David Deutsch - AI, America, Fun, & Bayes

Dwarkesh Patel: Okay, today I'm speaking with David Deutsch. Now this is a conversation that I've been eagerly wanting to have for years, so this is very exciting for me. So, first let's talk about AI. Can you briefly explain why you anticipate that AIs will be no more fundamentally intelligent than humans?

David Deutsch: I suppose you mean AGI. Yes. And by fundamentally intelligent, I suppose you mean capable of all the same types of cognition as humans are in principle.

Dwarkesh Patel: Yes. So that would include doing science and doing art and in principle also falling in love and being good and being evil and all that.

David Deutsch: So the reason is twofold. One half is about computation hardware, and the other is about software. If we take the hardware, we know that our brains are Turing complete bits of hardware, and therefore can exhibit the functionality of running any computable function program for any computable function. When I say any, I don't really mean any, because you and I sitting here having a conversation, we could say we could have any conversation. We can assume that maybe in 100 years time we'll both be dead, and therefore the number of conversations we could have is strictly limited. Also, some conversations depend on speed of computation. If we're going to be solving the traveling salesman problem, there are many traveling salesman problems that we wouldn't be able to solve in the age of the universe. So when I say any, what I mean is that we're not limited in the programs we can run apart from by speed and memory capacity. All limitations on us, hardware limitations, boil down to speed and memory capacity. Both those can be augmented to the level of any other entity that is in the universe. If somebody builds a computer that can think faster than the brain, then we can use that very computer or that very technology to make our thinking go just as fast.

So that's the hardware. As far as explanations go, can we reach the same kind of explanations as any other entity? Usually this is said not in terms of AGIs but in terms of extraterrestrial intelligences, but also it's said about AGIs. What if they are to us as we are to ants and so on? Again, part of that is just hardware, which is easily fixable by adding more hardware. So let's forget about that.

The idea is, are there concepts that we are inherently incapable of comprehending? I think Martin Rees believes this. He thinks that we can comprehend quantum mechanics, apes can't, and maybe extraterrestrials can comprehend something beyond quantum mechanics which we can't comprehend, and no amount of brain add-ons with extra hardware can give us that because they have hardware that is adapted to having these concepts which we haven't. The same kind of thing is said about certain qualia, that maybe we can experience love and an AGI couldn't experience love because it has to do with our hardware, not just memory and speed, but specialized hardware. I think that falls victim to the same argument.

The thing is, this specialized hardware can't be anything except a computer. If there is hardware that is needed for love, let's say that somebody is born without that hardware. That bit of the brain that does love, or that does mathematical insight or whatever, it's just a bit of the brain and it's connected to the rest of the brain in the same way that any other part of the brain is connected to the rest of the brain, namely by neurons passing electrical signals and by chemicals whose concentrations are altered. Therefore, an artificial device that computed which signals were to be sent and which chemicals were to be adjusted could do the same job, and it would be indistinguishable. A person augmented with one of those who couldn't feel love could feel love after that augmentation. So those are the only two relevant things. That's why I think that AGIs and humans have the same range in the sense I've defined.

Dwarkesh Patel: Okay, interesting. I think the software question is more interesting than the hardware one immediately, but I do want to take issue with the idea that the memory and speed of human brains can be arbitrarily and easily expanded. We can get into that later. Let's start with this question. Can all humans explain everything that even the smartest humans can explain? If I took the village idiot and asked him to create the theory of quantum computing, should I anticipate that if he wanted to, he could do this? For frame of reference, about 21 to 24% of Americans on the National Adult Literacy Survey fall in level one, which means they can't even perform basic tasks like identifying the expiry date of a driver's license or totaling a bank deposit slip. Are these humans capable of explaining quantum computing or creating the Deutsch-Joseph algorithm? If they're not capable of doing this, doesn't that mean the theory of universal explainers falls apart?

David Deutsch: There are people who—these tasks that you're talking about are tasks that no ape could do. However, there are humans who are brain damaged to the extent that they can't even do the tasks that an ape can do. There comes a point when installing the program that would be able to read a driver's license or whatever would require augmenting their hardware as well as their software. If some of the people you're talking about—if it's 24% of the population, then it's definitely not hardware. I would say for those people it's definitely software. If it was hardware, then getting them to do this would be a matter of repairing the imperfect hardware. If it's software, it is not just a matter of them wanting to or wanting to be taught. It is a matter of whether the existing software is conceptually ready to do that.

For example, Brett Hall has often said that he would like to speak Mandarin Chinese. He wants to, but he will never be able to speak Mandarin Chinese because he's never going to want it enough to go through the process of acquiring that program. There is nothing about his hardware that prevents him learning Mandarin Chinese, and there's nothing about his software either, except that he doesn't want to go through that process. He does want to learn it, but he doesn't want to go through the process of being programmed with that program. If his circumstances changed, he might well want to. For example, many of my relatives a couple of generations ago were forced to migrate to very alien places where they had to learn languages that they never thought they would ever speak and never wanted to speak, and yet very quickly they did speak those languages. Was it because what they wanted changed? In the big picture, perhaps you could say what they wanted changed.

If your driving license blind people wanted to be educated to read driving licenses in the sense that my ancestors wanted to learn languages, then yes, they could learn that. There is a level of dysfunction below which they couldn't, and I think those are hardware limitations on the borderline between those two. There's not that much difference. It's like the question of could apes be programmed with a fully human intellect? I think the answer to that is yes. Programming them would not require hardware surgery in the sense of repairing a defect, but it would require intricate changes at the neuron level. Transferring the program from a human mind into the ape's mind, I would guess that is possible. Although the ape has far less memory space than humans do and doesn't have certain specialized modules that humans have, neither of those things is a thing that we use to the full anyway. When I'm speaking to you now, there's a lot of knowledge in my brain that I'm not referring to at all, like the fact that I can play the piano or drive a car is not being used in this conversation. So I don't think the fact that we have such a large memory capacity would affect this project, although the project would be highly immoral because you'd be intentionally creating a person inside deficient brain hardware.

Dwarkesh Patel: Suppose it's hardware differences that distinguish different humans in terms of their intelligence. If it were just up to the people who are not even functionally literate—

David Deutsch: Wait, wait. I said that it could only be hardware at the low level, either at the level of brain defects or at the level of using up the whole of our allocation of memory or speed or whatever. Apart from that, it can be hardware.

Dwarkesh Patel: Is software analogous? Is hardware synonymous with genetic influences for you, or can software be genetic too?

David Deutsch: Software can be genetic too, though that doesn't mean it's immutable. It just means it's there at the beginning.

Dwarkesh Patel: The reason I suspect it's not software is because these people also happen to be the same people who—let's suppose it was software and something that they chose to do or if something could change. It's mysterious to me why these people would also choose to accept jobs that have lower pay but are less cognitively demanding, or why they would choose to do worse on academic tests or IQ tests. Why they would choose to do exactly the sort of thing somebody who is less cognitively powerful would do. It seems the more parsimonious explanation is just that they are cognitively less powerful.

David Deutsch: Not at all. Why would someone choose not to go to school, for instance, if they were given the choice and not to have any lessons? There are many reasons why they might choose that, some of them good, some of them bad. Calling some jobs cognitively demanding is already begging the question because you're referring to a choice that people make, which I think is a software choice, as being by definition forced on them by hardware. It's not cognitively deficient, it's that they don't want to do it. If there was a culture that required Brett Hall to be able to speak fluent Mandarin Chinese in order to do a wide range of tasks, and if he didn't know Mandarin Chinese he'd be relegated to low-level tasks, then he would be choosing the low-level tasks rather than the cognitively demanding task. But it's only culture that assigns a hardware interpretation to the difficulty of doing that task.

Dwarkesh Patel: It doesn't seem that arbitrary to say that the kind of jobs you could do sitting down on a laptop require different cognitive abilities, probably more cognition, than the ones you can do on a construction site. If it's not cognition that distinguishes, or if there's not something like intelligence or cognition that is measured by both these literacy tests and by what you're doing at your job, then what is the explanation for why there's such a high correlation between people who are not functionally literate and people who are doing, let's say, programmers? I guarantee you people working at Apple, all of them are above level one on this literacy survey. Why do they just happen to make the same choices? Why is that correlation?

David Deutsch: There are correlations everywhere. Culture is built in order to make use of certain abilities that people have. If you're setting up a company that is going to employ 10,000 employees, then it's best to make the way the company works—for example, to make the signs above the doors or the numbers on the dials all be ones that people in that culture who are highly educated can read. You could, in principle, make each label on each door a different language. There are thousands of human languages, let's say 5,000 languages and 5,000 doors in the company. You could, given the same meaning, make them all different languages. The reason they're all the same language, and what's more not just any old language but a language that many educated people know fluently, is why. You can misinterpret that as saying there is some hardware reason why everybody speaks the same language. No, there isn't. It's a cultural reason.

Dwarkesh Patel: If the culture was different somehow, maybe if there was some other way of communicating ideas, do you think that the people who are currently designated as not functionally literate could be in a position to learn about quantum computing? If they made the right choices, or the choices that could lead to them understanding quantum computing?

David Deutsch: I don't want to evade the question. The answer is yes. But the way you put it again rather begs the question. It's not only language that is like this. It's all knowledge. Just learning—if someone doesn't speak English, quantum computing is a field in which English is the standard language. It used to be German, now it's English. Someone who doesn't know English is at a disadvantage learning about quantum computers, but not only because of their deficiency in language. If they come from a culture in which the culture of physics, mathematics, logic, and so on is equivalent and only the language is different, then if they just learn the language they will find it as easy as anyone else. But if a whole load of things are different—if a person doesn't think in terms of logic but thinks in terms of pride, manliness, fear, and all sorts of concepts that fill the lives of prehistoric or pre-enlightenment people—then to be able to understand quantum computers they would have to learn a lot more than just the language of the civilization. They'd have to learn a range of other features of the civilization. On that basis, the people who can't read driving licenses are similarly in a different culture which they would also have to learn if they are to increase their IQ, their ability to function at a high level in intellectual culture in our civilization. But if they did, they would be able.

Dwarkesh Patel: If it's those kinds of differences, then how do you explain the fact that identical twins separated at birth and adopted by different families tend to have—most of the variance that exists between humans in terms of IQ doesn't exist between identical twins. In fact, the correlation is 0.8, which is the correlation you would have when you took the test on different days depending on how good a day you were having. These are people adopted by families who have different cultures, often in different countries. A hardware theory explains very well why they would have similar scores on IQ tests correlated with literacy and job performance. I don't know how software would explain why being adopted by different families—

David Deutsch: The hardware theory explains it in the sense that it might be true. It doesn't have an explanation beyond that, and nor does the software theory.

Dwarkesh Patel: There are actually differences at the level of brain that are correlated with IQ. Skull size has a 0.3 correlation with IQ. There are a few more like this. They don't explain the entire variance in human intelligence or the entire genetic variance, but we have identified a few actual hardware differences that correlate.

David Deutsch: Suppose on the contrary that the results of these experiments had been different. Suppose the result was that people who are brought up in the same family and differ only in the amount of hair they have or in their appearance in any other way, none of those differences make any difference to their IQ. Only who their parents were makes a difference. Wouldn't that be surprising? Wouldn't it be surprising that there's nothing else correlated with IQ other than who your parents are? Yes. Now, how much correlation should we expect? There are correlations everywhere. There are these things on the internet, joke memes that make a serious point, where they correlate things like how many adventure movies have been made in a given year correlated with how much the GNP per capita. That's a bad example because there's an obvious relation. But you know what I mean—the number of films made by a particular actor against the number of outbreaks of bird flu. Part of being surprised by randomness is the fact that correlations are everywhere. It's not just that correlation isn't causation, it's that correlations are everywhere. It's not a rare event to get a correlation between two things. The more things you ask about, the more you are going to get correlations.

What is surprising is that the things that are correlated are things that you expect to be correlated and measure. When they do these twin studies and measure the IQ, they control for certain things. Identical twins reared together or apart. But there are infinitely more things that they don't control for. It could be that the real determinant of IQ is, for example, how well a child is treated between the ages of three and a half and four and a half, where "well" is defined by something that we don't know yet. You would expect that thing, which we don't know about and nobody has bothered to control for in these experiments, to be correlated with IQ. But unfortunately, that thing is also correlated with whether someone is an identical twin or not. So it's not the identical twinness that is causing the similarity, it's this other thing—say an aspect of appearance or something. If you knew what this thing was and surgically changed a person, you would be able to have the same effect as making an identical twin would have.

Dwarkesh Patel: As you say, in science to explain any phenomenon there's an infinite amount of possible explanations. You have to pick the best one. It could be that there's some unknown trait which is so obvious to different adoptive parents that they can use it as a basis for discrimination or for different treatment. I would assume they don't know what it is, but aren't they using it as a basis to treat kids differently at the age of three?

David Deutsch: Not by consciously identifying it. It would be something like getting the idea that this child is really smart.

Dwarkesh Patel: I'm just trying to show you that it could be something that the parents are not aware of. If you ask parents to list the traits in their children that cause them to behave differently towards their children, they might list 10 traits, but there are another thousand traits that they're not aware of which also affect their behavior.

David Deutsch: We first need an explanation for what this trait is that researchers have not been able to identify, but it's so obvious that even unconsciously parents are able to reliably use it as a way to—

Dwarkesh Patel: It wouldn't have to be obvious at all because parents have a huge amount of information about their children which they are processing in their minds, and most of it they don't know what it is.

Dwarkesh Patel: Okay, let's leave this topic aside for now. Let me bring us to animals. If creativity is something that doesn't exist in increments, or the capacity to create explanations—you can use a simple example, go on YouTube and look up cat opening a door. You'll see a cat develops a theory that applying torque to this handle, to this metal thing, will open a door. It climbs onto a countertop and jumps on top of that door handle. It hasn't seen another cat do it. It hasn't seen a human get on a countertop and try to open the door that way. But it conjectures that this is a way, given its morphology, that it can access the door. That's this theory. The experiment is, will the door open? This seems like a classic cycle of conjecture and refutation. Is this compatible with the cat not having at least some bounded form of creativity?

David Deutsch: I think it's perfectly compatible. Animals are amazing things. Instinctive animal knowledge is designed to make animals easily capable of thriving in environments that they've never seen before. If you go down to the level of detail, animals have never seen the environment before. Maybe a goldfish in a goldfish bowl might have, but when a wolf runs through the forest, it sees a pattern of trees that it has never seen before. It has to create strategies for avoiding each tree and for actually catching the rabbit that it's running after in a way that has never been done before. The way to understand this is because of a vast amount of knowledge that is in the wolf's genes. What kind of knowledge is this? It's not the kind of knowledge that says first turn left, then turn right, then jump. It's an instruction that takes input from the outside and generates a behavior that is relevant to that input. It doesn't involve creativity, but it involves a degree of sophistication in the program that human robotics has not yet reached anywhere near that.

When it sees a wolf of the opposite sex, it may decide to leave the rabbit and go and have sex instead. A program for a robot to locate another robot of the right species and then have sex with it is beyond present day robotics, but it will be done. It clearly does not require creativity because that same program will lead the next wolf to do the same thing in the same circumstances. The fact that the circumstances are ones that it's never seen before and it can still function is a testimony to the incredible sophistication of that program, but it has nothing to do with creativity.

Humans do tasks that require much less programming sophistication than that, such as sitting around a campfire telling each other a scary story about a wolf that almost ate them. Animals can do the wolf running away thing. They can enact a story that's more complicated even than the one the human is telling, but they can't tell a story. Telling a story is a typical creative activity. It's the same kind of activity as forming an explanation. I don't think it's at all surprising that cats can jump on handles. I can easily imagine that the same amazingly sophisticated program that lets it jump on a branch so that the branch will get out of its way will also function in this new environment that's never seen before. But there are all sorts of other things that it can't do.

Dwarkesh Patel: That's definitely true, which was my point: it has a bounded form of creativity. If bounded forms of creativity can exist, then humans could be in one such. I'm having a hard time imagining the ancestral circumstance in which a cat would have genetically gained the knowledge that jumping on a metal rod would get a wooden plank to open and give it access to the other side.

David Deutsch: I thought I just gave an example. I don't know what kind of environment the ancestor of the domestic cat lived in. If it contained undergrowth, then dealing with undergrowth requires some very sophisticated programs, otherwise you will just get stuck somewhere and starve to death. I think a dog, if it gets stuck in a bush, has no program to get out other than just shaking itself about until it gets out. It doesn't have a concept of doing something which temporarily makes matters worse and then allows you to get out. I think dogs can't do that. But it's not because that's a particularly complicated thing. It's just that its programming doesn't have that. An animal's programming easily could have that if it lived in an environment in which that happened a lot.

Dwarkesh Patel: Is your theory of AI compatible with AIs that have narrow objective functions, but functions which, if fulfilled, would give the creator of the AI a lot of power? If I wrote a deep learning program, traded over financial history, and asked it, "Make me a trillion dollars on the stock market," do you think this would be impossible? If you think this would be possible, then it seems like AI is getting somewhere.

David Deutsch: If you want to be powerful, you might do better inventing a weapon or a better mousetrap because it's non-violent. You can invent a paper clip, to use an example often used in this context. If paper clips hadn't been invented, you could invent a paper clip and make a fortune. That's an idea, but it's not an AI because it's not the paper clip that's going out there. It's your idea in the first place that caused the whole value of the paper clip. Similarly, if you invent a dumb arbitrage machine which seeks out complicated trades that are more complicated than anyone else is trying to do, and that makes you a fortune, the thing that made you a fortune was not the arbitrage machine. It was your idea for how to search for arbitrage opportunities that no one else sees. That's what was valuable. That's the usual way of making money in the economy: you have an idea and then you implement it. That it was an AI is beside the point. It could have been a paper clip.

Dwarkesh Patel: The models that are used nowadays are not expert systems like the chess engines of the 90s. They're something like AlphaZero or AlphaGo, which is just a blank neural net. They were able to let it win at Go. If such a neural network that was kind of blank, and you just arbitrarily throw financial history at it, wouldn't it be fair to say that the AI actually figured out what the right trades were, even though it's not a general intelligence?

David Deutsch: I think it's possible in chess, but not in the economy. The value in the economy is being created by creativity. Arbitrage is one thing that can sort of skim value off the top by taking opportunities that were too expensive for other people to take. You can make a lot of money that way if you have a good idea about how to do it. But most of the value in the economy is created by the creation of knowledge. Somebody has the idea that a smartphone would be good to have, even though most people think that's not going to work. That idea cannot be anticipated by anything less than an AGI. An AGI could have that idea, but no AI could.

Dwarkesh Patel: There's definitely other topics I want to get to. Let's talk about virtual reality. In *The Fabric of Reality*, you discuss the possibility that virtual reality generators could plug in directly into our nervous system and give us sense data that way. As you might know, many meditators, people like Sam Harris, speak of both thoughts and senses as intrusions into consciousness that have a sort of similar quality. They can be welcome intrusions, but they are both things that come into consciousness. Do you think that a virtual reality generator could also place thoughts as well as sense data into the mind?

Dwarkesh Patel: Yes, but that's only because I think that this model is wrong. It's basically the Cartesian theatre, as Daniel Dennett puts it, with the stage cleared of all the characters. So that's pure consciousness without content, as Sam Harris envisages it. But I think that all that's happening there is that you are conscious of this theatre and you're envisaging it as having certain properties, which by the way it doesn't have, but that doesn't matter. We can imagine lots of things that don't happen.

One can interpret one's thoughts about this empty stage as being thoughts about nothing. One can interpret the actual hardware of the stage that one is imagining as being pure contentless consciousness. But it's not. It has the content of a stage or a space or however you want to envisage it.

David Deutsch: Okay. And then let's talk about the Turing principle. This is a term you coined. It's otherwise been called the Church-Turing-Deutsch principle. Would this principle imply that you could simulate the whole of the universe, for example, in a compact efficient computer that was smaller than the universe itself? Or is it constrained to physical processes of a certain size?

David Deutsch: Again, no, it couldn't simulate the whole universe. That would be an example of a task where it was computationally able to do it, but it wouldn't have enough memory or time. The more memory and time you gave it, the more closely it could simulate the whole universe, but it couldn't ever simulate the whole universe or anything near the whole universe, probably because if you wanted to simulate itself as well, there are logical reasons why there are limits to that. But even if you want to simulate the whole universe apart from itself, just the sheer size of the universe makes that impossible. Even if we discovered ways of encoding information extremely densely, like some people have said maybe quantum gravity would allow a totally amazing density of information, it still couldn't simulate the universe. Because of the universality of the laws of physics, that would mean the rest of the universe also was that complex, because quantum gravity applies to the whole rest of the universe as well.

But I think it's significant to separate being limited by the available time and memory from being limited by computational capacity. It's only when you separate those that you realize what computational universality is. I think that universality, like Turing or quantum universality, is the most important thing in the theory of computation, because computation doesn't even make sense unless you have a concept of a universal computer.

Dwarkesh Patel: What could falsify your theory that all interesting problems are soluble? I ask this because, as I'm sure you know, there are people who have tried offering explanations for why certain problems or questions—like, why is there something rather than nothing? Or how could mere physical interactions explain consciousness?—they've offered explanations for why these problems are in principle insoluble. I'm not convinced they're right, but do you have a strong reason for in principle believing that they're wrong?

David Deutsch: No. This is a philosophical theory and could not be proved wrong by experiment. However, I think I have a good argument for why they aren't, namely that each individual case of this is a bad explanation.

Let's say that some people say, for example, that simulating a human brain is impossible. I can't prove that it's possible. Nobody can prove it's possible until they actually do it, or unless they have a design for it which they prove will work. Pending that, there is no way of proving that it's not true that this is a fundamental limitation. But the trouble with that idea that it is a fundamental limitation is that it could be applied to anything. For example, it could be applied to the theory that you have recently, just a minute ago, been replaced by a humanoid robot which is going to say for the next few minutes just a prearranged set of things, and you're no longer a person.

Dwarkesh Patel: I can't believe you figured it out.

David Deutsch: Yeah, well, that's the first thing you'd say. So there is no way to refute that by experiment, short of actually doing it, short of actually talking to you and so on. So it's the same with all these other things. In order for it to make sense to have a theory that something is impossible, you have to have an explanation for why it is impossible.

We know that, for example, almost all mathematical propositions are undecidable. That's not because somebody has said, "Oh, maybe we can't decide everything because thinking we could decide everything is hubris." That's not an argument. You need an actual functional argument to prove that that is so. And then, that being a functional argument in which the steps of the argument make sense and relate to other things and so on, you can then say, "Well, what does this actually mean? Does this mean that maybe we can never understand the laws of physics?" It doesn't. If the laws of physics included an undecidable function, then we would simply write f of x, and f of x is an undecidable function. We couldn't evaluate f of x. It would limit our ability to make predictions, but then lots of our ability to make predictions is totally limited anyway. It would not affect our ability to understand the properties of the function f, and therefore the properties of the physical world.

Dwarkesh Patel: Is a system of government like America's, which has distributed powers and checks and balances, incompatible with Popper's criterion? The reason I ask is the last administration had a theory that if you build a wall, there will be positive consequences. That theory could have been tested, and then the person could have been evaluated on whether that theory succeeded. But because our system of government has distributed powers, Congress opposed the testing of that theory. So it was never tested. If the American government wanted to fulfill Popper's criterion, would we need to give the president more power, for example?

David Deutsch: It's not as simple as that. I agree that this is a big defect in the American system of government. No country has a system of government that perfectly fulfills Popper's criterion. We can always improve. I think the British one is actually the best in the world, and it's far from optimal. Making a single change like that is not going to be the answer. The constitution of a polity is a very complicated thing, much of which is inexplicit.

The American founding fathers realized they had a tremendous problem. What they wanted to do, what they thought of themselves as doing, was to implement the British Constitution. In fact, they thought they were the defenders of the British Constitution and that the British king had violated it and was bringing it down. They wanted to retain it. The trouble is that in order to gain the independence to do this, they had to get rid of the king, and then they wondered whether they should get an alternative king. Whichever way they did it, there were problems.

The way they decided to do it, I think, made for a system that was inherently much worse than the one they were replacing, but they had no choice. If they wanted to get rid of a king, they had to have a different system for having a head of state. They wanted to be democratic. That meant that the president had a legitimacy in legislation that the king never had. The king used to have it in medieval times, but by the time of the Enlightenment, the king no longer had full legitimacy to legislate. So they had to implement a system where his seizing power was prevented by something other than tradition. So they instituted these checks and balances.

The whole thing they instituted was immensely sophisticated. It's an amazing intellectual achievement. That it works as well as it does is something of a miracle. But the inherent flaws are there. One of them is this: the fact that there are checks and balances means that responsibility is dissipated, and nobody is ever to blame for anything in the American system, which is terrible. In the British system, blame is absolutely focused. Everything is sacrificed to the end of focusing blame and responsibility down to the government, past the law courts, past the parliament, right to the government. That's where it's all focused. There are no systems that do that better.

But as you well know, the British system also has flaws. We recently saw with the sequence of events with the Brexit referendum, and then Parliament balking at implementing some laws that it didn't agree with, and then that being referred to the courts. There were the courts and the parliament and the government and the prime minister all blaming each other. There was a sort of mini constitutional crisis which could only be resolved by having an election and then having a majority government. That's how it usually is in Britain, although we have been unlucky several times recently in not having a majority government.

Dwarkesh Patel: This could be wrong, but it seems to me that in an expanding universe, there will be a finite amount of total matter that will ever exist in our light cone. There's a limit. That means there's a limit on the amount of computation that this matter can execute, the amount of energy it can provide, perhaps even the amount of economic value it can sustain. It would be weird if the GDP per atom could be arbitrarily large. Does this impose some sort of limit on your concept of the beginning of infinity?

David Deutsch: What you've just recounted is a cosmological theory. The universe could be like that, but we know very little about cosmology. We know very little about the universe in the large. Theories of cosmology are changing on a timescale of about a decade. So it doesn't make all that much sense to speculate about what the ultimate asymptotic form of the cosmological theories will be.

At the same time, we don't have a good idea about the asymptotic form of very small things. We know that our conception of physical processes must break down somehow at the level of quantum gravity, like 10 to the minus 42 seconds and that kind of thing. But we have no idea what happens below that. Some people say it's got to stop below that, but there's no argument for that at all. It's just that we don't know what happens beyond that.

What happens beyond that may be a finite limit. Similarly, what happens on a large scale may impose a finite limit, in which case computation is bounded by a finite limit imposed by the cosmological initial conditions of this universe. That is still different from it being imposed by inherent hardware limitations. For example, if there's a finite amount of GMP available in the distant future, it's still up to us whether we spend that on mathematics or music or political systems or any of the thousands of even more worthwhile things that have yet to be invented. So it's up to us which ideas we fill the 10 to the 10 to the 10 to the 10 bits with.

My guess is that there are no such limits, but my worldview is not affected by whether there are such limits. As I said, it's still up to us what to fill them with. And if we get chopped off at some point in the future, then everything will have been worthwhile up to then.

Dwarkesh Patel: The way I understand your concept of the beginning of infinity, it seems to me that the more knowledge we gain, the more knowledge we're in a position to gain, so there should be an exponential growth of knowledge. But if we look at the last 50 years, it seems that there's been a slowdown or decrease in research productivity, economic growth, productivity growth. This seems compatible with the story that there's a limited amount of fruit on the tree, that we pick the low hanging fruit, and now there's less and less fruit and harder and harder fruit to pick. Eventually the orchard will be empty. Do you have an alternative explanation for what's going on in the last 50 years?

David Deutsch: Yes, I think it's very simple. There are sociological factors in academic life which have stultified the culture. Not totally and not everywhere, but that has been a tendency in what has happened. It has resulted in a loss of productivity in many sectors, in many ways, but not in every sector, not in every way.

For example, I think there was a stultification in theoretical physics starting in, let's say, the 1920s, and it still hasn't fully dissipated. If it wasn't for that, quantum computers would have been invented in the 1930s and built in the 1960s. That is just an accidental fact, but it just goes to show that there are no guarantees. The fact that our horizons are unlimited does not guarantee that we will get anywhere, that we won't start declining tomorrow. I don't think we are currently declining. I think these declines that we see are parochial effects caused by specific mistakes that have been made and which can be undone.

Dwarkesh Patel: I want to ask you a question about Bayesianism versus Popperianism. One reason why people prefer Bayes is because there seems to be a way of describing changes in epistemic status when the relative status of a theory hasn't changed. I'll give you an example. Currently, the many worlds interpretation is the best way to explain quantum mechanics. But suppose in the future we were able to build an AGI on a quantum computer and be able to design some clever interference experiment, as you suggest, to have it be able to report back being in a superposition across many worlds. It seems that even though many worlds remains the best or the only explanation, somehow its epistemic status has changed as a result of the experiment. In Bayesian terms, you could say the credence of this theory has increased. How would you describe these sorts of changes in a Popperian view?

David Deutsch: What has happened there is that at the moment we have only one explanation that can't be immediately knocked down. If we did that thought experiment, we might well decide that this will provide the ammunition to knock down even ideas for alternative explanations that have not been thought of yet. Obviously it wouldn't be enough to knock down every possible explanation, because for a start we know that quantum theory is false. We don't know for sure that the next theory will have many worlds in it. I think it will, but we can't prove anything like that.

I would replace the idea of increased credence with a theory that the experiment will provide a quiver full of arrows, or a repertoire of arguments that goes beyond the known arguments, the known bad arguments, and will reach into other types of arguments. The reason I would say that is that some of the existing misconceptions about quantum theory reside in misconceptions about the methodology of science.

I've written a paper about what I think is the right methodology of science where that doesn't apply, but many physicists and many philosophers would disagree with that, and they would advocate a methodology of science that's more based on empiricism. Of course I think that empiricism is a mistake and can be knocked down in its own terms, but not everybody thinks that. Once we have an experiment such as my thought experiment, if that was actually done, then people could not use their arguments based on a fallacious idea of empiricism because their theory would have been refuted even by the standards of empiricism, which shouldn't have been needed in the first place. That's why I think the repertoire of arguments would become more powerful if that experiment were done successfully.

Dwarkesh Patel: How far do you take the principle that open-ended scientific progress is the best way to deal with existential dangers? To give one example, many people have suggested something like gain-of-function research. It's conceivable that it could lead to more knowledge in how to stop dangerous pathogens. But in Bayesian terms, you could say it seems even more likely that it can or has led to the spread of a man-made pathogen that would not have otherwise been naturally developed. Would your belief in open-ended scientific progress allow us to say, "Let's stop gain-of-function research"?

David Deutsch: No, it wouldn't allow us to say, "Let's stop it." It might make it reasonable to say, "Let us do research into how to make laboratories more secure before we do gain-of-function research." It's really part of the same thing. It's like saying, "Let's do research into how to make the plastic hoses through which the reagents pass more impermeable before we actually do the experiments with the reagents." It's all part of the same experiment. I wouldn't want to stop something just because new knowledge might be discovered. That's the no-no in my view. But which knowledge we need to discover first—that's the problem of scheduling, which is a non-trivial part of any research and of any learning.

Dwarkesh Patel: Would it be conceivable for you to say that until we figure out how to make sure these laboratories are safe to a certain standard, we will stop the research as it exists now? And meanwhile, we will focus on doing the other kind of research so gain-of-function can restart, but until then it's not allowed.

David Deutsch: Yes, in principle, that would be reasonable. I don't know enough about the actual situation to have a view. I don't know how these labs work. I don't know what the precautions consist of. When I hear people talking about, for example, lab leak, I think, well, most likely lab leak is that one of the people who works there walks out of the front door. The leak is not a leak from the lab to the outside. The leak is from the test tube to the person, and then from the person walking out the door. I don't know enough about what these proportions are or what the state of the art is to know to what extent the risk is actually minimized.

It could be that the culture of these labs is not good enough, in which case it would be part of the next experiment to improve the culture in the labs. But I am very suspicious of saying that all labs have to stop and meet a criterion, because I suspect that the stopping wouldn't be necessary and the criterion wouldn't be appropriate. Again, which criterion to use depends on the actual research being done.

Dwarkesh Patel: When I had Tyler Cowen on my podcast, I asked him why he thinks human civilization is only going to be around for 700 more years. I gave him your rebuttal, or what I understand of your rebuttal, that creative optimistic societies will innovate ways of safety technologies faster than totalitarian static societies can innovate destructive technologies. He responded, "Maybe, but the cost of destruction is just so much lower than the cost of building. That trend has been going on for a while now. What happens when a nuke costs $60,000? Or what happens if there's a mistake like the kinds we saw many times over in the Cold War?" How would you respond to that?

David Deutsch: First of all, I think we've been getting safer and safer throughout the entire history of civilization. There were these plagues that wiped out a third of the population of the world, or half, and it could have been 99% or 100%. We went through some kind of bottleneck 70,000 years ago, I understand, which they can tell from genetics. All our cousin species have been wiped out. We were much less safe then than now.

Also, if a 10-kilometer asteroid had been on target with the earth at any time in the past two million years or whatever it is in the history of the genus Homo, that would have been the end of it. Whereas now it'll just mean higher taxation for a while. That's how amazingly safer we are now.

I would never say that it's impossible that we'll destroy ourselves. That would be contrary to the universality of the human mind. We can make wrong choices. We can make so many wrong choices that we'll destroy ourselves. On the other hand, the atomic bomb accident sort of thing would have had no zero chance of destroying civilization. All they would have done is cause a vast amount of suffering. I don't think we have the technology to end civilization even if we wanted to. I think all we would do if we just deliberately unleashed hell all over the world is cause a vast amount of suffering, but there would be survivors, and they would resolve never to do that again.

I don't think we're even able to do it accidentally, let alone that we would do it. As for the bad guys, I think we are doing the wrong thing largely in regard to both external and internal threats, but I don't think we're doing the wrong thing to an existential risk level. Over the next 700 years or whatever it is, I don't want to prophesy because I don't know most of the advances that are going to be made in that time. But I see no reason why, if we are solving problems, we won't solve problems.

I don't think it's like Nick Bostrom's jar with white balls and one black ball, where you take out a white ball and a white ball and a white ball, and then you hit the black ball and that's the end of you. I don't think it's like that because every white ball you take out reduces the number of black balls in the jar. Again, I'm not saying that's the law of nature. It could be that the very next ball we take out will be the black one that'll be the end of us. It could be. But I think all arguments that it will be are fallacious.

Dwarkesh Patel: I do want to talk about the fun criterion. Is your definition of fun different from how other people define other positive emotions like eudaimonia or well-being or satisfaction? Is fun a different emotion?

David Deutsch: I don't think it's an emotion. All these things are not very well defined. They can't possibly be very well defined until we have a satisfactory theory of qualia at least, and probably a more satisfactory theory of creativity, how creativity works, and so on.

I think that the choice of the word "fun" for the thing that I explain more precisely, but still not very precisely, as a creation of knowledge where the different kinds of knowledge—inexplicit, unconscious, conscious, explicit—are all in harmony with each other. I think that is actually the only way in which the everyday usage of the word "fun" differs from that. "Fun" is considered frivolous, or seeking fun is considered as seeking frivolity. But I think that isn't so much a different use of the word. It's just a different pejorative theory about whether this is a good or a bad thing.

Nevertheless, I can't define it precisely. The important thing is that there is a thing which has this property of fun that you can't compulsorily enact it. In some views, "no pain, no gain," you can find out mechanically whether the thing is causing pain and whether it's doing it according to the theory that says you will have gain if you have that pain. So that can all be done mechanically, and therefore it is subject to the criticism. Another way of looking at the fun theory is that it's a mode of criticism. It's subject to the criticism that "this isn't fun," meaning this is making and privileging one kind of knowledge arbitrarily over another, rather than being rational and letting content decide.

Dwarkesh Patel: Is this placing a limitation on universal explainers then, if they can't create some sort of theory about why a thing could or should be fun, why anything could be fun? It seems to me that sometimes we actually can make things fun that aren't. For example, take exercise. "No pain, no gain." When you first go, it's not fun, but once you start going, you understand the mechanics, you develop a theory for why it can and should be fun.

David Deutsch: Yes, that's quite a good example because there you see that fun cannot be defined as the absence of pain. You can be having fun while experiencing physical pain, and that physical pain is not sparking suffering but joy. However, there is such a thing as physical pain not sparking joy, as Marie Kondo would say. That's important because if you are dogmatically or uncritically implementing in your life a theory of the good that involves pain and which excludes the criticism that maybe this can't be fun, or maybe this isn't yet fun, or maybe I should make it fun and if I can't that's a reason to stop—if all those things are excluded because by definition the thing is good and your pain and suffering doesn't matter, then that opens the door not only to suffering but to stasis. You won't be able to get to a better theory.

Dwarkesh Patel: Why is fun central to this instead of another emotion? For example, Aristotle thought that a sort of widely defined sense of happiness is what should be the goal of our endeavors. Why fun instead of something like that?

Dwarkesh Patel: Well, that's defining it vaguely enough so that what you said might very well be fun. The point is, the underlying thing is as far as, you know, going one level below, we're really to understand that we'd need to go about seven levels below that, which we can't do yet. But the important thing is that there are several kinds of knowledge in our brains. And the one that is written down in the exercise book that says you should do this number of reps and you should power through this, and it doesn't matter if you feel that and so on, that's an explicit theory. And it contains some knowledge, but it also contains error. That's like all our knowledge is like that. We also have other knowledge which is contained in our biology, it's contained in our genes. We have knowledge that is inexplicit, like our knowledge of grammar is my always my favorite example, because we know why certain sentences are acceptable and why they're unacceptable, but we can't state explicitly or in every case why it isn't or why it is.

David Deutsch: So there's explicit and inexplicit knowledge, there's conscious and unconscious knowledge. All those are bits of program in the brain. They are ideas. They are bits of knowledge in this if you define knowledge as information with causal power. They are all information with causal power. They all contain truth and they all contain error. And it's always a mistake to shield one of them from criticism or replacement. Not doing that is what I call the fun criterion. Now you might say that's a bad name, but it's the best I can find.

Dwarkesh Patel: So why would creating an AGI through evolution necessarily entail suffering? Because the way I see it, or it seems to me, your theory is that you need to be a general intelligence in order to feel suffering. But by the point an evolved simulated being is a general intelligence, we can just stop the simulation. So where's the suffering coming from?

David Deutsch: Okay, so the kind of simulation by evolution that I'm thinking of—there may be several kinds—but the kind that I'm thinking of, and which I said would be the greatest crime in history, is the kind that just simulates the actual evolution of humans from pre-humans that weren't people. So you have a population of non-people, which in this simulation would be some kind of NPCs, and then they would just evolve. We don't know what the criterion would be. We just have an artificial universe which simulated the surface of the earth, and they'd be walking around, and some of them might or might not become people.

Now, the thing is, when you're part of the way there, what is happening is that the only way that I can imagine the evolution of personhood or explanatory creativity happened was that the hardware needed for it was first needed for something else. I have proposed that it was needed to transmit memes. So there would be people who were transmitting memes creatively, but they were running out of resources. Before it managed to increase their stock of means, in every generation there was a stock of means that was being passed down to the next generation. Once they got beyond a certain complexity, they had to be passed down by the use of creativity by the recipient.

So there may well have been a time—and as I say, I can't think of any other way it could have been—where there was genuine creativity being used, but it ran out of resources very quickly. But not so quickly that it didn't increase the meme bandwidth. Then in the next generation, there was more meme bandwidth. After a certain number of generations, there would have been some opportunity to use this firmware for something other than just trying blindly transmitting memes, or rather creatively transmitting memes, but they were blind memes.

So in that time, it would have been very unpleasant to be alive. It was already very unpleasant to be alive when we did have enough resources to think as well as do the memes. But I don't think there would have been a moment at which you would say, "Yes, now the suffering begins to matter because it's not just blind memes." I think the people were already suffering at the time when they were blindly transmitting memes because they were using genuine creativity. They were just not using it to any good effect.

Dwarkesh Patel: Gotcha. Would being in the experience machine be compatible with the fun criterion? So you're not aware that you're in virtual reality, but you're still doing the things that would make you have fun—in fact, more so than in the real world. Would you be tempted to get into the experience machine? Would it be compatible with the fun criteria? I guess there are different questions, but I'm not sure what the experience machine is.

David Deutsch: Oh, sorry. Just so I understand—is it just a virtual reality world in which things work better than in the real world or something?

Dwarkesh Patel: Yeah, so it's a thought experiment by Robert Nozick. The idea is that you would enter this world, but you would forget that you're in virtual reality. So the world would be perfect in every possible way that it could be perfect, or not perfect, but it would be better in every possible way it could be better. But you would think the relationships you have here are real, the knowledge you're discovering here is novel, and so on. Would you be tempted to enter such a world?

David Deutsch: Well, no, I certainly wouldn't want to enter any world which involves erasing the memory that I have come from this world. Related to that is the fact that the laws of physics in this virtual world couldn't be the true ones because the true ones aren't yet known. So I'd be in a world in which I was trying to learn laws of physics which aren't the actual laws, and they would have been designed by somebody for some purpose to manipulate me, as it were. Maybe it would be designed to be a puzzle that would take 50 years to solve, but it would have to be by definition a finite puzzle, and it wouldn't be the actual world. And meanwhile, in the actual world, things are going wrong and I don't know about this. And eventually they go so wrong that my computer runs out of power. And then where will I be?

Dwarkesh Patel: The final question I always like to ask people I interview is what advice would you give to young people? So somebody in their 20s, is there something that you would like to—some advice you would give them?

David Deutsch: Well, I try very hard not to give advice because it's not a good relationship to be in with somebody to give them advice. I can have opinions about things. So for example, I may have an opinion that it's dangerous to condition your short-term goals by reference to some long-term goal. And I have what I think is a good epistemological reason for that: namely, that if your short-term goals are subordinate to your long-term goal, then if your long-term goal is wrong or deficient in some way, you won't find out until you're dead. So it's a bad idea because it is subordinating the things that you could error correct now, or in six months time, or in a year's time, to something that you could only error correct on a 50-year time scale, and then it'll be too late. So I'm suspicious of advice of the form "set your goal," and even more suspicious of "make your goal be so and so."

Dwarkesh Patel: Huh, interesting. But why do you think the relationship between advice-giver and advice-receiver is dangerous?

David Deutsch: Oh, well, because it's one of authority again. I tried to make this example of "advice" that I just gave—I tried to make it non-authoritative. I just gave an argument for why certain other arguments are bad. But if it's advice of the form "a healthy mind in a healthy body" or "don't drink coffee before 12 o'clock" or something like that, it's a non-argument. If I have an argument, I can give the argument and not tell the person what to do. Who knows what somebody might do with an argument? They might change it to a better argument which actually implies different behavior. I can contribute to the world arguments. I make arguments as best I can. I don't claim that they are privileged over other arguments. I just put them out because I think that this argument works. And I expect other people not to think that they work. I mean, we've just done this in this very podcast. I put out an argument about AI and that kind of thing, and you criticize it. Now, if I was in the position of making that argument and saying that, therefore, you should do so-and-so, that's a relationship of authority, which I think is immoral to have.

Dwarkesh Patel: Well, David, thanks so much for coming on the podcast. And thanks so much for giving me so much of your time. Fascinating.

David Deutsch: Thank you for inviting me. Thank you.