Hmm regarding the donations to multiple charities—sorry for offtopic but I just want to post something vaguely in agreement.
I see that if the function which determines which charity is the best, may be exploitable, donating to top 5 charities or so which are substantially different (not cloning each other) would reduce the pay-off for finding an exploit in this function.
I think of it this way. If i were to make a charitable bot, would it be most effective for this bot to donate to just 1 charity that it determines based on some function? Big red bold NO because this bot will be deliberately gamed (on top of failing consistently) It’s why we need Google to return >1 search result. The first result may not work for the user at all.
Adding jitter/dithering to inexact calculations to improve resulting behaviour is common practice. If you have a target-shooting AI in a game, when it’s shooting at it’s best prediction of the target position, it is easy to evade. If it jitters some, it may score far more hits.
edit: Matter of fact I did write a shooting AI that did just this, and it made the AI more effective. If said AI was sentient but not clever enough to understand my rationale, this AI would either justify it with faulty logic or try to get rid of the fuzziness and jitter i’ve added (and then this AI would ponder why it is not becoming more effective).
Humans of course would, too, implement this sort of jitter and added fuzziness to their decisionmaking they’d pick from others when growing up, usually without understanding of the rationale behind it; backing the heuristics with faulty logic or dismissing it based on the faultiness of the logic.
It’s why we need Google to return >1 search result. The first result may not work for the user at all.
I dispute the analogy. With Google, you get the full benefit of finding what you’re looking for as long as at least one of the results is right. The cost of multiple results is the time spent reading them, which is not at all the same as the cost of splitting a donation.
Well, the analogy is not perfect but there is a cost to clicking on the other links, too.
When you are splitting between top 5 charities—the difference between expected utility of them drown in the error of your heuristics used to pick top—so the ‘cost’ of splitting probably boils down to the cost of e.g. paypal processing the payment, and other entirely trivial costs (which you can avoid by e.g. randomizing between top 5)
Hmm regarding the donations to multiple charities—sorry for offtopic but I just want to post something vaguely in agreement.
I see that if the function which determines which charity is the best, may be exploitable, donating to top 5 charities or so which are substantially different (not cloning each other) would reduce the pay-off for finding an exploit in this function.
I think of it this way. If i were to make a charitable bot, would it be most effective for this bot to donate to just 1 charity that it determines based on some function? Big red bold NO because this bot will be deliberately gamed (on top of failing consistently) It’s why we need Google to return >1 search result. The first result may not work for the user at all.
Adding jitter/dithering to inexact calculations to improve resulting behaviour is common practice. If you have a target-shooting AI in a game, when it’s shooting at it’s best prediction of the target position, it is easy to evade. If it jitters some, it may score far more hits.
edit: Matter of fact I did write a shooting AI that did just this, and it made the AI more effective. If said AI was sentient but not clever enough to understand my rationale, this AI would either justify it with faulty logic or try to get rid of the fuzziness and jitter i’ve added (and then this AI would ponder why it is not becoming more effective).
Humans of course would, too, implement this sort of jitter and added fuzziness to their decisionmaking they’d pick from others when growing up, usually without understanding of the rationale behind it; backing the heuristics with faulty logic or dismissing it based on the faultiness of the logic.
I dispute the analogy. With Google, you get the full benefit of finding what you’re looking for as long as at least one of the results is right. The cost of multiple results is the time spent reading them, which is not at all the same as the cost of splitting a donation.
Well, the analogy is not perfect but there is a cost to clicking on the other links, too.
When you are splitting between top 5 charities—the difference between expected utility of them drown in the error of your heuristics used to pick top—so the ‘cost’ of splitting probably boils down to the cost of e.g. paypal processing the payment, and other entirely trivial costs (which you can avoid by e.g. randomizing between top 5)