tamera comments on Relaxed adversarial training for inner alignment

[tamera]

I’m not sure what’s going on with the types in this equation, at the start of the formalization section:

$$\begin{array}{cc}\text{Adv}\left(M\right)=\text{Amp}\left(M\right)(& \text{“What's a pseudo-input on which}M\text{is particularly}\\ & \text{likely to recommend an unacceptable action?”})\end{array}$$

I’d think that the left side represents a pseudo-input, while the right represents an action. Am I missing something?

[evhub]

Actions are just language outputs—and since we ask for an action that describes a pseudo-input, hopefully we should be able to interpret it that way.