Fei-Fei Li: The Race to Build World Models for AI (45-Minute Podcast)

World Labs co-founders Fei-Fei Li, Justin Johnson, and Ben Mildenhall joined a16z's Martin Casado to unveil their latest World Model, Atlas, and discuss the direction of Spatial Intelligence they are pursuing.

The Core of Atlas: New View Prediction

At the heart of Atlas is New View Prediction technology. This model takes an image or view of a specific scene as input and predicts how that environment would look from different spatial and temporal positions. This integrates Generation and 3D Reconstruction within a single model, exploring whether view prediction can serve as a useful primitive for understanding the physical world.

Technical Limitations and Future Challenges

The teams discussed the model's technical bets, the current scope of what it can capture, and its limitations. They emphasized the importance of Dynamics, Editability, and Simulation in the development of world models, revealing that these factors are key determinants of the model's practicality.

Application Areas and Data Constraints

The discussion covered potential applications in creative work, architecture, and robotics. Fei-Fei Li identified the lack of access to Real-world Training Data as one of the biggest constraints on current developments in robotics and related fields.