I think there's a cultural disconnect here. I remember Bostrom 2014 where value learning was supposed to be impossible until after the model is already superintelligent and doesn't care. "RL imparts…

Posters: jd_pressman. jd_pressman

3 likes recorded by the source corpus.