Chinese models are sometimes better, even if they're distilled (dualuse.dev)

It is well known that while Chinese models compete on some dimensions in public benchmarks, they suck once evaluating against private benchmarks.

At the risk of hand-waving a complex personal belief, I don't really believe in copyright, but for us to train models on all of human writing, without asking or paying, then turn around and scream 'foul' at someone training a model on the first models seems... genuinely insane? I don't quite have an easy to way to verbalize this level of staggering hypocrisy. I suppose nothing should surprise me anymore, but here I sit, surprised.