Inside NVIDIA: what a world model is and how it connects to LLMs

By Nate · September 24, 2026.

The email is a preview of an article in which Nate recounts that he traveled the previous week to NVIDIA's facilities to interview Ming-Yu Liu, vice president of Cosmos Lab, the company's laboratory dedicated to so-called 'world models'.

The article's argument starts from a central idea: to understand why NVIDIA — a company most people associate with chips — is building this type of model, you have to begin from the premise that NVIDIA is, above all, a developer tools company. According to the text, these world models would make it possible to simulate how a factory operates, predict what happens when a robot moves, or generate the actions a machine should carry out.

The body of the email cuts off just as Nate begins to describe his conversation with Liu about Cosmos ('Cosmo…'), so the detailed content of the interview — the specific examples, verbatim quotes or technical explanations — is not available in this excerpt and cannot be summarized without risking making up information.

Broadly, and as industry context, 'world models' are an area of AI research aimed at having a system learn an internal representation of how the physical environment works (causality, physics, motion), something considered relevant both for robotics and for the evolution of language models toward more agentic capabilities. NVIDIA has publicly presented its Cosmos platform as a family of such models geared toward robotics and autonomous driving, although the specific details offered by Ming-Yu Liu in this particular interview do not appear in the email fragment received.