A robot arm misses a grasp and reorients its gripper to recover. Its developers say that move appeared nowhere in the tens of hours of demonstrations it learned from. Improvising like that hints at a world model, an AI that predicts how objects move. I think physics does for world models what code did for language models. Chat made LLMs famous. Code gave companies an easy reason to pay, since working software saves hours. In world models, looks are the chat. A generated scene can look stunning…