Google DeepMind has unveiled Genie 3, a next-generation world model designed to simulate interactive environments in real time, marking a major milestone in the pursuit of artificial general intelligence (AGI). Unlike its predecessors, Genie 3 doesn’t rely on hard-coded physics engines. Instead, it learns how the world behaves by observing and remembering its own generated simulations. This allows it to create consistent, photorealistic or imaginative 3D environments from simple text prompts, with simulations lasting several minutes at 720p resolution and 24 frames per second.
The model introduces “promptable world events,” enabling users to alter simulations dynamically—such as changing weather conditions or introducing new characters—without restarting the environment. Genie 3’s memory architecture allows it to reason over time, maintaining physical consistency and enabling agents to learn through experience, much like humans do. DeepMind demonstrated Genie 3’s capabilities using its generalist agent SIMA, which was tasked with navigating and interacting within these generated worlds.
Researchers believe Genie 3 could revolutionize training for embodied AI agents, such as robots and autonomous vehicles, by offering a rich curriculum of simulated scenarios. These environments could replicate complex real-world settings like warehouses or natural landscapes, helping agents develop skills before deployment. While Genie 3 remains in limited preview, DeepMind plans to expand access to select researchers and developers.
The model’s implications extend beyond robotics. It could reshape education, gaming, and creative prototyping by offering immersive, customizable simulations. DeepMind’s leadership, including Demis Hassabis, has likened Genie 3’s potential to building a real-life Holodeck—a reference to Star Trek’s fully interactive virtual reality chamber. Though not yet publicly available, Genie 3 is being hailed as a foundational step toward AGI, where machines can perform a wide range of tasks with human-like adaptability.
