Google's Genie 4 Conjures Explorable Worlds in Real Time, With a Short Memory

Google DeepMind has unveiled Genie 4, an updated "world model" that builds explorable virtual environments from a text description, according to Ars Technica. The model is meant to improve on last year's Genie 3, and Ars reports that Google may eventually offer access to people who pay for it.

The main limit is memory. DeepMind says Genie 4 worlds stay consistent for only a couple of minutes before the model loses track of how things began.

What is a world model?

A world model creates a virtual space without pre-built graphics or hand-coded interactions. Ars Technica explains that you describe what you want, and the model generates the experience one frame at a time, in real time, as you move through it.

The output can look like a video game. Ars reports that this is why Genie 3 unsettled the gaming industry: its reveal last year sent gaming stocks lower, the outlet said.

What is new in Genie 4?

Ars Technica's headline says Genie 4 promises more realistic physics and more realistic humans. The article also describes a change in how the model stays on track. When you prompt a world into existence, Genie 4 uses Google's Gemini 3.8 Flash model as a guiding agent. The agent breaks the goal into steps and checks each output frame, which is meant to keep the simulation from drifting away from the original idea.

As with Genie 3, you can change things midstream. You can alter the weather, add objects or move the camera, and Gemini helps steer the simulation after each change. Ars compares this to reasoning models that write extra internal prompts to get to the right answer.

What are its limits?

Consistency is the big one. DeepMind notes that Genie 4 worlds hold together for a couple of minutes at most before they forget their starting conditions. That limits how long a single simulated scene can be used without drifting.

The claims about realism and usefulness come from Google and DeepMind. The Ars report does not describe independent testing.

Why does it matter?

For most people, this is a research preview of where AI-generated media is heading, not a product to use today. DeepMind frames world models as a key step toward artificial general intelligence, or AGI, according to Google.

DeepMind also points to practical uses. It says the technology could supply endless training environments for physical robots and speed up prototyping in creative fields. Both depend on the model simulating what a user actually asks for, which is why improvements such as the Gemini-guided checking matter.

For gaming and creative work, Ars notes that world models have become a major focus at AI labs worldwide since Genie 3. Whether that translates into tools for developers or players remains to be seen.

Can anyone use Genie 4?

Not yet, as far as the report shows. Ars Technica says that if you pay Google enough, you might get to use Genie 4 before too long. Pricing, tiers and a release date were not disclosed in the report.

Around Prague, based on reporting by Ars Technica