Google DeepMind unveils Genie 3 that can create instant, interactive video game worlds

Google DeepMind is stepping up its efforts in AI-generated environments with the release of Genie 3, a new version of its “world model” AI that can create interactive 3D spaces in real time. Unlike previous versions, Genie 3 promises a more immersive and consistent experience, with the ability to remember objects in the environment and allow users (or AI agents) to interact with these worlds for longer periods.

World models are AI systems designed to simulate interactive spaces for applications like training robots, building games, or creating educational content. Instead of painstakingly designing these virtual spaces by hand, users simply provide a text prompt, and the AI generates a playable world on the fly, similar to how image generators create visuals from text descriptions.

Genie 3 builds on DeepMind’s earlier work with Genie 2, which could generate interactive scenes from static images. However, Genie 2 had major limitations: its worlds only allowed about 10–20 seconds of interaction and lacked visual consistency. Genie 3 addresses these issues by enabling a few minutes of continuous interaction and visual memory, meaning if you look away from an object and look back, it will stay in place—like writing on a chalkboard or patterns on a wall.