Genie 4: Google lets you steer AI worlds with sound in real time

What this is about

With Genie 4, Google DeepMind has unveiled the next generation of its world model. A world model is an AI that doesn't just generate a finished video. It simulates an environment you can move through and interact with. Think of a video game, except the model generates everything live. According to a report from 36Kr, which draws on a post on WeChat, the announcement came overnight before October 29, 2026. Genie 4 is so far open only to a small group of testers. Apart from the official demos, almost all hands-on reports come from X.

Google CEO Sundar Pichai shared the news and, according to the report, called Genie 4 "the most realistic and interactive world model to date." Some users even cried, "Game engines are dead, long live world models." That's an exaggeration, as the limitations below show.

What's new compared with Genie 3

The coherent duration of a single world grew from 1 to 2 minutes. The bigger deal, though, is sound: Genie 4 is Google's first real-time world model with native, synchronized audio and video generation. Wind noise, explosions and even human voices are computed by the model in real time to match the picture.

The second major change is editability. Genie 3 already had "promptable world events," which are events inserted by text, such as a rider or a hot-air balloon on a prairie. According to the report, that felt fairly basic. Genie 4 responds much better: you can change the weather, move the camera, and new elements blend into the scene. In a demo with a couple riding a motorcycle, the tester typed "explosion in the background" and then "giant alien tentacles that can destroy the world." Soon after, things blew up and tentacles wreaked havoc. Another demo shows a ride on a black dragon at 140 km/h, where a castle fires fireballs and arrows on command. The two-minute, uncut footage is generated live while a person steers.

Hybrid World Models: Gemini as supervisor

For more complex changes, Genie 4 alone isn't enough. DeepMind calls the overall system "Hybrid World Models." Genie 4 generates picture and sound in real time, while Gemini works in the background to keep the world running by logical rules. Gemini Omni is meant to keep the world closer to reality, including the look of the scene and the movement of characters and objects. Gemini 3.8 Flash handles more demanding jobs. It watches the live image Genie 4 produces, breaks a task into steps, checks whether the result matches expectations, and corrects course when things change along the way.

Physics and limits

Genie 4 is said to simulate physics such as aerodynamics and collisions much more realistically. As proof, Google showed a slow-motion test: a marble bust falls onto a watermelon, which bursts, spraying pulp and juice. Critics noted, however, that the bust should have tipped over according to the laws of physics, and it doesn't. In a demo at a rural Japanese train station, the conductor speaks English. According to the author, the generated speech still sounds unnatural, and the facial expressions aren't fluid.

Google itself admits this is a very early product. If the model runs too long, or you switch perspectives and commands often, objects and details in the scene can "drift," meaning they change in unintended ways.

Use cases: games, film, robotics

In gaming, users with test access created an open world in the style of GTA and dubbed it "GTA Waymo." In film and TV, you could lock in camera angles and character movement in advance in the virtual world.

According to the author, robotics is especially interesting. Robots struggle with tasks like folding laundry, partly because high-quality training data is expensive: humans have to demonstrate the movements over and over, including motion and force. Genie 4 could simulate scenarios virtually first. Google tested this with the robotics company Apptronik: "Gemini 4 Argon" (the name the source uses) worked with Genie 4 to simulate a robot playing checkers. Given the general instruction "Play checkers with the boy opposite you, move a red piece," Argon broke the task into 22 steps, which Genie 4 rendered: lower the head, reach out a hand, grab the piece, lift the arm, move, set it down.

Market context

The report points to a lot of movement around world models. Runway launched GWM Worlds 2 with multiplayer interaction, AMD reportedly acquired Fei-Fei Li's World Labs for $8.2 billion, and Amazon is said to be planning $7 billion for the startup Decart. These claims come from the source and have not been independently verified here.

Our take

Genie 4 shows how quickly interactive video generation is advancing: sound, live intervention and a link to language models as planners. Still, two minutes of consistency is a long way from a real game, and the physics and speech errors are visible. You can't try it yourself yet, and the source doesn't give pricing, access details or technical specs. The robotics idea is promising, but it was shown only in simulation, not proven on real robots. The source is also an enthusiastic media report, and independent tests are still to come.