Jim Fan, Director and Distinguished Research Scientist at NVIDIA, joins Radical Partner Molly Welch to explore what's holding back embodied AI and robotics. While large language models have advanced rapidly, Jim explains why progress in the physical world depends less on ever-larger models and more on solving fundamental problems: exploration-driven learning, scalable data pipelines, and synthetic simulation. Drawing on his journey from OpenAI to leading NVIDIA's Project GR00T, they discuss the robotics data bottleneck, why foundation models matter, and what must happen before robotics reaches its "GPT-4 moment."