I seem to remember an EA forum post (or maybe on a personal blog) basically formalizing the idea (which is in Holden’s original worldview diversification post) that with declining marginal returns and enough uncertainty, you will end up maximizing true returns with some diversification. (To be clear, I don’t think this would apply much across many orders of magnitude.) However, I am struggling to find the post. Anyone remember what I might be thinking of?
I was thinking about this recently too, and vaguely remember it being discussed somewhere and would appreciate a link myself.
To answer the question, here’s a rationale for diversification that’s illustrated in the picture below that I just whipped up.
Imagine you have two causes where you believe their cost-effectiveness trajectories cross at some point. Cause A does more good per unit resources than cause B at the start but hits diminishing marginal returns faster than B. Suppose you have enough resources to get to the crossover point. What do you do? Well, you fund A up to that point, then switch to B. Hey presto, you’re doing the most good by diversifying.
This scenario seems somewhat plausible in reality. Notice it’s a justification for diversification that doesn’t rely on appeals to uncertainty, either epistemic or moral. Adding empirical uncertainty doesn’t change the picture: empirical uncertainty basically means you should draw fuzzy lines instead of precise ones, and it’ll be less clear when you hit the crossover.
What’s confusing for me about the worldview diversification post is that it seems to run together two justifications for, in practice, diversifying (i.e. supporting more than one thing) that are very different in nature.
One justification for diversification is based on this view about ‘crossovers’ illustrated above: basically, Open Phil has so much money, they can fund stuff in one area to the point of crossover, then start funding something else. Here, you diversify because you can compare different causes in common units and you so happen to hit crossovers. Call this “single worldview diversification” (SWD).
The other seems to rely on the idea there are different “worldviews” (some combination of beliefs about morality and the facts) which are, in some important way, incommensurable: you can’t stick things into the same units. You might think Utilitarianism and Kantianism are incommensurable in this way: they just don’t talk in the same ethical terms. Apples ‘n’ oranges. In the EA case, one might think the “worldviews” needed to e.g. compare the near-term to the long-term are, in some relevant sense incommensurable—I won’t to try to explain that here, but may have a stab at in another post. Here, you might think you can’t (sensibly) compare different causes in common units. What should you do? Well, maybe you give each of them some of your total resources, rather than giving it all to one. How much do you give each? This is a bit fishy, but one might do it on the basis of how likely you think each cause is really the best (leaving aside the awkward fact you’ve already said you don’t think you can compare tem). So if you’re totally unsure, each gets 50%. Call this “multiple worldview diversification” (MWD).*
Spot the difference: the first justification for diversification comes because you can compare causes, the second because you can’t. I’m not sure if anyone has pointed this out before.
*I think MWD is best understood as an approach dealing with moral and/or empirical uncertainty. Depending on the type of uncertainty at hand, there are extant responses about how to deal with the problem that I won’t go into here. One quick example: for moral uncertainty, you might opt for ‘my favourite theory’ and give everything to the theory in which you have most credence; see Bykvist (2017) for a good summary article on moral uncertainty.
It might be helpful to draw a dashed horizontal line at the maximum value for B, since you would fund A at least until the intersection of that and the curve for A, and start funding B from there (but possibly switching thereafter, and maybe back and forth). Basically, you want to start funding B once the marginal returns from A are lower than the marginal returns for B. It doesn’t actually matter for whether you fund B at all that B hits diminishing marginal returns more slowly, only that A’s marginal returns are eventually lower than B’s initial marginal returns before you exhaust your budget.
If you’re including more than just A and B, and A’s marginal expected returns are eventually lower than B’s initial marginal expected returns before you would exhaust your budget on A, then we can still at least say it wouldn’t be optimal to exhaust your budget on A (possibly you would exhaust it on B, if B also started with better marginal returns, or some completely different option(s)).
I’m not sure if you’re disagreeing with my toy examples, or elaborating on the details—I think the latter.
It isn’t clear what you meant by crossover point. I assumed it was where the curves intersect, but if you did mean where A’s curve reaches the maximum value of B’s, then it’s fine.
If the uncertainty is precisely quantified (no imprecise probabilities), and the expected returns of each option depends only on how much you fund that option (and not how much you fund others), then you can just use the expected value functions.
Right. You’d have a fuzzy line to represent the confidence interval of ex post value, but you would still have a precise line that represented the expected value.
I have two related posts, but they’re about deep/model uncertainty/ambiguity, not precisely quantified uncertainty:
Even Allocation Strategy under High Model Ambiguity
Hedging against deep and moral uncertainty
It sounds closer to the first one.
For precisely quantified distributions of returns, and expected return functions which depend only on how much is allocated to their corresponding project (and not to other projects), you can just use the expected return functions and forget about the uncertainty, and I think the following is very likely to be true (using appropriate distance metrics):
(Terminology: “weakly concave” means at most constant marginal returns, but possibly sometimes decreasing marginal returns; “strictly concave” means strictly decreasing marginal returns.)
If the expected value functions for N different projects are all identical, positive, non-decreasing and (weakly) concave (and each only depends on the resources used for the corresponding project), then the even allocation is optimal, and if they’re also strictly increasing and strictly concave, then the even allocation is the only optimal allocation. And if you fudge the functions a bit so that they’re very similar but non-identical, then these conclusions are approximately true, too.
More precisely, for any given positive, increasing and (weakly) concave function, as the expected value functions approach the given function,
the maximum expected value approaches the expected value of the even allocation portfolio,
the minimum distance between an optimal portfolio and the even allocation portfolio approaches 0, and
under a strictly increasing and strictly concave function, the maximum distance between an optimal portfolio and the even allocation portfolio approaches 0.
Maybe you can claim a stronger result by making this more uniform over the function being approached (uniform convergence), or without specifying any function to be approached at all and just having the functions approach each other, assuming the functions come from (a subset of) a compact set of possible functions.
If you could precisely specify the expected value functions (and so had precise credences) as fi , then you could tractably optimize with the budget B.
Maximize the following over the allocation weights wi,0≤wi≤1,∑Ni=1wi=1:
N∑i=1fi(wiB)I think you can just use a greedy allocation strategy, starting from 0 initial allocation, and allocating each marginal dollar (or X dollars) where it maximizes the expected marginal return until you exhaust your budget.
I think this would also be a convex optimization problem, so fairly well-behaved, and there are other specific optimization methods that would be practical to use (e.g. gradient ascent (+ noise and from random initial points to escape flat regions), maybe a greedy random walk if there aren’t too many different projects).
If your credences aren’t precise, you could check how much the optimal weights and expected values change under different specifications of the fi.
You might be thinking of this GPI paper:
It has the point that with sufficient background uncertainty, you will end up maximizing expectation (i.e., you will maximize EV if you take stochastically dominated actions). But it doesn’t have the point that you would add worldview diversification, though.
I don’t remember a post but Daniel Kokotajlo recently said the following in a conversation. Someone with maths background should have an easy time to check & make this precise.
> It is a theorem, I think, that if you are allocating resources between various projects that each have logarithmic returns to resources, and you are uncertain about how valuable the various projects are but expect that most of your total impact will come from whichever project turns out to be best (i.e. the distribution of impact is heavy-tailed) then you should, as a first approximation, allocate your resources in proportion to your credence that a project will be the best.
Mathematical theorems you had no idea existed, cause they’re false...
This looks interesting, but I’d want to see a formal statement.
Is it the expected value that’s logarithmic, or expected value conditional on nonzero (or sufficiently high) value?
tl;dr: I think under one reasonable interpretation, with logarithmic expected value and precise distributions, the theorem is false. It might be true if made precise in a different way.
If
you only care about expected value,
you had the expected value of each project as a function of resources spent (assuming logarithmic expected returns already assumes a lot, but does leave a lot of room), and
how much you fund one doesn’t affect the distribution of any others (holding their funding constant),
then the uncertainty doesn’t matter (with precise probabilities), only the expected values do. So allocating in proportion to your credence that each project will be best depends on something that doesn’t actually matter that much, i.e. your credence that the project will be best, because you can hold the expected values for a project constant while adjusting the probability that it’s best.
To be more concrete, we could make all of the projects statistically independent and either return 0 or some high value with some tiny probability, and the value or probability of positive return scales with the amount of resources spent on the project, so that the expected values scale logarithmically. Let’s also assume only two projects (or a number that scales sufficiently slowly with the inverse probability of any one of them succeeding). Then, conditional on nonzero impact, your impact will with probability very close to 1 come from whichever project you fund succeeds, since it’ll be very unlikely that multiple will.
So, I think we’ve satisfied the stated conditions of the theorem, and it recommends allocating in proportion to our credences in each project being best, which, with very low independent probabilities of success across projects, is roughly the credence that the project succeeds at all. But we could have projects with the same expected value (at each funding level, increasing logarithmically with resources spent) with one 10x more likely to succeed than the rest combined, so the theorem claims the optimal allocation is to put ~91% into the most likely to succeed project, but the expected value-maximizing allocation is to give the same amount to each project.
I think the even allocation would still be optimal if the common expected value function of resources spent on each project was non-decreasing and (weakly) concave, and if it’s strictly increasing and strictly concave (like the logarithm), then the even allocation is also the only maximizer.