Summary
Welch Labs examines how World Labs, the company co-founded by AI pioneer Fei-Fei Li, built its Atlas world model. The video reports that World Labs was founded in January 2024 and that AMD announced an $8 billion acquisition deal in September 2026. It traces the path to Atlas through three decisions by the team: what the model should predict, which architecture to use and how much to bet on scale.
The first model, Marble, was trained to predict 3D representations directly, and the team later pivoted away from that. Atlas adopts an autoregressive diffusion transformer and was trained from scratch. Interview clips with Li and the World Labs team explain the reasoning. The video also revisits Li’s earlier work on the ImageNet dataset, which helped set off the deep learning era after AlexNet in 2012.
To explain the architecture, the video visualizes attention patterns inside the open-source Wan video diffusion model. It shows how bidirectional attention heads track objects such as a bouncing ball across frames, and why that approach differs from autoregressive generation.
📺 Source: Welch Labs · Published November 20, 2026
🏷️ Format: Deep Dive







