Google's Project Genie turns prompts into explorable, physics-aware worlds

Project Genie is Google DeepMind's research prototype for generating interactive worlds from a prompt. You sketch a world with text and images, pick a first or third-person view, then move through it while the system generates the path ahead in real time based on what you do. You can remix existing worlds or download a run as video.

The interesting part is the stack underneath. It runs on Genie 3 as the world model, Nano Banana Pro for image generation, and Gemini for language understanding, simulating physics and interactions for clips up to 60 seconds. Google is direct about the limits: worlds can lack photorealism, physics can be wrong, character controls can feel unresponsive, and latency is real. Access is limited to Google AI Ultra subscribers in the US.

This is a prototype, not a product, and the post treats it that way rather than overselling it. Read the announcement on the Google blog.

Why it matters

If you work on games, simulation, or robotics, the signal is that real-time world models have left the paper stage and are something you can poke at. The stated weaknesses are the useful part: they tell you exactly which problems are still open before you build on this.