Google’s Genie 3 drops you into a 3D world you can explore and play with as it happens

What you need to know

• Genie 3 lets you create interactive 3D worlds just by typing a short prompt, and you can actually walk around and mess with them.
• These virtual scenes are meant to last minutes, not hours, so don’t expect epic quests or massive open-world vibes.
• While earlier versions couldn’t keep things stable for long, Genie 3 holds it together way better, so stuff stays put even if you leave and come back.

Imagine being able to create an entire, interactive world just by typing a few words. That’s essentially what Google DeepMind's new AI model, Genie 3, can do.

Genie 3 is a world model AI that understands how environments work and can simulate them. Unlike Genie 1 and 2, though, it can create diverse, interactive worlds on the fly from a quick text prompt, Google said in a blog post.

Instead of just spitting out video or images based on text prompts, Genie 3 can now build interactive 3D scenes that you can actually move around in on the fly. You type something in, and within seconds, you’re dropped into a 720p, 24 FPS virtual world that responds to your actions and even holds its shape for several minutes.

Big upgrade from Genies 1 and 2

Earlier versions could only hold things together for 10 to 20 seconds before the scene fell apart. On the other hand, Genie 3 steps it up by keeping objects and spaces intact for over a minute, so if you walk out and come back, everything’s still in place.

In the demo below, a pair of virtual arms rolled blue paint onto a wall. After a few broad strokes, the view shifted away, then back, revealing the paint still exactly where it was left. So, it’s not just flashy visual output, but an actual, dynamic space.

What really sets Genie 3 apart is that it isn’t manually coded to obey real-world physics. Instead, it’s trained in such a way that logical consistency and environmental persistence just emerge.

Instant edits, no reloads

It can also react in real time to new text inputs, like changing the weather mid-scene or adding new elements like animals or objects, all without reloading anything.

Genie 3's interactive, real-time setup makes it a great testing ground for AI agents to learn by trial and error.

Of course, Genie 3 isn’t without its limitations. It’s still in research preview and only available to a small group of academics and creators. The interaction mechanics are pretty basic for now, and it can’t really handle multiple agents running around at the same time.

It also doesn’t create accurate real-world replicas or readable in-world text. And while the tech is impressive, these scenes are built to last minutes, not hours, so don’t expect a full-blown open-world game experience just yet.

Still, Genie 3 is a big step forward in AI simulation. While it’s not ready for public use, it’s giving us a glimpse of where things are headed, especially in the push toward more general forms of artificial intelligence.