DeepMind's Yao Shunyu: RSI is a system problem, and post-training recipes will soon be decided by models themselves
_AndrewZhao · x · 2026-09-19
Key points from an AGI House interview with Google DeepMind researcher Shunyu Yao:
- Positioning RSI: RSI is a spectrum; its key difference from RL/RFT is that it iterates on its own training recipe. In the endgame, RSI is a system problem, not a model problem, spanning harness and hardware like CUDA.
- Real progress: While many startups have appeared, he cares more about private/internal deployment — Gemini 3.8 Flash's core post-training recipe was decided by the model itself. This is becoming a practical problem, and a strategy for all labs, not just Google.
- Scaling insight: Scaling up model size doesn't require scaling up team size; RSI is essentially a bug/missing block in current models.
- Safety stance: Nothing is absolutely off-limits for models; auto-safeguarding is hard, and limiting models that exceed human capability is a positive research direction he studies.
- Q&A: On single-agent PRMs, he said percentage/LLM-based verification is fundamentally hard and unscalable with no clear breakthrough; for hard-to-verify productivity tasks, he suggests studying how humans build RM systems vertically.
More from AGI Musings
- Alignment researchers debate: the RL that grants agency also breeds misalignment — davidmanheim · 2026-11-04
- Bindu Reddy: AI fear-mongering has put the whole technology at risk in the US — bindureddy · 2026-11-04
- AI welfare vs AI safety: 'domesticate AIs like dogs' argument sparks debate — repligate · 2026-11-04
- Intent, not taste, is the real differentiator in the AI era — and it can't be bought — edelwax · 2026-11-04
- "11 biggest AI lies": viral Reddit takedown of solo billionaire and utopia promises — Own_Garage_1478 · 2026-11-04
- EpistemicHope: models with problem-solving agency are increasingly dangerously misaligned — EpistemicHope · 2026-11-04