Two Minutes. Then Google's Genie 4 Forgets Its Own World.
DeepMind's research preview builds photorealistic worlds with real-time audio and talking humans, but the simulation holds together for about two minutes before physics and memory fray. Google isn't selling this as a toy. It's building the training ground robots and agents need, and the two-minute wall shows exactly how far that work still has to go.

You type a sentence. A second later you're standing in it: rain on a Tokyo rail platform, a crowd moving with individual quirks, a stranger who turns and answers when you speak. For about two minutes, this works. Then the platform forgets where the stairs were, the crowd starts repeating, and the stranger says something that doesn't track with the conversation you just had. That collapse point isn't a bug Google is embarrassed about. It's the whole story.
Google DeepMind released Genie 4 as a research preview on Tuesday, Oct. 27, pitched as a world model that generates explorable, interactive environments from a text prompt or image. You steer with a keyboard and rewrite the scene as it runs. Everyone covering this is calling it a video generator that got good. That framing misses the mechanism DeepMind actually built.
The Architecture Is the News
Buried under the photorealism headlines: Genie 4 isn't generating video and hoping it holds together. A second model sits in the loop: Gemini 3.8 Flash watches the generated frames, decomposes the goal into steps, checks whether the scene actually did what was asked, and re-plans when it did not. That's not a creative tool. That's a closed-loop rehearsal environment.
DeepMind's research director said the quiet part out loud: "What we'll increasingly see in the future are these hybrid world models, where you have a combination of Gemini writing code and text, and Genie ingesting that in real time and simulating reality at different levels of abstraction."
Translation: Gemini writes the plan, Genie renders the physics to test it, and nobody needs a robot arm in a lab to find out if the plan breaks. DeepMind is already running that experiment. The company describes stronger physics, and an experiment where Gemini 4 Argon writes a robot plan that Genie 4 simulates as it renders. Not Hollywood pre-viz. Policy training.

Two Minutes Is a Confession
Coherent scenes now last up to about two minutes, double the 60-second limit of Genie 3. Doubling sounds like progress until you ask what two minutes actually buys an agent that needs to learn cause and effect across a multi-step task.
DeepMind's own blog is unusually blunt: while it can maintain coherent environments over short sessions, 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 throughout the scene. Complex physical interactions, real-time audio accuracy, rendering legible text, spatial accuracy, and long-term consistency are active areas of ongoing research.
A company admitting its simulator can't yet be trusted for anything that requires memory past a coffee break. An agent trained on a world that forgets its own starting conditions is an agent learning causal reasoning from a simulation that has none past two minutes.
Access tells the same story. Google has started rolling Genie 4 out as a research preview for trusted testers, with no price, no date, no public benchmark. That's not a launch sequence for a product. That's a lab keeping its best training tool close while it figures out whether the coherence window scales.
What to watch: whether the two-minute ceiling moves past five or ten minutes in the next few quarters, who gets API access first (robotics teams inside Google versus outside researchers), and whether OpenAI or a rival lab ships a competing world model that claims longer coherence. If two minutes is still the number a year from now, that's the field telling you scaling compute alone won't fix causal memory.
- Genie 4: A new generation of world simulation — Google DeepMind
- Google's Genie 4 world model adds audio, more realistic worlds, and real-time steering
- Google's Genie 4: Interactive Worlds Now Talk Back, and You Can Steer Them Live
- Genie 4 simulates worlds you can steer in real time
- Genie 4: Native Audio and a Two-Minute World Model
- Google Launches Genie 4 AI World Model, Could Disrupt Game Development
- Google's Genie 4 World Model Adds Audio, More Realistic Worlds - Slashdot
- World Models: A Comprehensive Survey of Architectures, Methodologies, Reasoning Paradigms, and Applications
- Google DeepMind unveils Genie 4 world model, targeting robots and film production
- From Masks to Worlds: A Hitchhiker's Guide to World Models
- Simulus: Combining Improvements in Sample-Efficient World Model Agents
- Google DeepMind Rolls Out Genie 4 World Model in Research Preview
- DeepMind Genie 4 Adds Native Audio and Gemini-Powered Reasoning to Real-Time World Simulation
- Google DeepMind debuts Genie 4 world model with native audio
- Google DeepMind unveils Genie 4, an AI world model promising audio and live controls · Digg
- Google DeepMind's Genie 4 Builds Interactive Worlds With a Gemini Planner
- Google DeepMind Unveils Genie 4 World Model With Native Audio And Real-Time Steering: 10 outlets compared
- Google's Genie 4 Conjures Explorable Worlds in Real Time, With a Short Memory
- DeepMind's Genie 4 brings audio and sharper realism to AI-generated worlds
- Google built AI agents a world to practice in. The physics still aren’t quite right. - The New Stack
- Google Deepmind's Genie 4 puts a modern spin on agent designs that predate LLMs
- Google DeepMind unveils Genie 4, the world model that once tanked gaming stocks - Startup Fortune
- Techmeme