Google’s Project Genie Opens a Door to Worlds That Spring From a Single Sentence or Snapshot

DeepMind has been working on a new experimental tool called Project Genie, which is a fairly cool prototype that you can test in a web app via Google Labs. To gain access, you must first obtain a Google AI Ultra subscription, which is currently only available to anyone aged 18 and over in the United States. The system employs a general-purpose world model known as Genie 3, as well as some fancy tools such as Nano Banana Pro (for creating some very nice graphics) and Gemini (which is intended to function as the operation’s brain).

First, you can begin sketching out a world by simply describing the desired location, adding some information for a character, and selecting where you want to be in relation to all of this, first person, third person, or whatever perspective you choose. Then there are a variety of possibilities for getting across the planet, including walking, flying, driving, and even riding specialized vehicles. Nano Banana Pro is always running in the background, ready to generate a short preview image from anything you’ve typed or uploaded, so you can always go back and make changes, change the colors, add or delete items, and so on. Until you’re satisfied with the scene, at which point you’ll be allowed to step in.

That’s when Genie 3 takes over, and the world begins to emerge before your eyes, at a smooth 720p and roughly 20-24 frames per second, which is quite quick if you ask me. The truly brilliant element is that Genie 3 is continually producing what will happen next, so no matter how you move the character, the scene remains consistent. Paths begin to form as you progress, physics act realistically (most of the time, at least, but I’ll get to that in a minute), and memories from earlier stages of your journey are kept so you can always find your way back to where you were. Sessions can last up to 60 seconds each, but there will be some lag as Genie 3 catches up.

The examples available online might give you an indication of what is possible, with the majority of them being completely mind-blowing. Want to envision an unending ocean with a surfer riding massive waves? Done. Or perhaps you’d prefer to witness an alpinist ascend massive snowy peaks? Easy breezy. Perhaps you’d rather to watch a paper airplane soaring through stormy skies or a bird flapping its wings above a rainforest canopy? Users have previously completed all of these tasks, as well as more mundane tasks such as constructing backyard racetracks for toy cars, creating claymation gardens monitored by ladybugs, and organizing midnight hunts with torches. There’s also a curated gallery where people have showcased their finished work for all to see, as well as a randomizer button that allows you to find fresh and surprising masterpieces.

Obviously, there are certain limitations. You won’t always receive a world that’s completely realistic because physics might be erratic at times, characters may respond more slowly or with less precision than you’d like, and details may not always match what you specified. Some rather advanced capabilities for Genie 3 are being discussed but have yet to be realized, such as the ability to initiate events during a session to change the weather or add items. Don’t get me started on the content restrictions, which include preventing exact copies of previous games.