Episode 38: Animal Learning and Popper’s Epistemology (part 2)

• Links to this episode: Spotify / Apple Podcasts

Unofficial, reviewed transcript. It may still contain mistakes; check the podcast when wording matters.

Speakers are labeled by first name where identified. Unidentified-speaker labels may refer to different people. Bracketed question marks indicate uncertain wording.

Transcript

[00:00:11]  Bruce: Welcome to The Theory of Anything Podcast. We are continuing discussing about animal intelligence. It's actually the first episode on animal intelligence that we did. We then took a break. I mean, like, you're gonna, people who are listening are gonna hear them in order. But there's actually like a six week plus break in between the first episode and this one that we're recording now. And we did the Julian Barbour interview in between, and it took us six weeks to get that interview together because things kept going wrong. So, six weeks is a long time. And so I've kind of changed, I discovered things in that six week period.

[00:00:48]  Bruce: For one thing, I started reading another one of Richard Byrne's books, The Thinking Ape, and came across a lot of cool stuff that wasn't part of my original presentation that I was gonna do on animal intelligence. So I think today we're gonna do an aside episode that's gonna be a bit of a hodgepodge of cool stuff I came across that seemed like it was worth discussing. So that I hadn't known at the time I had done the first episode. So one of them is this quote that from Michio Kaku, who's the guy who does like the physics of Star Wars and things like that, pretty famous guy. He says, at the present time, our most advanced robots have the collective intelligence and wisdom of a cockroach, a mentally challenged cockroach, a lobotomized mentally challenged cockroach. I thought that was hilarious.

[00:01:43]  Bruce: And I actually think that that's accurate, right? In our last episode, we talked about the fact that our current AI abilities cannot account for animal behavior or animal intelligence, that animal intelligence is itself a mystery to us. Never mind AGI, which is an even bigger mystery to us. We can't even explain animal intelligence. And so that's one of the reasons why I find animal intelligence so fascinating. Even if it doesn't turn out to be a path to AGI, which it may not, it's just its own interesting thing all on its own. And it also may be a path to AGI for all we know. We'll talk about that. Byrne's theory in future episodes, he thinks that there's a relationship between human intelligence and certain kinds of animal intelligence, specifically in the great apes, which would make some sense. I mean, obviously, those are our closest relatives in the animal kingdom. Although I don't know that for sure, right? And he doesn't know that for sure either.

[00:02:42]  Bruce: We'll just look at the evidence. And then I came across, Sadia, you and I talked in the last episode about the Baldwin effect. And you kind of took an interest in this idea that a squirrel's automated process where it digs and then it tries to bury its food and then it tries to pat it down with its nose. That the only way that that could have made sense under neo-Darwinism is if every single part of that program somehow evolved where only the one move was valuable. And that doesn't even seem possible under neo-Darwinism. Now, the Baldwin effect was an attempt to explain that. You imagine instead that there's an ancestor to the squirrel that lives way back in time. And it learned to bury the nuts on its own. And then that population through imitation, and we'll talk about imitation in a second.

[00:03:39]  Bruce: It's a bit of a misleading term. That spreads through the population of that proto-squirrel. And then once they've learned it and it's been spreading by [unclear], then it makes perfect sense that every individual move in any order becomes valuable in terms of evolution implementing it in the genes. And so this solves the mystery of how do you even get software into the mind of an animal via genes to begin with. OK. Now, it turns out, like right after that episode, I was reading Objective Knowledge by Popper, and he talks about this. And he makes claims much stronger than mine. And so I wanted to actually give the quotes from Popper because I thought they were so fascinating. So Popper talks about specifically the Baldwin effect.

[00:04:32]  Bruce: So quoting Popper, the Baldwin effect, that is, the theory of a purely Darwinian development that simulates Lamarckism seems to me an important step towards the explanation of such developments. OK, never mind the context of this. I'm just showing he is specifically talking about the Baldwin effect. And just like I said, without the Baldwin effect, the claim that animals are evolving software programs in their mind is Lamarckism. That's what Popper say, right? That it simulates Lamarckism. It looks like Lamarckism, but it's actually not. That's what the Baldwin effect is about. OK, now I had argued that it's necessary to have evolution of software in the mind. Without the Baldwin effect, which implies learning came first, that you could not develop something like the squirrels program, where it automatically digs and pats it down. Popper goes much further than me on this. He points out that physical evolution itself doesn't work without learning algorithms like that.

[00:05:39]  Bruce: This was shocking to me, but it makes such perfect sense now that once I read what he had to say about it, I can't unsee it now. We take it for granted today that our brains contain software and can evolve algorithms. But keep in mind that Popper didn't have, he's not coming from our time period. So software is such a common thing for us, we think in terms of it. Software existed at the time of Popper, but it wasn't so common, and he wouldn't have used the word software. So he tries to use other words to explain what he's talking about. So he talks about evolution is being dualistic, that's the term he uses. And he makes a distinction between the behavioral improvements and the physical improvements. This is in Objective Knowledge, page 273.

[00:06:22]  Bruce: And he tries to use the example of an automatic pilot, which is something he would have been familiar with. As an example of this, you've got this automatic pilot that, you know, it may be mechanical back then, but it's something very much like software. And today it would be software. So he's trying to make this distinction. He doesn't have the word software to work with, but he's trying to talk about physical hardware development through evolution and software development in the mind. And then he says, this is a quote, According to the dualistic hypothesis, which is his hypothesis, a favorable change of a physical organ would in many cases be only potentially favorable.

[00:07:03]  Bruce: To make any difference, the improvement would have to be used. And this new use might depend on a complementary accidental change in the, he says, central propensity structure, which is what he's the term he's using for the software side in the mind. But the probability of two such accidental changes, which would be at the same time, both independent and complementary, must indeed be vanishing. Okay, this is the exact same argument that I used with the Baldwin effect for the development of software in the mind. He's saying it's also true for physical evolution. Okay, so physical evolution would fail unless you have something like the Baldwin effect, something equivalent to it, where the software develops first that allows the animal to take advantage of the new adaption.

[00:07:55]  Bruce: So quoting Popper again, a mutation which makes the skill structure more flexible may become more favorable, and by such mutation of the skill structure, the organ may acquire the propensity to learn in the sense of improving its skill by trial and error. Okay, he's talking about learning algorithms. It's incredibly clear that he is making the claim that evolution had to evolve learning algorithms first, or physical evolution would not work either. Now, he does put this out as a fairly radical view. So he's guarded about it, and I'm not including the quotes where he's guarded about. So keep that in mind, okay? He's not sure if it's the truth or not, but he's putting it out there boldly as, let's look at this, this seems like it's something we need to be looking at as a new theory about evolution. Now, Sadia, as you know, a lot of the Deutsch fans, they make the claim that animals are just automatons. So let's actually use that as an example just for a second.

[00:09:02]  Bruce: So let's take like the Boston Dynamics Robots. Now, most of the Boston Dynamics Robots that we see that we're so impressed with have no learning at all. They're just straightforward robotics that are built to do some sort of, built to do whatever they do, and that's it. I understand like the dog robot has some learning capacity. It uses some form of machine learning. But most of the robots, what I understand, don't use any form of learning. So let's say that you took a Boston Dynamics Robot and you decided to give it a more powerful engine or better hands. The software would then be out of sync from these improvements. They would actually make the robot not work because it would be trying to use its software for the less powerful engine. And it would overcompensate and it would be a failure. This is what Popper is talking about. He's saying that evolution can't just attach physical improvements and the animal will actually not, it should hurt the animal to have physical improvements.

[00:10:08]  Bruce: It should make the animal worse to have physical improvements unless the animal is capable of trial and error learning using a Popperian-style epistemological process. Does that make sense?

[00:10:20]  Sadia: Yes. Are you saying that the physical evolution would have to go hand in hand with a learning algorithm? It can't just be that some mutation caused some physical change, which the way we think about it that way or...

[00:10:34]  Bruce: That is correct. What we're talking about is that the Neo-Darwinian evolutionary process would have created intense evolutionary pressure to create learning algorithms very early in the tree of life. That might explain, I say this with some hesitance, that might explain why even single-celled animals have learning algorithms, but the ones they do, habituation seems so primitive, I'm not even clear how that fulfills what Popper is saying here. But it makes sense that somewhere very early in the tree of life, animals would have developed learning algorithms, powerful learning algorithms. And I think that this ties in well with why animals are so mysterious to us. We haven't figured out what an animal learning algorithm is like. We don't understand how to program it. It's something very advanced compared to what we're currently doing. Because it had to allow the animal to be able to take advantage of physical mutations that were advantageous to it.

[00:11:42]  Sadia: No, I'm just thinking so, you know, typically if we think about like simple organisms and animals, one might think of variation and selection in terms of, oh, just randomly just like, you know, like an amoeba or something that has, that is trying to make its way around. It's just the certain physical things are kind of varying and then certain things get selected. But here you're saying that, so what is it that's being varied, right? Like if to address the question of how an organism is actually learning, what that learning algorithm is, would the question boil down to what exactly is being varied? Am I right or on the right track?

[00:12:24]  Bruce: Okay, good, interesting question. I'm not sure I'm quite following you yet, but let me say...

[00:12:31]  Sadia: Can I just maybe, just to clarify it, so for example, in humans, right, we have ideas, we have thoughts, so we can think in terms of when we're learning, we're thinking that, okay, we're varying different, you know, we have ideas, we come up with conjectures and we do error correction. But in the case of animals, if there is some learning algorithm and it doesn't come down purely at the physical level, but there is some other algorithm at play, then how is that? Like, what's, ontologically maybe, what is it that?

[00:13:02]  Bruce: I'm with you. As it turns out, I was about to talk about that. I referred to in the last podcast to classical conditioning. Now, classical conditioning is the most famous form of animal learning. I made the claim, well, it's not my claim. I, Byrne made the claim, and I'm just quoting him, that there's other kinds of animal learning, such as insight. And we're going to get to that. But most animals don't have insight, according to Byrne. So most animals rely on classical conditioning. That is the learning algorithm of animals. Now, Byrne in The Thinking Ape, which I'm currently reading, he makes a distinction that is actually really apt, that I had completely missed, because this isn't my field and I honestly don't know what I'm talking about. I use the term classical conditioning to cover two kinds of learning algorithms that Byrne actually considers two separate algorithms.

[00:14:00]  Bruce: So he refers to classical conditioning, but he also refers to instrumental conditioning. Now, I was using the term classical conditioning to refer to both, because to me, as a layman, classical conditioning meant both. But classical conditioning is reacting to stimuli. It's the dog hears the bell and it salivates. Instrumental conditioning is the same thing, but it's based on the active action of the animal. So a rat discovers that when it pushes a lever, it receives food. So there's an actual action on the part of the animal. So he goes on to say that instrumental conditioning is also called trial and error learning. Okay. So now in the past podcast episodes, we've talked about reinforcement learning. Whatever animals are doing, it is something like reinforcement learning. Now we've talked about the fact that reinforcement learning, the actual theory we have of reinforcement learning in machine learning, can't explain animals.

[00:14:53]  Bruce: It's got too many giant gaps. For one thing, you have to actually insert the world space into the animal's head for it to use it if they were using actual reinforcement learning as we understand it today. Yet it's clear that what we call reinforcement learning and machine learning was inspired by instrumental conditioning, by this trial and error learning of animals. That's why it bears a similar name. We refer to reinforcement learning of animals. The two theories aren't the same. Animal reinforcement learning and computer reinforcement learning, they have the same name, but they are not the same thing. They have a different vocabulary. They don't work the same.

[00:15:36]  Bruce: What a machine does today requires way more trial and error than what an animal needs to do. And yet, I think that is the answer to your question still, Sadia, that instrumental conditioning, the trial and error learning, is still ultimately a variation of which actions produce results that are positive versus which actions produce results that are negative. There's a feedback cycle to the valences of the animal. Getting back, not trying to get into animal sentience, but the animal has some sort of system where it feels pain or whatever. It feels pleasure when it eats or mates or something like that. And it does its actions and it does its play, its explore program basically. Animals come with curiosity, according to Temple Grandin. That's actually not something she came up with. She's just quoting what the literature says about it. That's one of the valences that animals have is curiosity. Its actions is what's being varied and the reward system is the valences. Does that make sense?

[00:16:49]  Bruce: Okay. Now, Byrne points out that there's a divide between learned and genetically programmed that is not binary. And I did not call this out well in the last episode, so I felt like I needed to include this. And he explains this better in The Thinking Ape than in Evolving Insight, which is the book I had read up in the last episode. So when we talk about animal learning, we do need to be a little careful versus animal programming. We do need to be a little bit careful because there's actually something in between, and that in between is what we normally mean. Okay. Animals don't have completely open-ended learning processes, only humans do. So no animal has that. Okay. On the other hand, you can't explain animal behavior as programs in their head because that is Lamarckism, that entirely violates neo-Darwinian evolution. What's the middle ground on this?

[00:17:44]  Bruce: That is what animals actually do. Well, he calls this genetically channeled learning or channeling of animals' selective attention by genetic predisposition. Okay. By the way, Byrne says, and he gives good examples of this, that this is true for humans too. The fact that humans do have open-ended learning processes doesn't mean we don't also have genetically channeled learning processes. Byrne uses the example of chaffinches. In an experiment, chaffinches, so chaffinches, they learn a song. There's a certain song that they have to learn to be able to mate, the males anyhow. So they wanted to do an experiment and find out if they could get a chaffinch to learn a blackbird song. What they found is that the chaffinch will not learn a blackbird song. So one thing you might consider is maybe the song is genetically predetermined. Now we know that that's true for some animals.

[00:18:38]  Bruce: There are some animals that have a genetically predetermined song that they never have to learn. They get to the right age and it just comes out on its own. So they wanted to see if that was true for chaffinches. So they tried to teach the chaffinch to learn the chaffinch song backwards. And it learned it backwards. So based on these experiments, we know two things. You can't teach a chaffinch a blackbird song. You have to teach it something similar to a chaffinch song. But it's not predetermined. It's not genetically predetermined because you can teach it to be backwards instead of the song forwards. So how are the genes doing that? Well, we don't really know. This is not something that we fully understand. But it would have to be something like this. It would have to be that the genes are affecting what the chaffinch is paying attention to. This is why we have this selective attention by genetic predisposition. So how does that work? I don't know.

[00:19:38]  Bruce: Maybe it like pays attention to more notes than others, the ones that are relevant to the chaffinch song, or maybe ones that are in close to each other, or somehow there's this genetic predisposition to pay attention to some things and not others. That makes sense. We all have that, right? Nothing would be anything particularly mysterious about that. Then on top of that, the genes have given chaffinches perfect memory. So when they first hear the song from their father, presumably, a year later, they can still apparently recall the song in their head. Now, how do they know that? I mean, like, I'm making these claims, and it's not like we can go ask the chaffinches. Well, they've done experiments to work this out, right?

[00:20:20]  Bruce: The chaffinch can apparently, completely absent from anything other than the original stimulus from the father, can basically start practicing. It initially does a very poor job of trying to do the chaffinch song, and then it will make steady progress even if it has no other examples to listen to. So they know, since you can teach it to do it backwards instead, they know that it can't be pre-programmed, but they also know that it's going through a trial and error process because it doesn't just suddenly jump to doing it right. Basically, what are you left with? You're left with the idea that it has been given by genetics perfect memory of the first songs that it hears from its father, and then it can practice against that memory. Imagine that the genes just have those two things.

[00:21:07]  Bruce: It selects the attention to which songs it pays attention to, and it has this way that it can do trial and error from memory, and it can practice against its memory. That alone, those two things, would be enough to almost guarantee that the chaffinch is going to learn the chaffinch song. So the genes don't contain the song, but they do make sure the chaffinch learns that song. Do you understand the distinction here? Does this make sense of what I'm talking about in terms of genetically channeled learning?

[00:21:40]  Unidentified speaker: Fascinating.

[00:21:41]  Bruce: Yeah, it is. It's like super fascinating.

[00:21:44]  Cameo: Yeah. I wonder if humans have any of similar constraints on our abilities to learn.

[00:21:54]  Bruce: You know what? I'm going to give you some examples of that.

[00:21:56]  Unidentified speaker: Oh, great.

[00:21:57]  Bruce: You know, with humans, it's sometimes harder to come up with examples because anytime a human does something, you can almost always claim that instead they use their general purpose learning because the general purpose learning can also accomplish the same thing. It's hard to find examples where you know for sure that we weren't using our general purpose learning on that. But I will give you at least one good example of that. Instrumental conditioning. This is from The Thinking Ape, page 52. So Richard Byrne. Instrumental conditioning has been likened to evolution by natural selection. Like natural selection, it is an automatic process by which behavior patterns with beneficial consequences are selectively increased in frequency—amount—frequency and amount [among] the many types of behavior produced by the animal. The mechanism is quite different in the two cases between biological evolution and trial and error learning.

[00:22:54]  Bruce: However, in instrumental conditioning, reinforcement increases the probability of advantageous responses to what the animal perceives in the world. Unhelpful responses become extinct, but the animal lives on. Natural selection is much more drastic and slower to produce change since whole individual animals survive or die and the selection effect on genes is therefore much less efficient. Rapid changes of behavior in response to changing environmental circumstances only come about by learning. Instrumental conditioning functions automatically to record the significant results of exploration and experiment, ensuring the animal profits from its experience and need not repeat its errors. Okay, this is a quote straight from Byrne. Let me be clear about this and why I included this.

[00:23:40]  Bruce: Byrne is saying that animals let their ideas die in their place, their actions in this case, not actual ideas, not in the human sense, die in their place. This should be exciting to Popperians. The fact that animals very early on gained learning algorithms and that learning algorithms are a Popperian process, which is what I've been saying, but now... just throwing this in here, even Richard Byrne and even the field accepts that this is the case. I didn't know that when I did the last episode. That's exciting, right? I mean, animal learning is as much an Popperian epistemological process and is of interest to us as AGI, as machine learning, as epistemology itself, science.

[00:24:26]  Bruce: You know, this is something we should be paying attention to, because this is the point of having learning algorithms, is that the animal can then respond to its environment and it doesn't have to rely on only pre-program. Basically put, the idea that animals are just meat robots in the sense of Boston Dynamics automatons, we know that's not true, right? That doesn't tell us anything about sentience, and that's a different question. But they absolutely have to be able to have learning algorithms. We know that's the case, and it is a trial and error process. It is Popperian. All right. Off my soapbox on this one. Any thoughts on that?

[00:25:07]  Sadia: I have some questions, but I'll need to think a little bit more about that. Because now I was trying to think of things that don't have a brain or that type of memory associated with it, like amoeba or, you know, bacteria.

[00:25:22]  Bruce: There's a completely fair question there. Like I said, we know that amoebas have a form of learning called habituation. But it seems like such a weak form of learning. I don't know how to make sense of all this in the case of single-celled animals. Do they have additional learning process we don't know about? Or is habituation a more powerful process than it first appears? I don't know the answer to that question. That seems like that is a legitimately interesting problem, epistemological problem, that probably is worth looking at, right? I mean, this is why I'm trying to always get like people in different fields to think more about epistemology and us epistemologists to think more about other fields and not think that they've got nothing to teach us.

[00:26:06]  Sadia: And you know, there's one thing that related to this that's been on my mind... quite a bit lately about just memory in general. You know, like in the case of humans, we can make sense of it in terms of, okay, you know, associated with brain, with other animals too. Then there's also a sort of genetic memory. And I think I'm kind of curious to see what type of other memory is there? Like, what are the things that are capable of having memory? And is there like some sort of a strong connection between learning and like if anything has memory? You know, to what level does that go?

[00:26:42]  Bruce: I think that's a completely interesting question. I think someone should be looking at that question, right?

[00:26:49]  Sadia: I think it kind of stems from because I feel like this question doesn't just apply to life. I think it might actually go down to the level of universe, but I don't want to take away from. I think the memory and learning just in general, I think is really interesting. It might have some fundamental implications as to which might go beyond life. But anyway, sorry, I didn't mean to just want to throw that out.

[00:27:13]  Bruce: OK, so why did nature evolve learning algorithms instead of just pre-programming every animal behavior right into the genes? So Byrne answers this question. We've already answered it. So I won't spend too much time on this, but I just thought it was interesting that Byrne basically gives exactly the answer we've already discussed in the last episode and in this episode. He says, on one pole, pure genes, each baby chaffinch would have been genetically programmed to sing the chaffinch song. This does indeed happen in some songbirds. Each baby chicken could have been equipped with a recognizing device that could tell its mother from not only trees and foxes, but every other female chicken. He's talking about imprinting in this case. That would take a lot of specification and have to be different for every mother chicken.

[00:27:55]  Bruce: It is really not even remotely feasible. If every baby rat were equipped with an innate avoidance of natural poisons, then rats would never learn avoidance of new poisons. He's trying to explain why pure genes just does not work. It is an impossible solution. And insofar as the Deutsch fan community may have accidentally gone down that road, it's wrong. We know it's wrong. But moving to the opposite pole, pure learning. It is not obvious how many birds would ever acquire species-specific song, nor how chicks would survive the formative learning experiences with foxes and motor cars. Genetically-channeled learning is evidently a powerful and efficient tool for giving animals what they need to survive.

[00:28:35]  Bruce: Okay. So now this is actually, if the Deutsch fan community were willing to adapt themselves a little bit, and instead of talking about all the knowledge being in the genes, and instead say their ability to gather knowledge is controlled by the genes, that would actually be a true statement. Like they're not totally wrong here. They're kind of getting something. They're just not quite getting it right, right? But it's still a Popperian process. And this explains most animals don't need a general purpose learner. That's why only humans have it. Okay. Now, Cameo, you asked about do humans have genetically channeled learnings? The answer is yes. So let me give you some examples that are fairly obvious.

[00:29:15]  Bruce: So rats, they did experiments, and they tried giving rats like novel flavored liquid to eat. And then they would induce in the rat sickness afterwards, you know, make it sick to its stomach or something, you know, and then they would try that with like blinking lights. So the food is this liquid and it has blinking lights in it. So one of the foods doesn't have blinking lights and one does, and the one with blinking lights, they then induce sickness into it afterwards. The rat will learn not to drink the liquid that tasted novel, but it won't learn to avoid the food or liquid with a blinking light. Well, the reason why is because in nature, blinking lights isn't the signal that this is food you should be avoiding. So the rats, their learning algorithm has been attuned by their genes to care about novel flavors, not blinking lights. That makes perfect sense. You can classically condition a rat to avoid novel flavors.

[00:30:15]  Bruce: You can't classically condition a rat to avoid blinking lights. Now, why am I telling you all this and how does this relate to humans? Well, and this is a quote from Byrne. Humans too seem prone to this kind of food aversion learning as found in rats. Most people have experienced going off some unusual food that produced nausea, even when they know very well that the nausea was really caused by some infective illness. I've had that happen. Have you guys had that happen before?

[00:30:44]  Cameo: Yeah, absolutely. It's interesting.

[00:30:47]  Bruce: Okay. So that is actually, you've had this happen to you before. That is an example of genetically channeled learning in humans. Similar to the way it works for rats. Now, and this is an example of how our general purpose learner overlays our more animalistic learners. We've got both. If you really needed to survive, you would take that food that you're trying to avoid because you associate it with your sickness and you'd be able to override it. You'd be able to say, okay, look, I know I only hate this food because of the unfortunate circumstance that I got sick afterwards. I know it's got nothing to do with the food. And you would be able to maybe force yourself to eat the food. Okay. You would use your general purpose learner to overcome your animal learner. But if you don't have an incentive to do that, you probably won't. You know, you'll probably default to your genetically channeled learning just like an animal would. Okay.

[00:31:41]  Bruce: Then this one's really interesting. And I'm not even sure what to make of this one. He throws this one in as an example of human genetically channeled learning. But like this one's bizarre. Okay. But you think about it and it's true. Learning and genetic predisposition are intricately meshed in an animal's development. And the same is even apparently true for skilled human behavior. For instance, it would seem that an adult male and female are predisposed to talk to babies in a particular way. What's that you've got there? A nice red ball, is it? Isn't it pretty? Wouldn't it be fun to drop it? And so on. This speech is singularly low in novel information, which is largely obvious from context anyway.

[00:32:22]  Bruce: Yet it is flamboyantly rich and complex in syntax, the largely unconscious everyday grammar that allows people to structure a limited set of words into phrases and sentences with unlimited reach of meaning. This kind of talk also has an exaggerated intonation pattern, making both the syntactic structure and those individual sounds, the phonemes of speech, that distinguish meaning more obvious than normal adult speech. Syntactic forms are often repeated in successive utterances, and the tendency of adults to interrupt themselves or trail off into incomplete sentences is missing. It is, in short, an ideal vehicle for helping the child to learn phonology and grammar. Mothers and other caretakers who talk this way are unaware that their speech has this special function in language development. So, somehow, we have a genetically channeled learning that has led us to learn baby talk.

[00:33:16]  Bruce: Now, obviously, we learn it from each other, right? But how is it that this is so consistently done across all societies? And it has this very specific purpose in teaching syntax and grammar to the child, and has no purpose in actually communicating information. How does that happen? I don't know, right? I mean, that would be a really interesting thing to go try to study. But it seems like that is a very powerful example of how humans must have genetically channeled learning going on, even in some things that are really skilled behavior. Cameo, does that answer your question?

[00:33:50]  Cameo: Yeah, I think that that's an interesting answer. Curious.

[00:33:56]  Sadia: The only thing is that you would think that there would be, you know, you should be able to think of some sort of an evolutionary advantage to doing that. Personally, I don't see. I mean, maybe I'm just not thinking enough of what the baby talk. Like, for example, I'm not into babies. Like, I don't. All right. I'm going to say something that's going to sound horrible. I don't find babies cute at all. I think human babies are probably the ugliest creatures. All like kittens are cute. Dogs are cute. So, you know, so when I see a baby, I'm not doing baby talks. I usually stay away from them. But I guess what I'm saying is that it would just be a behavior that induced by you see something like, oh, you're so cute. So I find myself doing that to kittens and not to babies.

[00:34:41]  Bruce: He points that out that we will do it to stuffed animals.

[00:34:44]  Unidentified speaker: Yeah. Yeah.

[00:34:45]  Sadia: So it's not that I don't do that talk. It's just that I don't do that to human babies because I don't find them cute.

[00:34:50]  Bruce: But, you know, I don't know what the detailed explanation for why we do baby talk is. That is, and we're probably a long ways off from having a detailed explanation. But it's not too hard to see a high level sketch of how it might happen.

[00:35:05]  Sadia: I mean, it might just be that we see something whenever we find something cute.

[00:35:09]  Unidentified speaker: That's right.

[00:35:10]  Sadia: That's what I was about to say. Maybe just get into that mode, you know.

[00:35:14]  Bruce: We have a certain way that we're predisposed to talk to things that we find cute.

[00:35:19]  Sadia: Yeah.

[00:35:19]  Bruce: It's not too hard to see that the feeling of something being cute would be genetically controlled.

[00:35:24]  Cameo: I wonder though if we, maybe we equate cuteness with the dumbness. And so we're not so much talking to the thing in that voice because we think it's cute, but because we don't think it understands our words.

[00:35:42]  Bruce: That could be. I think that's a fascinating idea.

[00:35:46]  Sadia: Yeah. And there might be this kind of concept of associating it with innocence, which is dumbness. So we're also pretending to be dumb as if when we're in that mode, we're like, oh, you know, I'm so dumb right now, you know.

[00:36:00]  Bruce: And there's clearly a memetic element, right? We've all seen other people do baby talk. So this is the way humans are. You would never find something that's purely genetically controlled other than maybe reflexes. There would always be a strong memetic element in the case of humans. And yet, the reason why this meme continues to survive is probably in part because the genes do give us this experience of cuteness. And for all the reasons you guys just said and maybe others we haven't even thought of, this has become a meme that won't die out and gets reinvented very easily in any society.

[00:36:38]  Sadia: Unless for some weird reason we imagine a culture where it just became, you know, somehow a horrible thing to do to talk to babies.

[00:36:48]  Bruce: Right, right. Where they have like an overriding...

[00:36:51]  Sadia: In our case, the culture can, you know, we can create knowledge at other levels, which can trump over genetic dispositions we might have.

[00:36:59]  Bruce: If like we came out with studies that showed, that were for some reason really good studies, that if you talk, baby talk to your child, then you know, they're going to, something really bad is going to happen to them, right? We could override this genetically channeled learning, right? We could, we could stop doing it because we have a general purpose learner.

[00:37:20]  Sadia: That still doesn't take away from the fact that across the board, we do see that phenomena. There's something to be, you know, questioned there.

[00:37:28]  Bruce: Okay. So now, Deutsch, he, he makes the claim, particularly in the interview with Tyler, which I'm going to quote a lot from in future episodes. He makes the claim that humans can always explain why they do what they do, which is true, by the way, humans can always explain why they do what they do. So he thinks this shows that, that humans always learn via explanation. Now we've just, we've just talked about baby talk. Humans don't even know why they do baby talk. There's a good reason for it because it teaches grammar to children. They don't really know why. So we have an example here that seems like a counter example where humans may have an explanation, but it's got nothing to do with the actual truth as to why they're doing it, right? So can we make sense of Deutsch's claim? What can we say about it? Well, I wanted to bring this up that there was the split brain patient experiments. Are you guys familiar with split brain patients?

[00:38:24]  Sadia: You have to remind me.

[00:38:25]  Bruce: Okay, so in a case where someone has like grand mal seizures, the seizure starts in one hemisphere of the brain and it moves over to the other hemisphere and it takes over the whole body and the body begins to shake. So one of the things that they used to do, or maybe they still do it, I don't know, but they used to split the hemispheres of the brain so that it couldn't move from one hemisphere to the other. And this would significantly reduce the number of grand mal seizures that the person would have. Disney actually had a movie about this, about a kid whose life was being destroyed by grand mal seizures. And so he wanted to get the surgery to become a split brain patient and his parents wouldn't let him do it. So he sued his parents and to have the right to choose for himself to have the procedure.

[00:39:07]  Bruce: And of course, the parents, they weren't being mean or something. They were terrified of the danger of the surgery. But the kid wanted, the kid had done his own research and he had decided that the danger of getting the surgery was less than the danger of not getting the surgery. So he actually thought this through. So I don't know how true, it was a true story. But of course, with Hollywood, you just never know how much of it is actually true. But very good movie that used to be like, be on Wonderful World of Disney or whatever that show was back in the past. And so split brain patients, the two hemispheres can't talk to each other, except a little bit through the lower parts of the brain, like the brain stem. Basically, they can share information about feelings because valences and feelings actually evolved early in the evolutionary process according to our current best theories and are part of the brain stem, not part of the higher functioning neocortex.

[00:39:59]  Bruce: So the two sides of the brain can share a little bit through feelings, but they can't share anything else. Now, a lot of people know about the whole left brain, right brain thing. It's mostly a myth, but there's some truth to it. And there's this idea that one side of the brain is the rational process. The other side is the creative. Most of this is just myth, right? I mean, it's not that easy to split the sides of the brain apart. But one of the things that you can split apart is that typically the left side of the brain is entirely the part that does language, and the right side of the brain has no language at all. Although I understand that sometimes that's flipped in some people.

[00:40:36]  Bruce: So even that's not like genetic destiny, if that makes any sense. So when you have a split brain patient, remember that the left side of your brain controls the right half of your body and the right side of your brain controls the left half of your body. It even controls the left and right field of view in each eye. So they realized with the split brain patients that they could communicate to the right side of the brain that had no language and talk to it because it can like read language, it just can't talk. They can tell it things that then the left side of the brain, which is the part that knows how to talk, doesn't know about because it doesn't get synchronized across the connection between the hemispheres because that's been cut in these patients. They might give a message in the field of view where only the right side of the brain can see it that says, get up and walk across the room.

[00:41:28]  Bruce: The patient will get up and start to walk across the room because it saw this message. Then they will ask the person, why did you just go get up and walk across the room? The language center doesn't know the answer. What they do is they confabulate an explanation on the spot and they're convinced it's the truth. They'll say something like, oh, I wanted to go get a coke. Everybody except the left side of the patient's brain knows that's not true. Yet it just basically makes up an explanation on the spot. This is why I'm a little bit concerned about Deutsch's claim. Well, humans, you can always ask them what they're doing. They give explanations. Thus, we always learn via explanations. That might be true. I don't know. But I don't think we can say that for sure because of split brain patients. We know that humans confabulate explanations on the spot, always do.

[00:42:22]  Bruce: The explanation that a human gives for why they did what they did could be the real explanation, or it could be something that they are making up right now, but they think now it's the real explanation. And so that does make it more difficult to try to test. Well, do humans always learn via explanation or not?

[00:42:41]  Sadia: I guess one of the things is that we're also good at fooling ourselves sometimes too, and we don't even know sometimes. And we don't even know, like sometimes there could be some sort of a value operating at a lower, you know, some sort of a level that might do that, and we're just not aware of it.

[00:42:59]  Unidentified speaker: Yeah.

[00:42:59]  Bruce: Well, one of the things I think we know at this point, and this is a little hard for people to accept, but I think it's the truth, is that you don't really know what your own motivations are, because those are all subconscious. Not all of them, but most of them are subconscious.

[00:43:09]  Sadia: Although to be honest with you, you know, the more introspective, like the more you sort of think about why you do certain, you can dig down pretty deep within yourself. But from the example you gave, you do wonder if there are certain things that there might be some sort of a roadblock.

[00:43:28]  Bruce: Well, I agree that you can dig down inside yourself, but have you ever noticed how you actually do that? Because maybe you're different than me. But the way I assess my motivations and dig down inside myself is I sort of look at myself from the outside as a separate person and then criticize myself. And I go, OK, I don't think the motivation I'm claiming I have could be the right one because of the following reasons. This refutes that explanation. Therefore, it seems like my real motivation was that I was greedy or, you know, whatever. Right. And I think when you dig down into yourself, it may not actually be by digging internally. I mean, there's probably cases where you can. We don't know how the brain works entirely. And we do have access to some internal information about what's going on or consciousness wouldn't exist. And yet, even when you do dig down, a lot of times you may be doing it as a third party, coming to understand yourself as a third party.

[00:44:20]  Bruce: So, again, it's not clear. Right. I mean, like we're a long way.

[00:44:24]  Sadia: Deep, deeper. What do you mean? Like, for example, you know, sometimes I'm really irritated and I'm coming, you know, I'm doing something and I notice it. And then I ask myself what could be the reason. It could simply be just because I mean, I'm literally I'm going to sound sexist. I'm having periods. Sometimes that really is one of the, you know.

[00:44:44]  Bruce: But notice, notice that what you're doing there is you're assessing yourself as a third party. You're saying, I'm feeling irritated. OK, why am I feeling irritated? You know, your your husband could have done that. Right. And he may have even come to the same conclusion using the same logic.

[00:44:58]  Unidentified speaker: Right.

[00:44:59]  Sadia: Yeah.

[00:44:59]  Bruce: So that's what I mean, is that it's not always clear if you're introspecting or for lack of a better term, extrospecting. Right. Is that a word? Did I just make that word up?

[00:45:09]  Sadia: So what would be the difference? In the other case, I'm trying to see what I'm not even sure what the other thing.

[00:45:15]  Bruce: OK, so the introspecting might be, I do feel irritated. Like, you can tell you feel irritated, but not everybody knows that they feel irritated. Sometimes irritation is very conscious and sometimes it's not.

[00:45:29]  Sadia: But I think even that can be because we experience so many different states. You know, I feel like using that as a background, you can. Because one of the things I've experienced is that sometimes I don't know when I'm stressed out, or maybe even like at times in my life when I was depressed, it just manifested differently, that I was not happy about circumstances. But I think I feel like I've been able to see through that. But it's sometimes it takes time.

[00:46:01]  Bruce: I had an interesting experience. I took some drugs to test to see. I thought my wife, let's be honest, my wife thought I was ADD. I've worked with Cameo, maybe I am ADD or something very similar to that. Cameo, don't say anything. Actually, go ahead if you want to.

[00:46:20]  Cameo: I honestly wouldn't call you very ADD. I think you're on the other side of the spectrum.

[00:46:29]  Bruce: Fair enough.

[00:46:31]  Sadia: [Unclear] the same way about myself, by the way, but anyways.

[00:46:35]  Bruce: I went to a doctor and I said, can you give me something for ADD and we can see if it works or not. He decided to do that. He gave me this drug. I was looking at the side effects and it says, it'll make you feel nervous. The next day, I work as a project manager for software, which is a very nerve wracking job sometimes. I'm feeling nervous constantly. What I do when I feel nervous as a project manager is I have to stop and ask myself, why am I feeling nervous? Because that's usually a sign that I've missed something, that some subconscious part of my brain has-

[00:47:08]  Sadia: Something unusual, right?

[00:47:09]  Bruce: Yeah, there's something I need to address. I've forgotten it. I've missed it. There's a pattern I've missed. There's some part of me that's not conscious, that has caught it, and the conscious part of me hasn't caught up yet. I'm spending the whole day trying to figure out what's wrong with my project, when it finally occurs to me that I'm taking a drug that makes you nervous.

[00:47:29]  Sadia: In a way, I guess because we all experience all this variety of things, sometimes we're stressed, sometimes we're not, sometimes we're feeling ecstatic, I think that kind of gives us a background where I think there is something like... The only thing is that sometimes, like I said, in my case, I've noticed that the stress would manifest itself in ways that I couldn't even tell that that was stress. I was like, yeah, I'm not stressed out. But I've questioned that, was that really stress or...

[00:48:02]  Bruce: By the way, they've done tons of studies on this. This is an interesting subject of itself, which we probably should move on. But for example, some of our emotions seem to be built in, but some of them seem to be labeled consciously. So there's the same affect. I can't remember an example. I want to go find one of the examples I've read in the past. But the affect that you would call stress, and the affect that you would call nervousness, and the affect that you would call something else, they might actually all be physically the same thing. And the only difference is that you label them differently based on context. And they've done experiments to prove this. But there's other things like anger is never seen as stress or something like that. There's some that seem to be built in, and there's some that you have like one that counts for several different emotions. And then the difference between the labels is conscious.

[00:48:56]  Sadia: I remember coming across that there was one that was associated with feeling love, and I was surprised. Like they almost seem like pretty, quite the opposite. And as if the body just, you know, if you look at how the body response seemed like, it was the same, right, which was weird.

[00:49:16]  Bruce: So now in the last episode, we talked about the idea. So Bart had raised the idea that maybe David Deutsch was utilizing the word knowledge in a specialized sort of way to mean specifically the open-ended sort of knowledge. And I considered that possibility. I admitted that's a possible way to understand the word knowledge. There may even be some advantages to thinking of it that way, but that's not the way I happen to use the term. I wanted to clarify this a little bit. So there's an idea of essentialism, or sometimes I call this word essentialism to make it a little bit more clear. It's very tempting to equate a word with a concept because we often think in terms of words. We don't always think in terms of words, but we often do. I have been in so many debates with people over whether or not AI algorithms create knowledge. You know, whether or not knowledge is only limited to biological evolution and human knowledge, which is what Deutsch says. Is that true or is it false?

[00:50:13]  Bruce: In a lot of ways, it's a dumb debate, okay? Because the word knowledge is just a word. It has a very long history that goes back a very long time. And there's no chance it hasn't been used in lots of different ways, you know, over time, right? So to some degree, when we argue over does this create knowledge or not, we're being dumb. We're just being essentialists. And I know I have engaged in that a ton. So, you know, my apologies to those that I've argued with people over this. That's not really what we're going to care about.

[00:50:50]  Unidentified speaker: Okay.

[00:50:51]  Sadia: But just to kind of point out that, so because Deutsch has written so much about it in that book and said even stuff on Twitter, you can understand what he means when he says knowledge, at least, right? It's not like, you know, when, for example, when if Bart said that I think Deutsch means this, I think it's because he's picking up something in that book that suggests that, you know, what...

[00:51:15]  Bruce: Yes. Now, here's the thing, though. Having said that, not all definitions are created equal. A definition is in some sense a tacit theory. So we should allow people to use the words anyway they want, but we should criticize the underlying concept. Now, that's really what I'm trying to do with you. I actually do believe his choice to, I'll give a quote here in a second that I think confirms Bart's theory, by the way. I think his choice to limit the word knowledge the way he does is a mistake. Not because it's wrong. There's no way that could be possible. But because it is going to almost inevitably lead to confusion. So I really want to keep my criticisms around the concept, not the word knowledge, if that makes any sense. So, you know, is the output of learning knowledge or not? Okay. We've been talking about learning and haven't, I've maybe been using the word knowledge.

[00:52:09]  Bruce: I don't even remember if I use it or not. Is the output of learning... [unclear restart] ...knowledge or not? It doesn't matter what you call it. Okay. But if by knowledge, you mean adapted information that was created via a Popperian Epistemological process, then without a doubt, classical conditioning and trial and error learning creates knowledge. So does machine learning. If by knowledge, you mean adapted information that keeps itself instantiated. Now, that is what Deutsch puts as his definition of knowledge in Beginning of Infinity. Then this is identical to the definition above because the output of a trial and error algorithm is always tautologically the item that kept itself instantiated. Now, maybe that's not what he meant. Maybe he didn't mean what he actually said, and he needs to qualify that further. But in terms of what he's actually said so far, animal learning produces knowledge. So does machine learning. Period. End of story.

[00:52:58]  Bruce: Now, knowledge may mean something specifically like human-level explanatory knowledge. In which case, I have no doubt that animals never create knowledge because they do not create human-level explanatory knowledge. Nobody is in doubt over that. But biological evolution doesn't either. If we're going to say animals don't create knowledge because we mean human-level explanatory knowledge, then we need to also say biological evolution doesn't create knowledge either under that particular definition of knowledge. I'm more worried about the consistency than I'm worried about what the actual definition is. In other words, if by knowledge you mean something that tautologically doesn't include animals, then yeah, tautologically doesn't include animals. So what? No big deal.

[00:53:42]  Bruce: Now, let me make this last point though. If a lay person, if you were to ask them, what do you call the product of learning? My guess is that they'll think around and they'll come up with the word and that some large percentage of them will say, oh, that's knowledge, that the word is of what you produce from learning, it's knowledge. So the word knowledge has this penumbra of different possible meanings. That's how all words in the English or in the natural language are. And using it as what we call the output of learning is already in use, widely in use, and trying to fight it doesn't make sense. This is why I'm going to just call it knowledge. I'm not trying to have a beef with people who want to call it something different. If I'm going to have a beef with them, it's going to be over the implications of their theory of choosing to [unclear verb] it that way.

[00:54:33]  Bruce: OK, but this is why I'm going to just continue to say, yes, machine learning creates knowledge. I only mean it in this very loose sense that I'm using it, right? I'm typically thinking of it's using a Popperian process to create solutions to a problem. And yes, we know by observation that if you define knowledge in that way, that machine learning and artificial intelligence and animals all create knowledge. OK, off my soapbox on that one. Now, why is this related to Bart's theory that Deutsch was reserving the word knowledge for the two open-ended creation processes? I came across a quote that really seems to support that theory. So before, I was just sort of entertaining it for the sake of argument. And now, once I came across the quote, I'm like, oh, I probably ought to admit that Bart was probably right about this.

[00:55:23]  Bruce: So Deutsch says genetic programming, he doesn't actually say genetic programming. In context, [that] is what he’s referring to, though. Genetic programming may produce learning in that perhaps any useful change can be considered learning. But it is not artificial evolution. I think that all the learning and creation of new knowledge in the robot happened in the mind of the graduate student that created the evolution. This is that example of using artificial evolution to get your robot to walk. Okay. And Deutsch in Beginning of Infinity makes the argument that there's no knowledge created in that process. Okay. Now, what does Deutsch mean by it is not artificial evolution? Well, he can't mean it isn't a trial and error process, because by observation, we can see that it is. He can't mean that it isn't analogous to Popperian epistemology, because by observation, we can see that it is.

[00:56:15]  Bruce: It tries variants, it discards the less good ones until it finds the best variant that it can improve upon. He can't mean that the output isn't an example of adapted information keeping itself instantiated, because by observation, we can see that it is. It tries different programs and the genetic programming that are in the population pool of programs. They keep the useful parts and the parts that aren't as useful, they die out in the population until finally you end up with a collection of useful parts that do the trick of actually walking. That's how genetic programming works. If you don't believe me, you may have to go learn this, but that is precisely how it works.

[00:56:53]  Bruce: So this qualifies as knowledge creation under any of those three definitions, including the last one, which is Deutsch's stated definition for what knowledge is. Okay. Now, Deutsch goes on and he says, and the token that that's the case, that it isn't artificial evolution, is that once it has walked well, as I described, it never walks any better. That particular program will never, for instance, learn to walk upstairs. It's not because the hardware can't do it. It could be programmed to do so. It's just that the evolutionary system can't reach that particular knowledge. Okay. And I put some emphasis on that last statement. It's just because the evolutionary system can't reach that particular knowledge. Deutsch, it really seems to me that Bart is correct. Deutsch is reserving the word knowledge in the way he's using it in this statement for only things that were created via the one of the two open-ended knowledge creation processes, biological evolution or human knowledge.

[00:57:50]  Bruce: I don't know how else to read it. Like if you can tell me some other reading of this that isn't that, I can't see what it is. Now, let's assume that's what he means. Then understand that what he's really saying here is, the token that genetic programming today isn't open-ended like biological evolution, is that once it has walked, it never walks better because its evolutionary system can't reach that particular knowledge. Well, that is a true statement. No one doubts that statement. In fact, it's trivially true. I think that's what he's actually saying. I think the debates over what is knowledge, what isn't knowledge, I think they're a waste of time. Deutsch is saying something that is true. He's pointing out that genetic programming, artificial evolution in algorithms, isn't open-ended like biological evolution. That's what he's pointing out. He's not making a different point. Everyone agrees that this is not a point that needs to be debated. Do you follow my argument here?

[00:58:44]  Bruce: You see what I'm trying to say?

[00:58:45]  Unidentified speaker: Yeah.

[00:58:46]  Bruce: There's even a term for this. It's called the problem of open-endedness. I've mentioned this before. It's a, I wouldn't say well-known. It's a not well-known scientific problem that there are people who do a really good job working on this and they've got interesting stuff to say about this. Okay. And then just how I'm conceptualizing this, I'm differentiating between the knowledge itself, which is the solution to the problem and the process by which the knowledge was created. I'm saying that there are narrow evolutionary processes that have very small range of what they can produce. They will produce knowledge, but they only do within a small range. And then there's open-ended evolutionary processes that can, in theory, produce maybe anything, particularly in the case of humans. I don't know if biological evolution could produce jet packs, but have a much broader range of the types of knowledge that they can produce.

[00:59:38]  Unidentified speaker: Basically, they would also produce new problems as well.

[00:59:41]  Unidentified speaker: And then, yeah,

[00:59:44]  Sadia: it's like, it's almost as if like, we're not just trying to solve some particular optimization problem. It's a type of problem where, yeah, I mean, there may be certain things which might get optimized, but that's not where it ends.

[00:59:57]  Bruce: Right, right, right, right. So you're bringing up kind of a high level theory as to what the difference between narrow evolutionary processes and open-ended evolutionary processes are. And I agree with everything you just said. But you saying that isn't actually specific enough to start programming. And that's the problem, right? Is that we do know that there are differences between narrow evolutionary processes and open-ended processes. For one, we can by observation see that biological evolution doesn't have the same limits as the artificial evolution of a walking robot, right? We can by observation see that that's true. But we can also criticize the robot algorithm. We can say, look, there's no chance it's ever going to produce anything except exactly what it was meant to produce. Because of the very narrow range in which it's going to be exploring possible programs, okay? We even have an idea of how to open the range a little bit, but it's always just a little tiny bit more.

[01:00:51]  Bruce: You could probably figure out a reward system that allows it to walk upstairs instead of not walking upstairs. And that would be hard, but we could probably do it, right? But there's no way it's going to then turn around and start playing Go. And yet, we don't really have a clue how to go about programming it to just be open-ended. That's the problem. The problem is that when we talk about open-ended evolutionary processes, we're not even really sure what we mean by that. We have a really good idea of what a narrow evolutionary process is like and how to build one. And those do create knowledge. They create the output exactly the same as an open-ended one, right? So getting back to the example of the immune system, David Deutsch specifically says the immune system does not create knowledge. And yet it is clearly a trial-and-error evolutionary process. Okay. And it,

[01:01:44]  Sadia: I'm sorry.

[01:01:45]  Bruce: Well, I was going to say if it produces an antibody, trying to say that isn't knowledge, but the same antibody and maybe identical antibody produced by scientists is knowledge. To me, that's just confusing everything. What we should really talk about is, yes, they're both knowledge because they're identical. And instead they were created by different processes. One was open-ended and one wasn't. I think that's a much more clear way of talking.

[01:02:09]  Sadia: I was going to say that going back to the whole thing about animal learning, if you look at life as a process, not just one species, not focusing on one particular organism, but just overall the evolution of life, that's indeed open-ended. But when we look at an organism, a particular one, within its lifetime, the problems it's solving would fall in that narrow category.

[01:02:34]  Unidentified speaker: That's correct.

[01:02:34]  Sadia: It is learning, but it's just that it's restrictive. But given enough time and a whole population of those organisms will keep learning, there will be an open-endedness.

[01:02:47]  Bruce: Let me make a point here because as someone who's studying artificial intelligence, I find this fascinating. Let's use the example of the walking robot. It won't learn to walk up the stairs. In fact, if you change the reward structure, it may learn to walk up the stairs. It's just a lot harder to try to figure out a good reward structure for that. A dog can learn to walk up stairs. Stairs don't necessarily exist in real life, at least not specifically stairs. A dog can learn to walk up stairs. Its learning algorithm is far more open-ended than anything we currently do in artificial intelligence. That's just a fact. A dog can learn to play Jenga as we talked about in the last episode. It's not clear what dogs can and can't learn because there's clearly things they can't learn.

[01:03:35]  Bruce: There are limits. They're not open-ended in the same sense that we are. It's hard to say what the limits of dog learning are because their algorithm is so fascinatingly powerful compared to anything that we know how to program. I think we're going to have to accept that this is a lot like computational theory. Computational theory, there are at least three different levels of universality that exist. In our computational theory episodes, we talked about this. We also talked about this in the interviews on AGI. Thatchaphol talked about this in the interview. Sometimes we're hesitant to refer to lower levels of universality as universality. It maybe violates our intuitions a little. If the Turing machine is universal, then a finite state machine is not universal. But finite state machines are actually universal just at a lower level.

[01:04:27]  Bruce: So Deutsch talks about this in The Beginning of Infinity, that universality exists in a hierarchy. That you can have a certain level of universality, and then you can go to a level above that. So for example, using text versus using dot matrix. There actually are levels of universality. And then eventually you may top out, like with the Turing machine. That's the top level. There isn't one above it. Or at least we believe there isn't one above it.

[01:04:58]  Sadia: Would this be a good example to say, for example, that hypothetical paperclip maximizer has a type of universality, but it's not the open-ended one. But all it ever does is that it can just, given the raw materials, I guess, it can create. I don't even know if that's a good example, actually.

[01:05:15]  Unidentified speaker: I don't know.

[01:05:16]  Bruce: Yeah. You're trying to get at something that is clearly true, but it's hard for us to word. Animals exist somewhere in between the super narrow knowledge-creating algorithms that we know how to make, but apparently well below what humans can do.

[01:05:30]  Sadia: So it seems like the algorithm actually has, there are things that there is a room to play around that their algorithm gives them. And they may not have fully explored it till the right conditions, that environment somehow, for example, a human teaching that bird that you're talking about to sing backwards, right?

[01:05:53]  Unidentified speaker: That's right.

[01:05:53]  Sadia: So something has to activate that or something in the environment has to, but the potential is there, but it's still a limited one.

[01:06:03]  Bruce: Byrne uses some fascinating examples of this with the great apes. The great apes have a great deal more capacity than they actually use in the wild. Chimpanzees are the exception. Chimpanzees use tools in the wild, whereas like great apes don't, but they will in captivity, right? There are, there's a mental capacity that the great apes have that puts them on very similar levels as chimpanzees in the wild, but they never use it in the wild, okay? And he gives all sorts of fascinating examples of this. The whole sign language thing, the fact, and we talked about this with Koko, that Koko couldn't pass the Turing test. The very fact that they can learn sign language at all and communicate at all is not something that they can do in the wild. In the wild, they actually use genetically pre-programmed gestures, and they never make up new ones, ever.

[01:06:54]  Bruce: And yet in captivity, they can learn ones that aren't genetically pre-programmed. So their learning algorithm has capacities that they will never explore in the wild. Does that make sense?

[01:07:07]  Sadia: Yeah.

[01:07:09]  Bruce: And that's fascinating, right? This fits perfectly with Popper's theory that when you have a better learning algorithm, evolution can start taking advantage of it physically as well. Okay. All right. So we're almost done. Let me just kind of a hodgepodge here, but let me one last thing.

[01:07:24]  Sadia: Just add one other comment. Thinking about what Deutsch has normally said to me, it sounds like, because he'd almost say that if, let's say, there was an algorithm that allowed all sorts of different possibilities, even if nobody ever ran the algorithm, it seems to me that he thinks that the knowledge is there, right? That knowledge is there. So if you run it, he would say that no new knowledge has been created. Let's say you run an algorithm and there is some output which does something, right? So I feel like it's almost as if he almost takes information, like if the information which is basically, you know, as something that's already there in the program, and whether somebody runs it or not, he's considering it as knowledge. So [a book] also contains knowledge, for example, right? I've always wondered that if there was a book that nobody ever read, right? I mean, how can we even say that that has knowledge?

[01:08:23]  Sadia: Or what if suddenly all humanity died out? Would that book still have knowledge? I mean, is there any objective sense? Isn't the environment important? Isn't it important that whatever is in there somehow interacts with the environment and something is created, like the process of creation? Isn't that important for it to qualify as being knowledge?

[01:08:44]  Bruce: Okay, there's a lot there to unpack. Let me, and there's more than one thing. You've actually mixed several things together that need a little bit different treatment each. So, but these are fascinating questions. Let me take a stab at these. First of all, does Deutsch think of the knowledge as pre-existing if it's part of its repertoire? You didn't use the word repertoire, but I think that's what you're really getting at. You know, I originally did not think that was the case. And you and Bart have been arguing this to me for a while that in fact, this is what he is arguing.

[01:09:16]  Sadia: Oh, I pretty much, this is like, I don't even doubt it anymore. I think that's pretty much what he means. Like he says books have knowledge, right?

[01:09:22]  Bruce: Yeah. So you've convinced me. I think I had misunderstood Deutsch on this. Now, is this correct? No, I don't think it is.

[01:09:31]  Sadia: Yeah, of course.

[01:09:33]  Bruce: It's I think that you cannot think of the repertoire of a learning algorithm as pre-existing knowledge. I think that that would be a very poor way to try to conceptualize the problem space. Okay. So, yes, I now convinced that that is what he means. I'm, you know, 90% sure or whatever. I'm not really still completely sure. And I think that I've been arguing with you guys. No, that's not what he means. And I think I'm wrong. But I don't actually agree with that viewpoint. I think that that leads you to conundrums that are impossible to solve. So I think it's just-

[01:10:07]  Sadia: And you know what's the difference there? So if there is something, unless that thing actually causes some physical transformations in the world, I feel like at least what I'm thinking is that it may not qualify as knowledge. It actually has to cause some physical transformations. In that process, then we can say, oh, okay, now we can say that has-

[01:10:28]  Bruce: So now let me disagree with you on that one. So this is the second part. And I understand where you're coming from on this one. But this is one that Popper has covered extensively. Okay. So I happen to know what Popper's answer to your statement is. You're familiar with Popper's Three Worlds, right?

[01:10:45]  Unidentified speaker: Yeah.

[01:10:46]  Bruce: This is why he invented the Three Worlds, was to try to answer this very question that you just brought up. And also think about our interview with Chiara Marletto. This is why Deutsch created the Constructor Theory. He's trying to make sense of this question in an even deeper way than Popper did. And he's using Constructor Theory to try to answer it. So both Deutsch and Popper would disagree with you on the idea that a book that is never read doesn't contain knowledge.

[01:11:13]  Sadia: Exactly. And I know that. And I have actually changed my stance on these things. I used to agree with them, but not anymore. But yeah.

[01:11:19]  Bruce: Okay. So let me make an argument that they're right about this. To try to understand knowledge in terms of a conscious mind has to, or maybe we can say an animal mind too. I don't know if we count that as conscious or not. A mind has to process it. I think that leads to all sorts of conundrums that just don't make sense.

[01:11:42]  Sadia: I don't think it has to be a mind. All I'm saying is that like, you know, the new concept...

[01:11:48]  Bruce: You're saying, you're saying, oh, I see what you're saying. You're saying just it has to create some sort of physical transformation.

[01:11:52]  Sadia: There has to be something happening in the world for it. We can't just say that it's just by mere, you know, I think the thing, the way Deutsch thinks about it is that information, you know, this whole idea that information is physical. Yeah, true. I mean, like, for example, a seed, right? A seed has to potential, has the potential to grow. So there is definitely there. There is DNA and stuff like that. I'm not denying that. But I don't know. I really struggle like to when you go to the point of saying a book, you know, for book, a reader is important. If there is nobody to read, if the language is lost, book is just nothing without a preserved language, right?

[01:12:36]  Bruce: You're in the Four Strands email group, right? Vaden in that group, he actually pointed out to me that Popper's Three Worlds is basically what you just said, slightly different. I'll explain how it's different. That what he's saying is, is that the first world is the physical world. The second world is subjective knowledge, which we've been debating. I was like debating with Brett Hall what subjective knowledge actually means, and Popper's not super clear. But just based on historical context, I'm fairly certain that just means knowledge contained within a brain or an organism. The third world is the objective knowledge. It is the books, the theories, the things that, you know, humans or possibly animals create that contains knowledge. He says that the worlds have to interact through each other. So the third world can't change the physical world directly. It has to work through something in the second world.

[01:13:31]  Bruce: So he sees the world of objective knowledge as interacting with the first world through the second world. OK, so what does this have to do with what you're just saying? So he's actually trying to come up with a criteria for reality, according to [unclear]. He's trying to say that what makes objective knowledge real is the fact that it has the potential to change world one, the physical world, through the second world. So his criteria for reality is actually the same as what you're talking about. Now, here's where you differ from Popper on this. He considers a book not read to still be objective knowledge because it could affect the world. He uses the example of aliens coming after the extinction of humanity. And, you know, maybe through archaeological means or whatever, they learn to translate the book, and it then affects the physical world.

[01:14:18]  Bruce: So he determines not whether it did affect the physical world, but whether it could have, which is really just constructor theory. Okay, let me just, two last things that are a setup for next time. So we're going to talk about animal memes. And I'm kind of just going to act like it's a given that animals have memes. Now, Deutsch talks about the existence of animal memes in Beginning of Infinity. So this is not something that I would anticipate the Deutsch fan community to have problems with. However, Byrne makes a distinction that I feel like needs to be brought out. When we talk about animal memes, it's not really clear what we mean in a lot of cases. So Byrne studies imitation. And animal imitation for a long time was sort of a given. We assumed animals could imitate. And we had tons of examples of animals imitating.

[01:15:11]  Bruce: All the way down the biological chain, the tree of life. And so it was just sort of a given that animals could imitate. And that was where animal, the term meme didn't exist back then, but that's where what we would today call animal memes came from. Imitation was actually thought of as something that kind of ruined your ability to see if animals were intelligent. If the animal actually just learned by imitation, then it wasn't actually intelligent. So goes the thinking back then. So it was kind of seen as a bit of a spoiler. Well, it turns out that science has changed its mind on this subject since then. According to Byrne, he quotes all the sources and everything. And they realize now that imitation is incredibly rare in the animal kingdom and in fact is a fantastically strong sign of intelligence. And so it's almost like it's entirely flipped on its head at this point. So if imitation is really rare in the animal kingdom, then how do animal memes exist?

[01:16:16]  Bruce: Well, in some cases, for the animals that actually can do imitation, it may be imitation. But there's a second way that you can do animal memes that's got nothing to do with imitation and therefore may not even feel very meme-like once you deconstruct it. So basically what you do is you imagine an animal by trial and error learning discovers a useful behavior at a certain spot. So it's out, it's using its regular trial and error learning algorithm, the instrumental learning that we talked about earlier, and it discovers something useful. So let's say that others in the same species didn't learn that behavior [wording uncertain] when previously none knew it. Now there's tons of examples of this. There's birds where one bird pecks through to get to the milk, delivered by the milkman, and within a few days every bird is doing it. No bird was doing it before. So we say, oh, they all imitated each other.

[01:17:13]  Bruce: Now perhaps the other animals, what Byrne points out, is they've actually figured out ways to experiment, to see if that's actually what's going on or not. And what they would find is that a bird that had never seen another bird peck through to the milk or whatever, if it was given a chance, like put in a place where it had to learn it on its own, its chances of learning it were exactly the same as one that had seen a bird do it. So that really called into question if the birds were actually imitating. When their meme passes, are they actually imitating each other? So what they do is, I'll talk about Byrne's methodology in the next podcast episode, which is really important, but he always assumes that it's the worst case scenario. And then he tries to find examples that aren't the worst case scenario.

[01:18:01]  Bruce: So if an animal, imagine that an animal could learn, you know, each animal that picks this up that you can poke through the... and get to the milk, imagine that each of these animals are actually just learning it on their own through trial and error, but that the animals have an attention process, the genetically channeled learning that makes them think, oh, I saw another member of my species over next to that milk, so that I should probably check that out. It's got no understanding of why it should. Okay. But it flies over and it tries to find the same place. And then it does its own trial and error learning, and it figures out, oh, I can poke through the milk.

[01:18:42]  Sadia: So you're thinking in a way if we just made the milk bottle available to any other bird given in a confined space, it'll just learn on, you know.

[01:18:50]  Bruce: Right. So what's really, so what's really going on is that the animals don't have any real understanding of that bird got milk by poking through, so I can too. They really just have an understanding of, you know, hey, I should check out anything I've seen another animal check out.

[01:19:10]  Sadia: Any potential thing that might be, you know.

[01:19:12]  Bruce: Right. And then it does its normal exploration, playful process, curiosity. And maybe the odds go up if it sees a hole, right? That may be the case. It may be that so they like did experiments. What if it has a hole? So yeah, it would actually learn it faster if it saw a hole in it. That makes sense. By these means, animals can transmit memes. And they're not actually learning memes directly from each other. They're actually just have a setup in their mind for the genetically channeled learning so that they have a high chance of rediscovering the same knowledge that their fellow members of the species have picked up on. And this is still, I'm going to still call this a meme. Okay, because that it is still a meme. Yeah, they're only really using their learning processes. They're not actually learning from each other.

[01:20:02]  Bruce: To actually learn from each other social learning, that does require a great deal more intelligence than most animals have. So let me give you a couple of examples of this. Here's one that comes from Nicholas Christakis. A group of macaque monkeys learned from one another how to run a racket with visiting tourists. They will swoop down and steal hats, glasses, and cameras and so on and return the item only if bribed with food. It's a highly distinctive, if not unique, other species don't do this. And it's clearly a socially transmitted behavior. Scientists studying the site noted the members of the new group of immigrant macaques began to see that they could, too, [unclear verb] stolen goods for food. This really looks like imitation, but Byrne would almost assuredly say, well, actually, you could explain this through simple trial and error learning and a few simple rules.

[01:20:51]  Bruce: So he probably would rule this one out as being true imitation. Now listen to this one. Indeed, it seems that, another Nicholas Christakis example, it seems that the elephants have learned to raid in the middle of the night, especially on moonless nights. So elephants will raid human settlements for food. And to form bigger than usual groups for raiding parties. Elephants also manifest other qualities we associate with efficient learning. They seem to put more credence in the behavior of peers who are deemed more reliable, such as older, more experienced elephants. And they even put more credence in the raiding strategies that they observe to have been adopted by multiple contacts. OK, this one's a lot harder to try to explain without intelligence. And yet you probably could still explain it. Maybe a few elements really challenge me.

[01:21:36]  Sadia: But one comment on the whole imitation thing. So even if he said that, like in the case of those birds pecking the milk bottle, I mean, obviously, even if even if the birds, even if he said for a second that they were imitating, when did the very first one, didn't the first one actually learn that? Or I mean, what did they imitate then?

[01:21:56]  Bruce: It didn't imitate it just by its own trial and error process happened to figure it out.

[01:22:00]  Sadia: So but but that just one case alone shouldn't that also qualify as just, you know, in support of that? Even if he said, OK, well, the other ones just ended up learning from its trial and error.

[01:22:12]  Bruce: Yes. What you're trying to explain is why there's this sudden every animal that in that area suddenly learns that behavior.

[01:22:20]  Sadia: Yeah. I mean, why did even any of them ever learn to peck a milk bottle to get the milk? Anyways, there was some some somebody had to take the first step in, you know.

[01:22:31]  Bruce: Yes. Byrne's not denying that. Right. We can explain the first one through trial and error. That's not hard to explain. The question is, can you explain the sudden explosion afterwards through trial and error? Also, he's saying yes, that you can't explain it [wording uncertain], that you don't have to explain it as true imitation. OK. He's saying that trial and error plus a simple rule.

[01:22:54]  Sadia: ...going on if every—if every bird did the trial and error. But to me, it seems like just one case should be enough to say that there is something going on there. They're learning from trial and error.

[01:23:05]  Bruce: Oh, yeah. Keep in mind that Byrne has no doubts that animals learn. The only people in the world that have doubts that animals learn are fans of David Deutsch. Like nobody challenges that. It's well known that animals can learn, right? What we're really talking about this case is not whether they can learn or not, but whether they can imitate or not. Now listen to this one. One group of chimps developed a useless practice of putting a long blade of grass into one ear like an earring. This appeared to serve no practical purpose and was thus akin to a fad in humans. Starting from one inventor, it spread to seven other chimpanzees. Noticeably, this spread was observed in only one of the four groups of chimps, which were all isolated from one another. You cannot explain this one through the regular trial and error learning process.

[01:23:55]  Bruce: The first one learned it by trial and error, but the odds that there would be suddenly seven of them that do the same thing when it has clearly, it does not have any survival value. There's no benefit to this. This is literally chimp fashion.

[01:24:09]  Cameo: Was there any observed changing in social status by the initial chimp? Because that's why humans do fashion.

[01:24:21]  Bruce: Yeah, there wasn't. They just seemed to enjoy it. And so they started to do it.

[01:24:26]  Sadia: I mean, why can't it be just both? I mean, it's like, yeah, sometimes, I mean, it's not like their imitation could be one form that you learn. I mean, humans do too.

[01:24:35]  Bruce: I mean, so this example can't be explained. This is an example of imitation that can't be explained by, or at least we don't know how to explain it, by a simple rule plus trial and error. So in this case, we would say that the monkeys probably learned it by trial and error because we always assume the worst. We'd probably say that for the elephants, although there are other examples of elephants where Byrne thinks elephants actually can learn through imitation. And then we'd say, it looks like we have some evidence that chimps can actually learn through social imitation. So this is all just to clarify. When I talk about animal memes, I'm not necessarily talking about the kind of chimp intelligent kind. I'm probably mostly talking about the unintelligent bird poking a hole to get to the milk kind. But I'm usually just not going to differentiate. If there's a meme, there's a meme. That's it. Now, one more thing.

[01:25:32]  Bruce: We're done. But Cameo, in the last episode, we talked about the monkeys that get jealous because they see another monkey get paid a grape and they're only getting paid a cucumber. Do you remember that conversation?

[01:25:45]  Unidentified speaker: I do.

[01:25:46]  Bruce: You challenged it and you said, well, I'll bet you that they do that, whether there's another monkey there or not. So I decided to go look up the actual study. Frans de Waal is the one who did the study, by the way. Do you know who Frans de Waal is? He's like one of the most famous.

[01:25:59]  Unidentified speaker: Yes, absolutely.

[01:26:00]  Bruce: Yeah. Frans de Waal is a really smart guy. So he absolutely did think of that. The experiment was controlled where they did several different groups. They did with this other monkey there. They tried doing it where they would pay a monkey that isn't there. So they would just simply put the grape there. And then they tried doing it where there's another monkey that's not there. And they just tried the cucumber or the grape. In every case, the monkey could see that grapes were available. So they kept it the same for all of them. What they found was that, so like in the case where the monkey can simply see that there's the grapes, but there's no one else to be jealous of, the monkey might initially refuse the cucumber in hopes of getting the grape. But if it doesn't change within a couple tries, a few tries, it will go back to eating the cucumber. So you're kind of were right.

[01:26:50]  Bruce: They'll give similar behavior, but they'll adapt back if they can see that the grapes not coming. But if they have another monkey there, the odds of them just refusing to eat the cucumber go way up. It's different for different monkeys. So they have to do this statistically over a population. That's why Frans de Waal saw this as some sort of jealousy going on and interpreted it as, you know, the whole, well, I don't know if he did or if the media did interpret as the whole equal pay thing. And that's how it kind of got caught up in popular culture. So they did actually test that. So we know that there's something more going on. So they did test the difference with a control group. The monkeys do react differently if there's another monkey there that's getting paid differently than them. This would seem to support the whole, at least this supports the idea that the monkeys are actually feeling some sort of jealousy and that's why they're behaving differently. Okay.

[01:27:46]  Bruce: Even though it does make sense, you were right that they will give similar behavior at least for a short period of time, even if there's not another competing monkey there. Competing is the wrong word. There's another monkey getting paid differently. But they will not, they will switch back to accepting the cucumber if there's not another monkey where that doesn't tend to happen if there's another monkey who continues to receive the grape. So that is the outcome. Now, here's the thing that's interesting. In the popular media, this got picked up as a whole equal pay thing, right? So one of the researchers who read the study, he wasn't really convinced that this showed that monkeys cared about equal pay. Now, he's not challenging that it's jealousy.

[01:28:26]  Bruce: So he's not challenging that, but he's like really not convinced that this has anything to do with human equal pay concerns. So he did a study with humans doing exactly the same thing. And it turns out, I mean, they found some other reward system that humans would care about. It turns out that the humans would get upset if they weren't getting equal pay, but they would accept the lower pay because to give up the lower pay is to make the inequality even worse.

[01:28:55]  Cameo: Oh, interesting. That is interesting, actually.

[01:29:00]  Bruce: So you probably could use this experiment to be very suggestive that monkeys, you know, feel jealousy. But you can't use that as a basis for equal pay because humans do not behave the same way, because humans are rational and monkeys aren't. Right.

[01:29:15]  Cameo: Because humans have to survive and pay is our primary way to survive in the modern world. And even if you know someone else is getting more, if you can't get it, you still probably have to accept what you can get. That's really fascinating.

[01:29:31]  Unidentified speaker: It is.

[01:29:32]  Bruce: I thought you would love that.

[01:29:34]  Unidentified speaker: I do.

[01:29:34]  Bruce: Anyhow, that's all I had. This is kind of been a hodgepodge of different things that mostly stuff that I had discovered since the first episode and just felt like they were important enough and interesting enough that I wanted to insert them in on this one episode. And then we'll go back to what I originally had planned starting next week. What I'm going to do is I'm probably going to do I don't know how many episodes we'll do, but we'll do a number of episodes where I actually walk through Byrne's theory and compare it to Deutsch's interpretation of Byrne's theory. What I really found is that Byrne and Deutsch interpret Byrne's theory in almost opposite ways. It's very fascinating. But we will go through Byrne's theory in detail, talk about his methodology. I really would encourage people to read Byrne's books. I think that I think it would be eye-opening to a lot of Deutsch fans to read Byrne's books.

[01:30:29]  Bruce: And I'm just I'm really amazed what a good scientist he is. He is so skeptical of his own findings. You know, I mean, like he's a Popperian to the core without knowing who Popper is. I really highly recommend him.

[01:30:42]  Sadia: All right.

[01:30:42]  Unidentified speaker: Sounds good.

[01:30:43]  Unidentified speaker: All right. Well, we'll talk to you then.

[01:30:48]  Bruce: The Theory of Anything podcast could use your help. We have a small but loyal audience and we'd like to get the word out about the podcast to others so others can enjoy it as well. To the best of our knowledge, we're the only podcast that covers all four strands of David Deutsch's philosophy as well as other interesting subjects. If you're enjoying this podcast, please give us a five-star rating on Apple Podcasts. This can usually be done right inside your podcast player or you can Google The Theory of Anything Podcast Apple or something like that. Some players have their own rating system and giving us a five-star rating on any rating system would be helpful. If you enjoy a particular episode, please consider tweeting about us or linking to us on Facebook or other social media to help get the word out. If you are interested in financially supporting the podcast, we have two ways to do that. The first is via our podcast host site, Anchor.

[01:31:37]  Bruce: Just go to anchor.fm slash four dash strands, F-O-U-R dash S-T-R-A-N-D-S. There's a support button available that allows you to do reoccurring donations. If you want to make a one-time donation, go to our blog, which is four strands dot org. There is a donation button there that uses PayPal. Thank you.