Google DeepMind connected Project Genie 3 to Street View on May 19 at I/O 2026. The user-facing surface is small: drop a pin anywhere in the United States on Google Maps, describe a character and a visual style ("a claymation monster in 1920s noir," "your favorite animal in an underwater Golden Gate Bridge"), and Genie 3 generates a 720p 24fps interactive world starting from that location. You can walk inside it with text instructions, weather can be changed mid-stream, characters can be dropped in. The model runs autoregressively, frame by frame, conditioned on the world description and your actions. Environments remain coherent for several minutes, visual memory for objects you have already seen extends about one minute back. Access is gated behind AI Ultra at $200 per month, US users 18+, US Maps locations first with global rollout planned. The architecture is not publicly disclosed: no parameter count, no training dataset specification, no benchmark numbers. The acknowledged limitations are concrete: limited agent action space, multi-agent interactions struggle, imperfect real-world location fidelity, text rendering is poor, continuous interaction caps at a few minutes.
The structural move worth pausing on is what Genie 3 does with Street View as a substrate. Street View is captured memory: 280 billion photographs taken across 110 countries on all seven continents over almost two decades, each one a sample of a place at a moment. Until now, Street View was a viewer onto those samples. You could see them, you could not walk past them. Genie 3 takes the Street View imagery as a conditioning prior and runs an autoregressive generative continuation forward from it. The captured photographs become the boundary condition for a forward-running dream. Memory and generation become the same substrate type, with different training objectives: Street View was trained on faithfulness to what was there, Genie 3 is trained on plausibility of what comes next. When you drop a pin at the Golden Gate Bridge and ask to see it underwater with schools of fish, what you are doing is feeding the captured-memory prior into the generative-continuation engine with a redirect on the physics. The bridge is from Street View. The fish are from Genie 3. The relationship between them is not stable: as you walk, the bridge from Street View becomes a Genie 3 hallucination that drifts the moment your one-minute visual memory rolls past it.
The architecture matters for what this is and what it is not. Autoregressive frame-by-frame generation conditioned on prior frames and actions is closer to how an embodied policy navigates the world than how a diffusion video model produces a clip. Diffusion video produces a fixed-length artifact. Genie 3 produces a continuously running sequence that responds to your input as it arrives. That is structurally an internal world model in the cognitive sense, an engine that predicts the next instant given the current instant and what you just did. The fact that it pairs with SIMA, DeepMind's embodied agent, to train and evaluate agent policies inside the generated environments is the operational tell: Genie 3 is the substrate inside which an agent can practice action policies that would be too expensive to practice in the real world. The promptable world events (weather change, character drop-in, scenario mutation) are the controllability handle that distinguishes this from passive video generation. You are not watching the world unfold, you are steering it.
The release also surfaces a philosophical layer worth naming directly, even though it sits in territory that gets pattern-matched as theoretical. The implementation is empirical, the numbers are concrete (720p, 24fps, several minutes consistency, one minute visual memory, 280 billion image prior, US AI Ultra subscribers only). What is theoretical is what the shape implies about the relationship between place, memory, and generation when an engine like this becomes a consumer product. A pin on the map used to mean a location you could navigate to. Now it can also mean a starting point for a private generated continuation. The "place" Genie 3 produces does not persist when you stop attending to it. The same pin returned to tomorrow with the same character description does not yield the same world. What is being generated is not a virtual location, it is a forward-running attended experience of a location, anchored just enough in Street View imagery that you recognize the entry point and lose the recognition somewhere past the one-minute memory horizon. The release is gated, the rollout is cautious, the audience is small. But the underlying capability is now shipped and visible: a generative continuation engine on top of a captured-memory substrate, steerable by text, real-time interactive, several minutes consistent. That is a different kind of object than the products we have been pricing against, and the next phase of agent training, simulation-based research, embodied AI evaluation, and consumer-facing generative experiences is going to be built on top of objects with this shape.