Google and DeepMind demonstrate AI recursive self-improvement loop Dream-RSI

Dream-RSI replays past discoveries to test exploration strategies, reducing agent calls by up to 162x while maintaining or improving results in algorithm design and GPU kernel engineering.