Article 335: Google's Genie 4 Generates Audio and Lets You Redirect the World Mid-Simulation
Google DeepMind released Genie 4 in research preview this week, roughly a year after Genie 3 and with a list of upgrades that reads less like a version bump than a change of category.
The headline features: photorealistic worlds that hold a consistent scene for up to two minutes - double the predecessor - better physics, more humanlike characters, and real-time generated audio including speech, music and ambient sound that matches the simulation.
The feature that matters most is the one that changes the interaction model. You can type mid-simulation to change the weather, move the camera, or drop in a new person, and the world adjusts live. Free-form steering, not pre-generated levels.
Gemini is now the director
The integration point is the interesting one. Gemini plans the steps and checks the frames while Genie renders. In a research experiment tied to Gemini Robotics, a Gemini 4 Argon model broke a task into steps and Genie 4 simulated them in real time.
That is the loop the robotics people have wanted for years: a reasoning model decomposes a task, a world model plays it forward, and the result is visible before anything touches a physical arm. Real-world robot time is expensive and destructive; simulated robot time is neither, provided the simulation is faithful enough to be honest about failure.
DeepMind's Tejas Kulkarni framed the direction as hybrid world models - Gemini writing code and text, Genie ingesting it and simulating reality at different levels of abstraction. Trusted testers are already pointing at a use nobody planned: storyboarding films, where keeping creative flow matters more than photoreal final frames.
The consolidation nobody is calling consolidation
The world model sector is being eaten alive while the models get better. Fei-Fei Li's World Labs is being acquired by AMD. Amazon is in talks for Decart. Rivian bought Third Dimension to bring driving simulation in-house. Runway shipped its own playable world model within months of Genie 3.
World models are being absorbed into three parent bodies: chip companies that want to sell the future demand they simulate, hyperscalers that want the compute contract, and physical-world companies - cars, robotics - for which simulation is a safety requirement rather than a product.
The unsolved problem
Two minutes of consistency is genuinely impressive and still nowhere near the length of a scene in a film, a shift in a factory, or a training episode for a robot. Every one of those runs far longer than a world model can hold coherence, which means the useful engineering work is not just "better world model" but "how do you stitch many short coherent worlds into one continuous simulation without the physics quietly breaking at the seams".
Nobody has published a good answer to that. The demos will not show you the seam.
Key Takeaway: World models have moved from "look at this pretty generated landscape" to "this is the test harness for physical AI", and the labs that own a reasoning model plus a world model plus a robotics programme are the ones who benefit. Google now has all three. The consolidation into AMD, Amazon and Rivian suggests everyone else has concluded the same thing: a standalone world model lab is a feature, not a company.