1,300 H200s and a closed experimental loop: open model Neon beats GPT-6 Astra on materials analysis
OriolVinyalsML · x · 2026-09-16
Liam Fedus' team built high-throughput materials labs in Menlo Park, creating a loop where experiments generate fresh data, models learn from it, and then help decide what to try next. Using only 1,300 H200s plus months of experimental data, they mid-trained and RL'd an open-source model — dubbed Neon — that surpasses GPT-6 Astra on their analysis benchmark. Oriol Vinyals shared the post, calling it welcome progress in AI for science.
More from Research
- Four LLM mayors, one deterministic city: open-source loop-design experiment wins CoreWeave Hacks top prize — hackgoofer · 2026-09-16
- RL-trained LLMs predict and shift human opinions, succeeding 89.4% of the time — LividResearcher7818 · 2026-09-16
- New paper: Two routes to scalable credit assignment without weight symmetry, code released — aran_nayebi · 2026-09-16
- Adaptive BCI decoders concentrate task info into fewer neurons, Nature Communications study finds — hugo_larochelle · 2026-09-16
- Reddit Deep-Dive: Could a "Large Emotion Model" Solve AI Alignment? — No-Zookeepergame-390 · 2026-09-16
- Open-source climbing CV tool uses SAM3.1 and pose models to quantify your hand and foot placements — spillai · 2026-09-16