The world as a collection of machines
A continuation of the exchange with the writer Kelsey Piper (later a Vox journalist) that began over Piper’s Tumblr post on personal budgeting for donations. Having argued in the previous round that, given the wealth concentrated in the Bay Area, budgeting to free up donation money is penny-wise and pound-foolish, Alyssa here lays out the underlying picture: she thinks “the economy” is a bad category, and proposes replacing it with a model of the world as a collection of “machines” — bundles of devices, procedures, institutions, and Schelling points that reliably produce certain results. The framework drives a sharp claim about effective altruism: financial contribution is only non-replaceable when paired with unusual judgement that no billionaire or hectomillionaire already has. The closing section addresses Piper’s case for kidney donation as a costly signal of commitment.
(Before I jump into the messy detail, I’ve definitely been enjoying this conversation, and I think it’s really helped me organize my own thoughts.)
I think “the economy” is mostly just a bad category — it takes a huge number of dissimilar things and throws them together in the same box, to the point where measurements of “the economy” (GDP, unemployment, inflation, etc.) are at best rough guesses and at worst outright lies. Economics contains a fair amount of useful knowledge within it, but IMO it really needs an overhaul to about half of its ontology. This isn’t really that surprising, for a science at such an early stage — you could think of it like, say, chemistry in the 17th century. There are lots of observations and rules and procedures that basically work, but there are still central concepts like “transmutation” that need to be thrown out, and other ones like “valence electron” that haven’t been discovered yet. (Not that I know how to do that — I have guesses, of course, but this’ll be a major decades-long project just like the invention of modern chemistry was.)
I think a better metaphor is to see the world as a collection of machines. A “machine” isn’t a literal mechanical device, but a collection of devices, procedures, memes, writings, traditions, institutions, Schelling points, and so on that operate together to reliably produce certain results. Some machines work well; others work surprisingly badly; and a great many simply fail to exist or haven’t been invented yet. You could say that entrepreneurship, in a broad sense, is the creation of a new machine; FDR and Florence Nightingale were entrepreneurs by that definition. Machines can also be destroyed, and of course they constantly evolve in response to the forces around them.
The way you produce happy lives for a large number of people — a larger number than you could help directly with your own muscles — is to build a set of machines that, taken as a whole, reliably give people what they want. (What exactly they do want is a whole new complex topic, and a central question to FAI theory, but for now we can just say that eg. no one ever wants to get infected with malaria.) In some cases, these machines already exist and you can freely make use of them when setting up your own stuff. Eg. if your plan is to help people by setting up a gold-mining operation in Kenya, there already exists a very efficient machine to buy, sell, transport, refine, distribute, and price gold that you can take advantage of. You can more-or-less just bring big sacks of gold dust to downtown Nairobi, and hand them off there — you can trust that someone else will take care of utilizing them in the most efficient known way. However, this machine only exists because of a number of background conditions: fungibility: one ounce of gold is the same as any other ounce perfect information: it’s easy to tell if a bar is made of gold or not cheap shipping and distribution: the cost of transporting and distributing an ounce of gold is far less than the gold itself practical contract enforcement: there exist organizations which would be meaningfully punished if they just stole all your gold, so they don’t (a bunch of others I won’t get into) By contrast, if tomorrow you discovered a cure for cancer, by itself that would be more-or-less useless. There’s no machine for evaluating and pricing and manufacturing and distributing cancer cures. You’d have to build one yourself, and that’s a huge amount of work and requires lots of different skills — dealing with bureaucracies, hiring and managing employees, raising funding, conducting human trials, and on and on and on. If you don’t happen to have those skills, then people will keep dying of cancer. (One example I have personal familiarity with is Eric Lagasse’s work on liver regeneration — I watched Zvi Mowshowitz try to build a machine for distributing this to patients, and saw him fall flat on his face, despite being extremely smart and quite capable in other domains.)
There isn’t any limit on how powerful a machine can be — the easiest historical example is Gutenberg’s printing press, the important part of which wasn’t really a “press” so much as a new set of techniques for making and using metallic type. On the other hand, trying to build an arbitrarily powerful one faces two fundamental constraints. The first is that, to be very powerful, it has to be fundamentally dissimilar from anything that many other people are trying to do. If it were similar to ones that tons of other people were already building, eg. how to make a better lithium-ion battery, odds are someone else would have built it already. The second constraint is that the vast majority of really original ideas are terrible; if you just naively disregard existing constraints, then you’ll probably fail, because reversed stupidity is not intelligence. (Paul Graham and Peter Thiel talk about this at length in How to Get Startup Ideas and Zero to One , respectively, though it’s a counterintuitive enough idea that you have to sort of see it from many angles to understand it well, kinda like the proverbial elephant with the blind men.) So to succeed, you have to know something that other people don’t; to do that, you have to know how to recognize which things you don’t know; and knowing how to recognize which things you don’t know is just really really hard. Eliezer’s Sequences are the best attempt I’ve seen so far to teach it ( Artificial Addition is one particularly good example), and I like to think I’m pretty smart, and even so I don’t think I really understood it until having read them three or four times over a period of about six years.
In keeping with the analogy, any given machine, once built, usually only works within a given set of operating parameters. You can make your car put out 100 kW instead of 50 kW by pressing down on the gas harder, but you’ll never make it produce 10,000 kW, because it’s designed to top out at 200 kW or thereabouts. Similarly, any given charity or type of charity can only handle so much money before it clunks out. And charities (or any other machine) that can operate productively under a load of even one percent as much money as the developed world has — tens or hundreds of billions a year — are more-or-less nonexistent because of various scaling issues. You’ve probably read that humans are evolutionarily adapted to work in small groups, from a handful up to 100 or so; the further you go beyond that, the more you’re stretching the cognitive abilities of the poor saps who have to run the whole thing beyond their natural design limits. One of the very few well-understood ways around this is to avoid tackling the scaling problem yourself, by just redistributing the money to others in some simple, well-defined way; but precisely because this is one of a very few well-known ways around a critical bottleneck, it’s one that’s extremely popular, and you’d therefore need a huge amount of resources to substantially add to what’s already being done (IIRC, even ignoring existing aid altogether, there’s already over $300 billion per year in direct remittances to the very poor from friends and family).
Hence, under this framework, the two largest ways to contribute at the margin are: to build a new machine where the type of machine is relatively well-understood, and the bottleneck is that the existing machines can’t scale well and the type of labor required to build new ones is scarce; this covers both creating new charities to address tropical diseases, and most “ordinary” software entrepreneurship, as well as many other things to build a new machine where the type of machine isn’t well-understood, and the bottleneck is the skill and background knowledge to have the required insights into what blanks need filling in; Eliezer is one example of someone we know who’s AFAICT succeeded at this, but successes here are necessarily much rarer than in the first category By “build”, what I really mean is “contribute to building in a relatively non-replaceable way”; there are usually many different types of skills required, hence many opportunities to contribute. And it’s certainly true that one opportunity is “provide the initial rounds of funding”. However, in order for your financial contribution to be non-replaceable, you yourself must have the same types of unusual cognitive abilities as the people running the organization, the ones that make them able to succeed when most others couldn’t. If you yourself only have ordinary-programmer cognitive abilities, and not (for example) figure-out-which-organizations-aren’t-likely-to-get-torn-apart-by-internal-conflict cognitive abilities, then your funding will on average just go to the same place as the ordinary programmer’s; and so either you won’t fund the organization at all, or else lots of ordinary programmers will fund it too and so your funding won’t mean much on the margin. And you can’t outsource your judgement to an organization-evaluator, because if your ability to judge the judgement of organization-evaluators is the same as an ordinary programmer’s, then lots of ordinary programmers will follow the recommendations of the organization-evaluator too and you get the same problem. The ability to contribute by offering funding is, to a first approximation, only valuable insofar as the funder personally has unusual abilities, not possessed by any billionaire or any hectomillionaire or by more than a microscopic fraction of Silicon Valley career software developers, to judge which things need more money and which need less. (And if you do have that ability — not meaning Kelsey-Piper-you here, but hypothetical-abstract-you — and don’t already have a good chunk of change to contribute, why not just become an accountant? All the important-to-humanity organizations I’ve been closely involved with have been in desperate need of good accountants. Again, it’s not accounting itself that’s valuable here, but accounting combined with highly-unusual-for-accountants-judgement-of-which-organizations-to-contribute-to.)
Whew! Okay. Probably lots of things still to say about that, but it’s late so I’ll wrap it up there. On donating a kidney: in my mind, the main thing that kidney donation signals about a person is that they’re extremely responsive to particular types of social pressure. Kidney donation is something that people are widely encouraged to do through mass media, but quite rarely through in-person social groups, except in cases where a friend or family member is the one who needs the kidney (which I’m excluding here). Hence, because kidney donation involves significant pain and difficulty, they’ve shown that they’ll do what they’re encouraged to by those particular types of media even when there are other strong forces pushing against it. To me this is a bad signal, because (eg.) if someone is motivated to action by systems of moral rules, you can construct a better system of rules that’s more effective than the default one; if they’re motivated by in-person social signals, you can rearrange your social group to some extent to make its signals more effective than default signals; if they’re motivated by money, you can talk people giving them money into redirecting the money they’re trying to capture in such a way that their trying to capture it creates more utility; and so on; but you can’t really create your own mass media, and so the original, non-optimized forces exerted on them by mass media will likely still be acting on them even after they’ve integrated into the community. Costly signals of commitment are important, but they only help your group insofar as they’re signals of commitment to your group and not some other, nearby, more powerful group. As an analogy, if you were starting a new political movement in 1970, and you wanted to establish higher taxes on big corporations, you should not have your costly signal of commitment be to embezzle money from some specific defense contractor that the local KGB agent wants you to embezzle money from; because then your actions are causally connected to the KGB and not to backward-chaining from your original goals, and maybe the next thing the KGB will want is to bomb the Capitol or something else obviously destructive.