We are working toward next-generation intelligent worlds: AI-generated environments that many people or agents can enter, change and experience together.
Most interactive world-model demos start with one player, one viewpoint and one generated future. A shared world has to carry the same event across many viewpoints. If one participant opens a door, moves an object or crosses another participant's path, everyone must inherit that change from their own position.
The model must preserve identity, location and causality while several participants continue to act. That makes multiplayer interaction a model problem, a data problem and an evaluation problem at the same time.
Interaction data explains the shared world
To train this behavior, we need synchronized interaction data alongside video. Each training sequence must align what every participant observes with their actions, changes to the shared world state and the outcomes that follow.
A video shows what happened. Interaction data explains why.
Those synchronized records let us train action-conditioned models, test whether one action produces compatible consequences across viewpoints and measure how the system behaves in unfamiliar layouts. Replayable state streams also let us render the same event from different cameras and match it back to the action that caused it.
| Data layer | What it captures | Why it matters |
|---|---|---|
| Observations | What each participant sees | Aligns multiple viewpoints |
| Actions | What each participant does | Explains causality |
| Shared world state | What changes in the environment | Preserves consistency |
| Outcomes | What happens next | Supports training and evaluation |
A multiplayer engine provides ground truth
Game engines already maintain structured records of actions, objects, state changes and outcomes. Repeatable simulations and controllable environments make those systems useful for training and evaluating models that need to learn how a world looks and how it changes.
We build on Yahaha, a global multiplayer UGC creation platform powered by a proprietary game engine. Its creator tools, controllable environments and live multiplayer product experience give us a foundation for designing shared environments and turning synchronized gameplay into training data and evaluation signals.
Depending on the environment and research setup, our records can include player actions, shared state, camera parameters, object poses, depth, motion vectors and segmentation. Together, they give us a replayable account of what changed, who caused it and what every participant observed next.
Khora is the public proof
Khora is our scalable multi-agent world model for real-time multiplayer interaction in shared AI worlds. Its public demo lets eight players enter the same generated environment in real time. The latest release runs across multiple maps, supports deathmatch gameplay and renders at up to 32 FPS.
The playable system connects world-model research to concurrent player actions and observable outcomes. It gives us a concrete environment for testing consistency, generalization and real-time interaction with people in the loop.
Shared environments beyond games
Multiplayer world models and synchronized interaction data can support foundation-model developers, research labs, game companies and robotics teams. Simulation and embodied AI face the same underlying requirement: several agents must reason about actions, space and one another inside one evolving environment.
We build the shared environments, interaction data and evaluation systems needed for that work. Our long-term mission is to connect people through AI worlds that can be generated, experienced and changed together.
| Domain | Shared-world requirement | Ophilus foundation |
|---|---|---|
| Games | Real-time coherent multiplayer worlds | Interaction data and environments |
| Robotics | Agents learning actions and consequences | Controllable simulation |
| Embodied AI | Spatial and multi-agent reasoning | Synchronized observations and state |
| Foundation models | Generalization across environments | Training data and evaluation systems |
