Genie 4: A Real-Time Generated World, and Its Two-Minute Memory

Google DeepMind has released Genie 4. It is not a game but a "world model": everything you see, including any scene, is generated by AI in real time.

In the demo video, you are riding on a dragon's back through mountains, surrounded by other riders. The visuals are not pre-rendered; the model computes them instantly based on your actions. The WASD prompt in the lower-left corner tells you that you really are "playing" this generated scene.

Genie 4 can generate more than visuals. Graphics, physics, ambient sound, music, and character dialogue are all produced in real time within the same model. More important is the state between frames: the model remembers your actions in the scene for a period of time. DeepMind officially says visual memory lasts about two minutes. That means if you push a box away, it does not automatically reset after you turn around.

It is currently in closed beta. You can move with a keyboard and mouse, and you can also change the world using natural language. For example, you can change the weather, move the camera, or add a character. Gemini-series models handle these instructions and keep the scene coherent.

DeepMind's blog splits the core capabilities into four areas: shaping the world in real time with natural language; more realistic physics and character motion, including airflow, collisions, and crowds; native audio so characters can converse; and multi-step event simulation guided by Gemini.

The third one is particularly interesting: characters are not pre-recorded; they speak live based on your input. This is already close to the experience of "stepping onto a film set."

Some online commenters said the most surprising thing is the state between frames. Two years ago, AI video could not even keep fingers consistent; now a world can hold together for several minutes. Some said this means that in the future you may not need to download a game; it could simply be generated online in real time. Others said two minutes of memory is the current boundary: that may not be enough for a game engine, but it is enough for prototyping.

The official demo also showed Genie 4 working with Gemini 4 Argon to simulate a humanoid robot playing chess. This points toward robot training scenarios.

Genie 4 is still early. Over longer runs, visual details and object existence drift, and complex physics and text rendering are unreliable. It is first being opened to trusted testers, and later it will expand to Google AI Ultra subscribers and Gemini API partners. All generated content carries a SynthID watermark.

It is not yet a game engine, but it has already made "from one sentence to an explorable world" a reality. The next question may be: when a world can be generated at any time, what is left of the game itself?