Why Fei-Fei Li Is Betting on Spatial Intelligence

Video summary

World Labs’ Atlas unifies 3D reconstruction and generation, using new-view prediction to rebuild scenes from as few as three cameras.

World Labs co-founders Fei-Fei Li, Justin Johnson, and Ben Mildenhall explain Atlas, their next-generation world model, which predicts how a scene looks from new positions in space and time. Unlike video models built around next-frame prediction, Atlas uses camera poses and spatial context to combine image generation with precise 3D reconstruction. It can create Matrix-style bullet-time shots from just three iPhones, and the team says it can reduce capture requirements by 50 to 100 times. Atlas builds on Marble but moves beyond its Gaussian-splat output bottleneck, handling text, images, video, depth, and camera data together. They describe potential uses in film, games, architecture, and robotics, where faster real-to-simulation workflows could address robotics’ shortage of training data. The team says scaling is still compute-limited, dynamics are already present in Atlas, and future work will improve editability and interaction. They argue new-view prediction could become a foundational intelligence primitive, much like next-token prediction for language models.

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, Reconstruction, and Depth Maps
3. 8:15Spatial Intelligence & Why New View Prediction Matters: Fei-Fei Li on Camera Pose and Atlas's Path Beyond Marble
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: Stanford Quad and House Fly-Throughs from Sparse Images
6. 21:27Did You Know It Was Going to Work?: Scaling Laws, Viewpoint Prediction, and Compute Limits
7. 24:42Use Cases: Marvel’s Gaussian-Splat Views and Persistent 3D Design
8. 28:05Use Cases: Atlas Targets 3D Revision Work and Robotics Data Bottlenecks
9. 32:17Use Cases: Dynamic Atlas, Neural Simulation, and Robotics Planning
10. 35:21Video Models vs. World Models: Atlas Adds Motion Beyond Static Marble
11. 37:55Will We Get Walkable 4D Video? Ben’s Editability Goals and AI Completeness
12. 41:59Why Atlas’s New-View Prediction Could Be Spatial Next-Token Prediction

What the full analysis includes

12 chapters · 62 key points · 12 section summaries · mind map

The remaining 59 key points, section summaries and mind map are in the full analysis. Analyze this video to see everything →

This video is 44 min. The Free plan analyzes videos up to 30 min; plans from Plus cover videos up to 2 hours.