Yao Shunyu on AI R&D automation, data moats, and where startup opportunities lie
青稞AI · wechat · 2026-11-20
A WeChat account summarizes a talk by Yao Shunyu covering model development, AI-driven R&D automation, safety, and career strategy. Key points:
- Small models & model hierarchy: AI-for-AI will drive distillation from frontier models into task-fit sizes; value hinges on capability-cost matching.
- Non-text data: time series, clinical data, and structured tables remain an untapped opportunity for agents.
- RSI & R&D automation: the essence is running research loops (bug fixes, data cleaning, experiments) better — progress comes from solving concrete problems, not new concepts.
- Scale effects: small-scale failures don't invalidate a direction; combine theory with small experiments to predict scale behavior.
- Data as core moat: pretraining needs data reuse, memorization/overfitting, synthetic data quality, dynamic mixing; posttraining balances intelligence and alignment.
- RL engineering: correctness and stability before efficiency; co-design training and inference.
- Safety: Safety vs Security distinction; evals must cover multi-agent coordination, permission boundaries; open source needs defenses and pre-release evals.
- Business: interaction changes and newly-feasible tasks create opportunities more than raw model gains.
- Talent: bottom-up exploration, Trust and Taste matter; individuals should join the model industry or build in model-complementary areas like physical world and hardware.
More from AGI Musings
- If doomers believe in doom, quit your tech jobs and leave us alone, says critic — TinfoilTricorn · 2026-11-20
- _xjdr publicly recants: 'I'm a flops guy at heart now' as tszzl also admits being wrong — tszzl · 2026-11-20
- Nature study: breathing rhythm dynamically modulates human sensory perception — AnnaCiaunica · 2026-11-20
- Fans call for boycott of film using generative AI; designer says resisters face a hard road — AIandDesign · 2026-11-20
- Scott Aaronson vs Robin Hanson: Is AI Extinction Risk Real or Panic? — burny_tech · 2026-11-20
- Scott Aaronson debates Robin Hanson: will powerful AI kill us, or are we panicking? — burny_tech · 2026-11-20