Google's Dream-RSI reduces discovery-agent calls by 162x

Google: Google is a major technology company with dedicated AI research efforts spanning multiple labs and academic collaborations. Its researchers developed Dream-RSI to enable agents to replay and refine exploration policies from stored historical discovery trees instead of repeating costly executions. The work directly targets inefficiencies in agentic discovery loops across algorithmic and hardware optimization tasks. Dario Amodei: Dario Amodei is the co-founder of Anthropic, an AI safety-focused company. On September 12 he published the essay “We Must Pace the Frontier,” which identifies recursive self-improvement as one of two developments that shifted his view toward slowing AI progress to mitigate risks of uncontrolled systems. The Google research is presented as illustrating narrower, pragmatic applications of RSI in contrast to those broader concerns. Google DeepMind: Google DeepMind is Google’s primary AI research organization advancing foundational models and agent systems. Researchers from the lab joined colleagues from Google, the University of Maryland, and the University of Virginia to build the Dream-RSI framework for programmable, history-aware exploration in discovery agents. Their contributions emphasize structured replay simulators that reuse prior execution traces for policy improvement. AI Research: Google researchers introduced Dream-RSI to convert accumulated agent execution history into a reusable replay simulator for refining exploration policies without repeated online runs. Policy Discussion: The Dream-RSI work appeared days after Dario Amodei’s September 12 essay arguing that recursive self-improvement warrants pacing frontier AI development.