AI - Google Genie 4, worlds you can explore and change as you go
Google has announced Genie 4, and apparently generating a video is no longer enough. Now we need to walk into it, change the weather and have a conversation with the people inside 🤣
The examples are impressive, but the reactions range from “what is the point?” to Google having already built GTA 7. Somewhere between those two is a genuinely interesting research release, so let us have a look 😊
Genie 4
Genie 4 is Google DeepMind’s latest interactive world model. Rather than producing a fixed clip that you watch from start to finish, it generates a simulation that responds as you move through it and change what is happening.
You use keyboard controls to navigate and control the camera, then natural-language prompts to steer the scene. That could mean changing the weather, introducing a character, or setting an event in motion. The world also has native audio, including ambient sounds and characters who can respond with speech.
That distinction matters: this is not just another text-to-video announcement. But an explorable generated scene is not automatically a finished game, a downloadable 3D model or an accurate simulation of the real world either.
Official announcement
Google published Genie 4: A new generation of world simulation on 27 October:
This is Genie 4, our most realistic and interactive world model yet. 🌐
Explore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.
Sundar Pichai highlights real-time control, native audio, realism and Gemini reasoning:
What is new?
Google groups the improvements into four areas:
- Steer the world while it runs: change the scene with text instead of starting again with a new generation. Gemini interprets the request in context and helps keep events progressing between your inputs.
- More realistic movement and physics: Google highlights human motion, collisions and aerodynamics. These are research claims about improved simulation, not a guarantee of physically accurate results.
- Native audio and dialogue: environmental sounds and interactive speech are generated alongside the world, rather than simply adding a soundtrack afterwards.
- Multi-step events with Gemini: Gemini 3.8 Flash acts as a guiding agent in the live loop, observing frames, breaking goals into steps and adjusting its plan as the scene changes.
Google says Gemini Omni helps anchor the simulations in real-world appearance and behaviour. It is also experimenting with Gemini 4 Argon and Genie 4 for robotics scenarios, including a simulation of Apptronik’s Apollo humanoid playing checkers. That is a simulated example, not evidence that a physical robot has learned the task from Genie.
Google AI explains the Gemini side of the system, while Google’s main account shares the launch demo:
David Pantera, who works on world models at Google DeepMind, adds some context on how Gemini steers the scene:
What would we actually use it for?
For me, the more interesting part is not whether this replaces a game engine tomorrow. It is how quickly you could explore an idea before spending time building the real thing.
Google says trusted testers are trying it for film pre-visualisation, interactive experience prototypes, historical environments based on photographs, and early simulation scenarios. Being able to change a camera angle, lighting or action while exploring a scene sounds useful as a creative draft, even if it is nowhere near a finished production asset.
Justine Moore shared an early-access example generated from an environment and character prompt:
Rourke Heath highlights turning a starting image into an explorable environment, while Min Choi shares examples and prompts:
The research angle is interesting too. Chris discusses world models as places where agents might rehearse actions and improve their understanding of space and physical interactions:
That is a potential direction, not proof that Genie 4 solves those problems. A simulation used to test real-world decisions needs more than convincing pixels; the outcomes need to be reliable too.
The examples are getting wonderfully strange
Ethan Mollick has already gone from an office covered in DOOM to a Victorian tea party with a flying mouse. As you do 🤣
His second example includes a useful qualification: it is glitchy, even if the progress is remarkable. Also, a DOOM-themed generated scene is not the same thing as executing the actual DOOM game.
Aleksander Holynski, a Google DeepMind researcher, shared a rather different experiment: talking to a generated younger version of himself. An imaginative use of the technology, although obviously not a reconstruction of what someone’s younger self would really have said.
Salio and Bee are enthusiastic about how the worlds look and the possibilities for exploring them:
Scepticism, GTA 7 and the end of the games industry
Mr Ash is sceptical, to say the least 🤣 But there is a reasonable question underneath it:
Still don’t understand the point of Google Genie.
Can we turn them into games? Can we turn them into websites? Can we turn them into 3D models?
What are we supposed to do by moving around an imaginary world using WASD keys?
I think Google’s examples of creative drafts and research environments answer part of that. However, the announcement does not establish a general workflow for exporting a production-ready game, website or editable 3D asset. Those are different deliverables from an interactive simulation.
At the other end of the scale, Ziwen says Google has built GTA 7 before GTA 6. Bold claim 🤣🤣🤣
Nick Dobos goes further, predicting that multiplayer could overturn Steam, the console platforms and the game-engine business:
That is speculation, not something Google has delivered here. Games need persistent state, dependable rules, progression, authored experiences and plenty more besides an impressive environment. Adding multiplayer would not magically solve all of that.
Aakash Gupta offers a different perspective on world models versus game engines. His post also makes market and partnership claims; those are his commentary, not results from Google’s Genie 4 research:
Google itself acknowledges that visual details and object permanence can drift, especially over longer sessions, after complex edits or while navigating. It also lists physical interactions, audio accuracy, readable text, spatial accuracy and long-term consistency as ongoing research problems.
That is quite an important caveat if your game relies on the sword still being where you left it 😉
More reactions and the research discussion
TestingCatalog, Andrew Curran, Bee and RoundtableSpace cover the announcement and access plans. The important distinction remains preview now, broader access coming soon:
And naturally, we have already arrived at the simulation hypothesis. Vittorio and Demis Hassabis both lean into the rabbit hole, albeit from rather different positions 🤣
For more on the research, Raia Hadsell shared the team’s discussion on Release Notes:
My take
I find this more interesting than another video-generation demo because you can intervene while the scene is unfolding. Native audio and responsive characters make that a much more compelling idea than just wandering through a silent generated landscape.
But I am not ready to declare the games industry dead, or pretend a convincing-looking world must understand physics correctly. For now, this looks like an interesting creative and research tool with some very real limitations. Let us see what people can actually build with it once access opens up 😊
Useful links
#Blog #AI #Google #Google Deepmind #Genie 4 #World Models #Gemini