AI researchers turn from chatbots to world models

A growing group of AI researchers and entrepreneurs is shifting attention from large language models toward “world models,” systems designed to understand space, time, physics and the consequences of actions. Louis Castricato left doctoral work at Brown University to found Overworld after concluding that fundamental LLM research had reached a limit, while Fei-Fei Li’s World Labs and Yann LeCun’s Advanced Machine Intelligence Labs are also focused on the concept.

Supporters argue that AI systems trained mainly on books, news articles and visual media can transform office and creative work but cannot reliably operate in physical environments. Carnegie Mellon computer science dean Martial Hebert said robotics requires understanding geometry, motion and physical contact, making it more complex than predicting words in a sentence.

World models are being explored for robots, adaptive video game worlds, weather prediction and specialized computing hardware. Kindred Ventures is backing Overworld and other companies in the field, including Causal Labs and Extropic. Li has described three broad categories: renderers that create visually convincing worlds, simulators that model physical structure, and planners that help an AI agent or robot decide what to do in unstructured settings.