One sentence, and a world appears. On Tuesday, Google DeepMind unveiled Genie 4, its new “world” model. You enter it with the keyboard, you move around, the image is built as you go. You type “rain”, “top view” or “add a dog”, and the scene obeys, whatever the the request.
250 000
trees felled by Google in Muhos, Finland, for a data centre, according to Yle
Sound, and two minutes of memory
What is new is the sound, produced along with the image, and a consistency that holds for two minutes, twice that of Genie 3, released in August 2025. “These are early days”, tempers Diego Rivas, an engineer at DeepMind. Objects drift over time, the text on screen is sometimes unreadable.
For whom, and when
Access comes “in the next few weeks” for Google AI Ultra subscribers, every frame marked with SynthID. When a machine builds a world in a few seconds, what is a level designer still for?
