careful here. What normal people say when they say does category X have property Y? Is ‘does it have it often, most of the time, etc’. But when computer scientists or mathematicians ask that they mean ‘does every single instance of X have property Y’ which in the real world happens almost never except for very easily formalizable things. Yes, almost almost always, AI alignment is computable. Because you can compute it to arbitrary (up to very very high physical bounds) precision in pretty much all practical cases. But they will try to say it isn’t, because of some conceivable but implausible situation, or because it isn’t computable with infinite precision, or some other foolishness. I stand with the normies!

This reduces to the halting problem when the AI is not supposed to be in an infinite loop, so no

@cockathiel what do you mean by “computable”? Is this just in principle? (If so I think most things should be computable in principle or at least have very good computable approximations, because “computable” is an extremely broad and flexible category.)