This interview podcast features Noah Kravitz interviewing Ming-Yu Liu, NVIDIA's Vice President of Research, about NVIDIA's new Cosmos platform for world foundation models. The discussion begins with an explanation of world foundation models as deep learning-based simulators capable of predicting future events by simulating physics and human actions. Liu then details how these models differ from LLMs and video generation models, focusing on their use in training and verifying physical AI agents, such as robots and self-driving cars, to improve safety and efficiency. The interview concludes with a description of the Cosmos platform, which offers pre-trained models and tools to help developers build and customize world foundation models for various applications. A key takeaway is that Cosmos provides open-source pre-trained models to accelerate development in this field.
