From Seeing to Doing by Fei-Fei Li at NeurIPS

Visual intelligence evolves through a ladder of understanding, reasoning, and generation, fundamentally shifting from flat 2D pixel-based analysis to 3D spatial intelligence. This progression relies on the synergy between diverse, large-scale data and robust algorithms, as demonstrated by milestones like ImageNet and modern diffusion models. Beyond perception, spatial intelligence enables robots to interact with the world, moving from brittle, isolated tasks to complex, long-horizon activities in ecological environments. Rather than replacing human labor, AI serves as a transformative tool to augment human capabilities in critical sectors like healthcare, where it improves patient safety and mobility, and in creative fields, where it facilitates 3D environment generation. Ultimately, the future of AI lies in its ability to bridge the gap between digital representations and the physical, 3D reality humans inhabit, fostering a symbiotic relationship that enhances human agency and problem-solving.