Ah, okay. I don’t see any reason to be concerned about something that we have no effect on. Will try to explain below.
Regarding “subjunctive dependency” from the post linked in your other reply:
I agree with a version of “They are questions about what type of source code you should be running”, formulated as “what type of an algorithm results in max EV, as evaluated by the same algorithm?” This removes the contentious “should” part, that implies that you have an option of running some other algorithm (you don’t, you are your own algorithm).
The definition of “subjunctive dependency” in the post is something like “the predictor runs a simplified model of your actual algorithm that outputs the same result as your source code would, with high fidelity” and therefore the predictor’s decisions “depend” on your algorithm, i.e. you can be modeled as affecting the predictor’s actions “retroactively”.
Note that you, an algorithm, have no control of what that algorithm is, you just are it, even if your algorithm comes equipped with the routines that “think” about themselves. If you also postulate that the predictor is an algorithm, as well, then the question of decision theory in presence of predictors becomes something like “what type of an agent algorithm results in max EV when immersed in a given predictor algorithm?” In that approach the subjunctive dependency is not a very useful abstraction, since the predictor algorithm is assumed to be fixed. In which case there is no reason to consider causally disconnected parts of the agent’s universe.
Clearly your model is different from the above, since you seriously think about untestables and unaffectables.
Ah, okay. I don’t see any reason to be concerned about something that we have no effect on. Will try to explain below.
Regarding “subjunctive dependency” from the post linked in your other reply:
I agree with a version of “They are questions about what type of source code you should be running”, formulated as “what type of an algorithm results in max EV, as evaluated by the same algorithm?” This removes the contentious “should” part, that implies that you have an option of running some other algorithm (you don’t, you are your own algorithm).
The definition of “subjunctive dependency” in the post is something like “the predictor runs a simplified model of your actual algorithm that outputs the same result as your source code would, with high fidelity” and therefore the predictor’s decisions “depend” on your algorithm, i.e. you can be modeled as affecting the predictor’s actions “retroactively”.
Note that you, an algorithm, have no control of what that algorithm is, you just are it, even if your algorithm comes equipped with the routines that “think” about themselves. If you also postulate that the predictor is an algorithm, as well, then the question of decision theory in presence of predictors becomes something like “what type of an agent algorithm results in max EV when immersed in a given predictor algorithm?” In that approach the subjunctive dependency is not a very useful abstraction, since the predictor algorithm is assumed to be fixed. In which case there is no reason to consider causally disconnected parts of the agent’s universe.
Clearly your model is different from the above, since you seriously think about untestables and unaffectables.