Google DeepMind has unveiled Genie 4, the latest version of its world model, which the company says improves on Genie 3 in almost every way. Those willing to pay Google enough may get a chance to use it before too long.
The earlier version made waves last year. When DeepMind revealed Genie 3, gaming stocks tumbled. Since then, world models have become a major focus of AI labs around the world, but Google has kept its researchers working on a successor rather than resting on that earlier success.
It is easy to look at what Genie 4 produces and see a video game, which helps explain why Genie 3 caused such consternation in the gaming industry. DeepMind does not see it that way. According to Google, world models are a key element in the march toward artificial general intelligence, or AGI. A world model creates a virtual world without pre-rendered environments or programmed interactive elements. A user simply tells the model what they want, and it generates the experience frame by frame, in real time, as they explore.
DeepMind says the technology could supply infinite training environments for physical robots and open new avenues for rapid prototyping in creative fields. That promise depends on the model accurately simulating what a user asks for. Genie 4 includes several improvements meant to make its simulations more useful, and DeepMind credits Google's updated Gemini models for part of that progress.
When a user prompts a world into existence in Genie 4, the system relies on Gemini 3.8 Flash as a guiding agent. The agent breaks goals down and evaluates each output frame, which helps keep the simulation on track. The result is that users can move through varying environments without the model losing the original context.
There is still a notable limit. DeepMind notes that Genie 4 worlds remain consistent for only a couple of minutes before the model forgets how things started.
As with Genie 3, the new version lets users change parameters midstream. They can alter the weather, add elements or move the camera, and Gemini helps steer the simulation with each update. The approach resembles the way reasoning models generate additional internal prompts to zero in on the right output tokens.
