I don’t think that’s a fair restatement of my expanded observations. It depends on what you mean by “stuff”—I definitely disagree if you substitute “declarative knowledge” for it, and this is what “more stuff” tends to imply.
If “stuff” includes all forms of insight as well as declarative knowledge, then I’d more or less agree, with the provision that you must also know the right kind of stuff, that is, have meta-knowledge about when to apply various kinds of insights.
I quite like the frame of Eliezer’s that “intelligence is efficient cross-domain optimization”, but I can’t think of a simple test for measuring optimization power.
The demand for “the simplest credible way” sounds suspiciously like it’s asking for a shortcut to assessing optimization power. I doubt that there is such a shortcut. Lacking such a shortcut, a good proxy, or so it seems to me, is to assess what a person’s optimization power has gained them: if they possess knowledge or insights that I don’t, that’s good evidence that they are good at learning. If they consistently teach me things (if I fail to catch up to them), they’re definitely smarter. So each thing they teach me is (probabilistic) evidence that they are smarter.
Hence my use of “teach me something” as a unit of evidence for someone being smarter.
I don’t think that’s a fair restatement of my expanded observations. It depends on what you mean by “stuff”—I definitely disagree if you substitute “declarative knowledge” for it, and this is what “more stuff” tends to imply.
That’s reasonable. I don’t mean to reframe your position as something silly, rather I say that I do not have a definition of ‘smarter’ for which the below is true:
They were smarter than you then, in the topic area in which you learned something from them.
When you’ve caught up with them, and you start being able to teach them instead of them teaching you, that’s a good hint that you’re smarter in that topic area.
I agree with what you say here:
The demand for “the simplest credible way” sounds suspiciously like it’s asking for a shortcut to assessing optimization power. I doubt that there is such a shortcut. Lacking such a shortcut, a good proxy, or so it seems to me, is to assess what a person’s optimization power has gained them: if they possess knowledge or insights that I don’t, that’s good evidence that they are good at learning. If they consistently teach me things (if I fail to catch up to them), they’re definitely smarter. So each thing they teach me is (probabilistic) evidence that they are smarter.
..but with a distinct caveat of all else being equal. ie. If I deduce that someone has x amount of more knowledge than me then that can be evidence that they are not smarter than me if their age or position is such that they could be expected to have 2x more knowledge than me. So in the ‘my teachers when I was 8’ category it would be a mistake (using my definition of ‘smarter’) to make the conclusion: “They were smarter than you then, in the topic area in which you learned something from them”.
I don’t think that’s a fair restatement of my expanded observations. It depends on what you mean by “stuff”—I definitely disagree if you substitute “declarative knowledge” for it, and this is what “more stuff” tends to imply.
If “stuff” includes all forms of insight as well as declarative knowledge, then I’d more or less agree, with the provision that you must also know the right kind of stuff, that is, have meta-knowledge about when to apply various kinds of insights.
I quite like the frame of Eliezer’s that “intelligence is efficient cross-domain optimization”, but I can’t think of a simple test for measuring optimization power.
The demand for “the simplest credible way” sounds suspiciously like it’s asking for a shortcut to assessing optimization power. I doubt that there is such a shortcut. Lacking such a shortcut, a good proxy, or so it seems to me, is to assess what a person’s optimization power has gained them: if they possess knowledge or insights that I don’t, that’s good evidence that they are good at learning. If they consistently teach me things (if I fail to catch up to them), they’re definitely smarter. So each thing they teach me is (probabilistic) evidence that they are smarter.
Hence my use of “teach me something” as a unit of evidence for someone being smarter.
That’s reasonable. I don’t mean to reframe your position as something silly, rather I say that I do not have a definition of ‘smarter’ for which the below is true:
I agree with what you say here:
..but with a distinct caveat of all else being equal. ie. If I deduce that someone has x amount of more knowledge than me then that can be evidence that they are not smarter than me if their age or position is such that they could be expected to have 2x more knowledge than me. So in the ‘my teachers when I was 8’ category it would be a mistake (using my definition of ‘smarter’) to make the conclusion: “They were smarter than you then, in the topic area in which you learned something from them”.