Google DeepMind has launched Genie 3, its latest leap in building digital “world models”—AI-generated 3D spaces where both humans and AI agents can move, act, and interact in real time. The goal? To inch closer to human-like AI by giving models a more realistic sense of memory, space, and continuity.
Unlike previous versions, Genie 3 can now retain visual memory for up to a minute, meaning if you leave a scene—say, a chalkboard or painted wall—and return, it remembers the details. It also delivers longer experiences, with interactions lasting several minutes, up from just 10–20 seconds in Genie 2.
World models like Genie don’t use hand-crafted game assets. Instead, they generate entire 3D scenes on the fly from a prompt, offering a flexible platform for AI training, robotics, simulation, and even entertainment.
A standout new feature in Genie 3 is “promptable world events”. You can now change the environment—add characters, shift the weather—just by typing a command. The visuals are rendered in 720p at 24 FPS, providing smoother, more lifelike experiences.
The previous Genie 2, released last December, could create interactive spaces from a single image but struggled with visual stability and short interaction windows. Genie 3 takes a clear step forward, though it’s not available to the public yet. For now, it’s being tested in a limited research preview with select academics and creators.
Text elements in these generated environments will remain limited unless included in the original prompt, and broader interactive capabilities are still under review. The project is led by a former co-lead of OpenAI’s Sora video generation project, bringing high-profile experience into Google’s push toward more immersive AI environments.