Genie 4: Google’s World Model Builds Explorable Worlds With a Soundtrack
Trending Topics Jakob Steinschaden ● Covered by 8 sources
Google’s Genie 4 turns text prompts into explorable 3-D worlds, now with sound. It can also change scenes live, but the model still slips on details.
Based on reporting by Trending Topics, Jakob Steinschaden — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google DeepMind has shown Genie 4, the latest version of its world model, and the headline change is hard to miss: the system now generates audio natively. It builds interactive 3-D worlds in real time, lets users move through them like a game, and syncs in ambient noise, music and dialogue as it goes.
That is a step up from Genie 3, which arrived a little over a year ago. DeepMind says the new model can keep a scene coherent for up to two minutes, about twice as long as before according to TechCrunch. It also lets users steer a simulation on the fly by typing instructions such as changing the weather, moving the camera, or adding people and objects.
Google says Genie 4 also handles physics better, with aerodynamics and collisions, and that characters move more naturally and react to what is around them. Behind the scenes, Gemini acts like a director: it breaks tasks into smaller steps, checks the generated frames, and helps maintain continuity while Genie renders the world.
The use cases Google is pushing are broad, but the company keeps coming back to training and simulation. In robotics, the idea is to let machines practice in a made-up world before they meet the real one. DeepMind showed an Apptronik Apollo humanoid robot playing checkers in a demo, with Gemini planning the steps and Genie 4 generating the world in real time. Google also points to game prototypes, education, and filmmaking, where creators could step through scenes and storyboards faster. Access is still limited to trusted testers for now, with Google AI Ultra subscribers, early access partners through the Gemini API, and Google AI Studio users set to get access in the coming weeks.
Google is also putting SynthID watermarks on all generated video and audio. And it is not pretending the thing is finished: details can drift, objects can vanish after edits, and long-term consistency, accurate sound, legible text and complex physical interactions are still open problems. That’s the part that matters most. The demo is flashy; the real test is whether the model can stay sane long enough to be useful.
My take — AI-written commentary, not fact-checked reporting
This is the usual AI playbook: show the pretty demo first, admit the messy bits later. World models are clearly getting serious, but the gap between a convincing simulation and a reliable tool is still doing a lot of work. The industry loves to confuse those two when the screenshots look good.
Read more about this at: Trending Topics