2026-09-18 03:13 UTC
【by 2021, OpenAI considered the prevention of memorization important "for fair use [compliance] and minimizing copyright violations in model output.】
Replies (1)
-
@SnoopJ@hachyderm.io 2026-09-18 03:20
Lots of discussion of n-grams (n=16, 25). I'm not sure how convincing those arguments are in summary, but the sheer volume of the overlaps found (O(100k)) carries some weight, as does the fact that OpenAI's own technical experts did the analysis. Hard to evaluate, but there's smoke near all the fire.