Google's Genie 4 Makes AI Worlds Talk Back Live
By Daniel Rao | 28 October 2026
Google has just given its imaginary worlds a voice, and that changes what they are. Until now, DeepMind's Genie models built silent places you could wander through — impressive, eerie, and faintly unfinished, like a film with the sound switched off. Genie 4, announced on 27 October 2026, talks back. Its characters speak, its beaches make noise, and you can rewrite the scene while you are standing in it. The catch, and it is a big one, is that the whole world still only holds together for about two minutes before it starts to forget itself.
That tension — a genuinely startling leap, stapled to a frankly short leash — is the story of Genie 4. It is not a product you can buy, not yet. It is a research preview, in the hands of trusted testers, with a waitlist for everyone else. But it is also the clearest sign yet of where Google thinks this technology is going: not towards better video clips, but towards simulated worlds that behave, sound, and respond like real places, built on the fly, for robots to practise in and filmmakers to sketch in before a single camera rolls.
We checked DeepMind's announcement and the Project Genie waitlist form directly, and cross-read TechCrunch's briefing report against DeepMind's own wording before writing this piece. Where the two disagree — and on a couple of points they do, in emphasis if not in fact — we say so.
First, what a world model actually is
A world model is not a video generator, and the distinction matters more with every release. A video generator hands you a finished clip. A world model keeps generating, frame by frame, in response to what you do inside it. You move, it reacts. You turn a corner, it invents what is round the corner, and tries — this is the hard part — to remember what was behind you.
Genie 4 is DeepMind's fourth attempt at this problem, and the lab's announcement, written by Raia Hadsell, Diego Rivas and Hang Qi, describes it as a model that "unlocks unprecedented realism and improves real-time interactive simulation". The company says it is "natively multimodal", which in plain language means the sound is not dubbed on afterwards. DeepMind says music, ambient noises and interactive dialogue are generated simultaneously with the world itself. A beach scene, in the lab's own example, arrives with waves crashing. Wind moves through trees. People in the scene answer you.
That last point is the one to dwell on. Earlier world models were places. Genie 4 is, at least in ambition, a place with inhabitants — characters who converse and react to your inputs or to events in the world, in real time, according to DeepMind. Whether they are convincing conversationalists or merely fluent ones is something only the trusted testers can currently judge, and DeepMind has not released transcripts. But the direction of travel is unmistakable: from scenery, to simulation, to something that talks back.
The sound breakthrough is the headline, not a footnote
It is tempting to treat audio as a garnish. DeepMind insists it is the main course, and for once the marketing has a point. Sound is how a world convinces your brain it is a place rather than a picture. A silent street feels like a diorama. Add footsteps, traffic, a voice calling from a doorway, and suddenly you are somewhere.
DeepMind says Genie 4 is its first real-time world model to generate audio natively, and that every world carries the ambient sound of its own setting. The company is equally clear that this is early work: real-time audio accuracy is listed among the model's active limitations, alongside legible text and spatial accuracy. In other words, the voices may not always say quite what they should, and the soundscape may not always match the scene. That honesty is welcome, and it should temper the demo reels. A model that generates its own soundtrack will also generate its own wrong notes.
All video and audio Genie 4 produces will carry SynthID, Google's imperceptible digital watermark, DeepMind says. That is a sensible guardrail in a year when synthetic media is already hard enough to spot, and it is one of the few parts of this announcement that is not hedged at all.
You can now rewrite the world while you stand in it
The second big change is steering. With Genie 4, DeepMind says, you can use natural language to modify a simulation in real time — change the weather, move the camera, introduce a new person or object — and Gemini will adapt your prompt to fit the surrounding scene. Between your inputs, Gemini also monitors and steers the scene to keep the world progressing cohesively.
Think about what that means for the person using it. In a conventional game, the world is built before you arrive: artists model it, programmers script it, and your freedom is the freedom the designers chose to give you. In a Genie 4 world, the world is improvised around you, and your typed sentence is a stage direction the cast has to obey mid-performance. Change the weather and the light, the sound, the mood of the scene all have to shift with it, without the whole thing falling apart. When it works, that is a new kind of creative tool. When it does not, you get the familiar world-model failure: a door that becomes a window, a character who changes coat halfway across the room.
Navigation itself is deliberately ordinary. DeepMind says users steer with standard keyboard controls for movement and camera, plus text prompts for the scene. There is something reassuring about that. The most futuristic part of Genie 4 is not how you drive it; it is what happens when you do.
Gemini is in the director's chair
None of this works without the second half of the partnership. DeepMind says Gemini Omni anchors Genie 4's simulations in real-world appearance and behaviour, while other Gemini models supply the reasoning that keeps a scene coherent. More specifically, and this is a detail worth getting exactly right, DeepMind says "Gemini 3.8 Flash serves as a guiding agent in the loop": as Genie 4 generates a world, Gemini continuously observes the scene, breaks goals into a sequence of steps, and reviews incoming frames to track progress and verify that each step succeeded.
That is a neat division of labour. Gemini plans and checks; Genie renders. If the user intervenes, Gemini adapts its plan in the moment. DeepMind frames this as the path to worlds with reliable causality and multi-step reasoning — worlds where an action has a consequence, and the consequence sticks.
The lab is already pushing the idea further. DeepMind says it has been experimenting with Gemini 4 Argon and Genie 4 together, in work tied to Gemini Robotics. In one example the lab describes, Argon breaks a user-instructed task into steps for Genie to render — in the demo, Apptronik's Apollo humanoid playing a game of checkers, step by step, inside a simulated world. There is something quietly wonderful about that image: a robot that does not exist in that room, learning a board game that does not exist either, in a world that will evaporate after two minutes. It is also the most commercially serious demo in the whole announcement, for reasons we will come back to.
Two minutes is both a triumph and a tell
Here is the number that will follow Genie 4 everywhere. TechCrunch reports that Genie 4 can maintain a consistent scene for up to two minutes, double the limit of its predecessor, Genie 3. DeepMind's own announcement is more circumspect: it says the model can maintain coherent environments over short sessions, and warns that visual details and object permanence can still shift or drift over time, especially as a simulation runs longer, after multiple complex prompt edits, or when navigating through the scene.
Both things are true, and the gap between them is instructive. Two minutes is a reported figure, not a specification DeepMind chose to headline. The lab's own emphasis is on drift as an unsolved problem, not a solved one. Complex physical interactions, real-time audio accuracy, rendering legible text, spatial accuracy and long-term consistency are all listed as active areas of research. A world that forgets its own furniture after a couple of minutes is not yet a place you can live in, work in, or ship a game in.
But do not dismiss the doubling. The failure mode of every world model so far has been the same: the longer it runs, the more it hallucinates. Getting from one minute to two, while adding sound, dialogue and live steering at the same time, is not a cosmetic gain. It means the coherence machinery — Gemini watching the frames, checking the steps — is doing real work. The question is whether the curve keeps bending: four minutes, ten, an hour. Nobody at DeepMind is promising that curve, and nobody should read it into this announcement.
Genie 3, for context, arrived as a research preview a little over a year ago, according to TechCrunch. Genie 4 is therefore not a rushed sequel; it is a year's worth of work on exactly the problems that made Genie 3 a curiosity rather than a tool.
What the trusted testers are actually doing with it
DeepMind says it has been working with cohorts of trusted testers across creative industries, journalism, education, gaming and engineering, and their reported uses are more revealing than the lab's own pitch. Filmmakers, the lab says, are testing Genie 4 as a rapid "draft mode" for pre-visualisation: step into a live simulation of a scene and run multiple takes in real time, adjusting camera angles, lighting, dialogue and action on the fly.
"The real time aspect becomes most valuable in the areas where keeping a state of flow in the creativity space becomes incredibly useful," Diego Rivas, a Google DeepMind engineer, said during a press briefing, as reported by TechCrunch. "For example, in the ideation phase, in the storyboarding phase — anytime you're kind of in need of rapid ideation."
That is a refreshingly modest claim, and it rings true. Nobody storyboards a film in a world that lasts two minutes and then dissolves. But a director who can block a scene, move the camera, change the light and hear the dialogue in a single sitting — without booking a location, a crew or a render farm — has a genuinely new sketching tool. The same logic applies to game developers testing an idea before writing a single line of engine code, which DeepMind says its testers are already doing, and to photojournalists exploring how a single archival photograph might become a grounded simulation of a historical environment, or how a flash flood or wildfire might unfold.
It is worth noting what is missing from that list: a shipped game, a released film, a classroom product. Genie 4's tester stories are all, so far, stories about preparation — drafts, prototypes, sketches. That is the honest stage this technology is at, and DeepMind, to its credit, has not dressed it up as more.
Why robots may matter more than games
Ask DeepMind what world simulation is for and the answer goes bigger than entertainment. "Understanding and simulating the world is a crucial aspect of building AGI," the announcement says, and the lab positions Genie 4 as a way to create safe, infinitely variable simulations — including training environments for robots and agents.
This is the least glamorous and most consequential use case. A robot cannot learn to handle the real world by reading about it; it needs to practise, fail, and try again, millions of times, without breaking anything or hurting anyone. A simulated world that can generate endless variations — different rooms, different lighting, different clutter, a checkers game that never runs out of moves — is a training ground with no rent and no repair bill. That is why the Apollo checkers demo matters more than any beach scene: it is a proof that a reasoning model can direct a simulated body through a multi-step task, and check each step as it goes.
It also reframes the two-minute limit. For robot training, a two-minute scenario is not a flaw; it is a unit. A robot does not need to live in a simulated house for a week. It needs ten thousand two-minute kitchens, each slightly different, each testing something specific. If Genie 4 can generate those on demand, its commercial future may have very little to do with games at all.
A crowded race, and Google is not running alone
Genie 4 arrives into a field that has suddenly become crowded and expensive. TechCrunch notes that Runway released its second world model, GWM Worlds 2, a couple of months earlier — a playable model that responds dynamically to user inputs and includes audio generation. The same report notes the consolidation already under way: Fei-Fei Li's World Labs is being acquired by AMD, Amazon is in talks to buy Decart, and Rivian has acquired Third Dimension to bring simulation in-house for self-driving tests.
The pattern is familiar from every AI land grab: incumbents are buying as much as they are building. Intel has just bought the quantum startup its chief executive already backed, betting that owning the software layer matters more than inventing it in-house. AMD's move on World Labs is the world-model equivalent. Nobody with a balance sheet wants to be left renting the simulation layer from a rival.
Google's advantage is integration. It owns the reasoning model, the world model, the robotics research and the cloud they would all run on. Its disadvantage is the same as ever: turning research into a product people can actually use, at a price anyone will pay. On that last point, DeepMind has published nothing: no resolution, no frame rate, no list of countries, and no API pricing for Genie 4. Those absences are not oversights to fill in from rumour; they are facts about this launch, and they mean nobody outside the tester pool can yet judge what Genie 4 costs to run or how good it looks at full stretch.
The wider AI industry is also learning, sometimes painfully, that announcements are not results. Readers of TechPulse USA will remember OpenAI's claim that AI killed a conjecture behind hundreds of papers — a claim that arrived with fanfare and is still awaiting independent verification. Genie 4 deserves the same discipline: believe the announcement, test the model, and keep the two separate until the second catches up with the first.
Our take: the sound is real, the world is not — yet
Here is our position, stated plainly. Genie 4 is a genuine step forward, and the step that matters most is not photorealism or the extra minute. It is sound plus steering, together, in real time. A world you can talk to and rewrite while you stand in it is a different kind of object from a world you merely watch, and DeepMind is right to put that at the centre of this launch. The Gemini-as-director architecture is also the right bet: a world model that cannot plan cannot remember, and a world model that cannot remember cannot be useful.
But — and this is the part the demo reels will not tell you — a two-minute world with drifting details, unreliable audio and illegible text is not a platform yet. It is a laboratory instrument with a beautiful demo. The gap between "a filmmaker can sketch a scene" and "a filmmaker can rely on a scene" is the whole ballgame, and DeepMind has not claimed to have closed it. The testers' enthusiasm for draft mode is encouraging precisely because it is modest: sketches are where a two-minute world is already enough.
What would change our mind? A public hands-on, a longer coherence window demonstrated live rather than reported, and a price. Until then, Genie 4 should be filed where it belongs: the most interesting world model of the year, and still a research preview. Those two descriptions are not in tension. They are both true, and the second is what makes the first worth believing.
What happens next, and how to try it
For now, Genie 4 is available as a research preview for trusted testers. DeepMind says it is coming soon to Google AI Ultra subscribers in Project Genie and to early-access partners of the Gemini API and Google AI Studio, and TechCrunch reports that wider rollout is expected in the coming weeks. A waitlist for Genie 4 in Project Genie is open now via the form linked from DeepMind's announcement.
That rollout order tells its own story. Google is putting its most expensive subscribers and its developer partners first, not the general public — the same cautious sequencing it used for earlier frontier releases. It also means the first independent verdicts will come from people with a professional reason to stress-test the model, which is exactly the audience that will notice when a chair changes shape or a line of dialogue goes wrong.
The competitive ripples, meanwhile, will not wait for the waitlist. Oracle inviting Nvidia's rivals into its own AI cloud showed this week how quickly the AI stack is unbundling, with Google's own TPU chips now heading into a rival cloud. World models are the next layer up for grabs: whoever owns the simulation layer may end up owning the training ground for every robot and agent that follows. Genie 4 is Google's bid for that layer. Whether it holds together longer than two minutes will decide whether the bid is remembered as a milestone or a demo.
Sources: Google DeepMind's Genie 4 announcement (27 October 2026, by Raia Hadsell, Diego Rivas and Hang Qi); TechCrunch's Genie 4 report (27 October 2026, including its press-briefing quotes); The Verge's Genie 4 coverage, as summarised in secondary reporting; Wikimedia Commons for the file photographs used, credited above. No Genie 4 output is reproduced in this article.
By Daniel Rao
Daniel Rao is a technology correspondent for TechPulse USA, covering artificial intelligence, robotics and the companies building them.
Comments
Post a Comment