Jason Wei Maps AI for Science Into Two Routes: DeepMind-Style RL and Generalist Experimenters
shyamalanadkat · x · 2026-10-02
Researcher Jason Wei argues AI for scientific innovation will grow strongly over the next five years, in two flavors. The "DeepMind style" targets one important problem (e.g. protein folding) with a purpose-built RL environment; he notes LLMs have so far been a poor fit since these structured, specialized problems don't benefit most from natural-language pretraining — though frontier model gains may change that. The "generalist style" trains an AI broadly better than humans at running experiments, such as automating AI research given a set of evals. A substantive first-hand take on the field's strategic split.
More from AGI Musings
- Survey: 76% of Americans say torturing sentient AI is wrong, 63% want AGI banned — jacyanthis · 2026-10-02
- New paper explores training risk aversion into AI to make misaligned models negotiable — sethlazar · 2026-10-02
- Researcher joins new cohort to explore how legal theory can improve rule-based alignment and reward design — PeterHndrsn · 2026-10-02
- Stanford's Omer Reingold: mathematicians must stop mourning and face the LLM era — IAmSamFin · 2026-10-02
- Do AI labs have liability insurance? Redditor likens regulation pleas to Price-Anderson nuclear shield — BarnabyWoods · 2026-10-02
- Self-description vs. machine pain: why LLM utterances can't reveal inner experience — ctjlewis · 2026-10-02