Fei Fei Li: The Race to Build World Models For AI
Summary
In this episode, World Labs co-founders Fei-Fei Li, Justin Johnson, and Ben Mildenhall discuss Atlas, a new frontier model built on "new view prediction" to advance spatial intelligence. They explain how Atlas unifies 3D reconstruction and generation, allowing users to create spatially grounded environments from just a few images, drastically reducing the capture requirements for dense reconstruction.
Topics discussed
Introduction: Atlas and the concept of new view prediction
Defining Atlas capabilities: generation, reconstruction, and simulation
New view prediction as a fundamental primitive for base models
Spatially grounded inputs and controlled camera flythroughs
Multimodal architecture: integrating text, image, video, and 3D data
The challenge of spatial intelligence and pixel contextualization
Evolution from Marble to Atlas: scaling and modality bifurcation
The scaling ladder and the history of dense 3D reconstruction
Reducing capture requirements: from hundreds of photos to three
Bridging generation and reconstruction to fill spatial gaps
Context windows and the continuum of reconstruction and generation
Architectural challenges and the pre-training cycle
Emergent behaviors and the 'soccer ball under the table' demo
Creative workflows: persistent 3D state vs. ephemeral video
Industrial applications: architecture, construction, and design
Robotics: real-to-sim pipelines and environment randomization
World models as learned simulators and planners
Handling dynamics: training on static and moving scenes
Future directions: editability, control, and interactive interfaces
AI completeness and the conclusion of the interview
Listen ad-free on Castria