Google's Agents Learn by Dreaming

🔥 What happened Google Deepmind unveiled Dream-RSI: AI agents "dream" through past search runs to test better strategies without re-invoking the model. On a genomics task, attempts dropped from 550 to 317 and runtime from 3,587 to 2,931 ms. 💡 Why it matters Rival SimpleTES needed 51,200 runs — Dream-RSI needed just 317. On GPU kernels it cuts generations by up to 2.43x or boosts performance 2.09x within the same budget. The underlying model stays frozen; only the search strategy learns. ⚡ Our take This is recursive self-improvement without retraining — cheap, elegant, scalable. Anyone still running fixed search strategies is burning compute tomorrow.