4 ms·
This method assumes that the weaker model is aligned. I'm curious how the paper addresses that point. > "But what does this second turtle stand on?" persisted
by akprasad 3y ago
This method assumes that the weaker model is aligned. I'm curious how the paper addresses that point.
> "But what does this second turtle stand on?" persisted James patiently.
> To this, the little old lady crowed triumphantly,
> "It's no use, Mr. James—it's turtles all the way down."
- Ninjinka 3y agoI think the assumption is we can align models less intelligent than ourselves, the hard part is aligning models that are more.
- PartiallyTyped 3y agoRecursive bootstrapping?
- oersted 3y agoThe weaker model is just a stand-in for human supervisors. They are experimenting on GPT-2 supervising GPT-4 as an analogue to humans supervising the superhuman AGI.