DeepMind reveals Genie 3, a world model that could be the key to reaching AGI

Google DeepMind has introduced Genie 3, a foundational world model designed as a significant step toward artificial general intelligence (AGI). Unlike previous narrow models, Genie 3 is a real-time, interactive, general-purpose world model capable of generating diverse 3D environments from simple text prompts. It can produce several minutes of photo-realistic or imaginative simulations at 24 frames per second and 720p resolution, surpassing its predecessor’s 10-20 second limit. A key innovation is its ability to maintain physical consistency over time by remembering prior generated content, enabling it to simulate coherent, physically plausible worlds that reflect an intuitive understanding of physics without relying on hard-coded physics engines. Genie 3’s architecture is auto-regressive, generating frames sequentially while referencing earlier frames to maintain continuity and realism. This memory-driven approach allows the model to simulate dynamic scenarios where objects interact naturally, making it an ideal training environment for embodied AI agents. These agents can explore, plan, and learn through trial and