The line between reality and simulation just got blurry
Google DeepMind has released Genie 3, an AI model that creates interactive, navigable worlds in real time, just from a text prompt.
You type something in, and Genie 3 generates a 3D environment you can explore at 24fps, with visual memory lasting a few minutes.
It’s the first version in the Genie series to support real-time interaction, with improved realism and consistency.
The tech works frame-by-frame, using your previous movements and actions to keep the world coherent.
That means if you revisit a spot after a minute, it still looks how you left it.
You can also type in extra prompts to change the world, like triggering weather, adding characters, or setting up new events.
DeepMind has been testing Genie 3 with its SIMA agent to explore how AI systems perform tasks in these environments.
The agent sends movement commands while Genie 3 simulates the results, no pre-programmed goals or scripts.
There are still limits (obviously):
Agents can’t perform many actions yet
Multi-agent interactions are at an early stage
Real-world geography isn’t exact
Simulations last a few minutes, not hours
NPCs better step up
DeepMind is releasing Genie 3 as a limited research preview.
A small group of creators and academics will test it while the team gathers feedback and studies risks.
POV: You asked AI for a haunted forest and now you’re stuck in it. -MG