn · 2026-09-16
- Hand-computable AI? Markov chains plus polynomial feature fits for POS tagging — cephaloform · 2026-09-16
- Pichai maps 9B genetic variants, launches WeatherNext 3 in Google's AI-for-science push — sundarpichai · 2026-09-16
- Most benchmark 'model errors' are actually benchmark or grading errors, analysis finds — zainhas · 2026-09-16