Liam Fedus' team builds robotic materials labs, trains open-source Neon on 1,300 H200s to beat GPT-6 Astra
gordic_aleksa · x · 2026-09-16
Liam Fedus unveiled Neon, an open-source model trained in a closed loop with high-throughput materials labs in Menlo Park: the labs generate fresh data, models learn from it and suggest next experiments. Using only 1,300 H200s plus months of proprietary experimental data for mid-training and RL, Neon surpasses GPT-6 Astra on their analysis benchmark — a notable case of AI-driven materials discovery.
Related event: Periodic Labs' Neon Beats GPT-6 Astra on XRD with Just 1,300 H200s(66 posts)→
More from Models
- Cohere launches Confidential Computing in Model Vault: encrypted even during inference — cohere · 2026-09-17
- Ex-ChatGPT co-inventor launches Jev, a decision model claiming 20-200x speedup over LLMs — rohanpaul_ai · 2026-09-17
- llama.cpp merges support for DFM Mimir 1B, GGUF weights on Hugging Face — noctrex · 2026-09-16
- Researcher warns LLM evals are broken after reading 'truly horrible' model traces — IanArawjo · 2026-09-16
- OpenAI's 10,000-word essay on a self-evolving 'alien mind' sparks ownership clash with startup JoyIn — SarahAnnabels · 2026-09-16
- Jev is a smart classifier, not a frontier LLM — a contrarian analysis — DavideCrapis · 2026-09-16