AI and Existential Risk (Robert Wright & Connor Leahy)
Robert Wright's Nonzero
00:00
The Dangers of Good Alignment Research
RLA-CHF is a kind of proto-alignment sort of technique, but not robust. So one thing that could have happened was they see, oh, we can use this to constrain our things and like to use smaller, more constrained models That still do what we want. And then they use smaller models, they have more constraint, et cetera. Of course, they just took the biggest models they could and used it to teach you as many new skills as possible as fast as possible. They also used it to teaching them new things, how to do more math and how to reason better or whatever. Any good alignment research which gets you control will also make you
Transcript
Play full episode
