Google's Genie 4 lets you rewrite a virtual world as it plays, with sound

What Google showed

Google DeepMind, the company's AI research lab, has made Genie 4 public. As technews.tw reported, it is the next version of Genie 3, which came out last year and was big enough news that game company share prices dipped.

A world model is an AI that creates a virtual place you can explore, rather than a single picture or a fixed video clip. You describe what you want, and the model generates the experience as you move through it.

Changing the scene mid-play

Like Genie 3, Genie 4 accepts changes while the simulation runs. You can shift the weather, move the camera, add new elements or interact with characters, all by typing instructions in plain language.

Behind the scenes, Google uses its Gemini 3.8 Flash model as a guide. It breaks your goal into steps and checks each generated frame, which helps the world keep its original context as you change things.

Sound and talking characters

Genie 4 is the first DeepMind world model with native audio, meaning the sound is produced by the model itself. Scenes can have background noise and music, and you can ask for particular sound effects.

Characters can also hold conversations. They respond to your prompts or to events in the world. Google also says physics, such as airflow and collisions, looks more natural than before.

The limits and the rollout

The catch is memory. Google DeepMind says a Genie 4 world stays consistent for roughly 2 minutes, after which the model starts to forget how the scene was first set up.

Access is staged. Trusted testers come first, followed by Google AI Ultra subscribers in Project Genie and early partners using Google's developer tools. Google says the technology could also help build training environments for robots, though that depends on the simulations being accurate.