Fei-Fei Li explains Atlas: new-view prediction as a primitive toward AGI

量子位 · wechat · 2026-09-05

Fresh off World Labs' launch of Atlas, billed as the first multimodal world model, Fei-Fei Li and co-founders Justin Johnson and Ben Mildenhall sat down with a16z to unpack it.

Key takeaways:

Core primitive is new-view prediction: like next-token for LLMs and next-frame for video models, Atlas predicts new viewpoints. Inputs (images with 3D camera poses) form a "spatial context" from which it renders the world from arbitrary positions.
Generation and reconstruction unified: one model natively handles text, images, video, 3D depth maps and camera poses; Li calls this the first unification in computer vision's 50-year history.
Massive capture reduction: dense reconstruction once needed hundreds of photos; Atlas rebuilds scenes from 3-25 images (e.g., Stanford's main quad aerial view generated from ground-level shots), and bullet-time footage with just three iPhones.
Robotics: Li argues robots' bottleneck is data, not chips—Atlas can port real environments into simulation with domain randomization.
Roadmap: next steps are dynamics, editing and interaction; the team frames new-view prediction as a potential foundational primitive toward AGI.

Unlike the previous Marble model (bottlenecked by Gaussian splatting output), Atlas was designed multimodal from pretraining onward.

Original post →

Fei-Fei Li explains Atlas: new-view prediction as a primitive toward AGI

More from AGI Musings