I sometimes wonder. Maybe it’s the other way round....
iwdw & Dave—it’s a tempting idea, but I’d say that ultimately it’s wrong.
My liking of Wigginettes is a fact about me, not a fact about Wigginettes. I can’t spontaneously create a new Thingspace dimension, say ‘look, Wigginettes glow when you look through this dimension, hence Wigginettes is an objectively valid category’. My liking is based on two unrelated properties, A and B, and maybe that ‘creates’ a third property C, but that property only describes me. Yes, my liking can be described neurochemically if you like, but that’s information about my brain. It doesn’t tell you anything about Wigginettes.
I hope Eliezer will back me up on this. Remember that Thingspace should be based on Scientific Facts About Things—that’s the only way it can help us think about the world.( I guess you could argue that we each carry our own little Thingspace around with us, but, well, meh. Ideal Thingspace is (supposed to be) a direct, exhaustive map of the territory, not a map of my map.) Assigning the property ‘liked by Ben’ to Wigginettes (rather than me) is a mistake.
[Hence Eliezer’s word ‘arbitrary’ when talking about trying to give this category/utility function to an AI. I suppose the measure of an ‘arbitrary utility function’ is whether or not it requires a hack to be transferred to a machine?]
But my original question stands. I’m not drawing a boundary around any objective pattern in Thingspace. Is Wigginettes a wrong word? Maybe this comes down to whether words are based on Facts About Things or Human Utility, Sticking To The Facts. And surely the latter is the more useful of the two. Eliezer, where do you go to define your words, Thingspace, or your head?
But… your head is part of reality, is it not?
I sometimes wonder. Maybe it’s the other way round....
iwdw & Dave—it’s a tempting idea, but I’d say that ultimately it’s wrong.
My liking of Wigginettes is a fact about me, not a fact about Wigginettes. I can’t spontaneously create a new Thingspace dimension, say ‘look, Wigginettes glow when you look through this dimension, hence Wigginettes is an objectively valid category’. My liking is based on two unrelated properties, A and B, and maybe that ‘creates’ a third property C, but that property only describes me. Yes, my liking can be described neurochemically if you like, but that’s information about my brain. It doesn’t tell you anything about Wigginettes.
I hope Eliezer will back me up on this. Remember that Thingspace should be based on Scientific Facts About Things—that’s the only way it can help us think about the world.( I guess you could argue that we each carry our own little Thingspace around with us, but, well, meh. Ideal Thingspace is (supposed to be) a direct, exhaustive map of the territory, not a map of my map.) Assigning the property ‘liked by Ben’ to Wigginettes (rather than me) is a mistake.
[Hence Eliezer’s word ‘arbitrary’ when talking about trying to give this category/utility function to an AI. I suppose the measure of an ‘arbitrary utility function’ is whether or not it requires a hack to be transferred to a machine?]
But my original question stands. I’m not drawing a boundary around any objective pattern in Thingspace. Is Wigginettes a wrong word? Maybe this comes down to whether words are based on Facts About Things or Human Utility, Sticking To The Facts. And surely the latter is the more useful of the two. Eliezer, where do you go to define your words, Thingspace, or your head?
-Apologies for length, all.