World models: Genie 4, Marble and the shape of simulated places

A video model hands you a finished clip. A world model hands you a place that answers back. The distinction stopped being academic this autumn.

Genie 4 (Google DeepMind, Oct 27) generates a world you step into: it responds to input every frame, lets you rewrite the scene mid-simulation in plain English, and synthesises the sound live — wind, footsteps, dialogue — instead of laying a track over video. It is a research preview for trusted testers; the consumer Project Genie still runs on Genie 3 with 60-second sessions. DeepMind’s stated next step is hybrid: a language-model layer carries the world’s logic while the world model renders how it looks, sounds and reacts.

Marble (World Labs) takes the opposite route to the same ambition: persistent, navigable 3D worlds from text, images or video, exportable as Gaussian splats, meshes or video — places that stay put instead of improvising forever. A second model, Atlas, arrived in September, and AMD announced on Sep 28 that it is acquiring World Labs for $8.2 billion. Spatial intelligence is now a balance-sheet item.

NVIDIA Cosmos is the industrial branch: open world-foundation checkpoints for synthetic training data and simulated futures for robots and vehicles — the branch where the "world" being modelled is a factory floor, not a fantasy valley. And Decart’s Oasis remains the small-model proof: an open-sourced playable world with no game engine in the loop.

For a game developer the honest summary is: generated worlds already win on speed and variety for exploration and prototyping; engines still win on determinism, design control and long sessions. The interesting work of the next two years is the seam between them.