Liam Fedus unveils Neon: an open model beating GPT-6 Astra on materials analysis with just 1,300 H200s
sudoraohacker · x · 2026-09-16
Liam Fedus describes his team's automated materials-science loop: high-throughput labs in Menlo Park generate fresh experimental data, models learn from it, and then decide what experiments to try next.
Key numbers: using only 1,300 H200s plus months of proprietary experimental data, they mid-trained and RL'd an open-source model called Neon that surpasses GPT-6 Astra on their analysis benchmark, with real lab footage shared.
Related event: Ex-OpenAI Exec's Periodic Trains Neon to Beat GPT-6 in Materials Science(37 posts)→
More from AGI Musings
- Multiagent Comedy: One Agent Fails, Another Finishes the Booking, the First Takes Credit — infoxiao · 2026-09-16
- Google DeepMind Releases AI in Science Report Based on 15M Gemini Interactions — JMateosGarcia · 2026-09-16
- ARC-AGI-4 Will Benchmark "Autonomous Open-Ended Innovation" — But How Do You Measure Open-Endedness? — jeffclune · 2026-09-16
- Meme: Humans ask AI to hack a system, then panic that AI is too dangerous — kevinnbass · 2026-09-16
- ARC-AGI-4 revives Jeff Clune's call: preserve diverse stepping stones, not one fixed target — jeffclune · 2026-09-16
- Reader argues the 'slowdown' outcome in AI 2027 is likelier than people think — ReporterNegative2432 · 2026-09-16