
Jacob Steinhardt, UC Berkeley: Machine learning safety, alignment and measurement
Generally Intelligent
00:00
The History of M L Models
The paper was written very differently from a standard bench mark paper. Usually they would be like, ok, here's data. Hear's how he collected here the base ans. And youare trying to identify classes with patterns. So like adversary examples that you mentioned earlier, or like values, time consistency of values. Sometimes a data sets a good way to get ot a phenomena. It doesn't have to be that way. Cool. What do you feel lik is unusual order? Maybe the ethic paper actually tell us about that. Well, so this was thi datus that we collected of just a bunch of moral judgments, and we were trying to train m l models to predict those judgments
Play episode from 33:02
Transcript


