A simulation can look convincing while losing track of the room it just created. Genie 4 makes that distinction worth testing.

DeepMind’s October 27 announcement introduces a world model that generates an interactive scene as a user explores it. Native audio adds ambient sound and dialogue. Gemini models help interpret instructions and steer events, including changes made while the scene is running. DeepMind presents improved physics and human motion as advances in realism.

The access boundary matters as much as the demonstration. Genie 4 starts as a research preview for trusted testers. Broader access through Project Genie for Google AI Ultra subscribers and early developer access through the Gemini API and AI Studio are planned. An announced access route does not give a team a dependable production service today.

There is also a boundary inside the simulation. DeepMind describes continuing problems with object permanence, spatial accuracy and longer-term consistency. Contemporaneous reporting likewise describes coherence lasting only a few minutes. The launch establishes a preview capability; it supplies no general proof that a realistic scene predicts how a real machine or workplace will behave.

Consider a hypothetical training exercise in which a learner leaves a room, returns and must respond to a changed condition. If an object disappears, a doorway moves or an earlier instruction is forgotten, the exercise changes underneath the learner. The resulting score could measure the simulator’s inconsistency as much as the learner’s judgment.

A useful trial would therefore set a short, explicit scenario and track its state through navigation and prompt edits. Record which objects must persist, which events must happen in order and which failures invalidate the session. Compare those checks with the intended training or prototyping task.

Realism can make an experiment easier to understand. Continuity determines whether its result means what the team thinks it means.