Experts Have World Models. LLMs Have Word Models.
Summary
The piece argues that true domain expertise in AI comes from modeling adversarial, multi-agent environments rather than single-shot artifact generation. It contrasts chess-like domains with poker-like real-world settings, highlights Pluribus as a lesson in robustness, and contends that current LLMs fail due to training signals that reward static artifacts rather than adaptive behavior. It calls for a new training loop that optimizes outcomes in interactive environments with hidden state and self-interested agents.