Dylan Hadfield-Menell, UC Berkeley/MIT: The value alignment problem in AI
Generally Intelligent
00:00
The General Purpose Programming Language in Machine Learning?
In mission learning, data augmentation is a way to take your data set and make it different. I think it is quite possible to make a lot of money with what we would call more shallow metrics. Optimizing for engagement or watch time is really effective. It has a lot of negative externalities, which are hard to measure. And so i think that, a, i systems have a bias towards optimizing for what's easy to measure.
Play episode from 55:31
Transcript
