“In training AI systems, contrastive methods are problematic for large output spaces because they require a gigantic number of contrastive samples.”