Written By | Torch Media Africa
DeepMind's newest world model generates music, ambient sound and spoken dialogue in real time
Search
SCIENCE & TECH
Imagine typing a sentence and watching a living, breathing world appear on your screen — complete with the crunch of footsteps, the murmur of wind through trees, a soundtrack that swells at the right moment, and characters who turn to look at you and start talking. Now imagine changing your mind halfway through: you type "make it storm," and the clouds roll in without the scene ever freezing or restarting.
That is the pitch behind Genie 4, the latest world model from Google DeepMind, unveiled on 27 October 2026. Google describes it as a real-time interactive world model that generates audio natively — music, ambient noise and interactive dialogue are produced alongside the visuals rather than bolted on afterwards. According to AI Weekly's write-up of the announcement, it is billed as the first real-time world model to do so.
It is not yet something you can download. Genie 4 is a research preview for trusted testers, with a waitlist open. But the demo has already set the tech world buzzing, and for a very good reason: the last time Google showed off a Genie model to the public, video-game stocks tumbled in a single trading session.
What Genie 4 Actually Does
A world model is an AI system that does not just produce a picture or a video clip but simulates an environment that responds to you. Walk forward and the scene updates. Turn your head and the world you left behind is still supposed to be there. Genie 4 pushes that idea further in three ways, according to DeepMind's own post.
First, sound. Earlier Genie models were silent. Genie 4 generates music, ambient audio and dialogue as part of the same process that creates the pictures. DeepMind's blog calls human speech "the true frontier," and says characters can converse and react to a user's inputs or to events in the world in real time.
Second, live steering. Users can change a running simulation with plain natural language — altering the weather, moving the camera, or adding a person or object. Google says Gemini adapts each prompt so it fits the scene and the world keeps progressing rather than resetting.
Third, physics. Genie 4 simulates scenarios such as aerodynamics and collisions, and its characters move and respond to their surroundings. Google is careful to say that this is still an active area of research, and that physical interaction is not yet perfect.
The Gemini Engine Under the Hood
Genie 4 is not a lone model. DeepMind says it is stitched together with several members of Google's Gemini family. Gemini Omni anchors how the world looks and behaves, while other Gemini models handle the reasoning needed to keep the scene coherent and continuous.
In the real-time loop, a lighter model — Gemini 3.8 Flash — acts as a kind of on-set director. It watches the scene as it is generated, breaks the user's goals into steps, and checks whether each step actually succeeded. Think of it as a continuity supervisor who never blinks.
The research team credited in coverage includes Raia Hadsell, Diego Rivas and Hang Qi. Google has not disclosed technical figures such as resolution, frame rate, inference cost or maximum session length, according to AI Weekly — details that will matter enormously once developers want to build on top of the system.
From Game Demos to Robot Brains
The headline-grabbing use is entertainment, but DeepMind is pitching far wider ambitions. Testers are already using Genie 4 for previsualization, turning single archival photographs into explorable simulations, prototyping game concepts and generating scenarios for enterprise simulation.
The most eye-catching demo, though, involves robots. Google says it is testing Gemini 4 Argon together with Genie 4 and Gemini Robotics. In one demonstration, a simulated Apptronik Apollo humanoid played a game of checkers, with the task broken down into steps in real time. The idea is that a robot can practise inside a generated world — failing safely, thousands of times — before it ever touches a real chessboard, or a real factory floor.
There is precedent. Waymo adapted Genie 3 to generate rare driving situations — tornadoes, floods, wildfires, animals wandering onto the road — producing simulated camera and lidar data to test the Waymo Driver, according to reporting by Startup Fortune. Rare, dangerous scenarios are exactly what self-driving cars struggle to encounter often enough in real life.
The Genie That Spooked Wall Street
To understand why Genie 4's debut matters, rewind to the earlier models. Genie 3 launched in August 2025, generating navigable, text-prompted 3D worlds at 720p and 24 frames per second. Scenes held together for a few minutes, it modelled basic gravity — and it had no sound at all.
In May 2026, Project Genie, Google's consumer-facing website for the technology, was connected to Street View so users could wander through environments based on real places. Early access capped generations at 60 seconds with limited character control and originally required the $249.99 AI Ultra plan; Google later added a $99.99 monthly Ultra tier and cut the higher tier to $199.99, per Startup Fortune.
Then came the stock-market shock. In early February 2026, after Project Genie drew attention, game-industry shares plunged. Startup Fortune, citing Sherwood News and Stockopedia, reports that Unity fell as much as 22.5% to 24% in one session, Roblox dropped around 13%, Take-Two between 8% and 10.5%, CD Projekt 8.9% and Nintendo 4.79%.
Not everyone agreed the panic was justified. Unity CEO Matthew Bromberg argued that the output was unsuitable on its own for games that need consistent, repeatable player experiences, because the frames are probabilistic. Analyst Piotr Poniatowski of mBank called the sell-off "unjustified." By the following Monday, Unity, Roblox and Take-Two had recovered at least 3% of their losses.
The Fine Print: Drift, Glitches and Guardrails
For all the dazzle, Google is refreshingly upfront about what Genie 4 cannot yet do. Its own post says visual details and object permanence "can still shift or drift over time" — particularly in longer sessions, after multiple prompt edits, or while navigating. Real-time audio accuracy, legible text, spatial accuracy and long-term consistency are all listed as areas of active research.
That caveat is at the heart of the gaming-industry rebuttal. A traditional game engine guarantees that the door you opened a minute ago is still open. A world model that is generating everything on the fly is, at the moment, making an educated guess.
On safety, Google says Genie 4 includes guardrails aligned with its policies and went through internal safety evaluations before the preview, which is why it launched first to trusted testers. All generated video and audio carries SynthID watermarking, Google's invisible tag meant to help identify AI-made content. Given that the model can produce convincing speech inside photorealistic scenes, that provenance marker will be closely watched.
Expert Perspective
Outside analysts are likely to focus on two questions: how far the technology is from replacing traditional engines, and how much it matters beyond games. The reaction to Genie 3 provides a template. Markets initially read the technology as an existential threat; industry leaders such as Unity's Bromberg countered that probabilistic generation is a poor fit for experiences that must behave the same way every time.
Genie 4 does little to settle that argument — Google itself acknowledges drift — but it shifts the debate. Sound and dialogue make generated worlds feel like places rather than silent clips, and live steering turns the user from a viewer into a director. The nearer-term opportunity, based on what testers are reportedly doing, looks like previsualization, rapid prototyping and simulation rather than shipping finished games.
What This Means For You
If you are a gamer, do not expect to be playing a Genie-generated blockbuster next Christmas. The tool is in preview, no public release date or developer pricing has been announced, and the consistency problem is unsolved. But the prototypes that studios build with tools like this could shape the games you buy in the years ahead, and indie developers may find the cost of testing an idea falling sharply.
If you drive, or ride, the impact may come through the back door. Simulated worlds that conjure rare hazards help engineers test autonomous vehicles without waiting for a real-life tornado. The same logic applies to warehouse robots and home helpers.
If you subscribe to Google's AI Ultra plan, you are first in the queue: Google says Genie 4 is coming to Ultra subscribers through Project Genie, as well as to early-access partners in the Gemini API and Google AI Studio.
And if you simply scroll social media, brace yourself. Synthetic video with synthetic voices is becoming easier to make and harder to spot. SynthID watermarking is a welcome safeguard, but it only works if platforms and people check for it.
The Bottom Line
Genie 4 is not a finished product, and Google does not pretend it is. What it is, is a clear signal of where the race is heading. The AI industry's first wave taught machines to write; the second taught them to draw and make video; the third — world models — is teaching them to build places you can step into, talk to, and change with a sentence.
Whether that becomes a creative revolution, a robotics breakthrough, or another false alarm for the share price of game makers will depend on the dull but decisive question of consistency: can the world still be there when you look back? For now, the genie is out of the lamp — but it is still being kept firmly in the hands of trusted testers.
