Google DeepMind's Dream-RSI cuts AI discovery search cost up to 162x
yangyi · x · 2026-09-18
Google/DeepMind researchers introduced Dream-RSI, a recursive self-improvement loop for AI-driven discovery:
- The system lets agents replay their past discovery attempts, cheaply test thousands of alternative exploration strategies, and deploy better ones in the next round.
- In algorithm design, math optimization, and GPU kernel engineering, it maintained or improved discovery quality while slashing search cost—cutting agent calls by up to 162x in one scenario.
- Crucially, it improves exploration strategies, not the underlying model weights.
More from coding & agent
- Jev scores tool calls KEEP/DROP to cut coding agent context 156k→62k — but deletes can backfire — Moretti_a · 2026-09-18
- Cutting agent costs: cheapest model as coder, squeezing reviewer spend next — craigbalding · 2026-09-18
- Chinese coding models use 2-3x the tokens, so cheaper per token isn't cheaper overall — craigbalding · 2026-09-18
- AI plugin plays Old School RuneScape, auto-decides when to eat in combat — JasonBotterill · 2026-09-18
- Dev builds real-time insurance claim voice agent with Gemini 3.8 Live, open-sourced — Arindam_1729 · 2026-09-18
- Kimi Agent Swarm tested in the cloud: 16-hour autonomous training run with self-recovery — solyarisoftware · 2026-09-18