Two Clouds of Agent Heaven

Two unsolved challenges:

Real-time environmental interaction streaming – 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).
Continuous experience accumulation – the model is like a genius with memory who doesn’t learn on the job: after each bump, the bumps are thrown away with the context.

The debate 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).