michaelcohen comments on Formal Solution to the Inner Alignment Problem

[michaelcohen]

It may be helpful to point to specific sections of such a long paper.

(Also, I agree that a neural network trained trained with that reward could produce a deceptive model that makes a well-timed error.)