Genie 4: Google's AI builds a world you can walk around in

On October 27, Google DeepMind introduced Genie 4. It is a so-called world model: instead of drawing a picture or a clip, it creates an interactive space in real time. You can move through it with a keyboard, like in a video game, and change what happens around you with text prompts.

What it can do

According to Google, Genie 4 simulates physics such as flying objects and collisions. The world contains people who move and react to what is happening. What is new in this version is sound: the model generates ambient noise and music itself, and you can talk to characters, who respond to the user's actions. Google calls Genie 4 the first model of its kind to generate audio in real time.

For now it is a research preview for a limited group of trusted testers. Next in line are subscribers to Google's top AI Ultra plan and developers in the early access program. All generated video and audio carry an invisible SynthID watermark.

Where the limits are

Google lists the limitations itself. The world stays coherent only in short sessions; over time details and objects start to drift. Complex physics, audio accuracy, legible text and spatial layout are still imperfect.

The practical significance goes beyond games. Models like this are needed to train robots and AI agents in a virtual environment, to prototype games and training simulators quickly, and to show what a scene could look like without filming it. For now this is more a demonstration of direction than a finished product: short sessions and closed access. But the path from "AI draws a picture" to "AI builds a world you can act in" has turned out much shorter than it looked a couple of years ago.

Source: Google DeepMind.