Welcome one and all to the Cool Worlds podcast with me, your host David Kipping. This week we're doing something a little bit different. Now, usually, as regular viewers and listeners will probably realize, I speak to my guests in person. So, normally we're in the same space, the same room. I can see their reactions and feedback off them in real time. But today, I'm going to be doing a virtual interview. And normally I am very strongly opposed to virtual interviews because I do think you get so much more out of a real life conversation, but I guess today just wasn't available for an in-person conversation, fortunately. So the best I could do was have a virtual chat with them, and because of who it was, I decided it was worth doing that anyway. So the reason I mention this is because I am curious, before I introduce who the guest is, what your reaction to this style of interview is. If it doesn't bother you at all and you think this conversation is, you know, just as productive as any other ones we've had in person, that's actually useful for me to know. But if you prefer the in-person ones, then that's also good for me to know going forward. So I can really maybe put a harder line on that when I go forward with future guests.
So we're going to introduce our guest, that is Nick Bostrom, who of course is a philosophy expert. He's a professor of philosophy at Oxford University, but I think we probably know him in popular discourse for his writings on AI, the existential risks of AI and superintelligence, the simulation argument, and maybe less so in popular discourse, but I'm certainly aware of it, his great work on anthropic reasoning. So, anthropic reasoning is actually a topic I'm deeply interested in at the moment. I've been doing a lot of work and research that kind of touches on this, but applying it to astronomy and indeed cosmology has long been using anthropic reasoning with things like the cosmological principle, which is by extension the mediocrity principle, the Copernican principle, which comes into use in astrobiology. So these ideas have sort of bled down and have become a core part of some of the reasoning in science and astronomy as well. So I definitely wanted to spend some time trying to better understand Nick's take and whether his thoughts have evolved on that topic in this conversation.
But of course I think the thing that many of you probably have heard of with Nick's writings is the simulation argument, which of course is the basis of the movie The Matrix and many of the popular sci-fi adaptations, I'm sure. And in this idea, of course, it's a fairly disturbing one that we all might be living inside a giant computer simulation. And Nick was the person who really put down a kind of formal, rigorous argument to justify why this could be the case. And I've written a little bit about this in my own research, but Nick's really the expert on it, so I'm going to defer to him in this conversation. And I know that many of you, not all of you, but some of you, are really unsettled by the consequences of this argument. Some of you may be just completely ignorant and say, 'I don't like that, so I'm just going to completely throw it away. I'm not going to ingest it into my public, you know, my personal philosophy of the universe.' But I do think some of us, when you read Nick's writings and you take it seriously, it leaves you with an unsettling taste in your mouth at the end, like, oh my god, like, is this real? Could we all be living in some simulation? What does this mean for my life? It is an existential question. So I definitely wanted to talk about that in this conversation as well.
Of course, there's a lot more I could talk to Nick about because he's worked on so many interesting topics, and unfortunately we only had about an hour for this conversation. So, you know, please, I do apologize. I know there's probably many burning questions you have for Nick that I just simply didn't have time for in this conversation. I hope we can get him back on. If this conversation goes really well and lots of people view it and watch it, hey, maybe we can persuade Nick to come back on and go a little bit deeper into some of this stuff. So, if you like this episode, please do, you know, share it around widely, share it to your mom, share it to your friends, and get people to watch it because that really does make a difference. So without further ado, here is my conversation with Nick Bostrom.
It's really a pleasure to have you on at last, Nick. I've been following your work for a long time, so I'm really grateful for the opportunity to talk to you today.
Um, yeah, no, I think it'll be, uh, yeah, I mean, I've seen some of what you've done. It's like interesting stuff. So yeah.
And I thought a common place that we, obviously you are sort of the master of this, is anthropic reasoning, and it's something I've been getting increasingly interested in in my own research. It of course has a long history being applied in science and especially astronomy and cosmology. So I thought maybe we could start there and maybe you can just explain to the audience what is anthropic reasoning and what kind of problems is it really trying to solve.
Yeah, it's a piece of methodology for reasoning about a special kind of information, indexical information, which is information about who you are, what time it is, or where you are, sort of self-locating beliefs, which raise some distinct puzzles compared to the normal kind of information that just describes the structure of the world. So I think it maybe originally came up in the context of cosmology, where if you have a big universe, maybe the theory about what the structure of the world is predicts that there are a lot of observers observing many different kinds of environments and histories and so forth. And if we want to derive some observational predictions from that kind of theory for what we should observe, then we need some step from the model that says this is what the large-scale, perhaps infinite, universe says, there are all these observers observing different things. So what does that actually predict that we should find when we, you know, look into our telescope or whatever it is, given that there are so many different kinds of observations that would be made. So that's one place where this kind of anthropics could come in. And then there are other applications in evolutionary biology, in sort of various philosophical puzzles, for simulation theory, which we might get to.
Yeah, I'm curious if you've heard of the, I think when I was trying to look at the history of this, obviously there's many applications in cosmology, the Copernican principle, sort of one of the versions of this that we use in astronomy a lot, but there's the German tank problem, which I think was, I don't know if this is urban legend or if this is real, but there was the idea that the allies would look at the numbers on German tanks and if it was like 327 or something, it had like a digit printed on it, then they used that to estimate how many tanks the Germans were producing to then adjust their manufacturing processes. Have you heard of that case before or is that just sort of urban legend?
Heard of some similar thing like that. So that's a kind of a form of statistical reasoning that I guess is kind of parallel to anthropics in that you think of this thing you're observing as a random sample and presumably you observe a random tank from the population of tanks, right? And then you can make inferences from that.
Yeah. And I guess in discussions of anthropic reasoning there's two tools which often come up in these discussions. And one is the self-sampling assumption or SSA. And that's at least in astronomy, I think that's the one that we see being used most often. But then there's also the self-indication assumption, the SIA. I get the impression historically that kind of emerged later. But maybe we could just start by sort of explaining what's different about these two assumptions. And are there any kind of simple examples you can give us that might distinguish how they vary?
The self-sampling assumption is basically the idea that you should think of yourself as if you were a randomly selected sample from some reference class of observers. In the simplest case, this might be all observers. If we assume that certain things are observers, some are not. And then you should reason as if that was like a kind of uniform probability distribution of you making any particular observation or finding yourself in any of these slots. So if in the context of a cosmology, imagine some theory that somebody has come up with that says that there are sort of, you know, 90 blue planets with people living on them and 10 red planets with people living on them scattered across a vast cosmos. Now what should you expect to find when you look at the color of your planet? You should expect with 90% probability to find the thing that 90% of observers observe, and if you find the other thing, then that would be some probabilistic evidence against that cosmological theory being correct. So that's the self-sampling assumption.
The self-indication assumption is the idea that you should reason as if you were a randomly selected set of observers that include not just the actual observers that some theory predicts, but possible observers, which has the main consequence that you should update on the mere fact that you're alive. So that the fact that I'm alive makes it more likely that there are a lot of observers in the world. There are sort of more slots for me to have been born into. I'm born in, so probably a lot of people were born into the world according to the self-indication assumption, and that is relevant, for example, when if we look at something like the doomsday argument, which... but it also has various counterintuitive implications if one accepts. So I should say that something vaguely like the self-sampling assumption, the first one, I think needs to be accepted to get sort of reasonable results out of cosmological theorizing and other applications, although the exact version of that is still murky. The self-indication assumption is more controversial and some people accept it and some reject it. And it's less clear whether we should accept that.
That's interesting. Yeah. So the reason you connect back to the doomsday argument that you bring up there, and I'll give you a chance to maybe explain that as well. But my real question about doomsday is that it seems like if you apply this SSA, this self-sampling assumption, you do end up with these kind of catastrophic predictions for the fate of humanity. And I'll let you explain exactly how that works, but if you use the SIA instead, it seems to almost diffuse and cancel out some of those catastrophes. And if the self-indication assumption is on more precarious ground, does that give some wiggle room for the doomsday to be, because I think a lot of people have like a real resistance to doomsday, it is uncomfortable and they don't like it and they're looking for a way to diffuse it. So maybe my real question is here, what does the doomsday argument really teach us about these two approaches to anthropic reasoning?
Yeah, I think you're right. The self-indication assumption, the second, was kind of rolled in more or less to get rid of the doomsday argument. Maybe should I first explain just briefly what the argument is?
So maybe it's easiest to explain via an analogy. So suppose you have a big urn put in front of you and you're told it contains either 10 balls or 100 balls. And in each of these possible urns, they're numbered from one, two, three up to the total number of balls. So each ball has a number printed. So then somebody flips a coin and selects one of these two possible urns based on the outcome, right? So now you ask, what's the probability that the urn contains only 10 balls? So you say, well, 50/50. It was like a fair coin. Now let's suppose you reach into this urn and pull out the ball and it's ball number seven, right? Then based on normal uncontroversial Bayesian reasoning, you should update now. So now it's more likely that that urn only has 10 balls, right? Because it has a higher conditional probability of getting ball seven if there are only 10 balls than if there are 100. So you update and you get an elevated posterior probability of that. So that's all uncontroversial.
Now the doomsday argument is the claim that you should reason in an analogous way, but where instead of two urns with balls, you have two hypotheses about how many, say, observers or humans there will have been throughout human history by the time eventually the human species goes extinct. So let's say one hypothesis is that maybe there will have been, you know, 500 billion in total, you know, maybe we blow ourselves up with some new bioweapon, you know, in a hundred years or something like that. So that's hypothesis one. And then another hypothesis is, no, we'll survive, colonize, you know, solar system and live for a long time, maybe 500 trillion humans will have lived in total. Like obviously there are more hypotheses, but the structure doesn't depend on that. We can simplify it, just consider these two different ones.
And then we have some prior, like maybe 50/50, like in reality we'll take into account our guesses about nuclear war and bioweapons, and maybe we think probably it will last for a long time, but there's a 5% chance of it ending early. Doesn't really matter. But then the key step here is that now, just as in the case of the urns, you selected this random sample ball seven and updated on it. So too should you, according to the doomsday argument, update on the fact that your birth rank, your sequence in all the humans that have been born, is, you know, around 100 billion or so. That's the number of people who have lived before you. So that's kind of unsurprising. If there will be 500 billion in total, that's like a pretty typical type of number, right? But if they're going to be 500 trillion, then your birth rank would be extremely early. And so just as the beginning of the story.
So just as getting like a ball as low number as seven is kind of improbable if there were 100, let alone thousands of balls. So similarly, we should think it's unlikely that we would be early if there were in fact going to be very many humans. And so then you get this systematic update in favor of doom soon, like hypotheses that imply that there will be fewer people in total will have their probability increased by this line of reasoning. And so that is an unpalatable implication. And so then people have tried to refute this in different ways historically, and sort of the 90s and like a lot of papers written thinking they have come up with a clever refutation, and a lot of those just failed. They were confused or were easily rebutted. So it's like stronger than it originally seemed, this argument and this assumption, the self-sampling assumption, that you should think of yourself as a random sample. Might seem kind of unmotivated, but in fact it looks like in various other thought experiments and in some scientific applications, it looks like you need something like that to sort of get common sense implications.
So then one maybe bad reason to reject the doomsday argument is like we just don't like the implication that we probably will go extinct soon, but give it a low prior out of just preference. So that doesn't seem to have much epistemic relevance, but it also seems just counterintuitive, like it sort of seems a bit fishy that you could get such a strong implication about what's going to happen in the future and how big the risk of new biological weapons are and stuff like that, just from thinking about, you know, your birth rank. That it seems like how can that piece of evidence imply stuff about all these other things that seem so unrelated? That's a little suspicious.
And one way to bring it home is if you think of a more extreme version of the doomsday argument. So imagine if there were like Adam and Eve, like the world is created, there are the first two people alive and they know they are the first two people. And they can then reason like they might think, should we engage in carnal embrace? And like maybe they have been warned that there would be bad consequences if Eve gets pregnant and they fill the earth with progeny. And so maybe naively they would think there's like, if they have intimate relations, maybe there's a 10% chance that Eve gets pregnant, right? So they think, okay, that seems too risky given we've been warned what would happen. But then they have this kind of doomsday argument thing and they think like, well, if she does get pregnant and then the earth will be filled with our progeny, there will be billions of people, like what's the chances we would be the first two in that scenario? It's like almost like one in a billion, right? So therefore, she almost certainly won't get pregnant and so we can go forth and maybe that's what the snake said. And that seems even more counterintuitive because it just seems ridiculous that they could be so confident that she won't get pregnant even if they knew all the facts about biology and like how sort of mechanics of that works.
So then this might be a slightly stronger reason to reject the doomsday argument. It just seems to lead to intuitively incorrect results. Not that we don't like the results, but that we don't believe the results, like a more epistemically justifiable ground for rejecting it. But then you think, how could you reject it coherently while still getting sort of the scientific applications to work out? And then that's where the self-indication assumption would be helpful. So if you imagine Adam and Eve adopting the self-indication assumption, they think, well, the fact that we are born would be much more likely if there were many people born into the world, if there are billions than if we were the only ones. And so it turns out that this initial probability shift in favor of there being many people that you get if you accept the self-indication assumption, yeah, exactly, cancels out the doomsday argument shift in favor of there being fewer.
That they get when they update on discovering that they're actually the first. So you cancel out the doomsday argument, which is neat. But then you have a different set of problems.
Yeah, we'll get to those because I want to talk about the, you know, the presumptuous philosopher and things like that. But I guess just because it took a while for us to expand upon exactly what Doomsday is saying, and there are all these disagreements and tensions between these different ways of thinking about this problem. It's an old problem, but what do you personally really take away from all the ink that's been spilled about this problem? What do you think is the real lesson from the doomsday argument?
So if we live in a big world, which looks quite likely on independent grounds, like it's not then the case that there either just is like a small version of the human race or a long-lasting big version of it, but instead there might be possibly infinite or there is very, very large. Now if you have that picture, then even if you don't accept the self-indication assumption, just the self-sampling assumption itself would imply that a priori you are more likely to be born into one of the bigger species, the ones with more members, because those are all actually existing human beings or observers. So if you're a random sample from the... Yeah. Yeah. So if you're a random sample of all of these actually existing beings, it's more likely to be in a big subset than in a small subset, right? Like aside from more specific evidence than you get. So you already at that point have this shift in favor of say the human species being long-lasting that then gets canceled out when you update on us being early. And so that's one consideration there.
And I think, yeah, I think there is still a bunch of places where I think we are confused about how to reason about these really, well, about these, the methodology of anthropics, I think still has points of confusion, and also how to reason about these really big picture, like particularly you start to roll in infinities, which looks like you need, there are all kinds of paradoxical consequences that arise from that, and probably our current thinking is still a bit muddled on exactly how ultimately to resolve that.
Yeah, for sure. If you have an infinitely sized universe, then like it's like monkeys on a typewriter, like everything will happen to some degree. Then what do these statistics even really mean in...
So then you need a measure over that, right? And then you need to define that measure, and then you get the consequence that if you just moved people around without changing anybody's local environment or brain, instead you could sort of change the probability of being one or the other of these. And so there's like, yeah, some stuff that we seem to be confused about.
So there are actually some successes at least, I mean, I'm certainly aware of at least one clear success in cosmology of an actual prediction with anthropic reasoning, and it was with Steve Weinberg, a cosmologist. So maybe you already know about this, but he predicted the cosmological constant, I think back in the late 80s, 1987 I think it was. So this is before the actual measurement of the accelerating expansion of the universe, which of course happened and won a Nobel Prize for, but he predicted the cosmological constant based off sort of anthropic arguments, and he was really actually using, he didn't call it the SSA, but it was essentially the SSA, I think, if you read that paper. And what was remarkable was that he got within a factor of, I think, five to ten of the right answer, which, you know, given the scale was, you know, many, many, many orders of magnitude, it was quite impressive. And really the argument was, you know, the cosmological constant can't be too big, otherwise the universe would basically explode too fast and structure wouldn't form, and it can't be too small or negative, because if it's, especially if it's negative, it actually leads to the universe collapsing too quickly for intelligent life to develop. And so just based off those bounds alone, you can already kind of get an estimate. And at the time it was kind of controversial, everyone thought the universe would be static and unchanging, or not accelerating certainly, that it wouldn't be this extra force of dark energy. So it was a big deal to stick your neck out and predict that there'd be a positive cosmological constant. So I'm aware of that as one sort of success. Maybe you can comment on that. But are there other successes in anthropic reasoning that we can point to and say this is why this has at least some vindication for being a credible way of reasoning?
Yeah. I mean, I'm not sure I would... So it depends a little on how you define anthropic reasoning, and like there's like whether that could be claimed as a success for the more ambitious forms of anthropic reasoning, because obviously we know that we exist, everybody would agree on that. So clearly the universe, whatever properties it has, has to be such that we exist, like just as a sort of logical implication. Right, so we know all kinds of stuff. We know that there's water. So, you know, we know there's an Eiffel Tower. So our best theory of the universe has to be compatible with there being an Eiffel Tower and all kinds of everything we know the universe has to be, otherwise we can rule it out. I think where it would come in is if we try to compare or evaluate the plausibility of, say, multiverse theories, where some particular cosmological parameter seems fine-tuned to allow, say, intelligent or human life. We might think that's puzzling and look for an explanation, say the gravitational constant or the fine structure constant. And then one possible way of explaining that is by saying, well, look, there's actually this ensemble of universes, like reality creates all of these different ones. There is no fine-tuning at that level, it's just randomly generate a huge distribution with all kinds of values of the cosmological constant and fine structure constant. Then of course, if you do that enough, some will by just random chance happen to have exactly the value that allows observers to arise. And then the anthropics would come in and say, well, if that's your theory of the structure of the world, that theory actually predicts specifically that when we look out, we will find this exact, apparently but not actually fine-tuned value, because those are the only parts of this large ensemble that are observed.
So the theory gives high probability to our predictions, meaning it's kind of supported by our observations. If you didn't have that anthropics, then you might say, well, no, you need a theory that says that only this kind of universe exists with these particular parameters. You might look for some particular theoretical mechanism that would make it more likely that the fine structure constant has this precise value. And so like your plausibility judgment as between these two different explanatory approaches or theories about what the large-scale structure of the world looks like will depend on your view on anthropics. And I think the multiverse explanations are in fact supported because they could potentially explain these apparent observations of fine-tuning without actually requiring fine-tuning.
Okay. Well, actually, I appreciate that answer, and I think I learned something there. I mean, because I think I would have operated under a definition of anthropic reasoning as something like probabilistic reasoning conditioned upon our existence, something like that, but it sounds like you would want maybe a little bit extra on top of that. How would you, you know, just to distinguish, you know, why Weinberg in that case you wouldn't consider that a reason, what would be your clear-cut definition you would offer in your classes for what...
Well, that means because you wouldn't need the self-sampling assumption specifically. It doesn't matter that it needs to have that value in order for us to observe, like it could be for any other fact that we know is true. If we could derive from that the fact that the gravitational constant or the cosmological constant has to have a certain value, it would work just as well. Like if we observed some feature on Pluto, that it has, you know, a certain mass, that we could exist if Pluto had a very different mass, or like a larger or smaller, but nevertheless we happen to have looked at this, if that somehow implied that the gravitational constant has a certain value, then the inference would go through just as well. In order just to make that prediction, you wouldn't need any of this machinery.
Yeah, I actually use the example of not Pluto, but Neptune, Uranus quite a bit when I talk about the Copernican principle. I get a little bit frustrated sometimes when some of my colleagues speak about the Copernican principle on the Earth, and they say, well, you know, Earth has life, therefore life will be common in the universe. It's a very typical kind of circular argument that is sometimes used. But, you know, the Earth also has oxygen and water, but that doesn't mean that all the planets in the solar system should therefore have oxygen or water and necessarily even be abundant. But the fact is we have, our solar system does have a Uranus and a Neptune, which, as you kind of allude to, have no bearing as far as we can tell in our existence whatsoever. And so in that case, I'm quite comfortable to use anthropic reasoning in that sense, or, you know, whatever you want to call it, self-probabilistic reasoning, and it actually works out quite well because there are lots of Neptunes and Uranus-type planets in the universe. That's actually a good application of it.
Yeah. So you could say you could regard it as a sort of correction of the Copernican principle, like rather than assuming that what we observe is typical, it would be closer to correct to saying that what we observe is typical of the parts of the world that are observed, which might not be typical of reality as a whole. I mean, in fact, we know most of the world is just empty, ultra-pure vacuum, right, with no people there or looking, as far as we can tell. And we live in this very special little crust of biosphere, which seems very untypical. So, like, but it might be quite typical for the kinds of places that are observed by beings like us.
We'll be right back after this short message. Today's episode is supported by our friends at Incogni, a fantastic service that helps protect you and your data online, a vital counterweight to data brokers, skimmers, and bots. It's really kind of sad and ridiculous that there are now thousands of companies out there who are profiteering off harvesting your data and then selling it on without your permission. But the law does protect us, and each of us has the legal right to write to these companies and demand that they delete your information. The problem is it would practically take hundreds of hours to do this. So this is where Incogni comes in. They scour the internet for your information, and you can either confirm or deny possible matches. And then they use the law to remove your information from those databases. And this is often very sensitive information. I'm talking home addresses, medical history, and even social security numbers. Incogni has a new killer feature called custom removals, which lets you copy and paste any links where you find your personal information online, and then they take care of the rest. So to fight back, use my exclusive link, incogni.com/coolworldspodcast, to get 60% off by using my code coolworldspodcast. Once again, head to incogni.com/coolworldspodcast and use the code coolworldspodcast for 60% off. Now, back to the episode.
So this kind of bridges a little bit to the simulation argument, I think, through presumptuous, because I see the presumptuous philosopher as a kind of a bridge here to simulation, because that certainly implies a big world, a big universe with many, many, many, many observers in it. So maybe I just want, we'll use that opportunity to introduce presumptuous philosophers. So that's a kind of a challenge, I suppose, with the self-indication assumption, that if you have two theories, A and B, let's say they're cosmological theories, and one, you know, B produces a trillion observers or a trillion, trillion, trillion observers, and A produces just a trillion observers, and so the ratio of those two by the SIA, you would strongly favor the big universe B. And so the problem is, I suppose, that if cosmologists come along and they start making measurements that hint that A is the correct answer, and then it becomes strong evidence that A is the right answer, it becomes almost impossible for them to overwhelm that prior, because the prior prejudice for the big universe is so overwhelming that you just would ignore the evidence. So this seems like a formidable issue for the self-indication assumption, which, you know, if you get rid of that, it kind of diffuses one of the counterarguments to doomsday. So I'm kind of curious about how serious of a threat or a flaw you consider presumptuous philosopher to be to the self-indication assumption reasoning.
Yeah. So that's the cheribus, right? If the sort of doomsday argument and the Adam and Eve implication is that chilla, and so either way, it looks like you're running into a lot of rough waters with this reasoning, right? Whether you accept or reject the self-indication assumption. And so this is what motivated me in my earlier work.
This is back in the early 2000s, my doctoral dissertation and later the book Anthropic Bias was trying to get clear on the implications of these different assumptions that one might make, because usually up to that it was kind of frustrating. There were all these contributions to the literature that consisted in sort of being aware of one side and then reinventing the self-indication assumption, or being aware of that and then reinventing the other. But you really only confront the difficulty when you look at all of these implications at the same time. That's when you realize how hard it is to find anything that avoids them. Right? If you need to sort of feel the horns of the dilemma pinching you from both sides. And so I tried to propose a theory there with the relativization of the reference class. So the self-sampling assumption in my formulation, if you recall, was you should think of yourself as a randomly selected sample from some reference class of actually existing observers. And it kind of gets too complicated to go into here. But the idea I had that first of all you need to move from observers to observer moments, because just as you can think of yourself as a random observer, you can also think of your current observation as randomly selected from some sequence of observations made by you. And so that's a kind of a strengthening of the self-sampling assumption, right? The strong SSS. So now you're a random observer moment. And then the idea is that when you receive this new information, and in these thought experiments you observe your own birth rank, you're actually changing which observer moment you are, like you now have a different informational content, and so it might be that this new observation moment that arises when you get this new information has a different reference class that it should think of itself as randomly selected from than the same observer's previous observer moment before they got this information. And so the hope then with that is that you can avoid both of these counterintuitive implications: the Adam and Eve implication, doomsday argument on the one hand, and this presumptuous philosopher implication on the other hand, which does require something like the relativization of the reference class.
