Google DeepMind's new model holds a scene together for two minutes, takes text edits mid-simulation and generates its own sound. Rivals are consolidating around the same idea.
Photo by Vlado Paunovic on Unsplash
World models have spent years as impressive clips with no obvious use. Genie 4, from Google DeepMind, is an attempt to change that. It is a simulation you steer while it runs, and it is now showing up in the places where people go to try new software.
It is still a research preview, and Google says plainly that it is early. But the feature list reads less like a lab paper and more like a product spec.
Genie 3 arrived as a research preview a little over a year ago. Genie 4 makes four visible jumps.
First, it holds a consistent scene for up to two minutes, twice its predecessor's limit. Second, it generates audio natively: ambient sound, music and spoken dialogue come out of the same model that draws the world. DeepMind calls it the first real-time world model to do this.
Third, editing is free-form. Users move with standard keyboard controls and type prompts to alter the scene as it unfolds. They can change the weather, shift the camera or add a person or object. Gemini adapts each prompt to fit the surrounding scene and keeps the world moving between inputs.
Fourth, the physics and the people are better. DeepMind says characters move more naturally and respond to their surroundings. A Product Hunt post describes the practical difference: this is not a video you watch but a place you walk through and change.
A world you can edit while standing in it is a different product from a clip you watch.
The more interesting design choice is how Gemini and Genie work together. Genie draws the world. Gemini watches the incoming frames, breaks a goal into steps and checks that each step actually happened. If the user intervenes, it revises the plan.
Google's research director Tejas Kulkarni told reporters to expect more "hybrid world models" built on this split. In one robotics experiment, a Gemini model broke tasks into steps and Genie 4 simulated them live. One example prompt has a Universal Robots UR5 arm dispensing liquid into a lab plate.
That points to a use beyond entertainment. If robot behavior can be tested in a simulation before it touches the real world, a world model becomes a testing tool as well as a toy. Whether the simulations are faithful enough for that remains unproven.
Genie 4 lands in a market that is crowding and consolidating at once. Runway released its second world model, GWM Worlds 2, a couple of months earlier. It is playable and includes audio. Meanwhile, according to TechCrunch, AMD is acquiring Fei-Fei Li's World Labs, Amazon is in talks to buy Decart, and Rivian has acquired Third Dimension to bring simulation in-house for self-driving tests.
Taken together, that suggests world models are being treated as infrastructure rather than curiosities. Chipmakers, retailers and carmakers are all buying in.
The pitched uses are gaming, interactive learning and robotics. Trusted testers have added one more: prototyping films. A DeepMind engineer, Diego Rivas, said real-time generation matters most where creative flow matters, such as ideation and storyboarding. He also said that turning those prototypes into finished assets, perhaps with Gemini Omni, is still a way off.
The caveats are real. Details and objects can drift the longer a simulation runs. The team has not yet worked out how to make the audio accurate or in-world text legible. Rivas also said it is unclear what the end-to-end product experience will be.
Access is staged. Trusted testers have it now. Google AI Ultra subscribers in Project Genie and early-access partners of the Gemini API and Google AI Studio are next, with a waitlist open for Project Genie.
A two-minute limit is still short. If it keeps doubling, long enough sessions to play a level or rehearse a scene could arrive quickly. If drift does not improve, the use cases will stay narrow: quick ideation, short tests, demos. Genie 4 is the clearest sign yet that the field is trying to turn world models into products.
Breakout: Anthropic
Anthropic's S-1 shows revenue exploding. The more telling number is how many big customers now buy both the developer platform and its apps, a pairing that makes accounts harder to pry loose.
Breakout
A viral keynote clip sent preorders for Hugging Face's open-source Microduck past 20,000 in a day. Nothing has shipped yet, but the demand is real and worth watching.