Training models on large-scale data has given us powerful generative capabilities for text, images, and video. However, this success has not yet extended to training generalist embodied agents. This talk tackles this gap by focusing on a potential solution to this problem: scalable world models. We'll trace the idea of planning in predictive models, from its origins to modern efforts on building world models directly from pixels. I'll discuss the primary challenge of scaling these models and present our work, Genie, which enables us to learn world models without explicit action labels at scale, demonstrating a new path forward for training the generalist agents of the future.