+In November, I received an interesting reply on my philosophy-of-categorization thesis from MIRI researcher Abram Demski. Abram asked: ideally, shouldn't all conceptual boundaries be drawn with appeal-to-consequences? Wasn't the problem just with bad (motivated, shortsighted) appeals to consequences? Agents categorize in order to make decisions. The best classifer for an application depends on the costs and benefits. As a classic example, it's very important for evolved prey animals to avoid predators, so it makes sense for their predator-detection classifiers to be configured such that they jump away from every rustling in the bushes, even if it's usually not a predator.
+
+I had thought of the "false-positives are better than false-negatives when detecting predators" example as being about the limitations of evolution as an AI designer: messy evolved animal brains don't bother to track probability and utility separately the way a cleanly-designed AI could. As I had explained in "... Boundaries?", it made sense for _what_ variables you paid attention to, to be motivated by consequences. But _given_ the subspace that's relevant to your interests, you want to run an epistemically legitimate clustering algorithm on the data you see there, which depends on the data, not your values. The only reason value-dependent gerrymandered category boundaries seem like a good idea if you're not careful about philosophy is because it's _wireheading_. Ideal probabilistic beliefs shouldn't depend on consequences.
+
+Abram didn't think the issue was so clear-cut. Where do "probabilities" come from, in the first place? The reason we expect something like Bayesianism to be an attractor among self-improving agents is _because_ probabilistic reasoning is broadly useful: epistemology can be _derived_ from instrumental concerns. He agreed that severe wireheading issues potentially arise if you allow consequentialist concerns to affect your epistemics.
+
+But the alternative view had its own problems. If your AI consists of a consequentialist module that optimizes for utility in the world, and an epistemic module that optimizes for the accuracy of its beliefs, that's _two_ agents, not one: how could that be reflectively coherent? You could, perhaps, bite the bullet here, for fear that consequentialism doesn't tile and that wireheading was inevitable. On this view, Abram explained, "Agency is an illusion which can only be maintained by crippling agents and giving them a split-brain architecture where an instrumental task-monkey does all the important stuff while an epistemic overseer supervises." Whether this view was ultimately tenable or not, this did show that trying to forbid appeals-to-consequences entirely led to strange places. I didn't immediately have an answer for Abram, but I was grateful for the engagement. (Abram was clearly addressing the real philosophical issues, and not just trying to mess with me the way almost everyone else in Berkeley was trying to mess with me.)