Yep, I was thinking about outputs from you (in addition to outputs from others) while writing this.
Given that human experts seem to have systematic biases (e.g., away from ensembling), wouldn’t the ensemble also exhibit such biases? Have you taken this into account when saying “My guess is that the answer is yes”?
Yes, I’m including this sort of consideration when I say top human experts do would probably be sufficient.
Your point around companies not wanting to ensemble and wanting to use their own epistemic strategies is a fair one, I’ll think about it more. This is connected to a broader point around political difficulties in making an AI that’s actually good to defer to.
Yep, I was thinking about outputs from you (in addition to outputs from others) while writing this.
Yes, I’m including this sort of consideration when I say top human experts do would probably be sufficient.
Your point around companies not wanting to ensemble and wanting to use their own epistemic strategies is a fair one, I’ll think about it more. This is connected to a broader point around political difficulties in making an AI that’s actually good to defer to.