Google's Genie 4 Builds Worlds You Can Talk To

Google DeepMind has announced Genie 4, the latest version of its "world model": an AI system that generates a scene you can move through and interact with in real time, rather than a fixed video clip. You describe a place or upload an image, then walk through it with the keyboard while the model draws each new frame as you go.

The new version, announced on 27 October, adds three things Genie 3 lacked, according to DeepMind's announcement. Every world now comes with its own sound, from wind and waves to music. Characters can speak and answer back. And users can steer a scene while it runs by typing plain instructions, such as changing the weather, moving the camera or bringing in a new person or object.

Gemini works behind the scenes

Genie 4 leans on Google's Gemini models. DeepMind says Gemini Omni anchors the worlds in how real places look and behave, while Gemini 3.8 Flash acts as a guide in the loop: it watches the scene, breaks a goal into steps and checks each new frame to make sure the step happened. TechCrunch describes Gemini's role as a director, with Genie producing the pictures.

That pairing is where DeepMind sees the field heading. "What we'll increasingly see in the future are these hybrid world models," Tejas Kulkarni, a research director at Google, told TechCrunch: systems that combine the reasoning of language models with pixel-level simulation.

Who it is for

DeepMind pitches world models at three groups. For robotics, a robot's tasks can be tried out in simulation before it touches the real world; DeepMind says it has been testing Genie 4 with Gemini 4 Argon and its Gemini Robotics work. For enterprises, partners are testing it to generate scenarios quickly. And for creative work, testers in film, journalism, education, gaming and engineering are using it to prototype ideas and storyboards, TechCrunch reports.

Games are the sensitive one. When Google began rolling out Project Genie, its consumer app built on Genie 3, in late January, Unity's shares fell about 20% in a day and Take-Two and Roblox lost 10% to 12%, Mobilegamer.biz reported. Unity's chief executive, Matthew Bromberg, responded that Unity "remains the system of record for runtime, distribution, and long-term operations".

More from InnovationEthiopia Leads South Africa on EVs, 60% to Under 0.2%A new 16-country electrification index puts Ethiopia in a higher group than South Africa, after a ban on petrol and diesel car imports pushed EV sales up.

What it still cannot do

DeepMind is open about the limits. Details and objects can drift the longer a session runs, after several edits or as the user moves around. The company lists complex physical interactions, accurate real-time audio, legible text, spatial accuracy and long-term consistency as unsolved. "It's still early days," Diego Rivas, a DeepMind engineer, told TechCrunch.

Every video and sound Genie 4 produces carries SynthID, Google's invisible watermark for AI-generated content, and DeepMind says the model went through internal safety testing before the preview.

• Now: a research preview for trusted testers.
• Coming weeks: Google AI Ultra subscribers in Project Genie, plus early-access partners of the Gemini API and Google AI Studio.
• Not announced: a price, or a date for wider release.