Every DINOv2 story collected by The AI Daily, 2 so far, newest first, refreshed hourly.
Small-scale replication: good views beat next-token predictionA researcher replicated Balestriero and LeCun's “good views” paper at laptop scale, using 10,000 COCO photos, human captions and 11M-parameter encoders. Good views scored 0.320 versus 0.192 for bad views and 0.148 for next-token prediction; scale limits remain.
Shelf audit: YOLO detects products, but embeddings confuse sibling SKUsA developer building a shelf-audit tool reports that YOLO detects products well, but SKU identification with DINOv2, SigLIP2, and OpenCLIP often returns the wrong sibling SKU for same-brand variants like 1.25L vs 2L. Correct and incorrect scores overlap, making thresholds unreliable, and resizing crops to 224px erases small size text. Reference photos are few and not studio-quality.
The A.I. Daily