A deep dive into Meta AI's self-supervised vision foundation model — exploring how DINOv2 learns robust visual features from 142M curated images without any labels, and why it rivals weakly-supervised methods across classification, segmentation, depth estimation, and beyond.