Project Eden Gives World Models Their First Persistent “Save” Feature

Over the past year, "world models" have become a buzzword in AI, with many claims that a model can generate continuous video from a single prompt or react to user actions. The article questions whether producing a coherent video truly constitutes building a world.

Most existing approaches fall into two categories. The first, action‑conditioned video generation, predicts the next frame based on pixels, text, or motion commands, but lacks an independent representation of world state; objects disappear when they leave the camera view. The second, static 3D scene generation, creates a single navigable space but lacks temporal dynamics, physics, and state transitions, preventing true world simulation.

VAST argues that a genuine universal world model must solve two core problems: (1) determining the objective state of the world at any moment, and (2) continuously evolving that state with actions, time, and interaction.

Project Eden implements a three‑layer architecture to meet these requirements. The structured state layer maintains a compact, time‑persistent, globally queryable representation of the world, independent of any camera view. The conditional interface layer translates this global state into view‑specific constraints (semantic, geometric, event cues) for rendering, guaranteeing consistency across multiple viewpoints. The generative rendering layer then produces high‑fidelity visual frames using the supplied constraints, focusing only on texture, lighting, and fine‑grained dynamics.

This decoupling unlocks three system‑level capabilities. First, long‑term persistence: objects remain in the underlying state even when off‑screen, enabling true memory of the environment. Second, scene reuse with deterministic control: a single world state can be read, written, and edited, so all users see the same changes, allowing branching and replay. Third, native multi‑agent concurrency: a single shared state supports many agents, with rendering performed per‑agent, reducing computational cost from exponential to linear.

The article also details Project Eden’s data strategy. VAST uses a two‑tier pipeline: L1 mines massive internet videos, automatically extracting depth, camera pose, and geometry to create “dual‑state” data; L2 generates engine‑synthesized data with precise 3D state annotations, actions, and environmental changes. This combination balances breadth and logical precision.

Beyond a stronger 3D generator, Project Eden is positioned as foundational infrastructure for the next generation of interactive content and general AI research, offering a stable, evolvable world that agents can explore, learn from, and manipulate.

In conclusion, VAST’s vision is to let anyone freely create and immerse in interactive worlds, moving the competition from "who can generate better video" to "who can maintain a persistent, controllable world".