I’ve read a lot of Ben Hoffman’s work over the years, but only this past week have I read his actual myriad criticisms of the Effective Altruism movement and its organizations. The most illuminating posts I just read are A drowning child is hard to find, GiveWell and the problem of partial funding, and Effective Altruism is self recommending.
This post is me quick jotting down my current understanding of Ben’s criticism, which I basically agree with.
The original ideas of the EA movement are the ethical views of Peter Singer and his thought experiments on the proverbial drowning child, combined with an engineering/finance methodology for assessing how much positive impact you’re actually producing. The canonical (first?) EA organization was GiveWell, which researched various charities and published their findings on how effective they were. A core idea underneath GiveWell’s early stuff was “your dollars can have an outsized impact helping the global poor, compared to helping people in first world countries”. The mainstream bastardized version of this is “For the price of a cup of coffee, you can safe a life in Africa”, which I think uses basically made up and fraudulent numbers. The GiveWell pitch was more like “we did some legit research, and for ~$5000, you can save or radically improve a life in Africa”. Pretty quickly GiveWell and the ecosystem around it got Large Amounts of Money, both thru successful marketing campaigns that convinced regular people with good jobs to give 10% of their annual income (Giving What We Can), but their most high leverage happenings were that they got the ear of billionaire tech philanthropists, like Dustin Moskovitz who was a co-founder of both Facebook and Asana, and Jaan Tallinn who co-founded Skype. I don’t know exactly how Jaan’s money moved thru the EA ecosystem, but Dustin ended up creating Good Ventures which was an org to manage his philanthropy, and it was advised by Open Philanthropy, and my understanding is that both these orgs were staffed by early EA people and were thoroughly EA in outlook, and also had significant personal overlap with GiveWell specifically.
The big weird thing is that it seems like difficulties were found in the early picture of how much good was in fact being done thru these avenues, and this was quietly elided, and more research wasn’t being done to get to the bottom of the question, and there’s also various indicators that EA orgs themselves didn’t really believe their numbers for how much good could done. For the Malaria stuff, it did check that the org had followed thru on the procedures it intended, but the initial data they had available on if malaria cases were going up or down was noisy, so they stopped paying attention to it and didn’t try to make it so better data was available. A big example of “EA orgs not seeming to buy their own story” was GiveWell advising Open Philanthropy to not simply fully fund its top charities. This is weird because if the even the pessimistic numbers were accurate, Open Phil on its own could have almost wiped out malaria and an EA sympathetic org like the Gates foundation definitely could have. And at the very least, they could have done a very worked out case study in one country or another and gotten a lot more high quality info on if the estimates were legit. And stuff like that didn’t end up happening.
It’s not that weird to have very incorrect estimates. It is weird to have ~15 years go by without really hammering down and getting very solid evidence for the stuff you purported to be “the most slam dunk evidence based cost effective life saving”. You’d expect to either get that data and then be in the world of “yeah, it’s now almost common knowledge that the core EA idea checks out”, or you’d have learned that the gains are that high or that easy, or that the barriers to getting rid of malaria have a much different structure, and you should change your marketing to reflect it’s not “you can trivially do lots of obvious good by giving these places more money”.
Givewell advising Open Phil to not fully fund things is the main “it seems like the parties upstream of the main message don’t buy their main message enough to Go Hard at it”. In very different scenarios the funding split thing kinda makes sense to me, I did a $12k crowdfunding campaign last year for a research project, and a friend of a friend offered to just fund the full thing, and I asked him to only do that if it wasn’t fully funded by the last week of the fundraising period, because I was really curious and uncertain about how much money people just in my twitter network would be interested in giving for a project like this, and that information would be useful to me for figuring out how to fund other stuff in the future.
In the Open Phil sitch, it seems like “how much money are people generally giving?” isn’t rare info that needed to be unearthed, and also Open Phil and friends could really just solve most all of the money issues, and the orgs getting funded could supposedly then just solve huge problems. But they didn’t. This could be glossed as something like “turns out there’s more than enough billionaire philanthropic will to fix huge chunks of global poverty problems, IF global poverty works the way that EA orgs have modeled it as working”. And you could imagine maybe there’s some trust barrier preventing otherwise willing philanthropists getting info from and believing otherwise correct and trustworthy EAs, but in this scenario it’s basically the same people, the philanthropists are “fully bought in” to the EA thing, so things not getting legibly resolved seems to indicate that internally there was some recognition that the core EA story wasn’t correct, and prevented that information from propagating and reworking things.
Relatedly, the thing that we seemed to see in lieu of “go hard on the purported model and either disconfirm them and update, or get solid evidence and double down”, we see a situation where a somewhat circularly defined reputation gets bootstrapped, with the main end state being fairly unanimous EA messaging that “people should give money to EA orgs, in a general sense, and EA orgs should be in charge of more and more things” despite not having the underlying track record that would make that make sense. The track record that is in fact pointed to is a sequence of things like “we made quality researched estimates of effectiveness of different charities” that people found compelling, then pointing to later steps of “we ended up moving XYZ million dollars!” as further evidence of trustworthiness, but really that’s just “double spending” on the original “people found our research credible and extended us the benefit of the doubt”. To full come thru they’d need to show that the benefits produced matched what they expected (or even if the showed otherwise, if the process and research was good and it seemed like they were learning it could be very reasonable to keep trusting them).
This feels loosely related to how for the first several times I’d heard Anthropic mentioned by rationalists, the context made me assume it was a rationalist run AI safety org, and not a major AI capabilities lab. Somehow there was some sort of meme of “it’s rationalist, which means it’s good and cares about AI Safety”. Similarly it sounds like EA has ended up acting like and producing messaging like “You can trust us Because we are Labeled EAs” and ignoring some of the highest order bits of things they could do which would give them a more obviously legible and robust track record. I think there was also stuff mentioned like “empirically Open Phil is having a hard time finding things to give away money too, and yet people are still putting out messaging that people should Obviously Funnel Money Towards this area”.
Now, for some versions of who the founding EA stock could have been, one conclusion might just be “damn, well I guess they were grifters, shouldn’t have trusted them”. But it seems like there was enough obviously well thought out and well researched efforts early on that that doesn’t seem reasonable. Instead, it seems to indicate that billionaire philanthropy is really hard and/or impossible, at least while staying within a certain set of assumptions. Here, I don’t think I’ve read EA criticism that informs the “so what IS the case if it’s not the case that for the price of a cup of coffee you can save a life?” but my understanding is informed by writers like Ben. So what is the case? It probably isn’t true that eradicating malaria is fundamentally hard in an engineering sense. It’s more like “there’s predatory social structures setup to extract from a lot of the avenues that one might try and give nice things to the global poor”. There’s lots of vary obvious examples of things like aid money and food being sent to countries and the governments of those countries basically just distributing it as spoils to their cronies, and only some or none of it getting to the people who others were hoping to help. There seem to be all kinds or more or less subtle versions of this.
The problems also aren’t only on the third world end. It seems like people in the first-world aren’t generally able to get enough people together who have a shared understanding that it’s useful to tell the truth, to have large scale functional “bureaucracies” in the sense of “ecosystem of people that accurately processes information”. Ben’s the professionals dilemma looks at how the ambient culture of professionalism seems to work against this having large functional orgs that can tell the truth and learn things.
So it seems like what happened was the early EA stock (who I believe came from Bridgewater) were earnestly trying to apply finance and engineering thinking to the task of philanthropy. They did some good early moves and got the ear of many billions of dollars. As things progressed, they started to notice things that complicated the simple giving hypothesis. As this was happening they were also getting bigger from many people trusting them and giving them their ears, and were in a position where the default culture of destructive Professionalism pulled at people more and more. These pressures were enough to quickly erode the epistemic rigor needed for the philanthropy to be robustly real. EA became a default attractor for smart young good meaning folk, because the messaging on the ease of putting money to good use wasn’t getting updated. It also became an attractor for opportunists who just saw power and money and authority accumulating and wanted in on it. Through a mix of ambient cultural pressures silencing or warping the clarity of good meaning folk, and thru Rapid Growth that accepted ambivalent meaning and bad meaning folk, it lost the ability to stay truth and mission focused, and while it might still do some higher quality research than other charitable entities, it has forgone the next obvious step of propagating the information about what the actual blockers and constraints on doing good in the world are, and has become the general attractor of “thing that just tries to accumulate more resources because We Should Be In Charge of more resources”.
I don’t really get the “it was weird for GiveWell to tell Good Ventures to fund Open Phil instead” complaint. From where I am standing that’s a much better use of money and definitely seems like good advice.
I think the more correct story of what happened roughly with GiveWell and OpenPhil is:
GiveWell was started by some quite competent people, and they did some work trying to figure out how to help the world most effectively
They discovered some great global health interventions, which attracted a bunch of funding
They then realized that global health is probably not the best way to improve the world, so they started Open Phil to have a broader remit for areas to research
GiveWell was henceforth understaffed and stopped really doing anything interesting, and mostly just populated a bunch of spreadsheets every year that said “yes, nothing obviously crazy has happened that invalidated the models our founders came up with 10 years ago”.
The people at Open Phil then did a bunch of things trying to improve the world by changing the course of history, most centrally by influencing AI, but also by influencing science and global policy and other really important things
This has either succeeded enormously or failed horrendously depending on your viewpoint. Among other things, it has led to the founding of Anthropic which was founded by the wife of the founder of GiveWell, and where the founder now works, which is one of the most successful companies in history, and also has a huge amount of leverage over how AI will go.
This story feels like it explains most of the datapoints of what’s going on. I do think in-particular the last step likely involves people making some huge mistakes, but the nature and structure of it feels just very different from the story you (and I think Ben) are trying to tell.
You don’t need to posit systematic self-delusion to explain any of these things—whether failing to eradicate malaria, or failing to recommend 100% funding of the top charities. You cannot arrive at a price to eradicate malaria by taking the price to prevent one case of malaria on margin and multiplying by the number of malaria cases. The same is true of poverty, homelessness, hunger, et cetera. There are a lot of popular, well-documented ways of giving to any of these causes and do approximately no good at all. Just identifying places where money can actually improve things a little bit on margin is an enormous improvement over the pre-EA state of the art in charity. But this doesn’t scale forever. For any one of these problems, as you spend money effectively on margin, you solve the most tractable cases and necessarily move down the ladder to increasingly intractable cases. At some point the most cost-effective way to save a life will switch from “buy malaria nets” to some other intervention. Even within a single year of grants, it makes sense to spread your money out a little, to make sure a) you aren’t saturating any of the top charities you’ve identified beyond the point of diminishing returns, and b) you aren’t missing out on value by ignoring second-tier charities who might outperform expectations.
The malaria thing seems like the load-bearing part of the post, so I’d really like to know the details. The GiveWell website currently says:
Should I strongly doubt that and why?
This seems to be an accurate description of what happened with EA (or at least one of the several dynamics). I’ve seen this plenty in my 10 years being involved, but one illustrative example is people campaigning for Carrick Flynn, without knowing anything about him except that he was an EA, because “having an EA in congress is obviously good.” You also see this whenever EAs talk about whether people are “aligned” (a shorthand for “values-aligned with EA”).
More generally, I think people often defer to the “EA elite” about crucial topics (like say, how one should relate to Anthropic) that require actually a lot of thinking and modeling to reason well about, because they assume that the EA elite are smart and earnest.
I think it’s sad that the “let’s actually do the research and figure out which interventions actually work” movement turned into a deferral network. I think it’s bad, and I’ve publicly critiqued EA on these grounds in the past.
That said, I want to defend the dynamics that lead to this outcome, or at least elucidate those dynamics and how they can feel from the inside.
In the early days, being affiliated with EA was, in fact, an extremely strong signal of earnest altruistic intent, and a basic but in-practice-rare kind of intellectual seriousness.
Like, there were a bunch of specific ideas that seemed basically obvious and important to me, but which almost no one in the world seemed interested in. Things like “some charities are much more impactful than others, so obviously you should prioritize based on where you can do the most good” (something people I knew in 2014 specifically argued against), or “when you’re presented a solid counterargument, you change your mind”, and a more diffuse attitude of “there are people to help, and we can help them, so I want to dedicate my life to helping.”
These basic ideas seemed and continue to seem, broadly unappreciated and unshared by most of the world. But if a person was an EA, that meant that they “got it”. Against a backdrop of civilizational insanity, there were these people who also saw these obvious things, and acted on them.
Such that, if I knew someone was any EA, but I knew literally nothing about them, I would be glad to let them crash on my couch, or introduce them to a professional connection, or boost their projects, etc, sight-unseen.[1]
I think it’s pretty natural for that to turn into a deferral network. Of course I want to empower the smart, earnest, thoughtful people who “get it”. Even if I don’t understand all their impact models in detail, it seems like a really appealing way to do good is to team up with those people and help them with their plans. If there’s a whole movement of people like that, the more I can help the movement, the better!
The sort of people who become EAs want to have a positive impact on the world, and EA-the-movement-and-community will often look to them like a big channel for having impact. eg A central way that I can help the world is through boosting EA.
This is especially convenient if I’m a college-age EA without much career experience, who’s not very well equipped to try for an ambitious project in bioengineering, or malaria-eradication, or government-reform, or making progress on AI alignment (!), or whatever. But a thing that I can do, while still in college, is EA movement-building. If movement building is the main thing that the young and excited new entrants to the movement can do, there’s a structural incentive for the movement to be functionally about propagating itself.
Plus, there’s something additionally appealing about boosting the EA movement generally, which is that it allows me to abstract over a bunch of hard-to-answer questions about cause prioritization, and the details of specific plans. Trying a specific object-level plan is like buying a specific stock, but investing in the EA movement as a whole is like buying an index fund—you’re diversified, .
So there’s a very natural inclination to believe in “EA” as a good thing in the world, which your own hope of having impact can route through. And in this way we end up with an EA that is largely self-recommending.
In early 2015, I read an okcupid profile, where someone declared themselves as an EA early in their bio (as I did as well). I sent her a message, and we hopped on a video call immediately, and talked for 2 hours. A few months later we were living in the same house, and she’s a close friend to this day.
Today, just saying you’re an EA wouldn’t be nearly enough signal to make it obviously the case that I want to talk with you for hours. This is in part because my opportunity cost has gone up, and in part because my standards for conversation have risen as I’ve learned more and it’s gotten harder to tell me something interesting that I don’t already know, and in part because the signal of being an EA has weakened as the franchize has expanded.
I think this is sad, but also… not surprising? (For different reasons than the ones you describe.) Specialization of labor and gains from trade also operate in the epistemic realm, and I don’t think we were ever not going to end up with a deferral network.
I think the main place where I wish EA had outperformed more was the endpoint of that deferral network. Like, EAs / rationalists have been major early adopters of prediction markets and forecasting tools, and I think there was the potential to use them more centrally. (Like, encouraging people to trade in the cause prioritization markets instead of ‘come to their own conclusions’ would have, I think, been better at helping people realize when and how they should or shouldn’t be deferring.)
It seems like 3 things are simultaneously true:
1) It’s not possible to eradicate malaria for $5,000/life saved (that’s the marginal cost, approximately, ballpark estimate with lots of wiggle room in the number). This generalizes to all other currently known interventions to make people’s lives better at low cost—it’s relatively cheap now, but one should expect that saving the last life that would otherwise have died from malaria, or helping the last person who can be helped with some other intervention that is currently near the best marginal cost, will cost a lot more than $5,000. I feel like Givewell et. al. are clear about this, or at least this is the understanding I came away with when reading their stuff, rather than being a surprise I discovered from other sources.
2) If we did assume that we could save all the savable lives for $5,000 apiece per year (again, this is false, but as a way of ballparking how much is definitely not enough to solve all problems), taking the number of 10 million deaths per year of communicable disease (referenced in Ben’s blog post, and noting that “communicable disease” and “all problems” are not the same) at face value, we get $50 billion per year. As I understand it, the amount of philanthropic wealth to eventually give away in places like the Gates Foundation and Good Ventures and affiliates is on the order of tens to hundreds of billions of dollars, not trillions. So spending $50 billion/year is not sustainable, and even at the marginal cost of $5,000 per life saved, there’s still going to be a long term funding gap.
3) Having recently done a deep dive into Givewell’s recommendations for a presentation to a local community group, I notice that their recommendations and methodologies are still developing. The basic ideas like “there are opportunities in poor countries that don’t exist in rich countries” and “malaria is bad, and helping with it on the margin is cheap” have stayed the same, but they flag lots of places where their ways of analyzing things have changed over the past 5-10 years, and generically for any page that hasn’t been updated in a few years, they put a disclaimer at the top that the analysis may not reflect the present situation or Givewell’s best understanding. Also new cause areas in EA are coming up every few years, and there are a lot of “more research needed” tags around most conclusions. My takeaway is that we know a lot more now about how to use money to do good, than we did 10 years ago, and I expect that 10 years from now, we’ll have similarly improved our knowledge-base, and may have uncovered better opportunities than are currently known. So it makes sense to hold some money in reserve pending that further research.
With this as background, if I try to put myself in the shoes of someone who has $10 billion and wants to do maximally good things with it, it makes sense not to spend it all immediately on malaria, even though that would save a bunch of lives. And it makes sense not to fund the crap out of all the marginally most effective interventions until all my money is gone this year, driving the marginal cost of various outcomes up to some multiple of its current level which makes it so everyone else feels like contributing to this effort is some multiple less useful than it was last year—and then my money runs out, all the orgs who got a bump in funding are scrambling for next year’s funds, the rest of the philanthropic community feels less interested, and chaos ensues. Clearly a suboptimal approach.
Just out of curiosity, I googled how much has been spent on Polio (almost, but not quite, eradicated), and estimates are on the order of $20 billion over the past 37 years (not adjusted for inflation) with the 2022-2029 budget being $7 billion (so the last few cases are going to be super expensive). So I would be unsurprised if solving all communicable disease had a cost on the order of hundreds of billions to trillions of dollars, and couldn’t be done tomorrow or this year even if I had $100 trillion to spend on it.
This seems to be a core of your argument:
Whereas to me, I don’t see evidence that difficulties were found in the cost effectiveness estimates (it was never claimed that marginal cost could just be multiplied by number of cases to get total cost) that advice from GiveWell makes sense, doesn’t seem weird, and doesn’t seem like the parties upstream of the main message don’t buy it. “Going hard” with $10-100 billion, when the total cost to solve all extreme problems is going to be trillions of dollars and tens of years at least and we’re constantly learning as we go, looks (to me) like approximately what the 10-billionaire philanthropists are doing. At least, a better strategy, if Givewell’s recommendations are basically sound, doesn’t immediately occur to me.
An analogy:
We are in the “you can save a drowning child for an affordable price” world. In this world (or a hypothetical one for the purpose of this analogy), 1,000 infants are being dumped in a large lake per day. Some of them are right by the shore, easy to get to like the drowning child thought experiment postulates, some are out in deeper water. I’m a strong swimmer, and could save any of those infants, but I can’t save all of them by myself, and if I try to save as many as I can today, I will exhaust and potentially injure myself, meaning I can save fewer tomorrow. I estimate I could save 100 today if I put forth all the effort I can, saving the ones that are easier to reach first. But there are some people who aren’t as good at swimming as I am, so maybe I should swim further out, save 10, and let the 90 I chose not to save that I could have, hopefully be taken up by others. But in that case, I’m still exhausted, and can do less tomorrow. So maybe I save 5, every day, a mix of easy ones and harder ones to demonstrate what’s possible and encourage others to join in, and spend some of my energy and time trying to figure out how to solve the underlying problem here, rather than just pulling people out of the lake. And then someone says “looks like it’s not true that it’s easy to save a drowning child, if it was true this guy would have saved more lives”. And what it looks like to “go hard” at this problem is not obvious, but “save a few, leave room for others, try and solve the underlying problem” is one plausible strategy. And because I can’t save them all “I need more people to help with this” is a thing I should say to anyone who will listen.
Fantastic response Thank you.
Thanks. Side note (I posted another comment about this just now, because it just clicked for me this morning): I think Ben Hoffman thinks (or thought, when he wrote his blog posts) that when you treat more malaria cases or do other philanthropy, marginal cost goes down. He says:
When in fact it’s the least generous, under the assumption that marginal cost goes up. If you think marginal costs will only go down from current levels as we scale, then it is indeed suspicious that nobody’s decided to just dump all their money into scaling.
things shouldn’t try to scale from MM->BB as a sudden step change and not assume it will break approximately everything about what was working at MM.
Only the GiveWell stream. But EA originated from a combination of GiveWell, the Oxford philosophers who ended up founding 80,000 Hours and Giving What We Can, and LessWrong.
Thank you, this is helpful to clarify and remind for me what Ben was trying to say.
Do you or Ben have a more detailed explanation of what happened here? What can/should “good meaning folk” do to prevent this? Should I personally be worried about something like this?
I think I’ve found a crux that makes things make sense today, that didn’t make sense to me yesterday as I was reading the first linked blog post. When trying to think about the existence or nonexistence of a funding gap, Ben says:
And my brain skipped a beat, and went “no, the opposite, that’s the least generous possible assumption. As we treat more, the next treatment becomes more expensive, not less. Maybe that’s what he meant to say?” And then the rest of the blog post just seemed sorta wrong. I was like “maybe he doesn’t get that this is marginal cost?”. But that’s not it, I realized as I was laying in bed thinking about this.
If you’ve been steeped enough in standard economics which talks about producing widgets, your mental association between “marginal cost” and “as you produce more, marginal cost goes down” is strong. Because usually, there are economies of scale which mean that’s what happens. From memory, I recall seeing something that said that every time the number of solar panels produced doubles, the cost per panel goes down by 20%, for example.
If you’re thinking “$5,000 is the current marginal cost, and therefore the most saving a life with a malaria net will ever cost, the more life-saving we do, the less it will cost”, then everything else he says follows. The right strategy for someone with a lot of money would be to scale this intervention to the max and drive costs down. And the fact that billionaire philanthropists aren’t doing this is suspicious, suggests that they don’t really believe the conclusions Givewell has put out. EDIT TO ADD: And the fact that marginal costs have been going up over time as we do more philanthropy, rather than being exactly what you would expect to happen, suggests that Givewell got its initial estimates wrong and what’s happening is as we learn more, we learn that everything is more expensive than we thought. There are not in fact any cheap drowning children to save, and never were.
If instead you think of the marginal cost as “the cost to pick the lowest hanging fruit, will go up when we pick that fruit” then you reach different conclusions.
Retracted, because after conversation on Ben’s blog, it’s not a matter of thinking marginal costs will go down with scale—rather, he treats “funding gap” more strictly than I would, something like “the amount of money charities can absorb and deploy at approximately present marginal costs per life saved”. And if you define the funding gap as “the amount of money that can be deployed at approximately present marginal costs, or better”, then “assume marginal cost does not go down at all” is indeed the most generous assumption.