Jacob Steinhardt, UC Berkeley: Machine learning safety, alignment and measurement
Generally Intelligent
00:00
Is There a Measure for Robustness or Capabilities?
A concrete thing would be like trying to decide what article is to give in a new steed. A solution that generally works right now pretty well is ou just pre train on a lot of unlabelled data. But i think that might break down if you had weirder distribution shifts that aren't sort of captured by your unlabelled, dated distribution. I think there's many faces, one for robustness. There's a few other ones, but i think those are cally the best we have right now.
Play episode from 25:50
Transcript
