Two Clouds of Agent Heaven

Two unsolved challenges: (1) streaming interaction with the environment in real time – the world doesn’t wait for the agent to think: speech is interrupted, the picture changes, mail arrives continuously. Paths: architectural separation of fast and slow plus accelerating the output itself (when decoding gives thousands of tokens per second, the “turn-based” and “real-time” boundary blurs); (2) continuous experience accumulation – the model is like a genius with a memory who doesn’t learn on the job: after each task, bumps and bruises are thrown away with the context. The dispute between the “small world hypothesis” (it is enough to train a large model once) and the “large world hypothesis” (you need to learn after deployment).

Related: [Co-evolution of Model and Agent], [Continuous Agent Evolution], [Full Duplex]