Dylan Hadfield-Menell, UC Berkeley/MIT: The value alignment problem in AI
Generally Intelligent
00:00
Unsupervised Learning and Manipulation
Unsupervised learning seems like a really good candidate for how to learn what people will do on line without actually learning what they should do. We don't even have sort of good theories of what it could look likef like a person that could enable manipulation. Olyy, what does it mean to measure manipulation, or prevent it or keep it from happening? Is really tough, and it' something you can look at on a case by case basis.
Play episode from 01:21:42
Transcript
