Effective Altruism at the Margins

With one Sam Bankman-Fried in the news again (now realizing that nominative determinism on that last half-surname is not working out as well for him as he might have hoped), I thought it might make sense to write about Effective Altruism, the philosophical approach towards charitable giving to which that crypto scammer was unfortunately a prominent proponent. This blog is about things that have influenced my thinking, after all, and I’d say that Effective Altruism, which I first encountered via Peter Singer’s book The Life You Can Save and organizations like Giving What We Can and GiveWell, has been a significant influence. Prior to that, the approaches to choosing between altruistic causes I’d encountered focused either on personal connection or purely internal measures of organizational efficiency, things like overhead ratios. There’s some relevance to that, an organization that spends 100% of its budget on fundraising and administration obviously doesn’t have any budget left for direct use on the actual mission. But overhead ratios don’t measure what an organization actually does, and neither fundraising nor administration (planning, research, coordination, logistics) are off-mission per se.

In breaking from that, EA was willing to bite some bullets regarding things like executive salaries in the nonprofit sphere. It may hurt effectiveness if the most talented nonprofit executives are stomaching too much of a pay-cut relative to industry. Of course, it’s easy to see how for someone pulling down big bucks in the nonprofit space could find this a convenient justification. Similar for the idea of “earning to give”. For an audience of overly-conscientious people wracked with guilt over the prospect of taking a normie job in the professional managerial class, I think it’s reasonable to point out that taking the obvious path career-wise and donating money might not be so bad in terms of having a positive impact on the world, over doing something more directly altruistic where you have less comparative advantage as your career. But that too can be taken as rationalization and basically turned on its head. The point of “earning to give” is that it’s okay(-ish) for very conscientious people to work normal jobs in finance or whatever. Not as a justification for people behaving the complete opposite of what you’d expect from “very conscientious” in those jobs.

Part of of the problem with Effective Altruism is a problem with utilitarianism in general. A comparative implies a superlative and there’s not really a sound point to stop. This is one of the ideas that’s been rattling around my head so long that I’ll try to state it pithily:

Idea #8: There is no good theory of moral sufficiency.

Zvi Mowshowitz wrote an interesting review of Michael Lewis’s new biography of SBF, and the bit where Zvi puts an interesting perspective on it is in the comparisons to concepts of unfriendly AI. Will MacAskill’s introduction of Effective Altruism’s “number go up” perspective put SBF on the path of a totalizing optimizer:

Even more than that, if you take such abstractions too seriously, if you follow the math wherever it goes without pausing to check whether wrong conclusions are wrong? If you turn yourself into a system that optimizes for a maximalist goal like ‘save the most lives’ or ‘do the most good’ along a simple metric? What do you get?

You get misaligned, divorced from human values, aiming for a proxy metric that will often break even on the margin due to missing considerations, and break rather severely at scale if you gain too many affordances and push on it too hard, which is (in part, from one perspective) the SBF story.

Yet SBF did not take such concerns seriously. […]

MacAskill set SBF on a maximalist goal using an abstracted ungrounded simplified metric, hoping to extract a maximal amount of SBF’s resources for MacAskill’s (on their face altruistic) goals.

The idea of an optimizer that attempts to max out some quantity in the universe is a central one in AI safety theory because it seems to be one that almost invariably goes wrong. Even innocuous or nominally altruistic goals in the hands of such an optimizer end up being extremely hostile in practice. There’s a concept of instrumental convergence, this common set of goals that this style of optimization implies regardless of the initial goal. An optimizer on whatever goal doesn’t want to be stopped; if it’s stopped, it won’t achieve the goal. It doesn’t want to have the goal changed; if it changes its objective, it won’t achieve the (current) goal. SBF was unlike this in many ways:

Imagine a world in which SBF’s motivations had even less anchors to human intuition, and also he had a much larger capabilities advantage over others (say he was orders of magnitude faster, and could make instantiations of himself?) and he had acted such that the house of cards had not come crashing down, and instead of taking the risks and trying to score object-level wins prematurely he had mostly instead steadily accumulated more money and power, until no one could stop him, and his inclination to risk all of humanity every time he felt he had a tiny edge under some math calculation.

But it would surely be apt to describe him as incorrigible.

Still, I think Effective Altruism has a lot to say at that comparative level. It really is good to help more people with the level of effort you choose to put in. There is a big divide between efforts focused on straightforward ways of helping some of the world’s poorest people and nonprofits that are very much not that (e.g. Make-a-Wish, Harvard University).

It reminds me of another influence, Bryan Caplan’s Selfish Reasons to Have More Kids, I found it very interesting in terms of how it approached that argument via marginal economics. It’s notably not making an argument that people should have as many children as possible, or that some specific number is the right number. Rather, it argues that raising children is often more rewarding and less painful than people estimate, therefore many people underestimate how many children they should have.

Effective Altruism makes a similar argument: Make a habit of donating to organizations that do more good for each dollar, that measure their impact and attempt to improve their effectiveness, and you can do more good more easily than you might expect. Perhaps you can, perhaps you should do more.