Li fei-fei's world model Atlas paves the way for AGI

On September 7, 2026, Li Fei-Fei and the World Labs team unveiled their latest multi-modal world model, Atlas, at a16z. Unlike traditional models that predict the next token or frame, Atlas focuses on 'new view prediction,' enabling it to generate images at any spatial and temporal position based on input images or text. This capability integrates generation and reconstruction within a single model, handling text, images, videos, 3D data, and camera poses. The team views 'new view prediction' as a fundamental primitive comparable to token prediction in large language models, potentially key to achieving AGI. Atlas significantly reduces data requirements for 3D reconstruction, generating high-quality results from just a few images. This advancement addresses limitations in earlier models like Marble, which struggled with Gaussian splats. Atlas's scalable spatial context window allows for precise spatial reconstruction, supporting applications in creative design, construction, and robotics. The model already has dynamic processing capabilities, with future improvements targeting dynamic effects, content editing, and human-machine interaction to enable intuitive 3D virtual world interaction.