EverWorld
Project lead · 2026 to present
Real-time interactive spatial video models. Explore a scene, act on its objects and talk to its characters, and your input shapes what happens next.
- Interactive video
- Spatial orchestration
- Agents
- Joint AV generation

Almost every video model today hands you a finished clip. EverWorld is built around the opposite contract: the scene is still being generated while you are inside it, and what you do decides the next one. You explore, act on objects and talk to characters, and the world answers in picture and sound.
It is planned as three versions, and each one changes who the person in the scene is.
Play (1.0): you are a participant. Interactive scene generation in a room-scale scene. Ring the bell, unfold the map, and the scene responds. Audiovisual generation keeps going, and your changes are still there when you look away and return. The work here is keeping state, geometry, sound and picture consistent while the scene changes.
Compose (2.0): you are a creator. Large-scale spatial orchestration. Describe a larger scene, then shape connected streets, docks and a lighthouse. Create interactive objects, edit the layout and rules, and step inside to try the space you designed. Physics, skills and events run through explicit scene rules, so what you build is something that executes, not only something that renders.
Evolve (3.0): you are a collaborator. Agent-driven scene evolution. Talk to your guide while exploring and change the route mid-conversation. Characters remember their experience, weigh possible actions and pursue their goals together, so the scene keeps evolving.
All three sit on one shared foundation: persistent state and memory, spatial perception and geometry, multimodal understanding, and joint audiovisual generation. Play establishes it; Compose and Evolve extend it rather than replace it.
The rule that runs through all of it is plan first, then act. Candidate actions are rolled out in isolated branches, evaluated, and only the selected one is executed in the live scene. Evolve reuses the Compose executor for this, and planning never modifies the scene the player is standing in.
Where it goes first: interactive livestreams, where the audience chooses together and turns the story in a new direction, and playable stories, where you enter a role and your actions and dialogue change the next scene.
The full vision is on the EverWorld site, and the architecture is laid out in its technical guide. If you want to work on the models, runtime, evaluation or interactive experiences, get in touch.