Google Genie 4 Builds Worlds From a Sentence (For Two Minutes)
Game studios spend years and hundreds of millions of dollars building one open world. Google just showed an AI that builds one from a single sentence.
The catch? It starts forgetting the world after a couple of minutes.
On October 27, 2026, Google DeepMind, Google's AI lab, announced Genie 4, its new AI world model. You type what you want, and Google Genie 4 generates a world you can walk around in with your keyboard, frame by frame, in real time.
And you can keep typing while you walk. Make it snow. Swing the camera around. Drop a new character into the street. Google says the scene changes live, without starting over.
No designer. No code. No level editor.
The new trick is sound. Genie 4 is the first Genie that makes its own audio: the wind in the trees, the waves on the beach, music, and characters who actually talk back and react to you.
Which is already more than your group chat does after someone asks who's paying.
Under the hood, Google says Genie 4 is built on its Gemini models. Gemini Omni makes things look and behave like the real world, and Gemini 3.8 Flash watches the scene and plans complicated events step by step. Google also says the physics got better (collisions, things flying through the air), and that the biggest jump is in how people move and react.
So why is Google really doing this? Robots. A robot that practices in a fake kitchen can drop a thousand fake plates before it touches a real one. Humanoids like Figure 03 need exactly that kind of practice. Google also names filmmaking, gaming, journalism and manufacturing.
Game companies are paying attention. According to reports, when Google opened Project Genie, an app built on the older Genie 3, back in January, video game stocks dropped.
Personally, I don't think Genie 4 is coming for your favorite game. But I think it's coming for the first month of every game and every movie, when a team spends weeks just sketching what a scene should look like.
But careful: you can't try it yet. Genie 4 is a research preview for a small group of trusted testers. Next in line, Google says, are people paying for its top Google AI Ultra plan (through the Project Genie app) and a group of early access developers on the Gemini API and Google AI Studio. Google gave no date. There's a waitlist.
And Google admits the weak spots. The longer a world runs, the more details and objects drift. Scenes hold together for about two minutes, and according to a DeepMind engineer, even that isn't guaranteed.
Two minutes. That's roughly a bag of microwave popcorn.
Every clip and sound Genie 4 makes also carries SynthID, Google's invisible watermark for AI content.
Google built a machine that creates whole worlds on demand. Now it needs one that remembers them.
Photos: Google