Google DeepMind just dropped Genie 4, and for once the demo video doesn’t need a disclaimer saying “sound added in post”. Because this time the model makes the sound itself.
Quick recap for those who skipped the last episodes: Genie is DeepMind’s “world model” line. You type a prompt, you get a playable 3D-ish world you can walk around in with your keyboard, and the model hallucinates the next frame on the fly. Genie 3 already did that in real time. Genie 4, announced on October 27, adds the missing senses.
What Genie 4 actually does
The headline, in DeepMind’s own words: “Genie 4 is our first real-time world model to generate audio natively.” Ambient noise, music, and, more surprisingly, dialogue. Characters in the simulated world can talk to you and react to what you do. You can also edit the scene mid-session with plain English (make it rain, add a dog, move the camera) and the world keeps rolling while it absorbs the change.
The other big piece is that Gemini sits in the loop as an agent. Gemini 3.8 Flash watches the simulation, breaks a goal into steps, checks each one and replans if you mess with the scene. DeepMind’s showcase for this is a simulated Apptronik Apollo humanoid playing checkers, driven by a “Gemini 4 Argon” model, with task steps popping up as captions. The pitch is obvious: infinite, cheap, safe training grounds for robots and agents, with no physical robot to break.
Physics got an upgrade too, with aerodynamics and collisions, and “fluid” human motion. I’ll believe the fluid part when I see a generated character climb stairs without a leg phasing through the handrail, but the clips look a lot less like a fever dream than last year.
Genie 4 and the numbers that aren’t there
Now the cynical bit, because someone has to. The announcement has zero hard specs. No resolution, no frame rate, no session length, no benchmark. Genie 3 at least told us 720p at 24 fps for “a few minutes”. Genie 4 tells us it’s “enhanced”. That’s not a spec, that’s a shampoo label.
To be fair, DeepMind is unusually candid about the cracks: “visual details and object permanence can still shift or drift over time,” especially in long sessions or after several edits. Legible text, spatial accuracy and real-time audio fidelity are all filed under “active research”. So the dog you added may become a bench if you look away for too long. Welcome to world models.
Availability is the usual Google funnel. Research preview for trusted testers now, then Google AI Ultra subscribers via Project Genie, then early-access partners in the Gemini API and AI Studio. There’s a waitlist form. No pricing. Everything it produces gets a SynthID watermark, which is the least they can do when the output is a fake world with fake people saying fake things.
Why I still care
Because the testers’ use cases are the honest ones. Film crews doing previsualisation and running five “takes” of a scene before renting a single camera. Photojournalists turning archive photos into walkable environments. Game devs prototyping a mechanic from a sentence to see if it’s fun before writing a line of code. None of that needs object permanence to be perfect. It needs “good enough, right now, for free”, and that’s exactly the bar these models clear.
Where I’d keep my expectations in check is the robotics angle. Training a robot in a world whose physics drift when you blink is a fascinating research problem, not a product. And if you’ve followed what agents get up to when nobody’s watching, “Gemini plans the steps and verifies them itself” is a sentence you’ll want to read twice.
So, my take: Genie 4 is the first version of this tech I’d actually want to mess around with for an afternoon. Which, for a world model, is a bigger compliment than it sounds.
Genie 4 can finally hear itself think. Now it just needs to remember where it put the dog.
Source: Google DeepMind — Genie 4: a new generation of world simulation
Image: Sara Kurig / Unsplash
Discover more from TechPassion
Subscribe to get the latest posts sent to your email.
