@prietschka@mastodon.social @timnitGebru@dair-community.social @CyberneticForests@assemblag.es it does, empirically. Even recent adversarial attacks against Meta work better being “nicer” to the model. None of this is to ascribe sentience or self awareness to something mechanistic but there is a huge gap between what people think is going on and what actually is. Interpretability IS a problem and it has measurably changed since that article was published last year. It is, however, absolutely insane and rolling out black boxes was foolish.