Computer Vision Interview Questions, issue 16, Jan 17, 2026
The Contrastive Hard Negative Trap
Senior AI Engineer interview at OpenAI, and the interviewer asks:
“Our CLIP model keeps confusing Golden Retrievers with Yellow Labs. To fix it, we’re going to manually curate hard negative batches, forcing these similar breeds into the same training step. Good idea?”
How aggressive batch difficulty pushes CLIP from semantic understanding into pixel-level cheating.
The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.