Is instrumentalism such a bad thing, though? It seems like instrumentalism is a better generalization of Bayesian reasoning than scientific realism
If you take instrumentalism seriously, then you remove external “reality” as meaningless, and only talk about inputs (and maybe outputs) and models. Basically in this diagram
from Update then Forget you remove the top row of W’s, leaving dangling arrows where “objective reality” used to be. This is not very aesthetically satisfactory, since the W’s link current actions to future observations, and without them the causality is not apparent or even necessary. This is not necessarily a bad thing, if you take care to avoid the known AIXI pitfalls of wireheading and anvil dropping. But this is certainly not one of the more popular ontologies.
If you take instrumentalism seriously, then you remove external “reality” as meaningless, and only talk about inputs (and maybe outputs) and models. Basically in this diagram
from Update then Forget you remove the top row of W’s, leaving dangling arrows where “objective reality” used to be. This is not very aesthetically satisfactory, since the W’s link current actions to future observations, and without them the causality is not apparent or even necessary. This is not necessarily a bad thing, if you take care to avoid the known AIXI pitfalls of wireheading and anvil dropping. But this is certainly not one of the more popular ontologies.