The podcast features an interview with Shlomi Fruchter and Jack Parker Holder from Google DeepMind, who discuss Genie 3, their latest world model capable of generating interactive environments from text prompts. They contextualize Genie 3 by comparing it to its predecessors, Genie 1 and 2, highlighting the advancements in resolution, real-time interaction, consistency, and the ability to simulate diverse environments. The discussion covers the potential applications of Genie 3, particularly in training AI agents and robotics, and touches on the challenges of achieving open-endedness and creativity in AI-generated worlds. The speakers also address the technical aspects of the model, including its autoregressive nature and emergent properties, as well as its limitations and future directions.
Sign in to continue reading, translating and more.
Continue