Fei-Fei Li on World Labs' Atlas: new view prediction as a world model primitive
a16z Podcast · rss · 2026-09-04
World Labs co-founders Fei-Fei Li, Justin Johnson, and Ben Mildenhall join a16z's Martin Casado to discuss Atlas, their latest world model. Core mechanism: "new view prediction" — given images of a scene, the model predicts how it looks from another position in space and time, unifying generation and 3D reconstruction in one model and raising the question of whether view prediction could be a primitive for understanding the physical world.
- The team explains the technical bets behind Atlas and what it can and can't yet capture
- Dynamics, editability, and simulation are flagged as key directions for world models
- Applications span creative work, architecture, and robotics; Li argues access to real-world training data is one of today's biggest constraints
More from Multimodal
- StreamTalk turns audio into SMPL-X body gestures, rig of your choice — multimodalart · 2026-09-04
- GPT Image 2 prompt trick: ads where your product is the only missing puzzle piece — aziz4ai · 2026-09-04
- Google's Atlas video model nails pixel-perfect camera control incl. fisheye and distortion models — teortaxesTex · 2026-09-04
- Creator uses MiniMax H3 as a local product-CG renderer before hitting the API — Hailuo_AI · 2026-09-04
- Spreadsheets as a Batch Interface for Multi-Video Generation — GeroldMeisinger · 2026-09-04
- MiniMax Design shows off Brutalist-style MV generated entirely by text with H3 — Hailuo_AI · 2026-09-04