Google adds sound to its AI-generated worlds
Genie 4 loses context after a couple of minutes; wider access is planned, but timing and developer pricing remain unknown.
By Saeed Ezzati7 min read
The audio edition
Listen to this newsletter
0:004:49
Read transcript
Google DeepMind has introduced Genie 4, its first world model with native sound—and the update highlights both the appeal and the hard limit of generated worlds. DeepMind says characters can speak in dialogue that adapts as a simulation changes, alongside ambient sound and audio specified in a prompt. The company also reports more realistic physics and more natural human movement, though those improvements have not been independently assessed. Unlike a pre-rendered scene, Genie 4 generates a world frame by frame as someone explores it. Users can change the weather, add objects, or move the camera while it runs. Gemini 3.8 Flash helps steer those changes: it breaks a goal into steps and evaluates frames to keep the experience connected to the original prompt. The idea is to create something people can interact with and reshape, rather than just watch. That distinction matters for DeepMind’s proposed uses, including robot-training environments and rapid creative prototyping. A convincing frame is not enough; the world needs to behave consistently as it changes. And there is a major caveat: DeepMind says a Genie 4 world stays consistent for only a couple of minutes before losing its initial context. Access is still limited to a small group of trusted testers. Google says wider access is planned, but has announced neither a release date nor developer pricing. Developers are expected to get access through the Gemini API and AI Studio. Consumer access is planned in phases through Project Genie for Google AI Ultra subscribers, whose plans start at a hundred dollars a month. Possible access for AI Pro and AI Plus is not a committed rollout. For now, the key test is not just whether these worlds sound and look convincing, but whether they can keep their rules and context long enough to be useful. That question of what a system can reliably demonstrate is also reshaping how some California professors grade. At UCLA, statistics lecturer Thomas Maierhofer still assigns homework and labs, but grades them for completion. The higher-stakes evidence comes from in-person, paper exams without phones or laptops. UC Berkeley mathematician Richard Borcherds says online exam scores rise substantially, which he attributes to students using AI. But he acknowledges he cannot identify who used it. His response is to base grades mainly on device-free exams, where students must show what they can do themselves. USC professor Morgan Polikoff has described a planned shift toward more sit-down, paper-and-pencil assessments. These approaches do not necessarily rule out AI as a learning aid; they draw a line between using it to learn and treating an answer as proof of independent understanding. The same distinction between a positive signal and proof of capability appears in Walmart’s workforce data. The retailer says associates who completed OpenAI training reported a 27 percent gain in confidence using AI. Walmart’s first AI experience survey drew more than 1.7 million responses, from a workforce of more than 2.1 million associates. The training and certification programs, offered with OpenAI and Google, are free to associates. But the 27 percent figure measures how capable workers feel, not demonstrated skill or productivity—and it does not show that training caused the change. Walmart separately reported that 77 percent of office-based associates felt confident using AI tools. Its 84 percent figure for U.S. frontline associates concerns confidence with digital tools broadly, not AI specifically. The numbers describe comfort with technology, not a direct comparison of AI ability across those groups. And in Google Research’s account of AI learning, the evidence question shifts from what a model can do now to what it can retain later. The team describes a “sleep” framework: a model pauses new inputs so recently encountered information can be consolidated into its weights, rather than remaining only in temporary context. The underlying paper was submitted in June; the October overview revisits that work. The authors report near-zero forgetting across sequential translation tasks, including unfamiliar languages, and near-perfect scaling on the BABILong benchmark up to 10 million tokens. That is not a claim that the model attends to one enormous prompt. The reported result comes from compressing information into weights through an offline consolidation process. The authors also report gains on other tests, including mathematics, against the baselines they evaluated. There are important boundaries: the architecture stays fixed, and people define the training rules. The framework automates a post-training cycle; it does not rewrite its own core design. Across these stories, the useful thing to watch is whether claimed gains hold up under direct tests—sustained world consistency, independent student work, demonstrated workplace skill, and retained model knowledge—not just confidence, polished output, or a striking benchmark number.
