Google DeepMind lets you generate playable worlds in real time

Most generative AI tools give you something static to look at, like a block of text, a polished image, or a short video that plays once and stops. But Google DeepMind is now testing what happens when the output is a dynamic space you can actually walk through with your keyboard.

Google rolled out access to Project Genie, an experimental research prototype web app. For now, it is available to Google AI Ultra subscribers aged 18 and older in the United States, with plans to expand to more regions later.

Worlds rendered on the fly

Project Genie is powered by a trifecta of Google models: Genie 3, which was first previewed in August, alongside Nano Banana Pro and Gemini. Rather than navigating a pre-baked 3D environment or a static snapshot, Genie 3 generates the world in real time based directly on your inputs.

The tool is structured around three main features: sketching, exploring, and remixing.

To sketch a world, you start with a text prompt or an uploaded image. You can specify a character, choose how you want to move (walking, driving, or flying), and set the camera to first-person or third-person. DeepMind integrated Nano Banana Pro to let users preview and fine-tune their starting scenes before diving in.

Once inside, you can navigate around while the model predicts and renders physics and camera movements step by step. If you run out of ideas, you can remix existing worlds from a curated gallery, hit a randomizer button, and download video captures of your playthroughs.

Big potential, real constraints

Because this is an early Google Labs experiment, DeepMind is upfront about its current boundaries. The prototype is far from a finished game engine replacement.

For one, each generated interactive session is capped at just 60 seconds. You also might notice latency when controlling characters, or find that the physics and visuals deviate noticeably from your prompt or real-world logic.

On top of that, some capabilities teased during the August Genie 3 showcase, such as typing out text prompts mid-session to trigger sudden environmental events, are not included in this release.

Why it matters

Generative media is clearly moving from static outputs toward interactive simulations. While current video generators show impressive photorealism, they do not understand agency. Project Genie proves that a neural model can simulate cause and effect dynamically based on user input, which drastically lowers the barrier for early game ideation and narrative prototyping.

For DeepMind, this is not just about making weird mini-games. Navigating diverse, unpredictable environments is essential for building general-purpose agents and training real-world robotics. Putting an interactive world model in front of paying subscribers gives Google real data on how people test physical boundaries.

We think pairing Genie 3 with Nano Banana Pro for pre-entry sketches is a smart UX choice that tames the randomness of pure prompt-based creation. That said, the 60-second limit and noticeable latency show how computationally brutal real-time world generation still is. It is not a daily productivity driver yet, but it offers a fascinating glimpse into how we might build software and virtual experiences down the road.