Google DeepMind previews Genie 4 for steering a simulation with words and sound

Google DeepMind began a restricted test of Genie 4, a world model that responds to natural-language scene directions and generates audio.

Google DeepMind introduced Genie 4 on October 27 as a research system for generating interactive worlds. Google says a user can steer a running scene with natural language, including changing the weather, redirecting the camera or introducing a character. The model generates audio with the scene, including ambient sound and dialogue. Google began a limited rollout to trusted testers and named two subsequent access paths, Google AI Ultra subscribers through Project Genie and early access partners through the Gemini API and AI Studio. The announcement gave neither group a firm date. [Google announcement](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-genie-4/)

What changes when the user can intervene

A conventional video generator turns a prompt into a clip. A simulation also has to respond to instructions that arrive while the scene is unfolding. If someone redirects the camera or adds a character, the system has to account for what was visible before the change and what happened off camera. Google says Genie 4 builds on Gemini Omni for the visual and behavioral qualities of the world and uses other Gemini capabilities to help maintain continuity.

The natural evaluation is a sequence of related commands. A user could enter a scene, change the weather and then return to a position visited earlier. Reviewers could check whether objects and characters are still where the earlier scene left them, whether the new conditions persist and whether the sound still matches the scene. One polished demonstration can show a possible result. A series of repeated tasks can reveal whether the behavior is stable enough for a production workflow.

Creative exploration and simulation have different requirements

A filmmaker or game designer can use an interactive world to compare camera placement and timing before producing a scene. A research team using it as a training environment has a different burden. It needs to know which interactions follow expected physical behavior, which can be repeated and how an agent performs outside the simulated environment. Google listed robot and agent training as a potential use. A team considering that use would need separate validation against physical or otherwise verified test cases before treating results as evidence of real-world performance. [Google announcement and linked demonstrations](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-genie-4/)