US Google AI Ultra subscribers can now access Project Genie, DeepMind's experimental prototype that creates playable 3D interactive worlds from text prompts or images. It's built on Genie 3 world model, Nano Banana Pro image generation, and Gemini.

What Project Genie Actually Does

You give a text prompt for the environment and character, tweak the Nano Banana Pro-generated image if needed, then Genie builds an explorable 3D world. Navigate in first- or third-person, jump, look around, remix worlds, or browse the gallery. Sessions are limited to 60 seconds due to compute demands.

Why this actually matters to YOU

Imagine typing "claymation candy castle in the clouds" and instantly exploring it. This could make game creation effortless for anyone, ignite creativity for kids and hobbyists, and eventually enable realistic simulations to train robots and AI agents in everyday scenarios.

Wait, What Is a World Model?

A world model is an AI that builds an internal understanding of an environment, predicts future states, and plans actions. DeepMind sees these as a critical step toward AGI because they enable AI to reason about the physical world the way humans do.

The Tech Behind Genie

Genie 3 is auto-regressive and remembers previous generations for consistency. Nano Banana Pro handles image creation and edits. Gemini adds reasoning. Together they create dynamic environments, though the tool remains experimental and sometimes inconsistent.

The Bigger Race for World Models

The world model race is heating up. Fei-Fei Li’s World Labs launched Marble last year. Runway recently released one. Yann LeCun’s AMI Labs is focused here too. DeepMind opened Genie 3 five months ago and now seeks broad user feedback to accelerate progress.

Looking Forward

Better world models could give AI realistic physics understanding and action planning. Short term: fun creative tools for everyone. Longer term: simulation environments to train embodied agents and robots before real-world use.

Bottom Line

World models like Project Genie are a necessary step toward AGI. By learning to understand, predict, and interact with environments coherently, they build the foundation for AI that can reason about the physical world like humans do.