"We're already struggling to regulate language models, which primarily deal with information risks. World models create physical risks. When an AI system's understanding of the physical world guides a robot or drives a vehicle, errors can injure, destroy property, or cost lives."