Fei-Fei Li and Justin Johnson discuss the limitations of current AI architectures, particularly language models' inability to grasp spatial and physical reasoning. They argue that true intelligence requires understanding 3D environments through interaction, not just text-based prediction. Their new model, Marble, generates explorable 3D worlds from text or images, leveraging Gaussian splats for real-time rendering and precise camera control. They emphasize that spatial intelligence—shaped by eons of evolution—is fundamentally different from linguistic intelligence and cannot be reduced to sequences. Transformers, they note, are better understood as set models, with sequence processing being an add-on. This insight informs their vision for world models that simulate physics and eliminate impossible configurations through structured reasoning, paving the way for applications in robotics, design, and simulation.