Dream-RSI reportedly improves how AI explores problems without changing model weights

A post says Google/DeepMind researchers introduced Dream-RSI, a system that replays past discovery attempts, cheaply tests thousands of alternative strategies and deploys a better strategy in the next round. It claims Dream-RSI matched or improved discovery quality across algorithm design, mathematical optimization and GPU kernel engineering, while reducing search costs. The post emphasizes that the system improves the exploration policy—how the agent searches for solutions—not the underlying model weights.

Separately, The Information reports that OpenAI researcher Noam Brown called improving new models’ ability to do AI research and development the company’s top priority. He described recursive self-improvement as the priority “by a pretty wide margin.”