Fei Fei Li: The Race to Build World Models For AI

Atlas, a frontier world model from World Labs, advances spatial intelligence by unifying pixel generation and 3D reconstruction through a fundamental primitive called "new view prediction." By treating camera poses as native inputs, the model enables high-precision 3D reconstruction and simulation from sparse data, effectively replacing traditional, labor-intensive dense capture methods. Co-founders Fei-Fei Li, Justin Johnson, and Ben Mildenhall argue that this approach mirrors the role of next-token prediction in language models, providing a scalable path toward AI that understands physical geometry, physics, and dynamics. Beyond creative applications, the model serves as a critical bridge for robotics, where learned neural simulators facilitate real-to-sim training and overcome current data bottlenecks. This architecture represents a shift toward persistent, stateful 3D environments that allow for interactive planning and complex spatial reasoning.