AI Agent UpdateGoogle DeepMind's Genie 4 builds living 3D worlds with sound — and lets you steer them mid-scene
Genie 4 generates interactive 3D worlds from a text or image prompt, adds native audio for the first time, and lets users change the weather, move the camera, or add characters while the simulation runs.
The News
Google DeepMind announced Genie 4 on October 27, and the headline feature is that its worlds now have sound. A world model is an AI that simulates an environment you can move through, rather than just describing one in text or showing a flat video. Genie 4 generates interactive, real-time 3D worlds from a text or image prompt — and for the first time in the Genie line, it generates native audio alongside the visuals: music, ambient noise, and interactive dialogue that belong to the scene rather than being pasted on afterwards.
The second new feature is steering. While a simulation runs, users can type prompts to change the weather, move the camera, or add people and objects, and navigate with keyboard controls. Two other Google models do the anchoring and guiding: Gemini Omni keeps the world looking and behaving like the real world, and Gemini 3.8 Flash acts as the in-loop guiding agent, planning steps and checking frames as the scene unfolds. TechCrunch reports scenes hold together for up to about two minutes, roughly double Genie 3's limit — still short, but a meaningful jump for a technology whose core problem is staying coherent over time.
DeepMind is not overselling it. Genie 4 is a research preview for trusted testers only, coming next to Google AI Ultra subscribers in Project Genie and to early-access Gemini API and Google AI Studio partners. All output carries SynthID watermarks, Google's invisible markers identifying AI-generated media. DeepMind itself flags drift over long sessions, audio accuracy, and illegible in-world text as current limits. The pitched uses are robot and agent training environments, game prototyping, and film storyboarding — places where a cheap, steerable simulated world is more useful than a perfect one.
Two minutes of a world that listens back is worth more than two hours of a world that just plays.
Why It Matters
World models are the training grounds for physical AI. A robot cannot safely learn to navigate a warehouse by crashing into real shelves thousands of times, but it can learn inside thousands of simulated warehouses — if those simulations look, sound, and behave like the real thing. Native audio matters here more than it might seem: a training world without sound teaches robots to ignore a sense humans rely on constantly, from footsteps to alarms. Genie 4's audio, steering, and two-minute coherence are steps toward simulations good enough to train in.
For Google, Genie 4 also shows how its model stack fits together. Gemini Omni grounds the world, Gemini 3.8 Flash guides it, and the Genie model renders it — three different models cooperating in real time. That composability is Google's answer to rivals building single giant models: a system of specialists, each doing one job well, is easier to improve piece by piece.
The limits deserve equal weight. Two minutes of coherence is not a world you can live in; it is a sketch. Drift over long sessions means the simulation slowly stops matching its own rules, and illegible in-world text breaks any scene that needs signs, labels, or documents. Genie 4 is a research milestone, not a product launch, and judging it as a product would miss what it actually is: proof that steerable, sounding worlds are now possible at all.
Analyst's Take
The right way to read Genie 4 is as infrastructure for the next decade of robotics and games. Today's two-minute, slightly drifting worlds become tomorrow's training environments, and training environments become capable robots. Google is not selling you a world today; it is showing you the factory that will make them.
Bullish, with the research-preview caveat attached: this is DeepMind at its best, turning a hard research problem — coherent, steerable, sounding simulation — into a measurable step forward. The gap between a two-minute demo and a useful product is still large, but it just got smaller, and in this field, smaller gaps compound fast.
By The Numbers
FAQ
What is a world model?
An AI that simulates an environment you can move through in real time, instead of just writing about it or showing a pre-recorded video. You give it a prompt — text or an image — and it generates a 3D world you can navigate and change.
What does 'native audio' mean?
The sound is generated by the model as part of the world, matched to what is happening in the scene, rather than added afterwards as a soundtrack. Music, ambient noise and dialogue are created alongside the visuals.
Can I try Genie 4 today?
Not unless you are a trusted tester. It is a research preview, coming next to Google AI Ultra subscribers in Project Genie and to early-access Gemini API and Google AI Studio partners. No public launch date was given.
What are its limits?
DeepMind lists three: drift over long sessions (the world slowly stops following its own rules), audio accuracy, and illegible text inside the world — signs and documents come out garbled.
Why does this matter for robots?
Robots learn fastest in simulation, where mistakes are free. A world that looks, sounds and behaves realistically — and can be steered to create new scenarios — is a far better training ground than a silent, static one.
👁
Source: Google DeepMind