In this episode of the Google DeepMind Podcast, Professor Hannah Fry interviews Shlomi Fruchter and Jack Parker-Holder, the creators of Genie 3, an interactive world model. They discuss Genie 3's capabilities, including generating diverse, visually interesting worlds from text prompts or images, its potential applications in agent simulation, education, and entertainment, and its advancements over previous iterations like Genie 1 and 2. The conversation explores the technical aspects, such as the autoregressive nature of the model, its understanding of physics, and the challenges of maintaining consistency and realism. They also touch on the safety implications and the future direction of the research, including the goal of creating a foundational model for simulated worlds, similar to what LLMs have achieved for language.
Sign in to continue reading, translating and more.
Continue