David Krueger–Coordination, Alignment, Academia
The Inside View
00:00
The Importance of Reward Hacking
So basically what we're doing is we're abstracting away the training process like the learning curves so there's this characteristic thing you'll see when you actually see reward hacking in practice sometimes where like the real reward and the proxy are both going up. And as soon as that happens your optimization is producing a sequence of policies, right? So then at some point the proxy goes up and the real reward goes down. That means that there was a policy before that happened and a policy after that happened. And that's that pair of policies that I'm talking about right so you're only talking about those two and not. The infinite number of policies in between or yeah because you don't actually visit
Transcript
Play full episode
