
Dylan Hadfield-Menell, UC Berkeley/MIT: The value alignment problem in AI
Generally Intelligent
00:00
How Do You Evaluate Your Processes?
I think we should be moving away from straight tative performance measures and moving in towards qualitative evaluations of systems. We don't teach that as part of the process of designing and building these deparell systems. I've started taking a pretty direct approach to when you build your system, you have this evolving set of quantitative metrics. And i think most a i researchers who succeed have something like this, either directly or indirectly, that they're doing. It's not magic.
Play episode from 52:15
Transcript


