From Words To Worlds

In a November 2025 essay, World Labs cofounder Fei-Fei Li argues that useful world models must preserve a coherent environment as people or agents act within it. An input action should produce a next state consistent with what existed before, including the scene's geometry and physical behavior.

That requirement connects generated environments to practical work. A creator exploring a generated scene needs it to stay coherent, and Li argues that world model outputs could help train robots across simulated environments where real training data is scarce. Li presents this as a research direction, with creative tools emerging now, robotics a mid-term horizon and the most transformative scientific applications further out.