I live for a high disagree-to-upvote ratio
huw
I think one thing that you gloss over a little bit is this chain of reasoning:
GiveDirectly demonstrate that extremely poor people generally spend new money responsibly
For the extreme poor, we can use this to measure purchasing choices as roughly contributing to life satisfaction
Elon Musk, who relies heavily on government subsidies to turn profits & who exclusively sells to wealthy nations, could be the most impactful person alive in WELLBY terms
Specifically, you might be extrapolating a relationship that apparently holds in poor countries to too wide of a net. I think there is good intuition behind this; people that are barely surviving are going to be far less prone to wasting a significant proportion of their income on speculative life satisfaction improvements. Whereas if you or I buy a Tesla I doubt it will have any impact on our life satisfaction at all (and not merely just a logarithmic one); similarly, if we developed a gambling addiction I’m sure that relationship would be negative.
It would be my intuition that the relationship between expenditure and life satisfaction doesn’t decay because it’s logarithmic, but because the relationship itself breaks down at higher and higher incomes.
(A smaller intuition that I have less of a clear reason for: It feels icky that a capitalist can double-dip by having a positive impact and making a monetary ROI. I suspect this hinges much more on what ideological ideas you’re bringing to the table; at the very least we can say that your method works in non-profit organisations)
I really like your accounting method, though, and I think it survives without most of the axioms you laid out. For instance, given that buying X product produces Y value, we have an exciting new method for accounting how to subdivide Y among its contributors! Then we can debate how to model Y, and for what business models this accounting method holds.
Sorry, to be clear, I was contesting that you can leap from (1) and (2), which I generally agree with, to (3). And to make that point, I was proposing that at higher income levels, the relationship from (1) and (2) may not (and in my opinion, does not) hold.
I did some more research & thought about your points. Although my own experience doesn’t suggest any relationship between conspicuous consumption and life satisfaction (arguably, for me, a negative one), a handful of low-citation papers (example) seem to suggest the relationship might be causal and not merely correlative (I agree it is clearly correlative). So I agree, in principle, it is probably true that in a clear-cut case such as Jeff Bezos (providing goods and services directly to customers in exchange for money where no counterfactual service likely would’ve appeared on a similar time frame), his work might have an immense impact on life satisfaction. (I tend to agree with your point (3) on if Elon switched to a more clearly-attributable business model).
I do have to wonder, though, if conspicuous consumption is fungible in a way that survival goods (food, medical, etc.) are not. I can’t shake the intuition that everyone would be just as happy if Amazon never existed; that when all of your peers have bought iPhones it becomes a table-stakes purchase and you have to get the iPhone Pro to stand out. I will think some more.
Self-guided mental health apps aren’t cost-effective… yet
Eddie, thank you (I’m a long time fan!)
Short-term impacts: Mmm, this has made me realise I wasn’t explicit about the assumptions I made there—I should either make that effect size bound a bit wider or model it as an exponential (or possibly a beta). I think this CEA is best interpreted as ‘if you built an evidence-based product, what would its cost-effectiveness be?’ but even that should probably have a wider bound. And there’s the new update in Linardon et al. (2024) that will be worth incorporating.
Adherence: Thank you! That roughly tracks with the decay curves from SimilarWeb, which is good validation. Although you raise a good point—decay probably depends a lot on whether you’re feature-gating after a trial period or not. Do you have a ballpark for the ratio of installs to DAU?
CPI: Those are lower CPIs than the estimates I had—good to know! Are those on Facebook, Tik Tok, elsewhere? I was also assuming organic traffic is negligible after the first hundred thousand or so, but do you still see an effect there?
Dev costs: Lovely! Having worked in industry, I definitely have the sense that there are good incentive reasons why headcounts might be unnecessarily bloated 🙃
Opportunities: I won’t ask what your roadmap looks like, but it’s very promising that you have this hunch. In my own experience as a user, I can definitely concur.
I’ll mull for a bit and update the OP with some adjustments. I might also shoot you a DM with some curiosity questions later. Thank you again! 😍
Cheers Barry!
Thank you for finding that update to Linardon, that looks very useful & I can pretty easily incorporate it. Will do ASAP and update the OP.
Guided self-help is interesting, of course, as a related flavour. I am much less familiar with the mechanics and cost structures. I suppose I also have general moral qualms about interventions that rely on low-cost labour to reach cost-effectiveness, but that’s a post for another day.
Yep, aware of Elephant in the Bednet. I deliberately chose the highest estimate of cost-effectiveness I could find for steelpersoning purposes, but for just evaluating cost-effectiveness head-to-head it would be inappropriate.
Will go through Torous’ work—thank you for the recommendation!
Yeah, I am very concerned about that aspect of the current market and a few other papers I reviewed suggested numbers in the range of 30% of users made download decisions based on privacy. I think there’s a lot here!
Thank you again, particularly for the extra research directions. Will analyse further :)
Hmm—good points. Getting Installs/DAU wrong could meaningfully affect the numbers, I guess longer-term retention per install is probably a better way of accounting for it. It was unclear to me whether to model retention as having a zero or nonzero limiting value, which would change some of the calculations.
Improving organic install rate would be promising if you could get it above 50%, I think (your apps sound very effective!). I suspect a lot of that is, as you say, about consistently building a good user experience and continuing to add value. (I see a lot of Daylio users complaining about the lack of updates & the increased ad load.)
I like the general framing & your identification of the need for more defensibility in these definitions. As someone more inclined toward GHD, I don’t do so because I see it as a more likely way of ensuring flow-through value of future lives, but I still do place moral weight on future lives.
My take tends more toward (with the requisite uncertainty) not focusing on longtermist causes because I think they might be completely intractable, and as such we’re better off focusing on suffering reduction in the present and the near-term future (~100 years). Similar thinking around wild animal suffering (although I wholeheartedly support improvements for farmed animals and think it could be tractable).
That feels different to your framing? But perhaps all that intractability is just a factor of my lower EEV estimates and/or the confidence in them?
Coherence may not even matter that much, I presume that one of Open Philanthropy’s goals in the worldview framework is to have neat buckets for potential donors to back depending on their own feelings. I also reckon that even if they don’t personally have incoherent beliefs, attracting the donations of those that do is probably more advantageous than rejecting them.
Volunteer led programs are often run by former clients turned group facilitators, and these can often be much more cost-effective than staff led programs.
Are these cost-savings due to not paying for labour, or because the staff are cheaper to train due to personal experience, or something else?
Thanks for the insights! I’m wondering how you think about self-funding charities, i.e. those which produce a product or service which they sell to users that can pay (or whose health insurers / governments can), and donate to users which can’t (or for whom it would be very cost-effective to give to). Have any applied to or gone through AIM before?
Yeah, I was asking w/r/t that counterfactual cost of labour. Perhaps this is a dumb question, but if that cost would’ve been negligible, as you say, then what explains the gap? Is the counterfactual cost of volunteer labour really that much smaller than staff labour?
it might best for entities to focus on what they are best at
Off the top of my head it seems like there would be reasonable exceptions to this. I think the necessary conditions would be in markets where a handful of companies control global development & distribution of a necessary thing that affects people across the globe (usually due to economies of scale & intensive development costs + IP control).
Insulin manufacturers, for example, might be justified in charging richer consumers (or more likely charging their governments & insurers), but would also be the best-placed actors to donate their products to poorer people. Specifically, they would be better-placed than separate organisations which use donations to buy & distribute insulin at low margins, because they’re still paying for the manufacturer’s margins. It equally doesn’t make sense for an insulin manufacturer to just target poorer people and run entirely on donations because of the cost-effectiveness they’d achieve by scaling globally.
But I’m not super sure on this. There are definitely advantages to having focus and I’m not sure really how many orgs fit this criteria, or, indeed, how many would practically do this given the overwhelming force of the profit motive.
Some feedback on your vignette—I can imagine a confounding effect from feeling like ‘a group of people wants to influence my decision-making’ or something similar. A purer form of your experiment might just include the cost-effectiveness numbers for both charities.
Possibly an infohazard, but would donating to sci-hub be the most cost-effective way to tackle this problem? Piracy had a massive effect on the cost structures of the entertainment industry; even if it didn’t remove the big players here it would force them to lower prices. (Moving to preprint servers is hard, given the way major journals control status in established fields).
The only other way out I see is regulation, and lobbying the EU in particular should be somewhat effective (they’re pro-regulation, they govern a significant part of this industry, they already previously supported then rescinded Plan S).
It feels like you’re arguing for a higher-than-necessary level of harm and suffering in the world, just because a high level of suffering already exists in this context? I can’t see an argument with this structure working anywhere else (and believe me, I think Sam should be punished).
It’s wild for a news organisation that routinely witnesses and reports on tragedies without intervening (as is standard journalistic practice, for good reason) to not recognise it when someone else does it.
I think this changed my mind, but at least for me I was more persuaded by your first point. I momentarily forgot that I really believe that white collar crime should have huge deterrent punishments that take little regard to the personal circumstances of the defendant; that ultimately a large punishment for Sam would proportionally create much less future harm. And that that’s not inconsistent with desiring & working toward an end to the hellish US prison system.
I think his take on GiveDirectly is likely to be very similar—he would point to the fraud and note that neither them or any of their evaluators took into account the harms caused by the beneficiaries of that fraud in their calculations. And I don’t think that that would be an unfair criticism (if delivered with a bit less snark).
I agree with both your points. I think the thrust of Leif’s argument, rather, is that no work was done to clarify the extent of those harms. They just say “we apologise to people counting on this” and quote statistics on how bad the militias in the area are.
On (2), I hope it was clear to anyone reading the article that Leif would like EAs to think in a negative-utilitarian way. I sincerely doubt he cares what proportion of the overall value of GiveDirectly’s work it was if a harm was done.
Specifically addressing your AI art point: In this case, you risk this fallacy being used to prop up technologies which solve problems they created. Which, I suspect, is part of the popular backlash against AI art in the first place. These justifications continue to be used by fossil fuel companies in developing ‘biofuelds’, ‘sustainable aviation fuel’, etc., and it’s not possible to falsify that some future iteration of the current harmful technology might exist; meanwhile the companies continue to pollute, often at greater and greater scales. There is a big difference between these companies developing sustainable fuels on the side, and redirecting 100% of their resources to that development. I suspect you might feel the same way about AI safety vs. general AI development.
Maybe we can amend the framing to exclude this somehow, because I really like the rest of this (the nuclear energy example felt particularly salient). To differentiate your examples, nuclear power intended to replace an existing harmful energy source, but AI art doesn’t replace… harmful manual artists? So I would perhaps frame it as occurring only when a promising new technology has potential harms today, but has some long tail of probabilities that could make it less harmful (rather than better) than the technology it replaces.