Google equips Genie 4 with Gemini as its "director": generating worlds and planning storylines simultaneously.

Beating AI News Flash: Google DeepMind has unveiled its next-generation world model, Genie 4. Users can generate explorable virtual worlds via text input, control movement with a keyboard, and adjust weather or add characters at any time. The model simulates collisions, aerodynamics, and character movements, and for the first time natively generates ambient sounds, music, and real-time dialogue. Characters not only speak but also respond based on the scene and user input. Genie 4 leverages Gemini Omni to enhance the realism of visuals and physical behaviors, while Gemini 3.8 Flash acts as the real-time controller. The latter continuously monitors the scene, breaks tasks into steps, checks if actions are completed, and determines the next move. Google also demonstrated a use case where Gemini 4 Argon planned a humanoid robot’s checkers game, with Genie 4 simulating the movements in real time. Each simulation runs for up to around 2 minutes, twice the length of the previous Project Genie experience. However, after extended scene runtime, objects’ positions and appearances may still change unexpectedly. In an official demo of a motorcyclist, a helmet suddenly appeared out of nowhere between shots. Genie 4 is currently available only to a small group of invited testers, and will later roll out to Google AI Ultra, Gemini API, and AI Studio. Google aims to use it for creating game prototypes, rehearsing film scenes, and more importantly, enabling robots and AI agents to practice complex tasks in virtual worlds.