A film does not care where you look. Every frame was settled before you sat down, and if you turn your head, the screen carries on without you. The newest system from Google DeepMind works the other way round. Until you move, the next moment of its world does not exist. You press a key to walk forward, and the street ahead is invented, as you go, to meet you.

That system is Genie 4, announced on 27 October as a research preview. You describe a place in a sentence, and it becomes somewhere you can walk around, with its own weather, its own people and, for the first time in the series, its own sound. I want to take it slowly, because it is genuinely strange and the claims around it are large.

What a world model is

Most of the AI you have met so far deals in words: you ask, it answers. A world model deals in consequences. In its post on the previous version, Genie 3, DeepMind describes such systems as using an understanding of the world to simulate aspects of it, so they can predict how an environment will evolve and how actions will change it.

A plain way to hold this: you already run one. When you reach for a cup, you know roughly what will happen if you knock it, before you knock it. Nobody taught you the equations. You built a working model of the world from years of watching it, and you consult it constantly without noticing. Genie does something similar, and then something you cannot: it shows you its prediction as moving pictures, and keeps going for as long as you keep acting.

That separates it from the video generators you may have seen. A video tool produces a clip and hands it over. A world model has to stay with you, answering each move as it comes.

What Genie 4 does

According to DeepMind's announcement, written by Raia Hadsell, Diego Rivas and Hang Qi, you navigate with an ordinary keyboard and mouse, and you can type instructions while the world is running: change the weather, move the camera, add a person or an object. Gemini, Google's language model, interprets those prompts and works them into the scene. The post claims sharper photorealism and more fluid movement, some simulated physics such as aerodynamics and collisions, and characters who react to what is around them.

The sound is new. Genie 4 is, in DeepMind's words, its first real-time world model to generate audio natively, which means the crunch of leaves and the roar of a crowd are made with the picture rather than laid over it afterwards. TechCrunch reports that this includes music and interactive dialogue.

There is also a quieter piece of machinery. A Gemini model acts as what The Batch calls a director: it breaks a goal into steps, then checks each new frame to see whether the step happened. Hold on to that detail. It matters later.

Why anyone would want this

The first answer is robots. Teaching a machine to act in the physical world is slow, expensive and occasionally destructive, so researchers have long rehearsed in simulation first. DeepMind describes early work with its Gemini Robotics team, including a simulated humanoid playing checkers, and its model page calls the result an infinite playground for robotics simulations. Software agents could practise there too.

The second answer is creative. DeepMind lists pre-visualisation for filmmakers (the rough sketch of a scene before anyone builds a set), game prototyping and concept design among the uses. Rivas told TechCrunch that the real-time quality helps most with rapid ideation and storyboarding, where staying in a creative flow matters.

The third answer is the large one. DeepMind's model page calls world simulation a key stepping stone on the path to AGI, the broad, human-like intelligence the field has long been chasing. The reasoning: a mind that can picture what happens next, before it acts, has something that a mind which only handles words lacks. It is a serious argument. For now, it remains an argument rather than a result.

What it does not do yet

Here the honest list is long, and to DeepMind's credit, much of it is DeepMind's own. Its page says visual details and object permanence can shift or drift over time, especially in longer runs or after several edits. Legible text tends to appear only when you asked for it. Real places cannot be reproduced accurately, audio accuracy is still being worked on, and interactions between several independent agents remain an open research problem.

How long a world holds together is less clear. DeepMind's page speaks of a few minutes of continuous interaction; TechCrunch reports scenes staying consistent for up to two minutes. Either way, these are short visits.

There is no technical report, and Google has published no quantitative evaluations, only demonstrations it chose itself. The Batch looked closely at one, in which a marble bust dropped onto a watermelon settles neatly on its plinth when it ought to have toppled off. Its editors wondered whether the Gemini director, checking that each step succeeded, had nudged the world towards the intended outcome. For a robot simulator, that is the question that matters most. A machine learns from the plans that fail, and a world that quietly helps every plan succeed teaches it nothing about falling.

Finally, almost nobody can try it. Genie 4 is open to trusted testers only. Google says access is coming soon to Google AI Ultra subscribers through Project Genie and to early partners of the Gemini API and Google AI Studio, and there is a waitlist. Everything it generates will carry SynthID, Google's invisible watermark.

I find myself less interested in whether Genie 4 can build a convincing world than in whether it can build one that says no. Any world worth learning from has to be willing to let you drop the cup.

Sources

Genie 4: A new generation of world simulation, Google DeepMind

Genie, Google DeepMind

Genie 3: A new frontier for world models, Google DeepMind

Google's Genie 4 world model adds audio, more realistic worlds, and real-time steering, TechCrunch

World Models Become Robot Playgrounds, The Batch (DeepLearning.AI)