Regarding (a), it doesn’t seem clear to me that conditional on Impact List being wildly successful (which I’m interpreting as roughly the $110B over ten years case), we shouldn’t expect it to account for more than 10% of overall EA outreach impact. Conditional on Impact List accounting for $110B, I don’t think I’d feel surprised to learn that EA controls only $400B (or even $200B) instead of ~$1T. Can you say more about why that would be surprising?
(I do think there’s a ~5% chance that EA controls or has deployed $1T within ten years.)
I think (b) is a legit argument in general, although I have a lot of uncertainty about what the appropriate discount should be. This is also highlighting that using dollars for impact can be unclear, and that my EV calculation bucketed money as either ‘ineffective’ or ‘effective’ without spelling out the implications.
A few implications of that:
There’s a ‘free parameter’ in the EV calculation that isn’t obvious: the threshold we use to separate effective from ineffective donations. We might pick something like ‘effective = anything roughly as or more effective than current GiveWell top charities’.
That threshold influences whether our probability estimates are reasonable. For instance depending on this threshold someone can object “A 1 in 1000 chance for $110B to be moved to things as effective as GiveDirectly seems reasonable, but 1 in 1000 for $110B to be moved to things as effective as AMF? No way!”
As noted in footnote 4, we assume that the donations in the ‘ineffective’ bucket are so much less effective than the donations in the ‘effective’ bucket that we can ignore them. Alternatively, we can assume that enough of the ‘effective’ donations are far enough above the minimum effectiveness threshold that they at least cancel out all the ineffective donations.
The threshold we pick also determines what it means when we talk about expected value. If we say the expected value of Impact List is $X it means roughly $X being put into things at at least the level of effectiveness of our threshold. We could be underestimating if Impact List causes people to donate a lot to ultra-effective orgs (and it might, if people try hard to optimize their rankings), but I didn’t try to model that.
Given the bucketing and that “$X of value” doesn’t mean “$X put into the most effective cause area”, I think it may be reasonable to not have a discount. Not having a discount assumes that we’ll find enough (or scalable enough) cause areas over the next ten years at least as effective as whatever threshold value we pick that they can soak up an extra ~110B. Although this is probably a lot more plausible to those who prioritize x-risk than to those who think global health will be the top cause area over that period.
Thanks for the feedback!
Regarding (a), it doesn’t seem clear to me that conditional on Impact List being wildly successful (which I’m interpreting as roughly the $110B over ten years case), we shouldn’t expect it to account for more than 10% of overall EA outreach impact. Conditional on Impact List accounting for $110B, I don’t think I’d feel surprised to learn that EA controls only $400B (or even $200B) instead of ~$1T. Can you say more about why that would be surprising?
(I do think there’s a ~5% chance that EA controls or has deployed $1T within ten years.)
I think (b) is a legit argument in general, although I have a lot of uncertainty about what the appropriate discount should be. This is also highlighting that using dollars for impact can be unclear, and that my EV calculation bucketed money as either ‘ineffective’ or ‘effective’ without spelling out the implications.
A few implications of that:
There’s a ‘free parameter’ in the EV calculation that isn’t obvious: the threshold we use to separate effective from ineffective donations. We might pick something like ‘effective = anything roughly as or more effective than current GiveWell top charities’.
That threshold influences whether our probability estimates are reasonable. For instance depending on this threshold someone can object “A 1 in 1000 chance for $110B to be moved to things as effective as GiveDirectly seems reasonable, but 1 in 1000 for $110B to be moved to things as effective as AMF? No way!”
As noted in footnote 4, we assume that the donations in the ‘ineffective’ bucket are so much less effective than the donations in the ‘effective’ bucket that we can ignore them. Alternatively, we can assume that enough of the ‘effective’ donations are far enough above the minimum effectiveness threshold that they at least cancel out all the ineffective donations.
The threshold we pick also determines what it means when we talk about expected value. If we say the expected value of Impact List is $X it means roughly $X being put into things at at least the level of effectiveness of our threshold. We could be underestimating if Impact List causes people to donate a lot to ultra-effective orgs (and it might, if people try hard to optimize their rankings), but I didn’t try to model that.
Given the bucketing and that “$X of value” doesn’t mean “$X put into the most effective cause area”, I think it may be reasonable to not have a discount. Not having a discount assumes that we’ll find enough (or scalable enough) cause areas over the next ten years at least as effective as whatever threshold value we pick that they can soak up an extra ~110B. Although this is probably a lot more plausible to those who prioritize x-risk than to those who think global health will be the top cause area over that period.