Re:AI researcher Timnit Gebru resigns from Google
by joshuamorton on Friday December 04, 2020 @05:33PM
(#25309198)
> Yann didn't say anybody about harms nor did the guy he was replying to (who talked of dangers)
This feels like unreasonable semantics. The dangers are precisely the danger of causing harm. The harms therefore are concrete results of theoretical dangers manifesting. They aren't different.
> In particular Yann did not suggest in any way a) there are no harms
I agree, and I've said as much.
> b) that harms are only related to biased training sets
He did, insofar as he suggested that the dangers were due solely to bias in the training set, which is implied when he says that if you train the same model on different data, everyone looks African. Like yes, that is true, but it doesn't reduce the harms (or dangers, if you want to be precise). It just creates a different set of biases with a different set of potential harms (which again, are "dangers").
This feels like unreasonable semantics. The dangers are precisely the danger of causing harm. The harms therefore are concrete results of theoretical dangers manifesting. They aren't different.
> In particular Yann did not suggest in any way a) there are no harms
I agree, and I've said as much.
> b) that harms are only related to biased training sets
He did, insofar as he suggested that the dangers were due solely to bias in the training set, which is implied when he says that if you train the same model on different data, everyone looks African. Like yes, that is true, but it doesn't reduce the harms (or dangers, if you want to be precise). It just creates a different set of biases with a different set of potential harms (which again, are "dangers").
I'm not seeing a strawperson.