Episode 130: The “Pseudo Deutsch Theory of Knowledge”
- Links to this episode: Spotify / Apple Podcasts
- This transcript was generated with AI using PodcastTranscriptor.
- Unofficial AI-generated transcripts. These may contain mistakes. Please check against the actual podcast.
- Speakers are denoted as color names.
Transcript
[00:00:00] Blue: Hello out there! This week on the Theory of Anything podcast, Bruce looks at what makes sense and doesn’t about popular theories on knowledge promoted by crit rats, i.e. critical rationalists on X. He looks at how much these theories align and don’t align with what David Deutch has actually said about knowledge. I enjoyed listening to Bruce here and I hope someone else out there does too.
[00:00:36] Red: Welcome to the Theory of Anything podcast, hey Peter. Hey Bruce, how you doing today? Good! So today, first let me just say that we’ve been talking about concepts versus theories and this is one of the things that we’ve kind of had an ongoing discussion about in the last few podcasts and we talked about how Popper thought theories were a hundred times more valuable than concepts. Now, I sort of agreed with Popper on that if we’re talking about his critical method, I completely agree with him, but I think that that ignores how important concepts and ideas are to how we make conjectures and I gave examples of how I feel Popper’s wrong if you take him quite literally as meaning that theories are a hundred times more valuable than concepts.
[00:01:23] Blue: Okay.
[00:01:24] Red: Now, we’re going to do, this is a little bit related, but it’s kind of a bit tangent that I need to go to before I can go back and finish talking about concepts versus theories.
[00:01:34] Blue: Okay, so just for my own sake and maybe some listeners out there too, I always try, you know, when I’m in my high school classroom, I always try to demonstrate epistemic humility to my students when I don’t know something, I look it up, I admit I don’t know it and we try to figure that out together. Can you just give me a one -sentence summary of what Popper says about concepts versus theories and then what Hofstetter says? Sure.
[00:02:04] Red: So Popper believes that concepts are purely instrumental. I can’t do it in one sentence. I can do it in one.
[00:02:10] Blue: Okay, that might be asking too much. I say one sentence. I know I’m going to give five, that’s as that goes.
[00:02:17] Red: Popper believes concepts are purely instrumental similar to how letters are in a word. They play a role. They’re important, but they are not what we’re really seeking. We’re really seeking theories, which he understands to be assertions about reality that can be either true or false. Okay,
[00:02:34] Blue: that’s good. That’s good. Okay, that’s clicking with me. All right. Okay,
[00:02:37] Red: by comparison, concepts cannot be falsified whereas theories can. Particularly, I’ll get to Hofstetter’s in just a second. So because of this, he thinks that theories are way more important than concepts. He thinks philosophers have spent way too much time arguing over concepts, arguing over language. He thinks boils down to arguments over language. And so he feels like it’s a giant waste of time and he does put some nuance in there where he says, well, sometimes the concept’s a theory and sometimes it contains a theory. Like he has these little exception cases. But basically, his argument is you should care about what’s true or not and then if you do, you’re talking about theories, not concepts.
[00:03:22] Blue: Oh, no, that’s a great summary. So arguing about concepts is something similar to arguing about words.
[00:03:29] Red: In Popper’s mind, yes. So I make a distinction here different than Popper does, where I make a very clear distinction between concepts and words. But even I would have to admit they’re like strongly related. And so that Popper sort of has a point. I feel like there’s a lot of exception cases. Like Popper’s kind of right. But there’s a bunch of exception cases he seems to have missed in my opinion. So Hofstetter argues that concepts, I don’t think he uses the term concepts, that fuzzy categories, I would call them concepts. Human ideas are creative in that they leap across these analogies that we make. We see similarities in a circumstance and we create a category in our mind, give it a name, maybe, maybe not. And then that allows us to reapply knowledge from some other area into a new area, but somewhat imperfectly. And then it still has to be tested. So Hofstetter’s idea then contrasts with Popper’s in that he thinks, what I’m calling concepts, these fuzzy categories that Hofstetter refers to, that are analogies. He thinks that that is the full basis for human creativity. Now obviously, if it is actually the full basis for human creativity, or even just one of the bases for human creativity, then Popper has misjudged the importance of concepts. In terms of the critical method, he would be correct. But in terms of the overall importance of concepts, concepts would be way more important in terms of how we are creative, which would be how we come up with conjectures. Now, Hofstetter does agree with Popper in an important way. The idea of an analogy, a concept as an analogy, or a fuzzy category, you can’t falsify them. They’re extensible.
[00:05:29] Red: Okay, so Popper and Hofstetter would agree that this fuzzy ideas that we have, that we would think of as a concept, that those can’t be falsified. So Popper’s correct that there’s a problem with concepts that they can’t be falsified, except that it’s not a problem, says Hofstetter. That’s actually the fact you can’t falsify them. That’s exactly why you can extend them out and then abstract them and come up with entirely new ideas. Okay.
[00:05:58] Blue: Okay, wait, I think I’m getting my mind around this. So to Hofstetter, the creativity is how we, human creativity is geared towards coming up with these fuzzy categories or concepts, and then it’s really just the last step is where you move it into a theory, and that’s when it can be falsified or not.
[00:06:21] Red: That’s right.
[00:06:22] Blue: Is that right?
[00:06:23] Red: That’s correct. Okay. Now, I have my own nuances on top of that for one of the things I claimed was, is that there’s not a great difference between concepts and theories, that you can treat a concept as a theory, you can treat a theory as a concept, you can move between them fairly fluidly.
[00:06:38] Blue: And is it possible that when we get the AI Hofstetter and the AI Carl Popper on the podcast to argue? I mean, I know Hofstetter is still alive, but he,
[00:06:47] Red: but we can’t get him on the podcast. He declined the request
[00:06:50] Blue: respectfully, but he doesn’t go on podcasts is what he said. He was really nice. But when we get the AI version, they might just agree on this and it might consider it in much of an issue. Is that possible?
[00:07:05] Red: It’s possible. I’m trying to find common ground between them.
[00:07:09] Blue: Okay.
[00:07:09] Red: So, okay, that’s
[00:07:12] Blue: good. No, that was helpful.
[00:07:14] Red: Okay. Now, throughout the podcast, especially in the last few podcasts, but really going way back, we’ve talked about something that I’ve called the pseudo -Deutsch theory of knowledge, which is a popular crit -rat interpretation of Deutsch’s theory of knowledge, especially around his two sources hypothesis. Okay. Now, in past podcasts, recently, particularly in episode 123, I compared the pseudo -Deutsch theory of knowledge to Campbell’s theory of evolutionary epistemology. And I mentioned that I consider the pseudo -Deutsch theory of knowledge in its current form, at least, to be a bad explanation. Now, I’m going to be referencing the pseudo -Deutsch theory of knowledge multiple times in the future. And I’ve never had a single podcast. I’ve criticized it all throughout the podcast, but because it’s a term I made up myself and so nobody knows what I’m talking about, I really need a single podcast where I can say, this podcast is about the pseudo -Deutsch theory of knowledge. I’m going to explain what it is. I’m going to explain how it gets used. I’m going to explain why I don’t like it as an explanation. Okay. So that’s what we’re going to do in this podcast as an aside before we continue with the concepts versus theory discussion. Okay. So that it’s easy for me to then say, the pseudo -Deutsch theory of knowledge, see episode blah, blah, blah. And when people, I’m criticizing it out on Twitter or something, I can say, you know what? I don’t buy the pseudo -Deutsch theory of knowledge. Here’s the episode where you can look up what I mean by that. Okay. So that’s what we’re going to do today. And it’s a little bit of an aside from the main discussion. Okay.
[00:08:50] Red: So I would like to pull my critiques of the two sources hypothesis and particularly the pseudo -Deutsch theory of knowledge, go over it in one podcast of why I believe that it is a bad explanation. And that way I’ve got something to point to, and it’s not just spread across eight or nine different episodes or something like that. Okay. So what is first the two sources hypothesis? Because even that’s my term. When I say it, people know what I’m talking about though. I try to pick terms that are like so intuitively obvious what I’m talking about that I don’t have to explain them. You never really fully succeed though. But when I say David Deutch’s two sources hypothesis, almost everybody knows what I mean, even though I’ve never defined it. Okay. It is David Deutch’s belief that there are only two sources of knowledge, biological evolution and human ideas. Now this idea that there’s only two sources does come from David Deutch. So this part is absolutely from David Deutch and he’s popularized it. And it’s also, I’ve argued on this podcast, known to be incorrect, even under his own constructor theory of knowledge. Now I’ve criticized this in detail in episodes 75 to 80 if you’re curious about what I’ve said about that. The two sources hypothesis is defended by a strategy that I call the pseudo Deutch theory of knowledge, which may or may not have come from David Deutch’s book, The Beginning of Infinity, depending on who you ask. I called this idea the pseudo -Deutch theory of knowledge because it is usually by its defenders attributed to David Deutch. They say, David Deutch says this and they kind of reference him. Okay.
[00:10:40] Red: For better or for worse, the defenders of this theory that I’m about to explain sincerely believe they are getting this straight from David Deutch’s book, The Beginning of Infinity. However, when I try to discuss it, I have found that there’s a different group of crit -rats who argue that David Deutch never actually argued this. And he never held or intended the theory I’m about to explain. Okay. So by calling it pseudo -Deutch, I’m taking no stance as to if it is or is not the theory that David Deutch intended in Beginning of Infinity. Okay. But it is a real theory, even if it did not come from David Deutch, even if it’s just a misunderstanding of David Deutch that got held by many crit -rats, it is now this common theory that you will come across within the crit -rat community that isn’t a good explanation and it needs to be criticized. Okay. So even if it wasn’t what Deutch intended, it still exists as a theory today and it still needs to be criticized. So I’m going to criticize that theory regardless of who is responsible for it or even if no one is responsible for it. Okay. So it may just be an evolved theory that nobody made up. Okay. So pseudo -Deutch is a series of arguments used to argue for the two sources hypothesis, which is the idea that only humans and ideas in biological evolution create knowledge, which for sure that part did come from David Deutch. That’s why this is a little bit confusing. But pseudo -Deutch goes beyond the two sources hypothesis in some important ways. And that’s what we’re going to be talking about.
[00:12:13] Red: Here is a steel man version of what I’m calling the pseudo -Deutch theory of knowledge in a nutshell so that you will, if you’re listening to this podcast, you have no idea what I’m talking about so far, you’re about to immediately recognize this argument. If you spend any amount of time in the crit -rat community, the argument’s about to lay out. You have absolutely heard them over and over again. Okay. So consider the idea of a robot that via a genetic algorithm learns to walk. Example that comes from beginning infinity directly. Did the genetic algorithm create the knowledge of how to walk or did the human programmer create all the knowledge? Now defenders of pseudo -Deutch point out how much knowledge pre -existed before we ran the genetic programming algorithm and point out how all that knowledge came from the human. So for example, and some of these examples come directly from beginning of infinity, Deutch points out that a human created a language with implicit knowledge about walking and that the genetic algorithm that utilizes that language. So there’s an implicit set of theories that’s in this language and the implicit set of theories are theories with reach. And so goes the crit -rat argument. This is all the real knowledge that caused the robot to walk. They may point out to other fairly obvious sources of knowledge that pre -existed the genetic algorithm that were also created by a human. They might point out that the computer was a bunch of knowledge, the robot itself, the operating system. Okay. The very genetic algorithm itself, a human created the genetic algorithm. So all of this is knowledge that pre -exists the genetic algorithm running. Okay.
[00:13:54] Red: And it all is without a doubt comes from humans. Okay. Not from AI. So their intuition, the crit -rat intuition here is that there is so much knowledge being created by the human that surely the genetic algorithm didn’t really cause the robot to walk and it was really the human that caused the robot to walk. Now a common additional argument made by the defenders of the pseudo -Deutch theory of knowledge is that only inspiration algorithms can create knowledge not deterministic mechanical so -called perspiration algorithms. They argue that real knowledge creation comes from quote inspiration whereas quote perspiration can’t create knowledge. Though they argue it may look like knowledge. It just isn’t really because it is just mechanically trying every variant. Okay. This is often put somewhat cynically. So you might hear a crit -rat say something like this. Oh, that genetic algorithm is obviously just a deterministic mechanical process of perspiration. So it can’t possibly be doing anything creative much less creating knowledge. Okay. So kind of a cynical attitude towards the idea that a mechanical deterministic process could ever be creative. Okay. Now this argument plays well with our intuitions that there is something very special about human knowledge creation that isn’t captured in any of our current known algorithms. For if it was, then we’d know how to build AGI. So let’s consider both of these pseudo -Deutch arguments a bit and we’ll see that there is something off with each argument. Consider the argument that all knowledge that made the robot walk came from the human.
[00:15:32] Red: Defenders of pseudo -Deutch count up all the knowledge that came from the human and then draw the conclusion that since so much of the knowledge came from the human that probably all the knowledge came from the human and the genetic programming algorithm must not have created any of the knowledge of how to walk. But this is from a critical rationalist perspective, exactly equivalent to counting white swans. In reality, it doesn’t matter how much knowledge the human brought to the robot. What matters is if any of it at all got created by the genetic programming algorithm. If only a teeny tiny bit of knowledge gets created by the genetic algorithm, then logically speaking, that means some of it got created by the genetic algorithm and not by the human. Now, this is as obvious as if a candidate had a billion dollars of donations and you donated one dollar out of that billion. It would not be possible to prove you didn’t donate that dollar by counting even hundreds of millions of dollars donated by somebody else. So it cannot possibly matter how much of the knowledge came from the human. Gotta say that again. It cannot possibly matter how much of the knowledge came from the human. So there is zero point in counting up how much knowledge came from the human. Notice that this argument that I just made is a nutshell version of critical rationalism itself. The right critical rationalist approach to this problem would be falsification, of course. Try to seek an example of knowledge that clearly the human cannot have been the source of. If we do this, it refutes the idea that all the knowledge came from the human.
[00:17:10] Red: Once you view the problem via this lens, the answer, to me at least, it seems fairly obvious. Prior to running the genetic algorithm, the robot couldn’t walk, and afterwards it could. Therefore, the genetic algorithm must have created at least something we can reasonably point to as knowledge growth. This is why we prefer falsification over verification, counting up white swans of knowledge that came from the human literally tells us nothing whatsoever about if the genetic algorithm is creating knowledge or not. Or put another way, the pseudo -Deutsch theory of knowledge attempts to verify that the generic programming algorithm didn’t create knowledge by counting all the knowledge the human brought to the table, when they should be trying to falsify their theory that all the knowledge came from the human. Now let’s consider the argument that deterministic and mechanical perspiration algorithms quote can’t create knowledge, whereas inspiration algorithms can. Actually, they don’t call them inspiration algorithms, they call them inspiration programs because they insist it’s not an actual algorithm. I’m going to use programming algorithm interchangeably, and there’s a good reason for why I do that. I actually want to get away from the distinction that a lot of crit rats make here that I think is misleading. But I will acknowledge I’m just using programming algorithm interchangeably here. So one obvious problem with this argument is that every single algorithm runs on a deterministic mechanical Turing machine. So all algorithms and all programs, if you want to say programs, are deterministic and mechanical by definition. This part of the argument is really an attempt to bias your intuitions via a bad intuition pump.
[00:18:51] Red: It is hard for many people to imagine that a mechanical process can be creative, since it is in their minds just mechanically trying every variant. This seems very uncreative to our intuitions. Okay, yet we know that the so -called inspiration program, once we know what it is, it will of course also run on a Turing machine and thus be mechanical and deterministic too. It’s physically impossible for it to be otherwise. Sorry, did you have a question?
[00:19:17] Blue: So I just wanted to make sure I understand what you’re saying. So these crit rats are criticizing the idea that knowledge can be created by animals in their genes as just an algorithmic process, which is hitting our intuition pump that creativity is not algorithmic, but that’s going against another aspect of the worldview, which is that consciousness is algorithmic. So it’s like there’s a contradiction there. Is that kind of what you’re saying? Yes. Now,
[00:19:58] Red: sometimes I try to get around the contradiction by making a distinction between algorithms and programs. So they’ll say,
[00:20:04] Unknown: oh,
[00:20:04] Red: no.
[00:20:04] Unknown: Which
[00:20:04] Blue: Deutsch mentioned on the interview.
[00:20:08] Unknown: He
[00:20:08] Red: does that a lot. Okay. And it’s a technically correct statement, but one that’s mostly misleading. So the thing that I’ve explained in the past in other episodes is that there’s no such thing as a program that isn’t made up of algorithms. And I don’t mean it has some algorithms, but also some programs. I mean, 100 % of every program is made up of 100 % of algorithms.
[00:20:30] Blue: What is the difference again, an algorithm terminates, but a program doesn’t? Terminates with an output and a program doesn’t.
[00:20:36] Red: Okay. But it’s a distinction that’s only important when you need that distinction. Okay.
[00:20:41] Blue: So
[00:20:42] Red: like if I were, the way a crit -wrap might say it, if they have the misunderstanding that has come from this, is they’ll say, well, the creativity program isn’t an algorithm, whereas all current AI algorithms are algorithms. Okay. I’ve heard them say this. Okay. Absolutely, you just made yourself look really dumb. Because Skyrim is not an algorithm in that sense because it’s an ongoing game you can play forever, but every line of code in Skyrim is inside of an algorithm. Okay. You’re misunderstanding the distinction when you say something like this. Okay. You take a program that doesn’t have a single final output and it runs forever. And you take every ounce of it and you slice it into things that are accomplished into modules and each of those will be an algorithm. Okay. So when we talk about the creativity program and try to say it’s not an algorithm, of course it’s an algorithm. You just slice it for one output and it counts as an algorithm. It’s that simple. Okay. Like it literally is a misunderstanding if you didn’t get that point. You have completely misunderstood the distinction between program and algorithm. So if you take the creativity program and you say, okay, I’m going to look at from the moment Einstein had the moment of inspiration until he finishes the paper, that would now count as an algorithm because there’s a final output. What determines if it’s a program or an algorithm or an algorithm is really just how I happen to slice it for whatever problem I’m trying to solve. Okay. Does that make sense? I think so.
[00:22:27] Blue: I’m hanging on so far.
[00:22:28] Red: There’s literally no other distinction. Right. It’s just a matter of utility, how I need to talk about it. Okay. So referring to the creativity algorithm is a perfectly legitimate thing and try to say it isn’t an algorithm. It’s a program that is absolutely a misunderstanding. Okay.
[00:22:44] Blue: So
[00:22:45] Red: that’s why I use the two interchangeably and they should be used interchangeably. And if you aren’t using them interchangeably, you have misunderstood. Okay.
[00:22:54] Blue: Okay.
[00:22:55] Red: Okay. In fact, wouldn’t any algorithm or program that solves a problem by trying multiple solutions via trial and error pretty much by definition be doing something identifiable as perspiration? Okay. Like whatever the algorithm or program is, if it’s solving some sort of problem, then it’s going to be doing something that it’s going to be running on a Turing machine. It’s going to be mechanical. It’s going to be deterministic and it’s going to look to us like perspiration. For that matter, humans and biological evolution use considerable perspiration when utilizing what we call their creativity. Think about the sheer volume of perspiration used by biological evolution. It’s an amazing amount of perspiration by any fair measure. Okay. How many variants of different theories? This is a human example now. How many variants of different theories or supporting ideas do you think Einstein went through during the eight years he spent developing general relativity? And why doesn’t eight years of work to tweak and perfect your ideas until it’s ready for prime time? Why doesn’t that count as perspiration? Of course, it’s also a kind of perspiration. Right? The whole idea that somehow perspiration and inspiration are distinct from each other and somehow opposites never really did make sense to me. Okay. What is really going on here is that the defenders of pseudo -Deutsch are actually relying on the fact that we have no idea how biological evolution and human intelligence really work. Nor do they ever feel a need to explain exactly what they mean by inspiration. It’s always left very vague. And this is the whole trick that they’re using. They know for sure that any algorithm we point to that is currently known will always for sure be, quote, mechanical
[00:24:44] Red: deterministic and use perspiration, i.e. use tryout different variants mechanically. But the human general intelligence program isn’t known. So short of discovering AGI, which is not happening anytime soon, they can always claim that the AGI algorithm is different in some way due to this always undefined inspiration that makes it different. Okay. From here, crit rats have a game that they play that I want to call out and explain to make it explicit and visible. Suppose you point to an AI algorithm that just did something really creative impressive, like say AlphaGo creating a whole new play style that was unknown by humans prior to AlphaGo starting to play and that now all humans have to adopt this superior play style because now it’s known and humans have now changed their play style. Okay. The crit rat can point to such an algorithm and say, oh no, that doesn’t create knowledge. See, it’s just a mechanical deterministic perspiration algorithm. There’s no inspiration at all. So clearly no knowledge got created. But this is an all -purpose argument that is guaranteed to be true of any algorithm you will ever point to. Now, suppose you tried to engage the crit rat, and I’ve done this, of course, by asking, you know, I have no idea what this whole inspiration algorithm thing is. Can you please explain it to me? And then we can check to see if AlphaGo does or doesn’t do it. At this point, the crit rat is going to say, oh, see, we know humans are universal explainers. So AlphaGo isn’t creative because it doesn’t use explanations. Now, at this point, I might ask, okay, what about explanation -based learning? That uses explanations. Does that then qualify as inspiration? Oh no, says the crit rat.
[00:26:31] Red: That is clearly just a mechanical deterministic process of perspiration. Can you please explain to me what a non -perspiration algorithm would look like? I ask. Well, says the crit rat. No one knows that, of course. But you can tell humans are different because whereas machine learning and AI use millions or even billions of variants, humans don’t have to. That’s the difference between inspiration and perspiration. Here, the crit rat is engaging in a rational fallacy that I’m going to call the arbitrary X fallacy. The idea of this fallacy is that you claim some quantity is the categorical difference between two things. In this case, between perspiration and inspiration programs. But you don’t specify by theory how much of that quantity is required to transition from one to the other. Because the boundary is an arbitrary non -theoretical quantity, it can be adjusted on the fly to get to whatever outcome you desire. For example, the implication of this crit rat argument so far seems to be that perspiration is about millions or billions of trials whereas inspiration is only a few trials. Thus, we now have, if we were going to hold ourselves to this, we now have a way to tell the difference between the two, between perspiration and inspiration, but also a potential way to falsify the view. Okay, because we’ve now gotten very explicit. In fact, let’s now make an argument that does falsify this view. All we need to do is point to an existing algorithm that utilizes very few variants or trials. So, at this point in the conversation, I might argue something like this to the crit rat.
[00:28:15] Red: Before the advent of deep learning, there was a large field that studied how to create learning algorithms that learned with very few trials. This was a big area of study. It probably still is. It’s just got sort of swamped with the excitement over deep learning where there are literally millions and billions of trials. In fact, there was a great deal of excitement even once deep learning happened when it was realized that large language models were excellent zero shot learners or one shot learners. Basically, you just explain to them what you want and they can give you a response or you give them one example and they can give you a response. Okay, they can learn from zero examples or one example, zero variants or one variant. Okay, now does this mean that large language models are doing inspiration instead of perspiration? It should mean that under the way the crit rat was just arguing. And thus, doing now count large language models as creating knowledge. Of course not, replies the crit rat. Okay, then is there some upper limit of trials that inspiration is allowed to have before it becomes perspiration? And what would that upper limit be theoretically based on? For example, if over the course of eight years when Einstein developed his theory, if we found he tried out more variants than this supposed upper limit, does that then mean general relativity was perspiration instead of inspiration and thus not count as knowledge? Of course not, replies the crit rat. My point here is that no crit rat really believes that the number of trials can play a meaningful role in determining what counts as inspiration versus perspiration.
[00:29:57] Red: The whole thing was actually a post facto justification for a difficult or perhaps impossible to define intuition. A crit rat today will gladly point to a low count variation algorithm and still call it perspiration, even if the count is so low that humans clearly try far fewer variants than the algorithm. This is the inherent problem with the arbitrary x fallacy. It allows you to pretend to a level of precision your theory doesn’t really have. It attempts to make itself look like it’s making a quantitative distinction in this case between inspiration and perspiration. But really it’s just an on the fly adjusting an arbitrary set value set to x, such that we can declare a distinction that doesn’t really exist. The simple truth is that what we’re calling inspiration and what we’re calling perspiration will of course end up with a great deal of overlap between the number of variants and trials involved. Why? Because what we’re calling inspiration must also be a mechanical deterministic process of trial and error via something like perspiration. The fact that we know exactly what perspiration is, but we never have any idea what counts as inspiration, this creates an obvious asymmetry, though it’s the opposite of the one Popper intended. That is to say it is easy to verify that any non -general intelligence algorithm is mechanical and perspiration, since all known algorithms are, and it may be impossible for it to be otherwise. But we have no idea what an inspiration program is or how to code one, so we can imagine it having any properties we wish at this point. It’s a sort of utopian program that we can attach our imagined hopes and dreams to.
[00:31:37] Red: If you aren’t required to actually explain what inspiration actually is and how it’s distinct from perspiration, it is easy to imagine that someday once we finally have the AGI program in hand, we’ll see that it uses quote no inspiration at all after all. It allows crit rats to hide their theory inside a future discovery that isn’t likely to happen anytime soon. Let’s call this rational fallacy the deferred falsifiability fallacy, or maybe the future proof fallacy. I kind of like both names. It’s the fallacy that your opponent theory is concrete and falsifiable now, while your theory is immune until some huge future breakthrough happens. There is no way to prove inspiration is distinct from perspiration until the AGI program is actually understood. So until then, pseudo -Deutsch defenders are free to imagine the two as distinct or even opposites without the need to put their theory to the test anytime soon. This is because their quote theory is equivalent to the following logical statement. There exists a currently unknown Turing machine program that is creative inspiration and does not utilize any perspiration. Since it’s a purely existential theory, it can’t be refuted prior to the discovery of AGI. Once every, even once it is discovered, it would always be possible to utilize the arbitrary X fallacy to post -factor declare a victory. Say that the inspiration program, whatever it is, utilized 1000 variants on average. Just claim that X equals 1000 and say C, perspiration was over 1000 variants. That’s what I meant all along. Or say the inspiration program used a million variants. Just set X to one million and say C, perspiration was over one million. That’s what I meant all along.
[00:33:21] Red: So until AGI is invented and maybe even after it’s invented, the pseudo -Deutsch theory of knowledge is a perfect immunization strategy. And that’s what I don’t like about it. Until AGI is invented, there will always be a human in putting knowledge into an AI program that allows us to claim all the knowledge came from the human. And we know exactly what a perspiration algorithm looks like. It’s all of them as far as we know. But we have no idea what an inspiration looks like. So there is no fear of using that idea to refute the theory. There is one last argument used by defenders of pseudo -Deutsch, the pseudo -Deutsch theory of knowledge that needs to be addressed. I call it the credit assignment argument. Take the idea of animal memes. Animals have knowledge that they acquire via memes rather than genes. This is not in doubt. Deutsch explicitly talks about how this is true in beginning of infinity. He literally says animals have memes. That’s one of the lines I’ve quoted from him before. Or take the immune system. Discuss this in more depth in the next episode. But the immune system is a near -perfect example of information that keeps itself instantiated by being useful. The information is even encoded in DNA. That’s how the immune system works. It uses DNA to encode the recipes. But because the DNA knowledge from the immune system is not passed through the germline, it violates the two sources hypothesis. And so it has to be dismissed as not knowledge. Likewise, animal memes are explicitly an example of animal knowledge not found in their genes. A meme is by definition not in the genes. Otherwise this not doesn’t count as a meme. Okay.
[00:35:00] Red: So how do defenders of pseudo -Deutsch the pseudo -Deutsch view handle a direct refutation of their theory like this? They simply reassign credit. The knowledge produced by the immune system is credited not to the immune system, the thing that actually created the recipes, but to the genes that built the immune system. Likewise, the animal meme is considered knowledge in the genes in the sense that the animal learning, the algorithm that is animal learning, is itself a created by knowledge that’s in the genes. It’s easy to see why this credit assignment argument is a bad explanation. It functions as an all -purpose escape hatch. If every counter example can be dismissed by noting that the knowledge producing process itself arose due to biological evolution, and this is always going to be true, then the two sources hypothesis becomes unfalsifiable. In fact, once the credit assignment argument is allowed, you don’t even need two sources of knowledge after all. You can take all human knowledge and say it is really knowledge in the genes, because human minds were produced by evolution. Thus, the argument collapses into a claim that all knowledge ultimately resides in biology in the genes, which drains the two sources hypothesis of any meaningful content. I’ve now mentioned the two most important arguments made by defenders of the pseudo -Deutsch theory of knowledge. The first is that they try to move from most knowledge comes from the human to all knowledge came from the human, as if it’s an inference. They try to do this by counting white swans, as if the more white swans you count, the more obvious it is that this inference must be correct.
[00:36:38] Red: When one points out to them that the robot couldn’t walk prior to the genetic algorithm, but could afterwards, this whatever it is that got created by the genetic algorithm, defenders of the pseudo -Deutsch simply move to the second strategy, denigrate the result as not real knowledge, based on an argument that only humans can create real knowledge via inspiration, whereas this is clearly perspiration. They also easily point to an algorithm and show it uses perspiration, since all of them do, yet they can never explain what inspiration actually is and what it would look like. If you try to concentrate the conversation around this fraudulent knowledge, what I’ve come to call it, they declare it uninteresting and refuse to discuss it, or perhaps they declare it, quote, organization of pre -existing knowledge. That’s what induction is, right? So if you manage to convince them that there is a counter -example, animal memes, or another clear -cut counter -example, if you manage to convince them that there is a counter -example, and animal memes is a good clear -cut counter -example here, they simply assign credit to biological evolution. This really is a perfect immunization strategy. There is no possible way to refute the pseudo -Deutsch theory of knowledge as it is currently formulated. Now that I’ve explained why I feel the pseudo -Deutsch theory of knowledge is a bad explanation, let’s take this idea of inspiration seriously and make an attempt to understand it better. It seems pretty obvious that human intelligence does include one or more processes that we might rightly call inspiration, and that isn’t currently well understood enough to program. So it is not that I’m claiming inspiration doesn’t exist, okay?
[00:38:24] Red: Instead, my criticism is that I see no reason at all to believe that inspiration won’t itself be a form of perspiration, collapsing the distinction between the two. In fact, I feel like it almost has to be that way, unless you’re arguing in favor of something like induction, okay? To see why I say this, other than the obvious point that all algorithms of trial and error are by definition mechanically trying variance by trial and error and thus look like perspiration, let’s consider Hofstadter’s explanation of what inspiration really is, which in his mind is analogies, okay? Of course, we don’t know that Hofstadter’s understanding of concepts is correct, per se. Surely, we can see that it is sometimes correct, but that isn’t really good enough. Probably the correct theory will turn out to be cubical type theory. That’s a real thing, but like I don’t know that it has any relationship here, or something like that, where analogies play an important role but aren’t the whole of the explanation, okay? But let’s pretend, okay? It’s often very useful to pretend, okay, and do a hypothetical. So as a hypothetical, we’re going to say that Hofstadter’s theory is correct enough for our purposes and use it to help us make sense of the problem that exists around the inspiration -perspiration distinction that seems to be tripping people up. Let’s ask the question this way. According to Hofstadter’s theory, what is inspiration? Hofstadter believes inspiration is the sudden leap of analogy that took place from water to waves, water waves to sound being a wave. We talked about that in the past in the previous episodes. That surely does qualify as inspiration, does it not?
[00:40:02] Red: Recall that Crispus noticed sound bouncing and his mind says, oh, maybe sound is like a wave, like water is a wave. There is no competing theory, per se. Nothing gets falsified. I’ve made a very strong point of that. It’s literally a case of an observation suggesting a new concept via building an analogy from something familiar, invisible to something invisible, okay? We can see a water wave, you can’t see a sound wave, but suddenly because you hear it bouncing, you think, I wonder if sounds a wave, okay? Like a water is a wave. At that time, there really only were water waves or waves of grain, there were no other types of waves. They did not exist way back in Crispus’ time. Later it is found that this is a hugely fruitful analogy and the abstract concept of wave is born. We covered this in the previous podcast. What is the algorithm that created this inspiration for Crispus? Suppose we had the algorithm right in front of us and we’re looking at how the inspiration algorithm works. What do we find? Do we find that it is induction, like various crit rats have told me in the past, a way to reject Hofzetter’s theory, not realizing that if Hofzetter’s theory is truly inductive, we just refuted Popper’s Pistemology. Or do we find that the inspiration program to be an algorithm of trial and error variation in selection? Now, if you believe Popper has this right, there is no such thing as induction. Then the only answer that remains is, it will be an algorithm of variation in selection. It will be evolutionary Pistemology, in other words.
[00:41:40] Red: This realization was even the basis for Campbell’s and Popper’s evolutionary Pistemology, an attempt at generalization of Popper’s epistemology that was supposed to cover everything, okay? They realized that unless we’re going to declare induction to be a real thing that can work and it doesn’t use any sort of variation in selection, that we’re going to have to start with the assumption that every single one of these processes is actually evolutionary epistemology. And that was where the idea of evolutionary epistemology was born. Okay, but what is being tried and varied? And on what grounds does it get rejected or selected? Okay, so we’re trying to say Hofzetter’s theory, it creates analogies, it does it through some sort of process of variation in selection or trial and error. But what is it selecting? Okay, we of course don’t know how this works. If we did, we’d be able to build an AGI. But we can make a pretty good guess here. The mind must find analogies by jumping across many different variations of different ways to link events or ideas together to be able to fit them into categories. The mind must try out different analogies by trying to find connections between otherwise unlike things. There are a huge number of ways in which we might connect any two things like this. So the mind must be trying out presumably via parallel processing a very large number of different variants at any given moment. Okay, there must be some really good heuristics that we’re using, otherwise this would be an exponential algorithm and it would be intractable. Okay, but I want to emphasize when you’re trying to figure out, okay, how do I actually conceptualize this into an analogy so that it fits into some sort of category?
[00:43:28] Red: The number of things and the number of potential categories is enormously large. So we have every reason to believe that the inspiration algorithm is a perspiration algorithm. Okay, but it must still be trying, so it must be trying out a very large number of variants to find an appropriate analogy for every new situation. For example, Crispus could have decided sound bounced not like a wave, but like a ball. And then he could have concluded, when I shout little bouncing particles ricochet off the mountain and return as echoes. Here are several other unfruitful analogies he may have come up with. Sound flies through the air like arrows through the sky. Sound bounces off walls like light off a mirror. Perhaps it’s a kind of light we cannot see. Sound is a spirit exhaled from the mouth, carried by the air’s breath. Perhaps sound is the shaking of a fiery element in the air, like heat trembling in the distance. When I speak, invisible strings made of threads connect my mouth and your ears, vibrating with my words. Presumably, there are many, many, many, many, many, many, many more variants possible here. Somehow, Crispus’ mind had to somehow eliminate all these variants and somehow judge that sound is a wave is a better analogy. Then there is, there must be some sort of criteria by which the subconscious mind decides, this is a pretty good analogy. Let’s make ourselves consciously aware of it now. Now this isn’t justified knowledge. The very fact that the subconscious mind makes the conscious mind aware of an analogy doesn’t guarantee the fruitfulness of the analogy. Presumably, even most of what we become consciously aware of, analogies that we become consciously aware of, are still unfruitful and they die out.
[00:45:19] Red: In short, if Hofstetter is correct, then we have every reason to believe that the inspiration program is very much a deterministic mechanical algorithm of perspiration that mechanically tries out different variants. In fact, it’s pretty much either that has to be the case or induction is correct and Popper’s epistemology is incorrect. Inspiration either is perspiration of trial and error or it’s induction. I don’t see how there’s any other option that’s truly possible here. This is another example of why I consider crit rats today to often be crypto -inductivists when they reject the idea that inspiration can’t possibly be a form of perspiration. Now, I have summarized my argument against the pseudo -Deutsch theory of knowledge. It’s built on two verification -based arguments. One, that there is so much human knowledge that goes into making the robot walk that it can’t possibly be that the genetic algorithm creates any knowledge. And two, that every known algorithm is perspiration and thus can’t be inspiration, which is supposedly not perspiration for reasons that are never explained. But I believe it is almost inevitable, barring induction turning out to be actually correct, that whatever the inspiration program actually is, it will also be a mechanical deterministic process of trial and error that will look very much like perspiration as well. In fact, it will be perspiration. So one final word of caution though. I think there are a few reasons why the pseudo -Deutsch theory of knowledge has become such a persistent aspect of the reason is just that Deutsch is incredibly important, almost guru status amongst Deutsch and Critt Ratz. A lot of them probably would not think to challenge him in any way on things he’s written on the book here.
[00:47:03] Red: However, there is another, and it’s probably a better answer here, and that’s that it really is an incredibly clever intuition pump. The pseudo -Deutsch theory of knowledge pumps your intuitions just perfectly towards seeing knowledge created by, say, the immune system as not knowledge, namely by pointing out how much more impressive is the kind of knowledge a human or biological evolution can create. But there is a third, even more important reason why the pseudo -Deutsch theory of knowledge has been latched onto so strongly, namely that it happens to be kind of correct. One of the things I’ve argued on this show is that bad explanations are far more likely to be true than good explanations. Okay, I’m going to say it again because I feel like it’s really important and I’m trying to drive it home. Bad explanations. Wow, that’s a provocative statement.
[00:47:49] Blue: Yes,
[00:47:50] Red: bad explanations are far more likely to be true than good explanations. Okay, this is because bad explanations have this very vague quality to them that makes them easily varied to match whatever the truth turns out to actually be. Okay, did you have something you wanted to add there? You’re right, it is a very provocative statement. No, no, I
[00:48:12] Blue: just have to digest that. It’s, that’s, I mean, it rings true in a sense. Yes.
[00:48:20] Red: Okay, and here’s the thing, because that’s true, because bad explanations are so easy to say post facto, see, that’s what I was saying all along. Bad explanations have an overwhelmingly good track record, where post facto, we can see that they were sort of right all along. Okay.
[00:48:38] Blue: So you’re saying that you do believe that there’s some core truth to this assertion that there’s something special about human created knowledge that’s not, that’s different than genes.
[00:48:53] Red: That is, yes.
[00:48:55] Blue: But you’re criticizing the ideas for why people think that as a bad explanation. That’s right. Yeah,
[00:49:03] Red: you know, okay, so let me just say that I think people wonder about me sometimes, because I have my own points of view, and I’m often arguing with people that agree with me. Okay. And there’s a really good reason why, and I’m kind of surprised people, and then they’ll immediately assume I’m in the other camp, because I’m agreeing, I agree with them, but I’m arguing with them, and so therefore they will treat me as if I come from the other camp. You’re
[00:49:33] Blue: embodying the spirit of critical rationalism is what you’re doing.
[00:49:37] Red: That’s right.
[00:49:37] Blue: That’s right.
[00:49:38] Red: Okay. And usually what I’m arguing over is I’m arguing, no, that’s a bad explanation. I care more about the fact, even if someone is agreeing with me, if they’re using a bad explanation, I want it cut out, right? I want to end that bad explanation and only use the good explanations.
[00:49:59] Blue: Yeah.
[00:49:59] Red: And that makes sense. I react against bad explanations, even ones that are arguing what I agree with. Okay. And isn’t that how it should really be? Like, we want to get rid of our bad explanations. We want to get down to our good explanations. We need to be able to criticize our explanations that are bad, and we need to say, okay, that one’s not good. Let’s go with this one.
[00:50:22] Blue: Yeah.
[00:50:22] Red: Okay. Makes perfect sense. Really? Okay. Yeah. So that is a huge part of the appeal of the pseudo -Deutsch theory of knowledge. It roughly captures an entirely correct idea that there are two very special knowledge creation algorithms, biological evolution and human ideas. Totally true.
[00:50:41] Blue: Yeah. I still think that was a big aha moment that I’ve never quite come away from when I first heard it put like that. I was like, yeah, that’s just, it just rings true.
[00:50:51] Red: Yeah. These are special. These two are special because they are the only known open -ended knowledge creation algorithms. So the result of these two sources of knowledge give far more impressive results than, say, the immune systems knowledge creation. Once the correct theory of knowledge is really understood well, we’ll easily be able to look back and say, well, Deutsch wasn’t entirely correct, but he was on the right path. And that will be a true statement. I promise you it will be.
[00:51:22] Blue: The immune system is never going to divert an asteroid or put us on the Dyson sphere. That’s correct.
[00:51:28] Unknown: I mean,
[00:51:28] Blue: I think we can pretty much safely say that.
[00:51:31] Unknown: We can say
[00:51:31] Red: that. Right. Okay. Here’s the thing though. This feature, this is a feature, not a bug, of bad explanations. And that’s why bad explanations are so dang popular. And why bad explanations may even represent good starting conjectures. So long as the defenders of the theory then go on to keep trying to reframe it to actually say something about the world that can be checked or tested as a real theory. Okay. So my argument against pseudo -Deutsch theory of knowledge should not be thought of as an argument that pseudo -Deutsch is a false theory. My argument needs to be understood explicitly as an argument that in its current form it’s a very bad explanation. One which if dropped and replaced with say Campbell’s theory, perhaps with Ken Stanley’s theories about open -endedness thrown in for good measure, that you’ll get everything out of pseudo -Deutsch without the misleading elements and without the inbuilt immunization strategies that turn off the means of error correction. That’s why you should drop the pseudo -Deutsch theory of knowledge and you should move to a more explicit theory that says the same thing, but it’s still falsifiable. Okay. And this is really what I’m arguing. Not that it is strictly incorrect. I think this critical rationalist way of thinking, I mean that we care more about how good the explanation is than if it happens to be vaguely true or not, is alien to how most people think. And even alien to how most crit -rats think today, I would imagine. But it is the view I’m advocating for on this podcast. My view is, is that bad explanations should be opposed on the grounds that they are bad explanations, regardless if they happen to be arguing the correct view or not.
[00:53:21] Red: So that is why I can at once oppose the pseudo -Deutsch theory of knowledge as a bad explanation while actually agreeing that it correctly captures the idea that biological evolution and human ideas are special in some way. Okay. That’s the end of my criticisms of the pseudo -Deutsch theory of knowledge as a bad explanation.
[00:53:40] Blue: Okay. Well, you’re in the weeds with this stuff, like no one else that I know of at least. And I just hope that we don’t live in a Lovecraftian universe where knowledge makes people who are knowledge leads to madness, because you might be halfway there, Bruce. I hope you’re, as the kids say, touching grass, too.
[00:54:07] Red: I hope so, too.
[00:54:09] Blue: Okay. All right. Take care.
[00:54:12] Red: Thanks. Bye -bye.
[00:54:21] Blue: Hello again. If you’ve made it this far, please consider giving us a nice rating on whatever platform you use or even making a financial contribution through the link provided in the show notes. As you probably know, we are a podcast loosely tied together by the Popper -Deutsch theory of knowledge. We believe David Deutsch’s four strands tie everything together. So we discuss science, knowledge, computation, politics, art, and especially the search for artificial general intelligence. Also, please consider connecting with Bruce on X at B. Nielsen 01. Also, please consider joining the Facebook group, the mini worlds of David Deutsch, where Bruce and I first started connecting. Thank you.
Links to this episode: Spotify / Apple Podcasts
Generated with AI using PodcastTranscriptor. Unofficial AI-generated transcripts. These may contain mistakes; please verify against the actual podcast.