The Shift to New View Prediction

World Labs co-founders Fei-Fei Li, Justin Johnson, and Ben Mildenhall argue that the next frontier in AI is 'spatial intelligence.' While Large Language Models (LLMs) rely on next-token prediction and video models rely on next-frame prediction, Atlas introduces 'new view prediction.' By treating the viewpoint and camera pose as a native input, the model can generate or reconstruct scenes from arbitrary perspectives in space and time. This allows the system to understand the physical world's geometry and structure rather than just hallucinating pixels.

Unifying Generation and Reconstruction

Historically, computer vision has been bifurcated into two distinct tracks: pixel generation (creating new content) and 3D reconstruction (mapping existing physical spaces). Atlas merges these into a single, multimodal architecture. By natively incorporating text, images, video, and 3D camera poses, the model can take sparse inputs—such as three photos of a scene—and generate a high-fidelity 3D representation. This represents a 50x to 100x reduction in the data required for traditional 3D reconstruction, which previously demanded exhaustive, dense coverage of a space.

The Path to Spatial Intelligence

Fei-Fei Li defines spatial intelligence as the ability to generate, reason within, edit, and interact with a space. Atlas serves as a foundational step by grounding pixels in 3D space. The team emphasizes that while current video models are impressive, they often lack spatial consistency. Atlas solves this by ensuring that every frame is anchored to a specific 3D camera pose, allowing for 'bullet time' effects or flythroughs that are governed by physical laws rather than stochastic generation. The team views this as a critical primitive for future applications in robotics, architecture, and creative simulation.

Scaling and Technical Evolution

Reflecting on their two-year journey, the founders noted that they did not start with Atlas. Their previous model, Marble, focused on Gaussian splats as an output representation. While effective, the team realized that bottlenecking the model through a specific output format limited its potential. Atlas represents a more unified approach where the model learns the underlying spatial context first, allowing it to output various formats—including RGB frames and 3D reconstructions—on demand. The team stressed that scaling these models requires significant GPU resources and a willingness to iterate through smaller experiments to build conviction in the underlying scaling laws.