Google’s Genie 4 Pushes World Models Closer to Reality—But Not Yet to Reliability

DeepMind’s Genie 4 adds real-time steering, improved physics and native audio to AI-generated worlds. Google pitches a leap for robots and creators, while acknowledging that simulations still drift and access will be tightly controlled.
Google’s Genie 4 Pushes World Models Closer to Reality—But Not Yet to Reliability

Google’s Genie 4 Pushes World Models Closer to Reality—But Not Yet to Reliability
Google DeepMind unveiled Genie 4 on Tuesday, positioning the new system as its most capable attempt yet to generate explorable, AI-made worlds in real time. The company says users can alter a scene while inside it—changing weather, objects, events or camera angles—rather than treating generation as a finished video clip. Sundar Pichai called it DeepMind’s “most realistic and interactive world model,” highlighting live control over environments and actions.

The release follows Genie 3, which startled parts of the games business last year and helped turn world models into a competitive priority across AI labs. DeepMind’s broader argument is not merely about making games: it sees systems that model environments as a route toward more capable AI, with potentially limitless simulated training grounds for robots and quicker concept development for creators.

Genie 4’s claimed advances are substantial. Gemini 3.8 Flash acts as a guiding agent, helping break down goals and assess frames as Genie generates the world. DeepMind says the result is better handling of collisions and aerodynamics, more natural human movement, and native sound—from ambient noise to dialogue that responds to what happens in the scene. Tejas Kulkarni, Google’s research director, described the direction as “hybrid world models,” combining Gemini’s code and text generation with Genie’s real-time simulation.

But the pitch comes with important limits. Genie 4 can keep a scene coherent for only up to two minutes, even if that doubles Genie 3’s reported limit. Objects and details may drift over longer runs; accurate audio and legible in-world text remain unresolved. Google engineers also say the end-to-end product is still taking shape.

For now, access is a research preview for trusted testers. Google plans phased availability through Project Genie, Gemini API and AI Studio, beginning with AI Ultra subscribers—whose plans start at $100 a month—rather than a broad public launch.

https://foxvector.com

Write a comment