Genie 4 makes Google’s world model steerable in real time, and the pitch is storyboards, not reality

TechCrunch reports that Google DeepMind has put Genie 4 into research preview. Compared with Genie 3, worlds are more photorealistic and now come with generated audio, including dialogue and music. Users can prompt mid-simulation to change the weather, move the camera or add people, and scenes hold together for up to two minutes, double the previous limit. Trusted testers have access now; Google says Google AI Ultra subscribers in Project Genie and early-access partners of the Gemini API and Google AI Studio will follow in the coming weeks.

The most interesting change is architectural. Gemini now acts as a director, planning steps and checking frames, while Genie renders the world. In one robotics experiment, a Gemini model broke a task into steps that Genie 4 simulated in real time. Testing robot behavior in simulation before it touches the physical world is one of the clearly useful things AI can do, and cheaper, richer simulation could make real robots safer.

Trusted testers describe grounded uses: prototyping films, storyboarding and rapid ideation. A tool that keeps a creative team in flow while sketching a scene assists the work rather than pretending to finish it. Google’s Veo video model has already found an audience among filmmakers, and Genie extends that into interactive space.

Genie 4 lands as the world-model market consolidates fast. TechCrunch reports that AMD is acquiring World Labs, Amazon is in talks to buy Decart, and Rivian bought Third Dimension for self-driving simulation. Runway shipped its own playable model with audio a couple of months ago. Nvidia is pushing open world models too, with Cosmos 3 for interactive simulation (our coverage). Whoever owns these systems will shape how games, training and robot testing get built.

Google is candid that objects drift over longer runs, audio accuracy is unsolved and in-world text can be illegible. Two minutes of consistency is progress, not permanence. That honesty matters because the core risk of world models is the gap between a convincing simulation and a correct one. If robot training or driver testing relies on worlds whose physics quietly drift, failures could get trained in rather than caught.

There is a cultural risk as well. One speculative example Google floated is stepping into a historical painting and chatting with its subjects. That could be a fine teaching tool. At scale it could blur history and plausible invention, with synthetic dialogue from real people in convincing audio. Provenance and labeling need to keep pace with realism, an area where Google already offers its SynthID checker (our coverage).

Keeping Genie 4 in preview is the right call. The acquisition spree shows the industry treats world models as infrastructure, and infrastructure needs verification standards before it needs more photorealism. More in our Google coverage.