Key Summary

Researchers from Google DeepMind, the University of Maryland and the University of Virginia published Dream-RSI on arXiv on the 14th, a method that recycles prior search traces to cut AI agent invocations by up to 162x on a Lasso regularization-path task, though the gain narrows to roughly 1.7x when the same model and settings are held constant. The efficiency comes from the orchestration layer — search ordering, parallelism and stopping rules — rather than retraining foundation-model weights, which points to lower inference spend per agent task without reducing demand for training compute. The 1.7x apples-to-apples figure is the number that matters for any compute-demand-destruction thesis on NVDA and hyperscaler capex; watch for third-party replication across broader benchmarks before repricing AI inference cost curves or Alphabet's agent-serving margins.