November 18, 2016
In this 2016 talk, The Future of Humanity Institute's Owen Cotton-Barratt discusses how effective altruists can improve the world, using the metaphor of someone looking for gold. He discusses a series of key effective altruist concepts, such as heavy-tailed distributions, diminishing marginal returns, and comparative advantage.
The below transcript is lightly edited for readability. Part 2 of the transcript can be found here.
The central metaphor that's going to be running through my talk is Effective Altruism as mining for gold. And I'm going to keep on coming back to this metaphor to illustrate different points. And gold here is a standing for whatever it is that we actually value. So, some things we might value include making more people happy and well educated. Or trying to avert a lot of suffering. Or trying to increase the probability that humanity makes it out to the stars. When you see gold, take a moment to think about what you actually value. For many people it won't just be one thing that they value. But do think about what you care about and put that in place of the gold. And then there's lots of observations we can make.
So, this is a photo of Viktor Zhdanov, and I learned about him by reading in Will MacAskill's book, Doing Good Better. He was a Ukrainian biologist, who was instrumentally extremely important in getting an eradication program for smallpox to actually occur. As a result he probably was counterfactually responsible for saving tens of millions of lives.
Obviously, we don't all achieve this. So, by looking at examples like this, we can notice that some people manage to get a lot more gold, manage to achieve a lot more of whatever we altruistically value, than others. And that's reason enough to make us question what is it that gives some people better opportunities than others? How can we go and find opportunities like that?
Elsewhere in this conference, there are going to be treasure maps and discussions of where the gold is. I'm actually not going to do that in this talk. I'm instead going to be focusing on the tools and techniques that we can use for locating gold, rather than trying to give my view of where the things are directly.
Another thing I just want to cover here is actually, I'm giving this metaphor, I want to say a little bit about why I'm even using a metaphor, because we care about these things. We care about a lot of these big, complicated, valuable things. Why would I try and reduce that down to gold? Well, it's because of where I want the focus of this talk to be. I want the focus to be on these techniques and the tools and approaches that we can use. And if you have complex values, that we're trying to put in the background, it's just going to keep on pulling attention. But a lot of the things that we might do to try to identify where valuable things are, and how to go and achieve them, are constant, regardless of what the valuable thing is. So, by replacing them with a super simple stand-in for value, I think it helps to put the focus on this abstract layer that we're putting on top of that.
So, the first thing I'm going to talk about is the fact that gold is, like literal gold is pretty unevenly spread through the world. There's loads of places with almost no gold at all, and then there's a few places where there's a big seam of gold running into the ground. This has some implications. One is that we would really like to find those seams.
Another is about sampling. For some quantities, say if I want to know roughly how tall people are, sampling five people, measuring their height and saying, “Well, the average is probably like that” is not a bad methodology. However, if I want to know on average how much gold there is in the world, sampling five random places, and measuring that, is not a great methodology, because it's quite likely that I'll find five places where there's no gold, and I'll significantly underestimate. Or possibly, one of them will have a load of gold and now I'll have a massively inflated sense of how much gold there is in the world.
So this is a statistical property that loosely gets called having a heavy tail on the distribution. This here on the left is a distribution without a heavy tail. There's a range of different amounts of gold in different places, but none of them have massively more or massively less than typical.
On the right, in contrast, is a heavy tail distribution. It looks similar-ish to the one on the left hand side, but there's this long tail of, getting up to very large amounts of gold, where the probabilities aren't dying off very fast. And then this has implications.
So here's another way of looking at these distributions. In this case, I've arranged going from left to right, the places in order of increasing amounts of how much gold they have. These are the percentiles, and then I've just put the amounts of gold on the vertical axis. And in this case, I've colored in beneath the graph. And that's because that quantity, that area, is meaningful. It corresponds to actually the total amount of where that gold is. So, in this case, on the left of the distribution that wasn't heavy tailed, I can see that the gold is fairly evenly spread across lot of different places. And so, if we want to just get most of the gold, what important is getting to as many different places as possible.
Solar power is like this. Sure, some places get more sunlight than other places, but the amount of solar power you generate depends more on how many total solar panels you have, than on exactly where you place them.
Over on the right, though, we have a distribution where you can see a lot of the area is that spike right at the right hand side. And so this just means that a lot of the gold, and if this is a true stand-in for something that we value, a lot of what's valuable comes in this extreme of the distribution of things, which are just unusually good.
So literal gold, I think, is distributed like this. Disclaimer, I'm not a geologist, I actually don't know anything about gold, but I understand that this is right. We might ask, is this also true of opportunities to do good in the world? So here's a couple of bit of support for this.
First is just general things, when we look into the world and it's pretty complex, we do see distributions with this heavy tail property coming up in a lot of different places. There are some, kind of theoretical reasons to expect certain types of distribution to arise and also empirically, if we go and look at something like income distributions around the world, again this is that percentile version and you can see the spike.
Okay, that was if we just go and look at the world. Obviously, there are lots of things which don't have this property as well. But the more that we look at things where there are complex systems and there are lots of interactions, that often increases the degree to which we see this property. And that is a big feature of lots of ways that we try and interact to improve the world.
I can also just try and look explicitly at opportunities to do good. And I can see a couple of reasons why I personally am convinced that we get some of this property. So, one is just convincing arguments. If I care about stopping people starving, and I do care about stopping people starving, I could ask should I be interested in direct famine relief and trying to get food to people who are starving today? I can compare this to something more speculative and I personally have been convinced by the arguments in this book, that it would be more effective to focus on doing research towards having solutions for feeding very large numbers of people in the event that agriculture collapses. It's pretty extreme. It's not something we normally think about, but I think that the argument basically checks out. And this tells me that one way of doing this, I'm just limiting myself here to trying to feed people, and one of the mechanisms looks much more effective than the other.
I can also look at data. So, this is data from DCP2, which has tried to estimate the cost effectiveness of lots of different developing world health interventions. The x axis is on a log scale, so these have been put into buckets, and each column is on average, 10 times more effective to than the one on its left. So, here the rightmost column is about 10,000 times more effective than the leftmost column. And this was, again, this was just within one area where we have managed to get good enough data that we can actually go and estimate these things. There's just a very wide range of cost effectivenesses.
So, the implications of this are that if we want to go and get gold, we really should focus on finding seams. In some cases, it might give us this surprising conclusion that getting evidence, say you discover that something is at the 90th percentile, that might make us less excited about it. Because before we knew anything, it might have been anywhere on the distribution. And if most of the possible value of it comes from it being up at the 99th percentile, then discovering it's only at the 90th percentile could actually be a bad thing. I mean, it's a good thing to discovery, but it makes us think less well of it. Now, that's if you've got a fairly extreme distribution, but it's interesting to see how you can get these kind of counterintuitive properties.
Another implication is that perhaps kind of naïve empiricism, “we'll just do a load of stuff and see what comes out best”, isn't going to be enough for us in judging this, because of this sampling issue. We can't go and sample enough times and measure the outcomes well enough to judge how it's actually going to be.
Okay. So, if we actually want to get as much gold as possible, we want to go to a place where there's lots of gold. We want to have the right tools for getting the gold out, and we want to have a great team who is going to be using those tools. I think that we can port this analogy over to opportunities to doing good as well. And we can roughly measure the effectiveness of the area or type of thing that we're doing, and the effectiveness of the intervention that we're doing to create value in that area, relative to other interventions in the area, and the effectiveness of the team or the organization who is implementing that, relative to how well other teams might implement such an intervention.
And if you have these things then the total value that you're going to be getting is equal to the product of these. I've represented it here by volume, and we want to be maximizing the volume. That means we're going to want to be trying to do fairly well on each of the different dimensions. At least, not terribly on any of the dimensions. And so, some implications there might be are that if we have an area and an intervention that we're really excited about, but we can only find a kind of bad, mediocre team working on it, it may be better not to just support them to do more of that, but to try and get somebody else working on it. Or to do something to really improve that team. Similarly, we might not want to support even a great team, if they're working in an area that doesn't seem important.
Okay. So, now in the next part I'm going to go into talking about the tools and techniques for identifying where in the world gold is. A nice property about literal gold is that when you dig it up, you're pretty sure that you can recognize, yes I have gold. We often have to deal with cases where we don't have this. We don't have the gold, so we have to carefully try to infer its existence, by using different tools. So this is like the dark matter of value.
And so that increases the importance of having good tools for trying to actually measure and assess this. It increases the importance of actually applying those tool diligently, as well. Iron pyrite also looks a lot like gold, so just because somebody says, “Hey, this is gold,” doesn’t mean we should always take people's word on this. It does provide some evidence, but we have motivation for wanting to have great tools for identifying particularly valuable opportunities. And being able to differentiate and say, “Okay, actually this thing, although it has some aspects of value, maybe it's not what we want to pursue.”
Okay. If you first go to an area where nobody has been before, then the seams of gold that are running through the ground have often been eroded a little bit, and you can have little nuggets of gold just lying around on the ground, and it's extremely easy to get gold. So you have some people go in, they do this for a bit, and they run out of all the gold on the ground.
And now, if they want to get more gold, maybe more people come along, they bring some shovels, and it's a bit more work, but you can still get gold out.
And then you dig deep enough and you can't just get in with shovels anymore and so you need bigger teams and heavier machinery to get gold out. You can still get gold, but it's more work for each little bit, for each nugget that you're getting out. This is the general phenomenon of diminishing returns on work that you're putting in, and I think that this actually comes up in a lot of different places and so it's worth having an idea about.
By the way, this is like several of the different things I'm going to be talking about, this is a concept, which I guess is native to economics. And in some cases, I'm fairly simply just porting across from economics and in some cases there's a little bit more modification on that.
But, for instance, I think that we get this in global health. I understand that 15 or 20 years ago, mass vaccinations were extremely cost effective and probably the best thing to be doing. And then the Gates Foundation has come in and they funded a lot of the mass vaccination stuff. And now, the most cost effective stuff is less cost effective than mass vaccinations. I mean, that's great, because we've taken those low hanging fruit. Or similarly, if in AI safety, writing the first book on super intelligence is a pretty big deal. Writing the 101st book on super intelligence is just not going to matter as much.
So, a minute ago, I talked about how we could factor the effectiveness of organizations into the area in which they were working, the intervention they were pursuing, and the team working on it. Now, I'm going to focus on that first one, trying to assess the area. And I'm going to give a further factorization, splitting that into three different things.
The first of these dimensions is scale. All else being equal, we would prefer to go to somewhere where there is a lot of gold, rather than a little bit of gold. And probably per unit efforts, we're going to get more gold, if we do that.
Second, tractability. We'd like to go somewhere where you kind of make more progress per unit work. So, somewhere where it's nice and easy to dig the ground, rather than trying to get your gold out of a swamp.
And third is uncrowdedness. This has sometimes been called neglectedness. I think that term is a bit confusing. It's a bit ambiguous, because sometimes people use neglectedness to just mean all things considered, this is an area which we should really put more resources in. What I mean here is just there aren't many people looking at it. All else being equal, we'd rather go to an area where people haven't already gone and picked up the nuggets of gold on the ground, than one where they have. And now the only gold remaining is quite hard to extract.
And so, ideally, of course, we'd like to be in the world where there's loads of gold, that’s easier to get out and nobody has taken any of it. But, we're rarely going to be in that exactly ideal circumstance. So one question is, how can we trade these off against each other? And I'm going to present one attempted way to try and make that precise. I've allowed myself one equation in this talk. This is it.
If you're not used to thinking in terms of derivatives, just ignore the ds here. But this on the left is the value of a little bit of extra work, so this is generally what we care about if we're trying to assess which of these different areas should we go and do more work on.
On the right is a factorization. So this is mathematically trivial, I've just taken this one expression and I've added in a load more garbage. And on the face of it, it looks like I've made things a lot worse. And I can only justify this, if it turns out that these terms I've added in, which cancel with each other, actually mean that the right hand side here is easier to interpret or easier to measure. And so I'm going to present a little bit of a case for why I think it is.
So this first term here, is measuring the amount of value you get for, say, solving an extra one percent of a solution. And that roughly tracks how much of a big deal the whole problem that you're looking at is, the whole area. And so, that, I think, is a pretty good precise version of the notion of scale.
The second one is a little bit more complicated. It's an elasticity, here, which is a technical term. It's actually a pretty useful and general term (go look it up on Wikipedia, if you're interested). Here it's measuring, for a proportional increase in the amount of work that's being done, what proportion of a solution does that give you?
And then the final term actually just cancels to one over the total amount of work being done. So that's very naturally a measure of uncrowdedness.
People have talked about this kind of scale, tractability, uncrowdedness framework for a few years without having a precise version. And that means that people have given different characterizations of the different terms, and I think there have been a few different versions of tractability, not all of them lining up with this exactly. But I think that this idea of it measuring how much more work actually gets you towards a solution, is fairly well captured by this version here.
And I think that all of these dimensions, again, matter. And again, that means we probably don't want to do or work on something, which does absolutely terribly on any of the dimensions. I'm not going to spend an hour helping a bee, even if nobody else is helping it and it would be pretty easy to help, because just the scale of it is pretty small. I don't think we should work on perpetual motion machines, even though basically nobody is working on it and it would be really fantastic if we succeeded. Because it seems like it's not tractable.
And there's nothing which is extremely not uncrowded, (that would be extremely crowded, just because there's only seven billion people in the world, as you can't force this that low, but this might give us a warning against actually working on climate change. Because at a global scale, that gets a lot of attention, as a problem.
I'm going to add some more caveats to that one. One is that this is going to be true while we think that there are other problems, which are just significantly more under resourced. And another is that, you might think that you have an exception, if you have a much better way of making progress on the problem of climate change, than typical work that is being done on it.
Even so, I think maybe we should think it's a bit surprising that I'm making a statement like “climate change is not a high priority area.” This just sounds controversial and we should be skeptical of this. But, I think that the term high priority is a little bit overloaded. And so I want to distinguish that a little bit.
If we have these two places where there's gold in the ground, and we say, “Where should we send people if we want to get gold?” The answer is going to depend. Maybe we send the first person to this place on the right, where there's only a little bit of gold, but it's really easy to get out. And then we send the next 10 people to the place on the left, just because there's more total gold there. The first person will already have gotten most of the gold on the right. And we want more people total working on this place on the left. Which of these is higher priority? Well, that just depends on which question you're actually asking.
These numbers are just made up off the top of my head, but we might have some kind of distribution like this on the left, where if we ask the question “How much should the world spend on this area, total?” we get one distribution, where maybe climate change actually looks very big on this.
And if we instead ask, how valuable is marginal spending? The graph might look actually quite different, because here it's significantly about how much is already being spent. You’ll see some black lines on the diagram on the left - they might represent how much is already being spent. And then, the graph on the right is a function of all sorts of things, like how much should be spent in total, how much is already being spent and of course, what the marginal returns are - what the curve looks like there.
But I think that both of these are actually important notions, and which one we use should depend on what we're talking about. If we're having conversation about what we as individuals or as small groups should do, I think it's appropriate to use this notion of marginal priority, of how much do extra resources help. If we're talking about what we collectively as a society or the world should do, I think it's often correct to talk about this kind of notion of absolute priority and how much resources ought to be invested in it, total.
Okay, for most of the things here, I've been extremely agnostic about what our view of value is. Just for this point, I'm going to start making more assumptions. I think quite a few people have the view that what we want to do is try and make as much value over the long term, as we can. Some people don't have that view, some people haven't thought about it. If you don't have that view, you can just treat this as a hypothetical: “Now I can understand what people with that view would think.” If you haven't thought about it, go away and think about it, some time. It's a pretty interesting question, and I think it's an important question, and is worth spending some time on.
But, if we do care about creating as much value in the long term as possible, in our gold metaphor, that might mean wanting to get as much gold out of the ground eventually, as possible, rather than just trying to get as much gold out of that ground this year.
And maybe we have some technologies which are destructive. So we can use dynamite and dynamite gets us loads of gold now, but it also blows up some gold and now we never get that gold later. And so that could be pretty good, if you are focusing just on trying to get gold in the short term. But it could be bad from this eventual gold perspective.
If we have different technologies that we can develop, maybe we can develop some that are also efficient but less destructive. And there are going to be some people in the world who do care about creating as much gold as possible in the short term. Then they're going to use whichever technology is the most efficient for that. And so one of the major drivers of how much gold is eventually extracted is the order in which the technologies are developed, and the sequencing. If we discover the dynamite first, people are going to go and have fun with their dynamite and they're going to destroy a lot of the gold. If we discover the drill first, then by the time dynamite comes along, people will go “Well, why would we use that? We have this fantastic drill.”
So philosophers like Nick Bostrom have used this to argue for trying to develop societal wisdom and good institutions for decision making, before developing technologies or progress which might threaten the long run trajectory of civilization. And also for trying to focus on differentially aiming to develop technologies which enhance the safety of new developments, rather than or before anything that's driving risk.
Continued in part 2.
- 1. Introduction to Effective Altruism
- 2. Prospecting for Gold (part 1)
- 3. Prospecting for Gold (part 2)
- 4. Crucial Considerations and Wise Philanthropy
- 5. The Moral Value of Information
- 6. The Long-Term Future
- 7. A Proposed Adjustment to the Astronomical Waste Argument
- 8. Three Impacts of Machine Intelligence
- 9. Potential Risks from Advanced AI
- 10. What Does (and Doesn’t) AI Mean for Effective Altruism?
- 11. Biosecurity as an EA Cause Area
- 12. Animal Welfare
- 13. Effective Altruism in Government
- 14. Global Health and Development
- 15. How valuable is movement growth?