On-policy distillation improves AI model training efficiency

Original: Great little video on modern on-policy distillation in post-training recipes.

Source: x.com ↗