Google DeepMind’s world-generation project has reached its fourth version, and the biggest change is that its worlds are no longer silent. Genie 4 creates interactive environments that users can move through and edit on the fly, and it now generates sound and dialogue alongside the visuals.
That distinction matters because Genie isn’t a video generator in the usual sense. Tools like Veo or Sora produce a fixed clip that you watch. Genie produces a scene that reacts as you explore it with keyboard and camera controls, more like a game engine improvising each frame than a finished film. Genie 3 established that idea last year, and version four builds on it with text commands that change the scene while you’re inside it. Typing a prompt can change the weather, add a character or object, or move the camera, with Gemini working those changes into what’s already on screen.
The audio is generated in step with the visuals, covering everything from wind and waves to music and full conversations, and characters can respond to events and to the user while the simulation keeps running. Google also says physics and human movement look more natural, though that’s the company’s own assessment for now.
The more interesting part is under the hood. Gemini 3.8 Flash acts as a supervising agent that watches incoming frames, splits a goal into steps, checks whether each step worked and adjusts if it didn’t. DeepMind is also testing Gemini 4 Argon with its robotics research to simulate tasks that play out over multiple stages. That points to where the real value of world models may lie. Simulated worlds are a cheap, safe place to train robots and AI agents, which is why companies from Google to Nvidia are investing heavily in this area.
Early testers are already trying it for previsualization in film, where directors can test angles, lighting and dialogue before shooting, and for game prototyping, where a developer can try an idea before writing any engine code.
Google is being unusually upfront about the limits. Details can drift over longer sessions or after complicated edits, objects may not stay consistent as you move around, and physics, audio accuracy, legible text and spatial accuracy all still need work. Everything Genie 4 produces will carry Google’s SynthID watermark.
Access is limited to trusted testers for now. A broader preview is planned for Google AI Ultra subscribers through Project Genie and for early-access partners via the Gemini API and Google AI Studio, but there’s only a waitlist and no date. Until consistency improves, Genie 4 looks less like a replacement for game engines and more like a very capable sketchpad, and a valuable training ground for the robots Google wants to build next.
