Google DeepMind’s Genie 3 Gets One Step Closer Toward Interactive Worlds That Feel Alive

DeepMind has released Genie 3, a system that creates interactive, ever-changing environments from a simple text prompt. This gets one step closer towards creating digital worlds that respond to your actions in real time, that hold their logic and detail for minutes at a time.

Genie 3 builds on DeepMind’s previous work, taking cues from Genie 1 and Genie 2 which were focused on creating environments for AI agents to navigate. Those models were the foundation, but they were more like static sketches compared to Genie 3’s dynamic canvas. This new system generates 720p worlds at 24 frames per second, so you can walk through them like you’re playing a video game made on the fly.

Unlike traditional game engines that have pre-programmed mechanics, this model learns how objects should behave. It’s like the system has an intuitive sense of the world, piecing together each moment based on what came before. DeepMind calls this an auto-regressive approach, meaning every new frame considers the entire history of your journey. If you loop back to a spot you visited earlier, the same building or tree is there, unless you’ve changed the scene. This memory gives the environments a rare sense of permanence, so they feel less like illusions and more like places you could actually be in.

Beyond navigation, Genie 3 has “promptable world events.” Type a few words and the world changes, enabling a user to change the environment on the fly, whether it be turning a sunny meadow into a stormy coast or adding a wandering animal to a desert. These changes create new opportunities for AI agents to train in these worlds, key to building systems that can handle the uncertainty of the real world, as well as a big part of the puzzle for more advanced AI.

DeepMind’s researchers see Genie 3 as a training ground for AI agents, especially those that act and learn like humans in complex environments. They tested it with SIMA, an agent designed to achieve goals in 3D virtual spaces. In Genie 3’s worlds, SIMA can chase objectives—like finding a hidden object or navigating a maze—while the environment responds to its every move. Because the system is consistent over longer periods, agents can do more complex tasks, learn from trial and error like humans do.

Looking ahead, Genie 3 could change how we train AI, from robots to autonomous systems. Its worlds are a sandbox for testing and refining skills in ways static simulations can’t. In education, it could refer to an immersive atmosphere in which students investigate historical contexts or practice complex activities. For creators, it’s a tool for developing ideas ranging from gaming levels to film sets without the need for a large production team.