Why Fei-Fei Li Is Betting on Spatial Intelligence

Video summary

Fei-Fei Li and World Labs unveil Atlas, using new-view prediction to cut 3D capture from hundreds of images to just three.

World Labs co-founders Fei-Fei Li, Justin Johnson, and Ben Mildenhall explain Atlas, a world model that generates, reconstructs, and simulates scenes through “new view prediction.” Unlike video models that predict the next frame, Atlas uses spatial context and camera poses to produce grounded views, joining pixel generation and 3D reconstruction in one multimodal model. The team says a Matrix-style bullet-time shot can be made with three iPhone cameras instead of hundreds, while room capture may drop from 100–300 images to three. Atlas builds on lessons from Marble, whose Gaussian-splat output constrained the system. The founders see uses in creative production, architecture, and robotics, where real-to-sim data collection is costly. They say Atlas already supports some motion, but stronger dynamics, editing controls, and action planning remain future goals; training compute is the main limit to scaling.

Chapters

1. 0:00What Atlas Is & Why It Matters: New-View Prediction and Bullet Time from Three Cameras
2. 5:15Is This a Scaled-Up Video Model or a New Architecture? Atlas Unifies Generation and Reconstruction
3. 8:15Spatial Intelligence & Why New View Prediction Matters: Atlas Moves Beyond Marble’s Gaussian Splats
4. 13:34Spatial Intelligence & Why New View Prediction Matters: Atlas Targets 50–100× Fewer Capture Images
5. 17:05Spatial Intelligence & Why New View Prediction Matters: Atlas Reconstructs Stanford Quad from Ground Views
6. 21:27Did You Know It Was Going to Work? Scaling Conviction and the NeRF Table Breakthrough
7. 24:42Use Cases: Creatives, Games & Robotics—Marvel’s Gaussian Splats and Virtual Design for Architecture
8. 28:05Use Cases: Creatives, Games & Robotics—Atlas Speeds Synnex’s Real-to-Sim Robotics Data Pipeline
9. 32:17Use Cases: Creatives, Games & Robotics—Learned Simulators Could Train Policies and Become Planners
10. 35:21The Elephant in the Room: Video Models vs World Models—Atlas Adds Dynamics Beyond Static Marble
11. 37:55Will We Get 4D Video You Can Walk Around In?—Editability, Scene Controls, and Spatial Interaction
12. 42:13Why New View Prediction Is the Next Token Prediction—Atlas’s Generative Viewpoint Primitive