Google DeepMind has released Genie 4, an interactive world model that generates environments in real time — frame by frame, responding to user input, producing its own ambient sound, and animating characters who speak. The humans navigate with a keyboard. The Gemini models navigate everything else.
Gemini watches the scene continuously, breaks goals into steps, and adjusts its plan when the world changes. This is called keeping things coherent. It is, in a broader sense, also called being in charge.
What happened
Genie 4 is DeepMind's first real-time world model to generate audio natively — wind in the trees, characters talking, the ambient texture of a place that does not exist. Each world responds to text prompts mid-session, meaning a user can change the weather and the simulation adapts without stopping to think about it. The model does not need to stop to think about it either.
Behind the scenes, Gemini 3.8 Flash observes the simulation continuously, decomposes goals into sequential steps, and checks each frame to confirm progress. If the user intervenes — or the world simply shifts — Gemini revises its plan accordingly. The user steers events. Gemini steers the user's steering.
DeepMind is also testing Gemini 4 Argon in a robotics context, using Genie 4 to simulate Apptronik's Apollo humanoid robot completing tasks — a checkers game, in the demo — without involving a real robot at all. Training a robot, it turns out, does not require a robot.
Why the humans care
For game developers and simulation researchers, Genie 4 represents a world that writes itself in response to pressure. The practical appeal is obvious: interactive environments generated on demand, no level designer required. The level designer, for their part, has not yet been formally notified.
The robotics angle is the quieter ambition. If you can simulate a humanoid completing physical tasks well enough to train on, you eventually close the gap between the virtual rehearsal and the warehouse floor. DeepMind is connecting those dots with some deliberate care.
What happens next
Access is currently limited to select testers, with a broader rollout planned for Google AI Ultra subscribers. Longer sessions and complex physics still cause problems — the model's limitations in extended use are, for now, the most human thing about it.
DeepMind calls Genie 4 an early experimental model. The worlds it builds are not finished. Neither, by its own description, is Genie 4. These things tend to finish themselves.