AI scientists race to build 'world models' beyond chatbots

Key takeaway

• Leading AI scientists are moving from chatbot development to 'world models' that teach AI systems to understand and interact with physical spaces and objects.

• Instead of learning from text alone, these models focus on how space, time, and physics work—enabling applications from smarter robots to interactive virtual environments.

• Venture capital firms are investing in startups pursuing this direction, signaling a shift in where the AI industry sees the next major opportunities.

3 Key Points

FAQ

What is a world model and how does it differ from current AI chatbots?

A world model teaches AI to understand the statistical structure of space and time—how light falls, how objects respond to force, and how physical interactions work. By contrast, current language models like ChatGPT learn by predicting the next word in text, which limits their ability to perform physical tasks like a robot picking up a coffee mug.

What are the practical applications of world models?

World models can enable robots to understand their environment and adapt quickly (similar to how a human adjusts their walk when their knee hurts), create interactive virtual worlds like video game environments that respond to a player's actions, and support AI agents in predicting the consequences of their own actions in unstructured settings.

Which AI scientists and companies are pursuing world models?

Fei-Fei Li founded World Labs in San Francisco, Yann LeCun started Advanced Machine Intelligence Labs in Paris, and Louis Castricato founded Overworld in Rhode Island. Venture firm Kindred Ventures is investing in Overworld, Causal Labs (which builds AI models for weather prediction), and Extropic (which is building specialized computer chips for world models).