How a Moonshot Led to Google DeepMind's Veo 3

Explore the technical journey behind Google DeepMind's groundbreaking Veo video generation models in this 48-minute conference talk featuring Dumi Erhan, co-lead of the Veo project, in conversation with host Logan Kilpatrick. Trace the evolution of generative video technology from its early research foundations in 2018 through the development of the state-of-the-art Veo 3 model with native audio generation capabilities. Discover the technical challenges involved in evaluating and scaling video models, including the complexities of maintaining coherence across long-duration videos and the limitations of physics-based evaluation methods. Learn about the significant leap from Veo 1 to Veo 2, the viral impact of Veo 3's audio capabilities, and how user feedback is actively shaping the future roadmap of AI-powered video creation. Examine the comparative complexity between image-to-video and text-to-video generation, explore new prompting methods that enhance user control, and understand the steerability challenges in video model development. Gain insights into the role of image data in capability transfer, the connection between video prediction and robotics applications, and the broader implications of world models like Genie 3 for the future of AI-generated content.