• Braun & Brains
  • Posts
  • Genie 3: Google Lets You Walk Inside Your Own AI-Generated Worlds

Genie 3: Google Lets You Walk Inside Your Own AI-Generated Worlds

PLUS: A quick scroll of the Genie 3 videos on my feed.

Author

Rachel Braun
February 02, 2026

Hello World, Happy Sunday! My entire feed has been:

  • The Epstein files

  • Moltbook (will write more on this next week)

  • Genie 3 videos

Because that last bullet point is the one that’s been most scroll-stopping for me, I wanted to share some of the fun. Below I dropped some videos I saw on X that I think you should check out.

Rachel Braun (@rachellbraun)

Project Genie is so sick - it lets you type a description of a world and walk around inside it while Google DeepMind’s Genie 3 model generates the environment in real time as you move. It remembers what it already built, so things stay consistent when you look back. Rn sessions cap at 60 seconds and physics can be off, but still sick. The bigger goal is giving AI agents unlimited training environments by letting them practice in any scenario you can describe.​​​​​​​​​​​​​​​​ (Video by Jake Steinerman)

substack.com/@rachellbraun/note/c-208314027

Before we get into that, let me explain what you’re actually seeing.

  • You type a text description of an environment and a character, and the system generates a 3D-looking world in real time at 720p and 24 frames per second that you can walk around in.

  • There is no pre-built 3D geometry. The model generates each frame autoregressively based on the previous frames and your inputs, with a visual memory that extends about 60 seconds back so the world stays consistent when you turn around.

Use cases are split between consumer and research.

  • On the consumer side it functions as a creative toy where anyone can build explorable scenes from a text prompt or a reference image.

  • On the research side it creates unlimited simulated environments for training AI agents. Instead of hand-building training worlds, you describe a scenario and drop an agent into it. That scales agent training far beyond what is possible with manually constructed simulations. (Google cares about this side of things)

What works:

  • Visual consistency is surprisingly good for something with no explicit 3D model, the style range is wide (photorealistic, cartoon, painterly, cinematic), and you can upload a reference image to guide the look.

What doesn’t:

  • Sessions cap at 60 seconds, physics are approximate so objects can clip or float, readable text in the world is unreliable, and multi-character scenes tend to break down. Access requires a Google AI Ultra subscription ($250/month), you have to be in the US, and you have to be 18 or older. It is still an experimental prototype, so not a shipped product.

With that, please enjoy the scroll.

Genie 3 Scroll

Reply

or to participate.