Genie 4: An AI World Where the People Talk Back
▶ Read as storyGenie 4 makes a world you can walk around in, and now it has sound, the people in it talk back, and you can change it by typing while you are inside. You commented GENIE, so here is every demo, with links.
None of this is filmed
01 / What a world model is
A video model makes a clip you watch. A world model makes the next moment of a world in response to what you do: arrow keys, camera, a typed sentence. Frame by frame, live.
DeepMind says it runs “end to end, with no game engine running in the background.”
Same prompt, Genie 3 on the left, Genie 4 on the right. Google DeepMind.
What changed from Genie 3
02 / Genie 3 vs Genie 4
| Genie 3 | Genie 4 | |
|---|---|---|
| Sound | Silent | Music, ambient sound and speech, made with the world |
| Characters | Less controllable | Talk and react to you in real time |
| Typed changes while inside | In the model, not in the public app | Yes, and Gemini fits your prompt to the scene |
| Planning | Not mentioned | Gemini 3.8 Flash plans steps and checks each frame |
| Session length | 60 seconds in the public app | About 2 minutes of consistent scene (TechCrunch) |
- Physics. DeepMind shows aerodynamics and collisions: a watermelon, a falling maple seed, a silk river.
- Type it, it happens. Change the weather, move the camera, add a person, set an event in motion.
- Watermarked. All video and audio from Genie 4 carries SynthID.
A typed prompt changes the scene mid-shot. Google DeepMind.
Every demo, with links
03 / What you came for
From Google DeepMind
- The launch film, 2:27, the best overview
- The Genie 4 blog post, with the dragon flight, motorbike, train station, piano duet, watermelon, maple seed and silk demos
- The Genie model page, with the magician and complex events demos
- Feature thread on X: photorealism, physics, controls, audio, reasoning
- Robot playing checkers: a humanoid robot simulated by Gemini 4 Argon and Genie 4. There was never a real robot
From people who tried it
- Talking to his younger self, a Genie 4 researcher in conversation with himself just out of college
- Walking back into an old photo, a favorite memory with his mom
- A city, then dragons, then an otter, Ethan Mollick building a world one sentence at a time
- An office where every computer runs DOOM, also Ethan Mollick
- One prompt, a full world, Justine Moore’s early access demo
- Recording takes on a live set, Bilawal Sidhu
- 10 examples with their prompts, Min Choi’s thread
- An honest hands-on, Matt Wolfe on where steering still slips
“Neogothic brutalist city inhabited by robots and horses.” Ethan Mollick in Project Genie.
Why Google built it
04 / Robots first
Officially, to train robots and AI agents. DeepMind calls simulating the world “a crucial aspect of building AGI.”
Gemini 4 Argon plans the moves, Genie 4 simulates the robot. Google DeepMind.
But look at the demos. A game you make by typing. A scene you test before you shoot it. An old photo you walk back into. That is where this lands for the rest of us.
How to get in
05 / Access
- Join the waitlist. Genie 4 in Project Genie, Google’s form.
- Watch for Project Genie. It is coming to Google AI Ultra subscribers. Genie 3 there needed the US, 18+ and the Ultra plan.
- Builders: watch the Gemini API and Google AI Studio. Early access partners first, pricing not announced.
Before you get excited
The fine print
- Trusted testers only for now. No public release date.
- It drifts. DeepMind says details “can shift or drift over time”, more so in long sessions or after many edits.
- A few minutes per session, not hours.
- Text and real places are weak. Clear text mostly appears only when it is in the starting description.
A year ago this was a silent world that fell apart after a minute. Now it talks back.
My take
Sources
Where this comes from
- Genie 4: a new generation of world simulation, Google DeepMind
- Genie model page, Google DeepMind
- Genie 3: a new frontier for world models, Google DeepMind
- Project Genie launch, Google
- TechCrunch on Genie 4
- Ars Technica on Genie 4
What people asked
Questions
What is Genie 4?
Google DeepMind’s newest world model. It makes an interactive world from a prompt, in real time, with sound and characters that speak, and you can change it by typing while you move around inside.
Can I use Genie 4 today?
Only trusted testers have it. Google says it is coming soon to AI Ultra subscribers in Project Genie and to early access partners in the Gemini API. You can join the waitlist now.
Is Genie 4 a video generator?
No. A video generator makes a fixed clip. Genie 4 makes each frame live from what you do, so you can walk around, look back and change the scene.