Summary
Nate B Jones visits NVIDIA to talk with Ming Yu, head of the Cosmos world model family, about what world models are and how close they are to being useful in the physical world. The conversation starts with a practical question: will a household robot fold laundry before it can cook perfect scrambled eggs?
Yu describes his path from generative AI researcher, who predicted a decade ago that visual content would be created by deep learning rather than computer graphics, to leading Cosmos development. The discussion covers NVIDIA’s role as a software and developer company as well as a chip maker, and how layers such as CUDA, libraries and models feed back into hardware design.
A major theme is physics. Yu explains that models improve at approximating physical behavior as data grows, but that physics engines like Omniverse and Isaac are used to generate synthetic data for interactions, such as objects colliding, that are hard to find in real footage. He also outlines the Cosmos 3 architecture, a mixture-of-transformers design that combines a language-based reasoning component with a generator and takes actions as an input and output, which matters for robotics. The interview gives a clear look at how world models differ from language models and where policy models for robots are heading.
📺 Source: AI News & Strategy Daily | Nate B Jones · Published September 24, 2026
🏷️ Format: Interview
