Chinese models are sometimes better, even if they're distilled(dualuse.dev)
dualuse.dev
Chinese models are sometimes better, even if they're distilled
https://dualuse.dev/posts/chinese-models-are-sometimes-better-even-if-distilled
4 comments
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.
It is well known that while Chinese models compete on some dimensions in public benchmarks, they suck once evaluating against private benchmarks.
I'm happy with them and would rather support those who share. Ai is taken from the history of human knowledge, so it seems fair we all get access to the fruits built on out shared knowledge.
Not sure how much I trust private benchmarks, preferences and incentives are unknown
Not sure how much I trust private benchmarks, preferences and incentives are unknown
In this case, the benchmarks were private and it still outperformed.