organizations working outside these areas, such as those working on existential risk and far future. My impression, however, is that OpenPhil has done a good job filling up the funding gaps in this area and that there are very few organizations that would meet the criteria I’m using for these recommendations.
[Disclaimers: speaking only for myself, although I do some work for Open Phil.]
I think that that many EAs are overestimating the degree to which this funding changes the marginal returns of individual donations, for a few reasons:
In a number of these cases the Open Philanthropy Project grants discuss intentions to take up a percentage of the grantee’s budget, and a preference not to exceed half of it; desire to avoid single-donor funding issues creates opportunities for small donors, as I discussed in this post
If a large donor limits itself to half of the grantee’s budget, then not only is there ‘room for more funding’ left for other donors, but it also implicitly acts as a delayed counterfactual 1:1 matching grant, as each small donor dollar allows for another large donor dollar (less the opportunity cost of Open Philanthropy’s ‘last dollar’ but insofar as one isn’t just topping up Open Philanthropy’s reserves then presumably one aims to do better than that), which could largely offset diminishing returns for the marginal donor
Where ‘room for more funding’ suggests a steep cliff of diminishing returns, in reality diminishing returns are normally much smoother, as additional funds enable reserves, marginal expenditures, openness to and pursuit of additional expansion, etc; see the linked articles by Max Dalton and Owen Cotton-Barratt
Concretely, I think small donors could ‘top up’ many of the AI grants in the Open Philanthropy grant database and get marginal cost effectiveness within a factor of 2-4 of the average cost-effectiveness of the dollars in the relevant grant
In cases where the topping up would work better with larger amounts (e.g. $100,000 or $500,000) because of transaction costs (e.g. working with academic labs, or asking for advice on how to do it), small donors can make use of a donor lottery to convert their donation into a 1/n chance of a donation n times as great for which the transaction costs are manageable
In my view the larger shift induced by Open Philanthropy is that the returns to using one’s labor, knowledge, and other resources to create opportunities that it will find competitive have gone up (since they are more likely to be able to grow later if successful). That is a boost for several of the organizations you mention, but can also apply to larger organizations whose activity tends to produce those opportunities through other channels than being a new organization (e.g. by building pipelines for new scientists or activists, research that better prioritizes options, demonstrations that technical projects can make progress).
So I don’t think that the arguments in the post are sufficient to establish this:
while I think some organizations may be more impactful per dollar overall, the marginal donation is not as useful as they are highly likely to have been able to fundraise it already with much less effort and there is less at risk (e.g., whether a program happens at all versus whether it is scaled up further).
I agree that CSH looks attractive for a donor who would otherwise give to AMF, that WASR and SI make sense for a donor who might otherwise give to The Humane League (as demonstrated by, e.g. Lewis’ EA Funds grants), and that providing access to donation methods for Canadian donors could pay for itself for those donors (with some caveats about distributional details, and due diligence).
However, I don’t think that increased Open Philanthropy funding provides adequate reason to dismiss the cause area of existential risk reduction for marginal funds (and in fact my own view is that the most attractive marginal opportunities lie in that area, directly or indirectly).
Thanks Carl, it’s good to know that there are RFMF opportunities in topping up AI grants.
My reasoning for not donating to AI projects right now is based much less on a RFMF argument and more on not knowing enough about the space. I think I know enough about opportunities in global poverty, animal welfare, and EA community building to recommend projects there with confidence, but not for AI. I expect it would take me a good deal of time to develop the relevant expertise in AI to consider it properly. I have thought about working to develop that expertise, but so far I have not prioritized doing so.
I don’t understand how that logic leads to thinking it’s a good idea to donate to the causes you’re thinking of donating to. Donating to a cause area because you can identify good projects within it seems like the streetlight effect.
If you think that AI stuff is plausibly better, shouldn’t you either want to learn more about it or enter a donor lottery so that it’s more cost-effective for you to learn about it?
1.) While I do think AI as a cause area could be plausibly better than global poverty or animal welfare, I don’t think it’s so plausibly better that the expected value given my uncertainty dwarfs my current recommendations.
2a.) I think I’m basically okay with the streetlight effect. I think there’s a lot of benefit in donating now to support groups that might not be able to expand at all without my donation, which is what the criteria I outlined here accomplish. Given the entire EA community collaborating as a whole, I think there’s less need for me to focus tons of time on making sure my donations are as cost-effective as possible, and more just a need to clear a bar of being “better than average”. I think my recommendations here accomplish that.
2b.) Insofar as my reasoning in (2a) is some “streetlight effect” bias, I think you could accuse nearly anyone of this, since very few have thoroughly explored every cause area and no one could fully rule out being wrong about a cause area.
3.) There is still more I could donate later. This money is being saved mainly as a hedge to large financial uncertainty in my immediate future, but could also be used as savings to donate later when I learn more.
Although I laud posts like the OP, I’m not sure I understand this approach to uncertainty.
I think a lot turns on what you mean by the AI cause area being “Plausibly better” than global poverty or animal welfare on EV. The Gretchenfrage seems to be this conditional forecast: “If I spent (lets say) 6 months looking at the AI cause area, would I expect to identify better uses of marginal funding in this cause area than those I find in animal welfare and global poverty?”
If the answer is “plausibly so, but probably not” (either due to a lower ‘prima facie’ central estimate, or after pricing in regression to the mean etc.), then I understand the work uncertainty is doing here (modulo the usual points about VoI): one can’t carefully look at everything, and one has to make some judgments on what cause areas look most promising to investigate on current margins.
Yet if the answer is “Probably, yes”, then offering these recommendations simpliciter (i.e. “EA should fully fund this”) seems premature to me. The evaluation is valuable, but should be presented with caveats like, “Conditional on thinking global poverty is the best cause area, fund X; conditional on thinking animal welfare is the best cause area, fund Y (but, FWIW, I believe AI is the best cause area, but I don’t know what to fund within it).” It would also lean against making ones own donations to X, Y etc., rather than spending time thinking about it/following the recommendations of someone one trusts to make good picks in the AI cause area.
If the answer is “plausibly so, but probably not” (either due to a lower ‘prima facie’ central estimate, or after pricing in regression to the mean etc.)
An additional point to take into account when it comes to examining the research on AI as possible space for donations: as a scientific domain the topic of AI risks and safety can easily fall under the public/academic funding, even under the assumption that it is currently underfunded. To this end, individual applicants (precisely those who would be conducting research by means of donations) can apply for individual PhD and postdoc grants. There are numerous opportunities of that kind across EU. Moreover, the funding agencies (e.g. in Germany, Belgium, Netherlands, etc.) will employ expert refereeing system (sometimes even asking the applicant to suggest suitable referees) to assess the project and its effectiveness (which I find very relevant from the perspective of EA). If we take this into account, then a number of other organizations that can’t be so easily funded via already existing institutional channels becomes much more urgent.
: one can’t carefully look at everything, and one has to make some judgments on what cause areas look most promising to investigate on current margins.
This is why EA rarely falls into what can accurately be described as a “streetlight effect”. We aren’t looking for one set of keys, we’re looking for a bunch of keys (threats to human welfare) and theres a bunch of us drunkards, all with differing abilities and expertise. So I’d argue if its dark somewhere, those with the expertise need to start building streetlights, but if the lights getting brighter in certain areas (RCTs in health) then we need people there too.
[Disclaimers: speaking only for myself, although I do some work for Open Phil.]
I think that that many EAs are overestimating the degree to which this funding changes the marginal returns of individual donations, for a few reasons:
In a number of these cases the Open Philanthropy Project grants discuss intentions to take up a percentage of the grantee’s budget, and a preference not to exceed half of it; desire to avoid single-donor funding issues creates opportunities for small donors, as I discussed in this post
If a large donor limits itself to half of the grantee’s budget, then not only is there ‘room for more funding’ left for other donors, but it also implicitly acts as a delayed counterfactual 1:1 matching grant, as each small donor dollar allows for another large donor dollar (less the opportunity cost of Open Philanthropy’s ‘last dollar’ but insofar as one isn’t just topping up Open Philanthropy’s reserves then presumably one aims to do better than that), which could largely offset diminishing returns for the marginal donor
Where ‘room for more funding’ suggests a steep cliff of diminishing returns, in reality diminishing returns are normally much smoother, as additional funds enable reserves, marginal expenditures, openness to and pursuit of additional expansion, etc; see the linked articles by Max Dalton and Owen Cotton-Barratt
Concretely, I think small donors could ‘top up’ many of the AI grants in the Open Philanthropy grant database and get marginal cost effectiveness within a factor of 2-4 of the average cost-effectiveness of the dollars in the relevant grant
In cases where the topping up would work better with larger amounts (e.g. $100,000 or $500,000) because of transaction costs (e.g. working with academic labs, or asking for advice on how to do it), small donors can make use of a donor lottery to convert their donation into a 1/n chance of a donation n times as great for which the transaction costs are manageable
In my view the larger shift induced by Open Philanthropy is that the returns to using one’s labor, knowledge, and other resources to create opportunities that it will find competitive have gone up (since they are more likely to be able to grow later if successful). That is a boost for several of the organizations you mention, but can also apply to larger organizations whose activity tends to produce those opportunities through other channels than being a new organization (e.g. by building pipelines for new scientists or activists, research that better prioritizes options, demonstrations that technical projects can make progress).
So I don’t think that the arguments in the post are sufficient to establish this:
I agree that CSH looks attractive for a donor who would otherwise give to AMF, that WASR and SI make sense for a donor who might otherwise give to The Humane League (as demonstrated by, e.g. Lewis’ EA Funds grants), and that providing access to donation methods for Canadian donors could pay for itself for those donors (with some caveats about distributional details, and due diligence).
However, I don’t think that increased Open Philanthropy funding provides adequate reason to dismiss the cause area of existential risk reduction for marginal funds (and in fact my own view is that the most attractive marginal opportunities lie in that area, directly or indirectly).
Thanks Carl, it’s good to know that there are RFMF opportunities in topping up AI grants.
My reasoning for not donating to AI projects right now is based much less on a RFMF argument and more on not knowing enough about the space. I think I know enough about opportunities in global poverty, animal welfare, and EA community building to recommend projects there with confidence, but not for AI. I expect it would take me a good deal of time to develop the relevant expertise in AI to consider it properly. I have thought about working to develop that expertise, but so far I have not prioritized doing so.
I don’t understand how that logic leads to thinking it’s a good idea to donate to the causes you’re thinking of donating to. Donating to a cause area because you can identify good projects within it seems like the streetlight effect.
If you think that AI stuff is plausibly better, shouldn’t you either want to learn more about it or enter a donor lottery so that it’s more cost-effective for you to learn about it?
My excuses in order of importance:
1.) While I do think AI as a cause area could be plausibly better than global poverty or animal welfare, I don’t think it’s so plausibly better that the expected value given my uncertainty dwarfs my current recommendations.
2a.) I think I’m basically okay with the streetlight effect. I think there’s a lot of benefit in donating now to support groups that might not be able to expand at all without my donation, which is what the criteria I outlined here accomplish. Given the entire EA community collaborating as a whole, I think there’s less need for me to focus tons of time on making sure my donations are as cost-effective as possible, and more just a need to clear a bar of being “better than average”. I think my recommendations here accomplish that.
2b.) Insofar as my reasoning in (2a) is some “streetlight effect” bias, I think you could accuse nearly anyone of this, since very few have thoroughly explored every cause area and no one could fully rule out being wrong about a cause area.
3.) There is still more I could donate later. This money is being saved mainly as a hedge to large financial uncertainty in my immediate future, but could also be used as savings to donate later when I learn more.
[Note: I work on existential risk reduction]
Although I laud posts like the OP, I’m not sure I understand this approach to uncertainty.
I think a lot turns on what you mean by the AI cause area being “Plausibly better” than global poverty or animal welfare on EV. The Gretchenfrage seems to be this conditional forecast: “If I spent (lets say) 6 months looking at the AI cause area, would I expect to identify better uses of marginal funding in this cause area than those I find in animal welfare and global poverty?”
If the answer is “plausibly so, but probably not” (either due to a lower ‘prima facie’ central estimate, or after pricing in regression to the mean etc.), then I understand the work uncertainty is doing here (modulo the usual points about VoI): one can’t carefully look at everything, and one has to make some judgments on what cause areas look most promising to investigate on current margins.
Yet if the answer is “Probably, yes”, then offering these recommendations simpliciter (i.e. “EA should fully fund this”) seems premature to me. The evaluation is valuable, but should be presented with caveats like, “Conditional on thinking global poverty is the best cause area, fund X; conditional on thinking animal welfare is the best cause area, fund Y (but, FWIW, I believe AI is the best cause area, but I don’t know what to fund within it).” It would also lean against making ones own donations to X, Y etc., rather than spending time thinking about it/following the recommendations of someone one trusts to make good picks in the AI cause area.
This is what captures my views best right now.
An additional point to take into account when it comes to examining the research on AI as possible space for donations: as a scientific domain the topic of AI risks and safety can easily fall under the public/academic funding, even under the assumption that it is currently underfunded. To this end, individual applicants (precisely those who would be conducting research by means of donations) can apply for individual PhD and postdoc grants. There are numerous opportunities of that kind across EU. Moreover, the funding agencies (e.g. in Germany, Belgium, Netherlands, etc.) will employ expert refereeing system (sometimes even asking the applicant to suggest suitable referees) to assess the project and its effectiveness (which I find very relevant from the perspective of EA). If we take this into account, then a number of other organizations that can’t be so easily funded via already existing institutional channels becomes much more urgent.
P.S. Great post, Peter, only now saw it.
To attempt to complement what Peter already said,
This is why EA rarely falls into what can accurately be described as a “streetlight effect”. We aren’t looking for one set of keys, we’re looking for a bunch of keys (threats to human welfare) and theres a bunch of us drunkards, all with differing abilities and expertise. So I’d argue if its dark somewhere, those with the expertise need to start building streetlights, but if the lights getting brighter in certain areas (RCTs in health) then we need people there too.