Playing a video game needs more than generating video. The world must persist, respond to controls, and allow interactions. Models using generated frames as memory lose spatial consistency over time.
PlayGen2 separates world simulation from rendering. It keeps a persistent 3D memory of geometry and entities, updates it with player actions, then renders the view from that memory.
The system runs in real time on a laptop, with about 160 ms latency. It transfers to new maps and supports multiplayer without retraining. This shows explicit 3D world state is a practical foundation for interactive learned game worlds.