Experts should be able to regularly win bets against nonexperts, and generally valued generally scarce commodities like money should be able to incentivize people to make bets. If you don’t know how to construct an experiment to distinguish experts from nonexperts, you probably do not have a clear idea what it is you want an expert on, and if you don’t have any idea what you are trying to buy, it’s unclear what it would even mean to intend to buy it.
Abstract example: 16th Century “math duels,” in which expert algebraists competed to solve quantitative problems.
Concrete example: LockPickingLawyer, who demonstrates on YouTube how to easily defeat many commercially popular locks.
If I needed to physically secure a closure against potentially expert adversaries, I’d defer to LockPickingLawyer, and if I had the need and budget to do so at scale, I’d try to hire them. Obviously I’d be vulnerable to principal-agent problems, and if the marginal value of resolving those were good enough I’d look for an expert on that as well. However, it would be cheaper and probably almost as effective to simply avoid trying to squeeze LockPickingLawyer financially, and instead pay something they’re happy with.
I expect such a policy to produce the opposite of good results. By default, it will Goodhart on legible metrics rather than whatever’s actually valuable. For instance, in the case of physical security, a lockpicker can very legibly demonstrate how easy it is to beat commercial locks—but this is potentially a distraction, if locks aren’t the main weak point of one’s physical security. Similarly with math duels: they’re very externally legible, but they’re potentially a distraction from more important skills like e.g. the ability to find high-value problems to work on.
I agree that streetlamp effects are a problem. I think you are imagining a different usage than I am. I was imagining deferring to LockPickingLawyer about locks so that I could only spend about 5 minutes on that part of the problem, and spend whatever time I saved on other problems, including other aspects of securing an enclosure. If I had $100M and didn’t have friends I already trusted to do this sort of thing, offering them $20k to red-team a building might be worth it if I were worried about perimeter security; the same mindset that notices you can defeat some locks by banging them hard seems like it would have a good chance at noticing other simple ways to defeat barriers e.g. “this window can be opened trivially from the outside without triggering any alarms”.
Holding math duels to the standard of finding high-value problems to work on just seems nuts; I meant them as an existence proof of ways to measure possession of highly abstract theoretical secrets. If someone wanted to learn deep math, and barely knew what math was, they could do a lot worse than hiring someone who won a lot of math duels (or ideally whoever taught that person) as their first tutor, and then using that knowledge to find someone better. If you wanted to subsidize work on deep math, you might do well to ask a tournament-winner (or Fields medalist) whose non-tournament work they respect.
I went through a similar process in learning how to use my own body better: qigong seemed in principle worth learning, but I didn’t have a good way to distinguish real knowledge (if any) from bullshit. When I read Joshua Waitzkin’s The Art of Learning, I discovered that the closely related “Martial” Art of Tai Chi had a tournament that revealed relative levels of skill—and that William C. C. Chen, the man who’d taught tournament-winner Waitzkin, was still teaching in NYC. So I started learning from him, and very slowly improved my posture and balance. Eventually one of the other students invited me to a group that practiced on Sunday mornings in Chinatown’s Columbus Park, and when I went, I had just enough skill to immediately recognize the man teaching there as someone who had deep knowledge and was very good at teaching it, and I started learning much faster, in ways that generalized much better to other domains of life. This isn’t the only search method I used—recommendations from high-discernment friends also led to people who were able to help me—but it’s one that’s relatively easy to reproduce.
Experts should be able to regularly win bets against nonexperts, and generally valued generally scarce commodities like money should be able to incentivize people to make bets. If you don’t know how to construct an experiment to distinguish experts from nonexperts, you probably do not have a clear idea what it is you want an expert on, and if you don’t have any idea what you are trying to buy, it’s unclear what it would even mean to intend to buy it.
Abstract example: 16th Century “math duels,” in which expert algebraists competed to solve quantitative problems.
Concrete example: LockPickingLawyer, who demonstrates on YouTube how to easily defeat many commercially popular locks.
If I needed to physically secure a closure against potentially expert adversaries, I’d defer to LockPickingLawyer, and if I had the need and budget to do so at scale, I’d try to hire them. Obviously I’d be vulnerable to principal-agent problems, and if the marginal value of resolving those were good enough I’d look for an expert on that as well. However, it would be cheaper and probably almost as effective to simply avoid trying to squeeze LockPickingLawyer financially, and instead pay something they’re happy with.
I expect such a policy to produce the opposite of good results. By default, it will Goodhart on legible metrics rather than whatever’s actually valuable. For instance, in the case of physical security, a lockpicker can very legibly demonstrate how easy it is to beat commercial locks—but this is potentially a distraction, if locks aren’t the main weak point of one’s physical security. Similarly with math duels: they’re very externally legible, but they’re potentially a distraction from more important skills like e.g. the ability to find high-value problems to work on.
I agree that streetlamp effects are a problem. I think you are imagining a different usage than I am. I was imagining deferring to LockPickingLawyer about locks so that I could only spend about 5 minutes on that part of the problem, and spend whatever time I saved on other problems, including other aspects of securing an enclosure. If I had $100M and didn’t have friends I already trusted to do this sort of thing, offering them $20k to red-team a building might be worth it if I were worried about perimeter security; the same mindset that notices you can defeat some locks by banging them hard seems like it would have a good chance at noticing other simple ways to defeat barriers e.g. “this window can be opened trivially from the outside without triggering any alarms”.
Holding math duels to the standard of finding high-value problems to work on just seems nuts; I meant them as an existence proof of ways to measure possession of highly abstract theoretical secrets. If someone wanted to learn deep math, and barely knew what math was, they could do a lot worse than hiring someone who won a lot of math duels (or ideally whoever taught that person) as their first tutor, and then using that knowledge to find someone better. If you wanted to subsidize work on deep math, you might do well to ask a tournament-winner (or Fields medalist) whose non-tournament work they respect.
I went through a similar process in learning how to use my own body better: qigong seemed in principle worth learning, but I didn’t have a good way to distinguish real knowledge (if any) from bullshit. When I read Joshua Waitzkin’s The Art of Learning, I discovered that the closely related “Martial” Art of Tai Chi had a tournament that revealed relative levels of skill—and that William C. C. Chen, the man who’d taught tournament-winner Waitzkin, was still teaching in NYC. So I started learning from him, and very slowly improved my posture and balance. Eventually one of the other students invited me to a group that practiced on Sunday mornings in Chinatown’s Columbus Park, and when I went, I had just enough skill to immediately recognize the man teaching there as someone who had deep knowledge and was very good at teaching it, and I started learning much faster, in ways that generalized much better to other domains of life. This isn’t the only search method I used—recommendations from high-discernment friends also led to people who were able to help me—but it’s one that’s relatively easy to reproduce.