From the time that Tripo was founded in 2023, our vision for the company has been to build the bedrock infrastructure for the next generation of interactive user-generated content (UGC) platforms. We operate on a first-principles engineering culture: rather than indexing against competitor roadmaps, the team takes a pragmatic approach by focusing entirely on core problem-solving. The team continuously monitors cutting-edge developments in the global research community, looking for new tools and methodologies that align with our vision.  

Back then, high-quality 3D asset generation was the mandatory foundational step. You cannot compute the dynamics of a world if you have not first defined the objects inside it. We put significant resources into building native 3D generative architectures that would produce structurally sound, usable assets instead of mere visual approximations that required massive manual cleanup. Today, that technology has matured and is actively transforming studio workflows. 

At the same time, we observed a growing interest across the AI research community in state-prediction models—the ability of models to understand and predict how environments evolve over time. For us, this trend validated our long-held roadmap. To be truly persistent and interactive, a world model must simultaneously address "State" (the objective physical and structural condition of the world at any given moment, independent of the camera perspective) and "Transition" (the continuous evolution of the world driven by actions and physics).

Currently, many research initiatives are limited by the conceptual framework of video generation models. The real world is 3D, and the 2D pixels in a video are merely a compressed, low-dimensional projection.

In pure action-conditioned video generation, the state of the world is implicitly hidden within a history of pixels. When an object leaves the field of view, its independent state is not saved: the model simply forgets it. When the camera pans back, it relies on guessing to hallucinate the object anew. It has time, but no true spatial memory. Conversely, static 3D scene generation has a strong spatial structure but no time evolution: it has space, but no dynamics.

For a truly interactive experience, the system must understand how objects physically transform and interact. Video models merely act as "optical simulators" hallucinating pixels, whereas a true world model must act as a "physical simulator" deducing physical realities.

The challenge with end-to-end monolithic video systems is that they face significant capacity bottlenecks. Forcing a single neural network to simultaneously compute complex state logic and generate high-fidelity pixel arrays is computationally crippling.

It wasn't until we hit upon the idea of decoupling the underlying state deduction from the final visual rendering that the basis for Project Eden, our world model, was born. Project Eden's architecture natively separates the "physics" from the "pixels" across three distinct layers:

To train this decoupled architecture, we had to solve a massive data bottleneck. If you only train on traditional physics engines, your model becomes rigid and lacks the infinite diversity of the real world. If you only train on 2D video, your model learns hallucinations instead of physical boundaries.

Our solution is a "dual-state alignment" training base. First, we utilize Tripo's accumulated 3D foundation models to reverse-engineer massive amounts of unlabeled internet video, extracting depth, camera poses, and geometric trajectories (L1 Data). Second, we use traditional game engines to generate synthetic data with absolutely precise 3D state annotations and action logic (L2 Data). By combining the boundless generative diversity of video with the strict structural consistency of game engines, Project Eden learns the true "latent general actions" of the physical world. As a result, it has already demonstrated unprecedented capabilities in environmental persistence, scene reusability, and multi-agent concurrency.

Returning to Tripo's overall vision from the beginning, we don't see Project Eden as just a game development tool, but as the underlying engine for the next generation of interactive content creation. For regular content creators, Project Eden bypasses the technical barriers and specialized expertise traditionally required for 3D world creation. 

Our long-term objective is to trigger the "ChatGPT Moment" for interactive media. For example, a user could input "a medieval tavern escape room containing a locked chest, with the key hidden inside a wine keg," and the system will instantly compile a fully realized, playable sandbox. Because all logic and physics run natively on the world model, the engineering cost of defining digital world rules drops to zero. Regular consumers can become active world-builders, laying the foundation for a universal next-generation UGC interactive platform. 

True world models are an indispensable milestone on the path toward Artificial General Intelligence (AGI). Language models taught AI to "speak," video models taught AI to "see" and "draw," while 3D foundational models paired with world models are teaching AI to "understand, generate, and deduce the physical world." The next step for AI will be to directly generate a truly interactive world, rather than just producing better content at a faster pace.

The AI Collective is built by volunteers across 180+ chapters in 40 countries.

Thank you to the thousands of volunteers around the world who make this work possible. We truly could not do this without you.