Ben, to put that point more generally, Eliezer seems to be neglecting to consider the fact that utility is sometimes a reason to associate several concepts, even apart from their probability of being associated with one another or with other things. An example from another commenter would be “I want a word for red flowers because I like red flowers”; this is entirely reasonable.
A nice point. But it there is something a bit weird.
We all agree that utility is subjective—you like red flowers, I like blue. As Bayesians, we also know that empirical information is somewhat subjective as well. We don’t all have access to the same empirical observations.
So, it appears that we cannot really “carve nature at the (objective) joints. The best we can do is to carve where we (subjectively) estimate the joints to be. Now, that is fine if we are using words only to carry out private inferences. But we also frequently need to use words to communicate.
It appears that even Bayesian rationalists who understand cluster analysis must sometimes argue about definitions. There may be arguments regarding how to delimit the population. There may be arguments about how best to quantize the variables. It is not completely clear to me that it is always possible to distinguish communication problems from single-person inference problems.
I think that the usefulness of “Sparg” is that it has a stronger correlation with “usefulness to Ben” than is indicated by the seperitate correlations of “usefulness to Ben” with “carrying size”, “contains Vanduim”, and “rock”.
Ben, to put that point more generally, Eliezer seems to be neglecting to consider the fact that utility is sometimes a reason to associate several concepts, even apart from their probability of being associated with one another or with other things. An example from another commenter would be “I want a word for red flowers because I like red flowers”; this is entirely reasonable.
A nice point. But it there is something a bit weird.
We all agree that utility is subjective—you like red flowers, I like blue. As Bayesians, we also know that empirical information is somewhat subjective as well. We don’t all have access to the same empirical observations.
So, it appears that we cannot really “carve nature at the (objective) joints. The best we can do is to carve where we (subjectively) estimate the joints to be. Now, that is fine if we are using words only to carry out private inferences. But we also frequently need to use words to communicate.
It appears that even Bayesian rationalists who understand cluster analysis must sometimes argue about definitions. There may be arguments regarding how to delimit the population. There may be arguments about how best to quantize the variables. It is not completely clear to me that it is always possible to distinguish communication problems from single-person inference problems.
I think that the usefulness of “Sparg” is that it has a stronger correlation with “usefulness to Ben” than is indicated by the seperitate correlations of “usefulness to Ben” with “carrying size”, “contains Vanduim”, and “rock”.