Arthur H

Language should be a mere readout of a world-model-based reasoning. The only domain where language and world overlap is programming, no surprise this is where LLMs shine.

Until agents can reason about their own action space as first-class objects, they remain capable, but fragile when confronted with novelty or complexity

If we go back to Jaynes, consciousness emerged when command-based behavior stopped being enough and we needed an internal model of their own decisions and possibilities In agentic AI, we may be at a similar point: lacking an introspectable model of the agent's environment