LLMs Can Think While Idle: Researchers from Letta and UC Berkeley Introduce ‘Sleep-Time Compute’ to Slash Inference Costs and Boost Accuracy Without Sacrificing Latency