Ability to generalize is something we would expect from deliberate optimizers. The point is that the strategies of accidental optimizers do not generalize well compared to deliberate optimizers.
Deliberate optimizers intend to optimize and they do in fact optimize by applying one or more strategies. We would expect those strategies work well when a different agent applies those strategies to optimize the same criterion. In that sense, if the observer is this different agent, then yes.
Great points. Two comments:
Ability to generalize is something we would expect from deliberate optimizers. The point is that the strategies of accidental optimizers do not generalize well compared to deliberate optimizers.
Deliberate optimizers intend to optimize and they do in fact optimize by applying one or more strategies. We would expect those strategies work well when a different agent applies those strategies to optimize the same criterion. In that sense, if the observer is this different agent, then yes.