From Labels to Grounding
ImageNet made the modern eye possible by teaching machines to name what appears in a photograph. But naming is not placement. Fei-Fei Li's spatial turn asks for pixels that are 'truly spatially contextualized and grounded'—a phrase from the a16z conversation that captures the difference between detection and worldliness. YOLO accelerated real-time recognition; CLIP joined vision to language. Neither closes the gap alone. The hard step is making representation answer to three-dimensional space, where objects persist, occlude, and move. Atlas is her wager that the gap can be crossed.



