Real-time world sim that talks back — DeepMind Genie 4 ships native audio with the scene.

Ambient sound, music, and character dialogue generate with the world. Mid-run, change weather, move the camera, drop in a person or object in plain English; Gemini steers so the edit fits and the scene keeps going.

What it is: an interactive world model — not a one-shot video clip. Gemini Omni grounds look/behavior; other Gemini models hold coherence. Early joint tests with Gemini Robotics (Apptronik Apollo, step-by-step checkers).

Still preview: trusted testers first, then Google AI Ultra (Project Genie) and Gemini API / AI Studio early partners. Long runs still drift on object permanence and detail; complex physics and readable text are open problems. Video + audio carry SynthID.

https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/