[I]f the expected concrete value of two interventions is similar, we should generally favor investing in interventions that have less evidence supporting them.
I like Askell’s talk and think this is an important point. Though when making the point without the full context of the talk, it also seems worth noting that:
As an empirical matter, one’s naive/early/quick analyses of how good (or cost-effective, or whatever) something is seem to often be overly optimistic.
Though I don’t know precisely why this is the case, and I imagine it varies by domain.
Additionally, there’s the optimizer’s curse. This is essentially a reason why one is likelier to be overestimating the value of something if one thinks that thing is unusually good. The curse is larger the more uncertainty one has.
For both reasons, if you see X and Y as unusually good, but you have less evidence re X, then that should update you towards thinking you’re being (more) overly optimistic about X, and thus that Y is actually better.
I think your comment is completely valid if we imagine that the two options “look to be equally good” even after adjusting for these tendencies. But I think people often don’t adjust for these tendencies, so it seems worth making it explicit.
(Also, even if X does currently seem somewhat less good than Y, that can be outweighed by the value of information consideration such that it’s worth further investigation of X rather than Y anyway.)
As an empirical matter, one’s naive/early/quick analyses of how good (or cost-effective, or whatever) something is seem to often be overly optimistic.
One possible reason is completely rational: if we’re estimating expected value of an intervention with a 1% chance to be highly valuable, then 99% of the time we realize the moonshot won’t work and revise the expected value downward.
That definitely can happen, and makes me realise my comment wasn’t sufficiently precise.
An expected value estimate can be reasonable even if there’s a 99% chance it would be revised downwards given more info, if there’s also a 1% chance it would be revised upwards by enough to offset the potential downward revisions. If an estimator makes such an estimate and is well-calibrated, I wouldn’t say they’re making a mistake, and thus probably wouldn’t say they’re being “overly optimistic”.
The claim I was making was that one’s naive/early/quick analyses of how good (or cost-effective, or whatever) tend to not be well-calibrated, systematically erring towards optimism in a way that means that it’s best to adjust the expected value downwards to account for this (unless one has already made such an adjustment).
But I’m not actually sure how true that claim is (I’m just basing it on my memory of GiveWell posts I read in the past). Maybe most things that look like that situation are either actually the optimiser’s curse or actually the sort of situation you describe.
Also related is the idea that the moral value of additional information is high when there is relatively low resilience in your credence that the current intervention is best. This leads to the (to me) rather unintuitive conclusion that if you have two research paths that both look to be equally good to look into for potentially improving the world, then, ceteris paribus, it may be better to invest in the research path for which you have less evidence that it is a good research path to follow. From Amanda Askell in the link:
I like Askell’s talk and think this is an important point. Though when making the point without the full context of the talk, it also seems worth noting that:
As an empirical matter, one’s naive/early/quick analyses of how good (or cost-effective, or whatever) something is seem to often be overly optimistic.
Though I don’t know precisely why this is the case, and I imagine it varies by domain.
See also Why We Can’t Take Expected Value Estimates Literally (Even When They’re Unbiased)
Additionally, there’s the optimizer’s curse. This is essentially a reason why one is likelier to be overestimating the value of something if one thinks that thing is unusually good. The curse is larger the more uncertainty one has.
For both reasons, if you see X and Y as unusually good, but you have less evidence re X, then that should update you towards thinking you’re being (more) overly optimistic about X, and thus that Y is actually better.
I think your comment is completely valid if we imagine that the two options “look to be equally good” even after adjusting for these tendencies. But I think people often don’t adjust for these tendencies, so it seems worth making it explicit.
(Also, even if X does currently seem somewhat less good than Y, that can be outweighed by the value of information consideration such that it’s worth further investigation of X rather than Y anyway.)
One possible reason is completely rational: if we’re estimating expected value of an intervention with a 1% chance to be highly valuable, then 99% of the time we realize the moonshot won’t work and revise the expected value downward.
That definitely can happen, and makes me realise my comment wasn’t sufficiently precise.
An expected value estimate can be reasonable even if there’s a 99% chance it would be revised downwards given more info, if there’s also a 1% chance it would be revised upwards by enough to offset the potential downward revisions. If an estimator makes such an estimate and is well-calibrated, I wouldn’t say they’re making a mistake, and thus probably wouldn’t say they’re being “overly optimistic”.
The claim I was making was that one’s naive/early/quick analyses of how good (or cost-effective, or whatever) tend to not be well-calibrated, systematically erring towards optimism in a way that means that it’s best to adjust the expected value downwards to account for this (unless one has already made such an adjustment).
But I’m not actually sure how true that claim is (I’m just basing it on my memory of GiveWell posts I read in the past). Maybe most things that look like that situation are either actually the optimiser’s curse or actually the sort of situation you describe.