New method extracts proprietary AI model knowledge ten times faster

Original: Extremely interesting that this beats the sample efficiency of OPD. I suspect some labs will try this at scale. Quite concerning for alignment since it seems to distill proprietary models quite well.

Source: x.com ↗