15 – Theories of AGI

Transcript

This transcript was generated automatically and may contain mistakes.

Dennis: Today I’m speaking with Ella Hoeppner. Ella is a software engineer and independent AGI researcher. We’ve been discussing all things AGI for a while now, and I wanted to have her on the podcast to present her theory of AGI, and also to discuss a disagreement we’ve had about the role of replicators in any theory of AGI. Since we both spend a fair amount of time criticizing each other’s theories, you may get more out of this episode if you familiarize yourself with my neo-Darwinian theory of the mind first, which I have laid out on both Christofer Lovgren’s podcast, I’ll link to that episode in the show notes, and in my book A Window on Intelligence. That being said, it’s certainly not required. I think you’ll get a lot out of this episode either way.

Ella and I also briefly touch on conventional narrow AI research, the role of neuroscience, and the hard problem of consciousness. I even catch myself having changed my mind on the urgency with which we should pursue whole brain emulation in addition to explaining creativity. I’ve said in the past that there are maybe a handful of people in the world who work on AGI whose approach has not been refuted, and Ella is one of them. And so with that, Ella is going to give us a glimpse into how the mind may work.

Okay, so I’m here with Ella Hoeppner. Ella, welcome to the podcast.

Ella: Hi, it’s great to be here. I’ve been a big fan of this podcast for a long time, so it’s great to be on.

Dennis: Oh, that’s great to hear. So let’s start with what got you into AGI research.

Ella: Yeah. So I think I have about the same backstory as a lot of other people in our space in that I read David Deutsch’s The Beginning of Infinity, and it blew my mind. You know, the entire stance that it took towards science and the philosophy of science and epistemology really was inspiring to me. And I’d had somewhat of an interest in AI before that, but it was in reading David Deutsch’s work that I really started to realize that the epistemology that he was talking about was really, really important to AI. And it sort of invalidated a lot of the approaches that many people are taking right now.

And any approach towards AGI that was actually going to work would need to, you know, be in line with sort of Popperian, Deutschian epistemology. And so that’s sort of what inspired me to begin working on it like I am now.

Dennis: So did you also read any Popper after you read Deutsch?

Ella: Yep. So after I read The Beginning of Infinity, I went back and read The Fabric of Reality. And then I read Conjectures and Refutations by Popper, as well as the book Objective Knowledge: An Evolutionary Approach. And then I’ve read a few of his essays from other books, but those are the only two books of his I’ve read in full.

Dennis: Yeah, I love Popper’s books. I mean, especially Objective Knowledge, I thought was fascinating.

Ella: Yeah, that was a great one.

Dennis: Yeah, my story is pretty much similar to you. I think we read all the same books in almost the same order.

Ella: That’s funny.

Dennis: Uh-huh. So you actually have developed a theory for what you think might be the best approach to building AGI.

Ella: That’s right.

Dennis: What is this theory called?

Ella: Yeah, so I call the theory CTP theory. And that name, that acronym CTP, it used to mean something, but it doesn’t really anymore. It’s just kind of a placeholder because I don’t have any better idea for a name. So it’s called CTP theory, but not for any good reason.

Dennis: And what did CTP stand for in the past?

Ella: So, yeah, so initially CTP stood for claim theory problem. And those were just kind of the name of the three sort of fundamental entities in the theory. But since then, I’ve developed the theory some more, and I realized that there doesn’t really need to be a distinction between claims and theories. And so instead, there’s just sort of a single sort of idea data structure that you could call an idea or a theory. It doesn’t really need a super technical name anymore because it doesn’t need to be distinguished from anything else since there’s just the one type.

And so the notions of claim and theory have just kind of gone by the wayside as the theory is developed. So and so the name is kind of very outdated at this point. But if anybody listening has a better suggestion for the name of the theory after hearing about it, please just email me. I’d be happy to hear any suggestions because I’d be happy to stop calling this thing CTP theory.

Dennis: Okay, we’ll put your email address in the show notes.

Ella: Yeah, sounds good.

Dennis: Okay, so tell us how did you get started on CTP theory? Was there something that sparked it? Did you just have a sudden epiphany? Was it more of a gradual thing? And what is it about?

Ella: It was more of a gradual process developing CTP theory. I had some initial ideas on AGI, which looking back on them now, were very flawed and kind of naive.

But what initially led me towards CTP theory as I think about it now was considering what Popper said about criticism and contradiction. So in an essay in Conjectures and Refutations, he talked about how all criticism consists of pointing out contradictions between theories or between a theory and some sort of statement of fact about evidence. And so I started realizing that contradictions between ideas are really central to Popperian epistemology. And so I started thinking about how ideas could be represented on a computer in a way that would allow for contradictions to be sort of seen between them in a way that a computer could see the contradictions between ideas and then work from there to resolve the problems.

And so I can go ahead and give an explanation of how CTP theory works. Basically, CTP theory says that, and I, a mind contains a set of ideas, and each idea, the way it’s represented on the computer is that it is basically some kind of, it’s an instance of a data structure. And all ideas in the mind consist of the same, data structure, but they each have their own data. And so just as an example, we can imagine that the ideas are all bit strings. And then one of the essential parts of CTP theory is that the mind has a way to detect direct contradictions between ideas in its set of ideas.

And what I mean by a direct contradiction is a little bit, I don’t think this terminology is used elsewhere, so this is kind of my way of thinking about it, but a direct contradiction means a contradiction between a statement like X is true and X is false, right? So very sort of self-evidently contradictory. You know, it’d be very easy to sort of kind of represent this computationally, whereas there are other ideas that are contradictory. An example that I like to use a lot is my cat is calico and my cat is male, right? All calico cats are female, so that those two statements are contradictory, but they’re not directly contradictory.

And so CTP theory says that the mind has a way of recognizing direct contradictions between the ideas in its mind. And so if the ideas are represented as bit strings, the method for detecting direct contradictions might be, say, two ideas are directly contradictory if they are bit strings of the same length and they have all the same bits except for the first one, say. And so that would give you a way to determine, you know, if two ideas are contradictory or not, directly contradictory, that is. So that’s kind of the basic way that CTP theory recognizes contradictions.

But obviously, that isn’t sufficient, because like I said, there are these more indirect contradictions, like my cat is calico and my cat is male, which wouldn’t be covered by that method. And so the way that— Go ahead.

Dennis: Before we go further, what is a bit string?

Ella: A bit string? Yeah, sure. So a bit string, the way I’m using the term, just means a list of bits. A bit is a zero or one, or true or false. And by a list, I just mean an ordered collection of them of some size. And so, you know, 0010. That’s a bit string, that’s a bit string of length four. You could have bit strings of any size. And yeah, so that’s kind of what I mean by a bit string. And I’m just using that as an example data structure. In the real CTP theory, you probably wouldn’t actually use a bit string, but that’s just a data structure that I find it easy to think about. So that’s what I’m going to use as an example.

Dennis: Okay. And you said you’re reserving the first bit. So that means the first character in the string, the first number in that string, you’re reserving that to be the truth or falsity value for that idea. Is that correct?

Ella: Yeah, that’s one way to think about it. And yeah, so you can think of the first bit in the string as sort of saying the rest of these bits are true or the rest of these bits are false. And then you sort of, whenever there’s two ideas, one of which says, you know, 001 is true and one of which says 001 is false, well, those are obviously in direct contradiction. And so, yeah, you can think of the first bit in the bit string as being sort of a true flag for the rest of the content of the bit string.

Dennis: Now, one thing we should point out, I think, is that it might seem strange at first to have just a string of zeros and ones. But I think it’s important to realize that zeros and ones in this configuration are computationally universal. So you could indeed express any possible idea through bit strings as long as these bit strings can be long enough.

Ella: Yeah, that’s right. Yeah, so one of the sort of interesting consequences of information theory is that actually almost any alphabet, or well, really any alphabet of characters can be used to express the same as any other alphabet. You know, so if you think of an English sentence, there is, you know, probably 40 characters that we use in English pretty regularly, not including the capital letters, of course. But you can imagine that you assign a, an eight-bit code to each of those letters. And then in that way, you could represent a string of English characters with just bits, zeros and ones.

And so the fact that I’m using a bit string here doesn’t really limit the capabilities of the theory at all. It’s a universal data structure. It can represent an infinite number of different things. And you could convert it in your mind, if you want to, any other kind of data structure because it is universal for representing data.

Dennis: Right on. Okay. Now, so what would it mean, for example, for an idea to contain the numbers 1011?

Ella: Right. So that’s a good question. And the answer is that it wouldn’t necessarily mean anything inherently. So in the same way that, you know, the series of syllables “hello” doesn’t really mean anything on its own. It only means things by sort of the way it takes on context in relation to other strings of syllables that we use. And so in the same way that, you know, language is really just arbitrary strings of symbols, but it takes on meaning in a proper context, a string like 0010 or whatever you said wouldn’t really mean anything on its own, but it would take on a meaning by the way that it relates to other ideas in the mind.

Dennis: Got it. So presumably, I’m guessing you would have another idea, itself encoded in binary code, which lets you interpret an idea like 1011?

Ella: Right. So I think this will become a little bit clearer if I kind of go a bit further with the explanation I was giving earlier on how the theory works.

Dennis: Yeah, sure.

Ella: So as I said, the mind has a way to detect direct contradictions between ideas, but we also need a way to detect indirect contradictions, of course. And this is where one of the central ideas in CTP theory comes in, which is that ideas aren’t just instances of data structures. They’re a little bit more than that, in that the mind has a way to interpret an idea, an instance of a data structure, as a function. And what I mean by a function here is the way the term is generally used in mathematics, which is just a mapping from some set of inputs to some set of outputs.

And when I talk about the mind interpreting an idea as a function, what I mean is that it essentially has a programming language built into it. And the bit strings in the mind are programs in that programming language, which define functions, which map from inputs to outputs. And so you can think of that, you can imagine the mind contains a, you know, a C compiler or something like that, or any programming language compiler that converts the bit strings into that programming language, and then it runs that code. That’s one way to think about it. And so when a function, you know, when an idea is interpreted as a function and it takes another idea as an input, it’ll produce a new idea as an output, right? So, ideas can be functions where their domain is ideas and their range is ideas.

Dennis: Sort of higher level ideas in a sense.

Ella: Yeah, yeah. They can exist at like any level of abstraction. They can return, you know, they can return other ideas that, interpreted as functions, and so on. But basically on an epistemological level, what this represents is the ability for ideas to imply other ideas, right? So if we have one idea A and that idea is interpreted as a function, it takes an idea B as an input, and then it produces a new idea, C, as an output. The ways this should be thought of on an epistemological level is that A and B together imply C. And so this is how the mind uses those ideas to produce new ones.

And this kind of gives the mind a way to detect indirect contradictions in the sense that if an idea’s implications, which is to say the new ideas that it produces when it’s used as a function or as an input to a function, if one of those leads to a direct contradiction with something else, then you can trace that back to the antecedent ideas, the ideas that were used as the functions or the inputs to the functions to produce the output. And those can be said to be indirectly contradictory with the other idea because even though they aren’t directly contradictory, they led to something that is directly contradictory.

And, in the example, my cat is calico and my cat is male, you can imagine that the idea my cat is calico, along with some other idea like all calico cats are female, would imply the idea my cat is female, right? And then the idea my cat is female would imply my cat is not male. And then you have a direct contradiction between my cat is not male and my cat is male. And so that’s the way that the mind in CTP theory would be able to detect both direct and indirect contradictions.

Dennis: Got it. Just to linger on that example of the calico cat, if I’m remembering correctly, we’ve spoken about this example before, but the reason there is an indirect contradiction there is because there’s a biological phenomenon that leads to the fact that all calico cats are female. Am I remembering that correctly?

Ella: Yes, right. Yeah. So there are no male calico cats in the world. Or actually, I believe that there are, you know, some tiny, tiny number of genetic mutants that are male and calico. But, you know, I’m not actually implying that the ideas in the mind in this example are true. You know, this is just something that a mind could believe, right? So it could believe that all calico cats are female, and thus it would be able to, you know, derive that its cat, or that my cat is female, if I thought that my cat was calico, for instance.

Dennis: Got it. And if somebody claimed that their cat was calico and male, you would then perceive, I don’t mean perceive as in sense data, but you would perceive a conflict there in your mind.

Ella: You’d perceive a contradiction. And to go a little bit further with CTP theory, contradictions between ideas are said to be problems. So the word problem in sort of critical rationalism more broadly is used in a lot of different ways. And it’s sort of hard to nail down exactly what that means on a technical level. But in CTP theory, when I use the word problem, I’m specifically referring to a direct contradiction between two ideas, along with kind of this set of, you know, ideas that are indirectly responsible for that direct contradiction. And so resolving a problem in CTP theory means basically getting rid of some of the ideas that caused the direct contradiction, right? So if you have the idea my cat is male and my cat is not male, you have to just get rid of one of those two in order to solve the problem.

And, you know, you can’t just get rid of, you know, if you have the idea my cat is female, and that idea came from my cat is calico and all calico cats are female, you can’t just get rid of my cat is female without getting rid of one of those ideas too, right? Because it’s a logical implication of one of those ideas. And so you have to get rid of one of those as well. And so the modifications that you make to a mind in terms of solving a problem will often involve removing sort of quite far disconnected ideas that are indirectly responsible for the contradiction.

Dennis: So in the example of the calico cat, you said one of those ideas has to go in order to resolve the conflict. Either the cat in question is not male or the cat is not calico. Unless, what you said earlier, which, there might be some mutant cats that occur every now and then in nature. Now, here we have an alternate explanation. And so in this case, the conflict goes away without dismissing one of the ideas. You can still keep both.

Ella: Yeah, so in that case, the idea my cat is female and, or sorry, my cat is calico and my cat is male, those aren’t the only two ideas involved in the problem here. There’s also the idea all calico cats are female, right? Because that’s necessary to derive the idea my cat is female from the statement my cat is calico. And so it’s actually at least those three ideas, right? So my cat is calico, my cat is male, and all calico cats are female, right? You have to consider all three of those ideas. And so if kind of adopting the new explanation that, you know, most calico cats are female, but there’s a small proportion of, you know, male calico cats, that would involve getting rid of the idea all calico cats are female.

And so kind of by definition, any way of resolving a contradiction is going to involve getting rid of at least some idea involved. And when I say getting rid of, what that could mean is sort of modifying the idea, which in other terms involves getting rid of an idea and then adding a new idea, which is kind of similar to the old one but varied in some way.

Dennis: Okay. So then how does the mind decide, how does the mind know whether to dismiss one of the ideas in question or to generate a new theory that resolves the conflict?

Ella: Yes. So I think CTP theory, as I’ve explained it so far, does a good job of explaining how the mind could detect contradictions and how the mind could go about figuring out how to resolve those contradictions by getting rid of some ideas. And there’s kind of, you know, if you look at my recent blog post on fourstrands.org or .com or something, I can’t recall, then that explains in sort of more technical detail exactly the process of figuring out which ideas the mind can think about removing.

But CTP theory isn’t yet a complete theory, right? If it was, then, you know, maybe you’d be talking to the AGI that I created instead of me. But it’s not at that stage yet.

Dennis: Maybe I am. How do I know?

Ella: Yeah, maybe I’m just not giving you the full theory. Who knows? But, so where CTP theory, kind of is currently is in the process of figuring out, how exactly does the mind decide which idea to get rid of? Because it, in a strictly logical sense, you could just get rid of an arbitrary idea in the set of responsible ideas. You could just arbitrarily say, okay, I’m going to get rid of the idea all cats are, or all calico cats are female. And if you get rid of that idea, then the contradiction is resolved, but you can’t just solve, on an epistemological level, we know that you can’t just, solve any, relevant or any, any problem by just getting rid of one of the, relevant ideas, right?

So if, an example that I used in my recent article is if you consider the scientists, who first noticed the, anomaly with the perihelion of Mercury’s orbit, which is to say the first people who noticed that Newton’s theory produced the wrong predictions about where Mercury should be. Those scientists would have had a contradiction in their mind, according to CTP theory, between Newton’s theory and the evidence that they were seeing. And so if we just kind of go with CTP theory the way I’ve described it so far, they might just arbitrarily say, okay, well, Newton’s theory is false. I’m just going to get rid of it. But obviously, that’s not actually how scientists think. And it’s not how they should think, because that would just be a completely fruitless way of viewing the world.

Dennis: And they would lose explanatory power for a lot of other problems that they have already solved.

Ella: Absolutely, yes. So losing explanatory power is kind of what CTP theory needs to account for, which it doesn’t yet. So yeah, and that’s kind of one way of thinking about it. I do have some ideas as to how CTP theory can go about doing that, which I can describe, but this is kind of the active area of where I’m thinking. And to go into what the possible solution might be here is something that I call the requirement system. And one way to think about this is, as you said, those scientists wouldn’t throw out Newton’s theory because they would lose a lot of explanatory power.

And the way that I am considering modeling this in CTP theory is that, you know, there’s a lot of things that the scientists might want to be able to do. They might want to be able to say, or to answer the question, you know, how fast is this object going to fall if I drop it from a height of three meters? You know, Newton’s theory would answer that question. And they might also want to answer the question, where is, you know, Neptune going to be in 100 years? And Newton’s theory can answer that question. So if you threw it out, then there would be sort of these unfulfilled desires in the mind, which is to say, you know, you would have a desire to know what is the speed of this object going to be, or where is Neptune going to be in 100 years, but you would have no way of fulfilling that requirement or that desire.

And so basically the solution that I’m thinking CTP theory will need to implement is that it has a way of representing requirements or desires. And when a suggestion to get rid of Newton’s theory or something, when a suggestion like that would lead to some desires or requirements becoming unfulfilled, then that is rejected, right? You have to find a way to get rid of some ideas, introduce some new ideas, whatever, to resolve the contradiction without leading to any requirements or desires becoming unfulfilled, is kind of the guess that I have currently.

And so to kind of connect that to the idea of explanatory power, explanatory power is a useful sort of high-level concept, but it’s not really clear how to implement the notion of explanatory power directly on a computer. And this notion of having requirements or desires in the mind is basically a way of representing kind of the quanta of explanatory power, which is to say, I want to be able to explain this and this and this. And they’re all individual kind of units of things that you want to be able to explain. And so you need to have the explanatory power to answer all of those individual questions.

Dennis: One thing I want to touch on real quick is that the ideas that are encoded in binary code in the mind, as per CTP theory, these are not credences. These are not generated by sense data. I do want to disambiguate a little bit from the traditional model of this inductivist model of how the mind works, just to make it clear for our listeners. You’re not thinking of establishing truth of ideas. You’re focusing explicitly, following Popper, resolving conflicts because you know that you can never establish the truth of an idea.

Ella: Yes, exactly. And to be clear, in CTP theory, there’s no kind of credence assigned to any theory or any ideas at all. There is simply the idea in the mind, and it can contradict other ideas, and that’s it. There’s no notion of credence or likelihood of being correct. There’s just ideas and contradictions between ideas.

Dennis: Okay, great. How do the senses play a role, though? I mean, if we’re talking about creating an AGI, eventually you would want that AGI to interact with the world. So presumably there would be some stream of input data coming in from the outside world. What would a mind, a creative mind in CTP theory do with that stream of data?

Ella: Yeah. So the initial question really is, what form would that stream of data take, right? How would that actually get into the mind in a way that the ideas in the mind could interact with it? And the answer here, and this is going to sound a little bit empiricist, but I promise it’s not, and I’ll explain why, is that the sense data will take the form of an idea in the mind. And so basically the way this would work is you’d have some sort of external sense organ, you know, maybe that’s using a camera to take a video of the world, and then that’s getting put into the mind. That information from that camera or whatever other sort of sense organ would be put into the form of an idea, and then that idea would be inserted into the mind.

And now that sounds like empiricism because I’m saying, oh, well, ideas in the mind are formed by sense data. But really even calling this an idea is a bit incorrect in an epistemological sense. What’s happening merely is that, you know, ideas are built so that they can work with a certain type of input, right? And if we, so the idea is that we want to put the sense data into that form so that ideas can use it as an input and then produce outputs from it. And so it takes the form of an idea and it sort of looks like an idea in the mind, but it isn’t on an epistemological level. It isn’t the same as, you know, some sort of conjectured theory. It really is just raw data, even though the mind treats it in the same way that it would treat an idea on a low level.

Dennis: So in some way, you have to impose the same constraints onto the sense data that you’re imposing on ideas. For example, sense data would always start, you would always treat the first bit in the sense data as a truth value?

Ella: Yes. So in this example, it would have to be put into the form of a bit string. But you can imagine that, you know, all of the bit strings of sense data just have a one as their first bit, right? Because you wouldn’t really want direct contradictions between, you know, different types of sense data. You would want that to happen on the level of interpretation, right? You would want the mind to see something and then see something else and interpret those things as being contradictory. And interpretation here basically means taking the sense data and using it as an input to some other idea and producing a new idea as an output.

And that idea, the idea that’s being used as a function, is sort of interpreting the sense data to produce some new statement. And so the way that you would translate the sense data into a bit string doesn’t have to be just the most straightforward way of doing it. You can do something like, you know, making sure the first bit is always a one just so that the information is still there, but it isn’t going to directly contradict anything in the mind.

Dennis: Okay, but then couldn’t you have ideas that look just like sense data and vice versa? So let’s say you get some sense data and it takes the form of 1011, the same as I said earlier, and your mind happens to contain an idea that is encoded in 1011. At that point, how would the mind distinguish between these two? Because one is still different from the other, isn’t it?

Ella: Yes. So it’s a good question. And I suppose the answer is just that I think that in practice, the mind wouldn’t really, I think that the mind would evolve into a way where it wouldn’t really end up producing the exact same form as some kind of sense data. And so I guess what I mean here is that you can imagine that, if you imagine that maybe all the bit strings that represent sense data start with some kind of special code, right, like a code. And this is something that you just program into the way sense data is put into the mind. You might say all sense data starts with a string of 10 ones, and that is sort of to mark it as sense data.

And then so the mind can simply evolve, you know, not to start all of its ideas with 10 ones, right? Like, maybe the mind decides, okay, well, the second bit is always going to be a zero in my ideas, right, or something along those lines. And then in that way, because you have sort of designed the mind to have this marker for sense data, it would be easy for the mind to tell the difference between sense data and some other form of idea.

And it’s still possible in theory that the mind can produce something with 10 ones at the start, but I think if, if that’s, if that were the case, then, the mind would be able to realize, oh, wait, these, you know, this is not really sense data, it’s just kind of masquerading as sense data. I shouldn’t, I shouldn’t use this format for representing my ideas in the future, because if you’re doing that, then it’s probably going to cause contradictions. If you’re, you know, tricking yourself into thinking that you’re seeing things that you’re not, that’s going to interact with your other ideas. And you’re going to say, wait, that elephant shouldn’t be floating in midair. Where did that idea come from? And then you trace that back and, you know, the mind would decide, okay, I’m not going to format ideas like this anymore, because that’s interfering with the sense data system.

Dennis: Got it. It could actually be a really neat explanation of where, you know, hallucinations come from or stuff like that, where maybe an idea mutated in such a way that it took the shape of sense data. And so now the rest of your mind interpreted it as sense data, as if it was coming in from the outside, when really it originated from the inside.

Ella: Yeah, it definitely could.

Dennis: It shines light on this idea, this Popperian idea that really, we don’t perceive anything directly anyway. Everything is subject to interpretation. We don’t have direct access to the real world, even though it exists. What we see around us is subject to error and subject to interpretation.

Ella: Yeah, exactly.

Dennis: Okay. Now, would this mean that, you said something there just now about how the mind would then decide, okay, I’m not going to, I forget exactly how you phrased it, but something along those, okay, I’m not going to take ideas of this shape into account or something. So basically what that means is the way the mind reasons about ideas is itself subject to change. Is that right?

Ella: Well, so it’s not that the mind would decide, I’m not going to take ideas of this form into account. It’s more just that if the mind was producing ideas of that form, that would probably eventually lead to some kind of contradictions because, you know, it wouldn’t be the type of sense data that the mind is used to seeing. And then in that case, it would, you know, trace the source of those, you know, sort of false sense data ideas backwards. And it would realize, oh, this, and when I say realize here, I’m kind of speaking anthropomorphically. I guess I can try to speak in a more technical way. And that might be more helpful.

So the mind would trace backwards to whatever idea is producing these sort of confusing false sense data ideas. And it would eventually, just by trying, you know, blind conjectures to solve the problem, it would come up with a conjecture that would involve not phrasing ideas in that way anymore, right? Not using the same format as a sense data idea, a true sense data idea, so that it isn’t confusing the sensory, you know, interpretation system anymore. So that those ideas aren’t, you know, so that they don’t look like sense data ideas anymore. And the other parts of the mind can distinguish between them.

Dennis: Okay. But I guess what I’m getting at is, would every mind do this reliably?

Ella: Well, you know, we can’t make any guarantees, you know, that any mind is going to solve any particular problem. But I suspect that if the mind had a sort of already, if you imagine that it already has a sort of sophisticated way of interpreting its sense data, and then it starts, you know, producing sort of these false sense data ideas that are accidentally getting interpreted as true sense data. I suspect that there wouldn’t really be any way, any easy way of solving that problem other than by tracing back to the idea that’s causing the confusing false sense data ideas and getting rid of that idea or forcing that idea to format its outputs in a different way.

And the reason for that is that if you tried to modify any other idea involved in the sort of sensory interpretation part of the mind, then you would probably end up failing to satisfy a lot of requirements or desires, right? If you decided to throw out one of those ideas, then you would probably end up losing a lot of explanatory power. There’s probably not any easy modifications that you could make there that would, you know, be satisfying solutions to the problem in the sense of resolving contradiction while also keeping the explanatory power that the mind wants.

And so the easiest solution to this problem would almost certainly be that the mind just stops, the mind just changes that idea that’s producing the confusing signals and forces it to format its outputs in a different way. So it’s not that the mind would always 100% reliably do that. It’s just that if the mind already has a mature, you know, well-adapted sense data or sensory interpretation part of its mind, then it’s going to be very hard to modify that rather than modifying the confusing sensory data producer. And so that would generally be, I think that we would expect the mind to find that solution just because that’s going to be the easiest way to resolve the problem.

Dennis: Okay, so you said that some open problems currently with CTP theory are how to decide which ideas to discard when there is a conflict. And this notion of explanatory power, it’s not quite clear how to quantify that. Well, how does, well, first of all, do you, do, are you working on any other open problems with CTP theory?

Ella: So that one’s, kind of the biggest one, the, the fact that, you know, we need a way to sort of phrase explanatory power, in a computational way. And I think the requirement system that I have, is a, is a good start. It’s just, I, I’m, I’m struggling to make that idea or, you know, get that system, describe it in a sufficiently detailed way where I can actually turn that into code. There’s some parts of that that I haven’t nailed down yet that I’m working on nailing down. So that’s definitely the biggest open problem in CTP theory.

And there are some other open problems, but in my mind, it’s sort of like all of the other problems kind of rely on this one. Because CTP theory as it is just doesn’t work, right? It doesn’t produce like half of an interesting mind. It just doesn’t work at all without a solution to this problem. I think there’s still valuable theoretical insights from the rest of CTP theory about, you know, how contradictions can be recognized. But the actual picture of how the mind evolves over time just is clearly and completely incomplete without something like an explanatory, a way to account for explanatory power. So that’s by far the biggest open problem, I’d say.

But some other open problems, just kind of little ones, are how exactly does the mind choose which ideas to feed to other ideas, right, as inputs? And it might just be that the mind can just kind of do this blindly, but that might end up being really inefficient. And the reason I don’t really consider that as important of an open problem is that, you know, once we resolve the problem of explanatory power, right, and once I nail down the details of the requirement system, I could just sort of try different things out and find out what the most efficient way to do that is, right? I can just sort of empirically figure out the best way to, you know, feed ideas to other ideas.

And so that’s just kind of a small sort of subsidiary problem to this greater problem of getting the theory to work in the first place. And there are other kind of small problems like that that I could, you know, go on about, but none of them are nearly at the level of importance of the explanatory power problem.

Dennis: Okay. I do want to linger on the problem of explanatory power for just a moment. Am I understanding correctly that you, so would, let’s say you have found a way to determine an idea’s explanatory power. Is the idea that, is the purpose of that determination that an idea that has higher explanatory power than another one would win the conflict? If there was, if there was ever an, if there was ever a conflict between an idea with high explanatory power and one with low explanatory power, you would favor the one with high explanatory power?

Ella: It’s not quite that. It’s more, if we go back to the example of, you know, the perihelion of Mercury and Newton’s theory. It’s not just that we would choose to, you know, keep Newton’s theory and throw out the, you know, explanatory ideas involved in the experiment, just because Newton’s theory has higher explanatory power. What it is instead is that you need to find some conjecture, right, some way of modifying the ideas in the mind, getting rid of ideas and introducing ideas that satisfies all of the explanatory power that Newton’s theory has, or, you know, find some different way of interpreting the experimental data, right, that satisfies everything, all the explanatory power that we want that to satisfy.

But that turned out to be impossible, right? The measurements about Mercury were actually correct, right? There wasn’t any problem in the experimental apparatus. And so the only solution was to find a replacement for Newton’s theory that had all the explanatory power, but didn’t produce the contradiction. And eventually, Einstein came along and gave us the theory of general relativity. And that theory had all of the explanatory power that Newton’s theory had, but it didn’t lead to the contradiction surrounding Mercury’s perihelion. And so it’s not as simple as just saying the idea with higher explanatory power wins, you know, in any conflict. It’s just that when there is a conflict, we judge solutions by whether or not they preserve all the explanatory power that the mind had before.

Dennis: Right. Okay. So in other words, as Popper would say, Newton’s theory lives on in Einstein’s as an approximation.

Ella: Yeah, exactly. And that is part of why we consider Einstein’s theory better, because it answers, it solves all the problems, I think it does all this, or maybe with some exceptions, but it solves all the relevant problems that Newton’s theory solved. And then some, it also solves some other stuff.

And just to expand on the reason why I see kind of the requirement system or the desire system that I was describing earlier as sort of a good way to handle this problem of explanatory power. I think that the notion of, say, the explanatory power of Newton’s theory can kind of be reduced to the question of what, you know, what ideas does Newton’s theory lead to that we want to preserve, that we want to be able to answer, right? And so that would be things like predicting the movements of planets, predicting the movements of, you know, objects on Earth under the influence of gravity, etc. And so those would be framed as sort of individual requirements, right? Like, I want a way to predict the movement of planets. I want a way to predict the movements of objects on Earth under the influence of gravity.

And so when each of those requirements is satisfied by another theory, like Einstein’s theory of general relativity, then there’s no need for Newton’s theory anymore, because those requirements are already satisfied by something else. And then you can resolve the conflict between, you know, Newton’s theory and the experimental data about the perihelion of Mercury by simply getting rid of Newton’s theory, because you already had this other theory, Einstein’s theory, which has all the explanatory power of Newton’s theory in the sense that it fulfills all the relevant requirements, but it doesn’t have the contradiction.

Dennis: Okay. Now, I do want to get a little bit into criticisms of CTP theory, just to play devil’s advocate and see what your response is. Maybe this will lead to criticism, maybe not, but one more thing that I want to clarify is, I think I remember us talking about this, and just for listeners, you and I have been talking about CTP theory for a while as part of the Four Strands, but, I think, do I remember correctly that you incorporate some notion of ideas being hard to vary in the Deutschian sense?

Ella: I, I did, I, I’ve considered in the past, in the past before I had this, new notion of the requirement system or the desire system or whatever you want to call it. I did in the past consider trying to come up with a way to sort of measure hardness to vary, or kind of another way to think about it would be adaptedness, how well adapted an idea is to, you know, whatever its purpose is, to try to come up with sort of a metric for that, and then use that as kind of a deciding factor when deciding which ideas to throw out. But that isn’t a part of the way that I currently think about CTP theory. The sort of the requirement system or the desire system has, kind of, sort of, encapsulated that, it’s sort of, so the idea of sort of explicitly incorporating a metric of hard to vary, no, I’m not, I don’t think that’s necessary anymore.

Dennis: Got it. Now, one thing, I’m getting a general sense, but maybe you can shed any doubt that I have, one, one sense I get is that what CTP theory is describing is a rational mind. It sounds a little bit like a science machine. I don’t mean this in a derogatory sense. It would be great if something like that existed. But what it seems to me is, okay, you have a conflict between ideas. And then what the mind does is it keeps trying to resolve the conflict, right? And you’ve introduced some notions of how it might go about this. Like maybe the notion of explanatory power plays a role here. There is the notion of direct contradictions and indirect contradictions. So it’s sort of iterating over all these things to find a better solution.

I just wonder if that is really what a creative mind does, because it doesn’t sound like, what you’ve described so far doesn’t really sound creative necessarily. So because, for example, a scientist who finds a problem with Mercury’s orbit, a problem for Newton’s theory, will have very different, first of all, he’ll have very different ideas as to how to resolve that conflict than someone who’s not an astronomer, say. But you could still accommodate that in CTP theory, I think. It’s just that those two minds have developed different ideas and different conflicts between those ideas.

But there will come a point at which even the most fascinated scientist, the most fascinated astronomer who’s so fascinated with this problem, there will come a point where he’s hungry and he gets up and makes a sandwich. So he puts the work aside for a moment. I don’t quite see yet how CTP theory accommodates this. And maybe it does. Can you elaborate?

Ella: Yeah, so the process that I’ve described of, you know, kind of searching through various ways of modifying the mind to, you know, find a solution to a problem, that I imagine that the mind would sort of have multiple instances of that process running in parallel at the same time, right? So it wouldn’t just devote all of its energy to the problem of how do I fix Newton’s theory or something like that, right? It might be solving that problem while, you know, another kind of, another process running in parallel is kind of trying to solve the problem, I’m hungry. How do I fix that, right? What can I do to fix the fact that I’m hungry?

And so it’s not that the mind would be devoting all of its energy to any one problem at any one time. And actually, you could say that the problem of how does the mind decide which problems to work on is kind of another open problem with CTP theory. I don’t really have a strong answer to that. The default answer would just be that it looks at all the problems in its mind, it picks a random one, and then it tries to solve it, right? And then maybe if it can’t solve it for five minutes, and it picks a different one and tries to solve that. And that at least I think would work. It might not be as efficient as some other, you know, more sophisticated algorithm for choosing the right problems. But I think that that would, that that’s just sort of the basic problem that I’m assuming for now, or the basic system that I’m assuming for now. And once the problems with the requirement system are more laid out, maybe more work can be done there. But that’s kind of the way that I imagine the mind working for now.

Dennis: Got it. How do you square this with the notion of interest, for example? I mean, a scientist who’s just fascinated with astronomy, say, will keep working on his problems even after five minutes, you know? So in the, in the model you just described, he may change to another idea within five minutes. Would there then be another idea, maybe, that’s, that overrides this mechanism and says, no, no, let’s, let’s stick with the problem we’re working on now, and then that would sort of manifest as a, as an interest, maybe?

Ella: Yeah, so I think that the notion of interest can kind of be, it’s somewhat related to the requirement or the desire system. So if you’re imagining, you know, how does somebody develop an interest in, you know, astronomy or, you know, the problem with Newton’s theory in the first place, if their mind had a set of ideas, right, that would produce the desire, I want to know how to predict the positions of, you know, planets, right? If something in the mind leads to that desire being created, then that is at least part of, you know, an interest in astronomy, right? The way that you would solve that unfulfilled requirement or unfulfilled desire is by, you know, learning about astronomy. And so, that’s kind of the way that I broadly think of people developing different interests in the framework of CTP theory.

But you’re also asking about, you know, it can’t just be, you know, he stops working on a problem after five minutes and does something else because, you know, that’s not how people work, right? We can decide to, you know, work on ideas for as long as we want. And that’s true. And the system that I’m, you know, proposing of just pick something random and work on it for five minutes, you know, that’s just kind of a first guess as to how the mind could do things. Like, as I said, that’s almost certainly not the best way to do things.

It might be instead that maybe the mind, based on, you know, how many important ideas, or really just how many ideas and how many desires or requirements are involved in a problem, maybe it spends more time on those problems, right? And so if you have a big interest in astronomy and you know a lot about astronomy, then there’s going to be a lot of conflicting ideas and requirements and so on in the mind. And then maybe the mind would devote more time to solving those problems as opposed to somebody who has an interest in gardening or something.

Dennis: Yes. Okay. That makes sense. And I guess what I’m picturing here is the problem that, I mean, you could, I understand that the switching every five minutes is just a first guess, you know, and we both understand that that’s probably not it because that’s not how people work. But I think any, I’m guessing, any particular answer, anything of the form every five minutes, or do this, and then after some condition do this, sounds to me like it might preclude creativity, because it’s still a reliable and mechanistic way of doing things. There’s some criterion that seems set in stone.

So I guess what I’m hinting at is the ways in which we go about, and I think you said something along those lines too, actually. So correct me if I missed something here. But it sounds to me like what a creative mind does, it seems to me like what a creative mind does, is the criteria by which it chooses what problems to work on are themselves ideas that arose through the creative process.

Ella: Yeah.

Dennis: So a mind in CTP theory probably won’t start out with an idea like that. Or if it does, maybe that would be analogous to an inborn idea in people. It would be subject to change and could be overwritten almost immediately. Do you see something like that being the case?

Ella: I see what you’re talking about. And I definitely do think that the process of deciding which problems to work on is itself a creative process, right? You have to come up with guesses and, you know, and there are different ways that you could choose to prioritize your time. And I see what you’re saying. I see what you’re saying. And I, this is something I’ve thought about before. And I think that the solution here is going to be, you know, imagine that you have the process. Well, there’s going to be some, you know, chunk of code in the implementation of CTP theory that, you know, decides which, what the problem’s going, or what problem the mind is going to work on next, what it’s going to devote time to.

And the way that I imagined this problem of, you know, needing to have creativity there could be solved is that, you know, the, the way that the mind chooses which, problems to work on is itself, influenced by the content of ideas in the mind, right? So, what I mean here is, well, perhaps a way to think about it is that, what to work on next, right, is itself a problem, right? It is itself, you know, it’s an unsatisfied requirement, right? If you don’t have an idea yet, it’s, I want to have a plan for how I’m going to spend my time in the next five minutes or something, what I’m going to, you know, spend my mental resources on.

And so I suspect that there is going to be a way, and I acknowledge that this is a problem, but it just doesn’t seem like a huge problem to me. I suspect that there’s going to be some way to simply have the mind, the part of, you know, the implementation that is responsible for choosing what to work on next, I think that that will just need to sort of take input from the surroundings, right, from the, or not the surroundings, from the ideas in the mind. And if it takes input from the ideas in the mind, then the mind can kind of decide what to do by shaping its ideas in the right way. So that’s kind of a vague description, and I don’t have the details exactly laid down, but I think that something along those lines is going to be the solution.

Dennis: Right on. Okay. One more quibble I have with CTP theory that came to mind as you were talking about it is this notion of direct contradiction. You said when we have two ideas and they have the same binary encoding except for the first digit, then that constitutes a direct contradiction, a direct logical contradiction, if I remember correctly.

But children aren’t born with a notion of direct logical contradiction, right? So a child wouldn’t know, I mean, let alone the fact that he can’t speak a language yet, but let’s assume that he could somehow, and you told a child, my cat is female and my cat is not female, and let’s say somehow just a, it’s starting to be a rather contrived example, but let’s say that the child understands what all of this means, but the one thing he doesn’t understand yet is that the statements are in direct logical contradiction.

That strikes me as true, that logical contradictions, what a logical contradiction is, how to detect it, what to do about it, those are all themselves ideas that are quite complex and sophisticated in themselves. I could imagine that people may have different ideas about what it means to be in direct logical contradiction with each other. How do you address that in CTP theory, where the concept of direct logical contradiction seems to be sort of given from the start and also set in stone?

Ella: Yeah, so it’s a good question. And I agree that sort of articulating what exactly a direct logical contradiction is and how you can detect one, that is a very complicated thing, right? Being able to articulate that and explain that in explicit terms. I agree that that requires a lot of knowledge. But just because a mind works that way on a low level, that doesn’t mean it’s going to be able to articulate that. So the mind isn’t going to start out with explicit knowledge about how to detect direct contradictions. It’s just that that’s going to be what’s happening on a low level.

And so the mind could develop a theory of what’s happening on a low level within itself, which is to say a theory of direct logical contradictions, etc., or it could develop an incorrect theory, right? So a mind working according to this theory could, you know, be an empiricist, right? It could be completely wrong about the way that its mind works. And so the fact that it sort of works according to this, you know, algorithm for detecting direct logical contradictions and, you know, coming up with conjectures in an attempt to solve them, that doesn’t imply that it’s automatically going to have the correct epistemology or, you know, know how to articulate exactly what a direct logical contradiction means, right? That’s just the way that it works on a low level. And the sort of high-level properties about, you know, what it knows how to articulate, what its theory of knowledge is, that stuff isn’t built in by CTP theory.

Dennis: Got it. Yeah, what you said about the empiricist mind being wrong about how it itself works, it strikes me that really much of epistemology is just minds getting to know themselves better.

But, okay, so what about the, what about the, I mean, you said it’s indeed very complex and sophisticated to have an explicit theory of what it means to be in direct logical contradiction with each other, sort of to put that into words. And I agree. But when I detect a problem, say, just a feeling of curiosity, there’s something missing there or there’s something wrong, sometimes I can’t even quite put it into words yet, and yet my mind is telling me, so to speak, that there is something interesting to be explored there. So I’m guessing that that actually could be in favor of what you just said, is that there’s some underlying mechanism, some lower level mechanism that detects problems in that sense, and then you, your mind still has to sort of make sense of it, right?

Ella: So, so in that case, it might be that, you know, there’s a contradiction between two ideas that aren’t, explicit ideas, right? They’re ideas kind of in your subconscious, and they are still, you know, directly logical contradict, or directly logically contradictory. You know, everything does come down to that according to CTP theory, but that doesn’t mean, you know, a mind is always going to know, is always going to know how to articulate or how to think about or how to model the contradiction. All it’s going to know, you know, all that it’s going to be built in with is the ability to notice direct logical contradictions.

That doesn’t mean that it’s going to have, you know, a very logical way of thinking about them. It doesn’t mean that it’s going to know how to articulate them or how to put them into words. But yeah, and so, so that kind of feeling of a curiosity that you can’t quite articulate, I think, is pretty easily explained as just, some sort of contradiction or some sort of unfulfilled requirements on a subconscious level.

Dennis: Okay. But then if, if it was true that the mind detects this contradiction on a low level, and it’s just these two bit strings that are the same, except for the first bit, then it, it seems weird that the higher level mechanism or the higher level mind or the I or whatever that consciousness is that then is actually aware of the problem, why couldn’t it just look at those two bit strings, not, you know, I mean, figuratively, why, if it’s really as simple as comparing two bit strings, how come that information isn’t accessible to the mind directly? Why does it have to, why is there an additional layer when, when there’s a direct contradiction readily available already within the mind?

Ella: Yeah, so I think I see what you’re getting at. And, the, the way that I, would think about this is that, the mind has, so just think about the part of the mind that is responsible for putting things into words, right? Like that part of the mind is going to consist of, you know, some set of ideas, right, that takes in an input, which is just, you know, obviously going to be a bit string, you know, of some form. And it’s going to return a bit string that is something that, you know, that the mind recognizes as an articulation of, you know, in explicit terms, right?

So obviously, as I said before, you could use a bit string to represent an English sentence, right? And so there’d be a part of the mind that was responsible for taking in ideas, right, of any form and returning ideas in a, you know, explicit recognized form where bits represent, you know, words or letters or, you know, something that is very easy to translate into spoken word, right?

And so that part of the mind would be responsible for turning thoughts into, you know, putting things into words, for articulating things. That part of the mind is going to be something that’s created via conjecture and problem solving, and it’s not going to be perfect, right? So there’s going to be certain types of ideas that it can take in and, you know, correctly articulate, but it might take in some ideas and then incorrectly articulate them or just not be able to produce any output, you know, just, you know, the program, you know, runs into an error and just, you know, it doesn’t produce any new idea.

And so I guess, to sum up what I’m saying is that the mind’s ability to articulate things and to think about things on an explicit level is itself something that is, created through conjecture and problem solving, and it isn’t perfect. And so that’s why the mind might be able to, you know, articulate some of its ideas, but some of its ideas, it doesn’t know how to articulate. It doesn’t have that knowledge yet. And so all that, you know, you get is just kind of a feeling, a curiosity, you know, that something’s wrong without being able to articulate it at all.

Dennis: Yeah, although I guess the question, I understand what you’re saying. I guess the question remains, though, why does the mind even evolve those higher level ideas about what it means to have a problem and how to interpret it and so forth, if that underlying information is readily available? Like there’s, according to CTP theory, there’s some underlying mechanism that compares ideas directly. I guess I don’t fully understand why the results of that mechanism wouldn’t be directly accessible. In fact, they have to be accessible at least to some degree, because otherwise you would never even notice any conflict. I mean, there has to be some sort of chain reaction set in motion for you to even experience curiosity.

Ella: Yeah, so they definitely are accessible. I’m just saying that they’re not, the fact that you can access them doesn’t mean that you know how to translate them into words, right? You know, so you, your mind will notice there’s a contradiction there and you’ll feel that, or maybe there’s some contradictions, you know, that don’t even make it to the level of consciousness, right? You don’t even, it’s hard to say, but maybe the mind, you know, just solves some, problems without even ever alerting your consciousness. So you don’t even get that, you know, feeling of curiosity or feeling that something’s wrong.

But, but even when you do get that feeling, the, the fact that the mind has noticed a contradiction doesn’t mean that it knows how to put into words, you know, all the, all the details of the contradiction. I don’t know if that, I don’t know if that, you know, is satisfying for the question you’re trying to ask, but, that’s, that’s sort of the way that I think about this.

Dennis: Right on. Is, would you explain the phenomenon that two people might not find the same problem in the same set of sentences? Let’s say, actually, there might be multiple answers to this, but what you just said, does that explain that somebody who, somebody might readily find a problem in something you say and someone else might not?

Ella: Yes, I think that that is well accounted for in CTP theory. And I think going back to the example of the calico cat will be nice here. So imagine that you didn’t know, right, that all calico cats are female. And I told you, I have a problem. You know, my cat is calico, but also my cat is male, right? You would say, what problem are you talking about? There’s nothing there. That’s just, you know, you’re just talking about your cat. I don’t see any problem. And the difference between you and I, you know, would be that I have this idea, you know, all calico cats are female, which is allowing me to derive an idea, right, that’s making me notice a direct contradiction, whereas you just don’t have that idea.

And so you were kind of saying earlier, it seems like, you know, CTP theory is almost just describing a science machine or a logic machine. And I see where you’re coming from, talking about things and all these low-level details can make things seem that way. But I want to clarify that I think CTP theory absolutely can account for, you know, people having completely different interpretations, for not seeing problems that other people see, for, you know, having wrong ideas and for behaving completely irrationally and so on. So the way that a mind behaves, the types of problems that a mind sees, depends completely on what the ideas within the mind are.

Dennis: Something that Popper stressed again and again is that the, I mean, I guess really the major breakthrough that Popper made is that he applied the notion of evolution, not just to biology anymore, but he applied it to knowledge generally. And so what Popper discovered is that our knowledge is the result of guesswork and criticism alternating between these two.

Ella: Yes.

Dennis: So this brings us to the thing that you and I have discussed privately and in the Four Strands group here and there, which is the notion of replicators. So I follow Dawkins and Deutsch, who was himself influenced by Dawkins here, that according to the neo-Darwinian picture, replicators are one of the essential ingredients of evolution. So basically we’re describing some, you know, for any evolutionary system, and it doesn’t have to be biological again, it could be anything, as long as it has a pool of replicators and they replicate imperfectly, which introduces variation, and there’s some selection mechanism acting on this pool of replicators, you get, complex adaptations can arise over time.

But you’ve said that you don’t think, and you follow Campbell here, I think is his name, that replication or replicators, there’s a difference there, actually, are not a necessary ingredient of evolution. Can you elaborate on that?

Ella: Yeah, sure. So just to, you know, disambiguate one thing really quickly, the term evolution means different things to different people. You know, some people are just, when they hear evolution, they’re just thinking of biological evolution. And in that sense, you know, replicators certainly are an essential characteristic or an essential part of biological evolution. And more broadly, you could say that they’re an essential part of neo-Darwinian evolution. So that includes genetics and memetics, right? And some other things as well. But the way that I’m using the term evolution is that neo-Darwinian evolution isn’t the only type of evolution.

And so kind of following from Campbell here, well, I think that maybe a good question to kind of articulate the differences between you and I is, you know, how does adaptivity arise? So, no, adaptivity can arise through, you know, the neo-Darwinian kind of evolution where you have a pool of competing replicators, you know, etc. But I think that there’s something that, I don’t think that that’s the only way that adaptivity can arise. So I think that the essential ingredients to the creation of adaptivity, and here I’m following from Campbell, is just blind variation and selective retention.

And so I guess a good analogy here would be that, you know, in computer programming, there’s this idea of, you know, the genetic algorithm or the genetic, genetic programming. And that is kind of an attempt to replicate the neo-Darwinian kind of evolution as a, you know, to create adaptivity in computer programs. But, sort of evolutionary genetic algorithms and genetic programming aren’t the only way of using blind variation and selective retention as a computer program to produce adaptivity. And there’s, for instance, there’s, another method called simulated annealing. I don’t know if you’re familiar with this, but it, it does, the same process of blind variation and selective retention, but it doesn’t have a pool of competing replicators.

Instead, the way it works is, there’s some details to it that I won’t go into, but you just have a single entity, right? So, you know, you have a, you could think of it as a gene pool of size one, right? There’s only one genome, right? And then it produces a variation of that genome. And then it simply decides between those two. And the way in which it decides between those two in simulated annealing can get kind of complex. But you can just imagine that for now that, you know, it has one, it creates a variation, and if the variation is better, it picks that variation, right?

And so I think that that process, right, of just blindly varying a single object, and then, you know, selectively retaining one of them, that can lead to adaptivity, right? And so there, there’s no notion of replicators. There’s just a single thing that becomes, that is varied, and then you, you know, narrow the pool down to one. And so I think that that, it might not be as efficient or as effective as neo-Darwinian evolution in all cases, but that is absolutely a type of evolution in the sense that it produces adaptivity through blind variation and selective retention.

Dennis: Yeah. One thing we’ve discussed before is that there is a difference between replicators and replication. And following Dawkins here, who’s said that he, so for example, when he speaks of the gene, he explains this in his book, The Selfish Gene, when he speaks of the gene, he doesn’t actually mean some specific string, in the genome. He, he’s speaking, and I forget what the, what the term for that is, so there’s a, there’s a biological term for like the, the part of the DNA string that specifies your eye color, say. He doesn’t consider that a gene. He considers the part of the DNA string that manages to stick around unchanged over many, many generations, he considers that a single gene. So this may code for many different things.

So likewise, we could argue, in a system that doesn’t employ replicators, there’s still replication in the sense that even if you discard at every stage what you previously had in favor of your variation, or maybe not even that, but you always only have one instance, right? And you keep the better one according to some criterion. You could still argue, following Dawkins, that those parts of the thing that you’re mutating, that you’re varying, that weren’t changed, that weren’t touched by the variation, if they managed to stick around over many, many instances, over many, many steps of this iterative process, we would then consider that the replicator. Does that make sense? Or we might not consider it a replicator, at least not an active replicator, but there’s still replication going on in the sense that something stuck around over many, many generations.

Ella: Yeah, so I see what you’re saying. And if we’re going to define sort of the neo-Darwinian notion of a replicator or replication in that broad of a sense, then maybe we don’t really have much of a disagreement. But when we’ve talked in the past, I’ve kind of gotten the sense that you think that AGI, right, however it’s created, must involve a population of sort of competing replicating entities, right? It must involve replicators that are, you know, there’s a bunch of copies of each replicator, and each of those copies is trying to, you know, create more copies. But occasionally they, you know, fail, they have an error and they, you know, produce something that isn’t quite an exact copy. And then that thing can go on to, you know, take over if that happens to be more adapted to, you know, the current environment. And so am I correct in sort of thinking that that’s your model of how AGI must work?

Dennis: I don’t think I necessarily think that it must involve replicators. But I do think that if we adopt replicators, the explanatory power of the theory goes up tremendously, and it gains reach because we can explain many other things as well. And it is consistent with other theories of evolution, like genetic evolution, memetic evolution. There is a question to be answered there. Why should the evolution in a single mind be so fundamentally different from evolution that we’ve observed elsewhere in the universe?

And so I follow Dawkins here, who says that, I guess maybe the fundamental disagreement between Dawkins and Campbell, if there is such a disagreement, maybe they’ve already talked about it, I’m not aware of it. I guess the fundamental disagreement comes down to, is what Dawkins says, that evolution works the same, roughly, modulo some parameters that change. Does evolution work the same everywhere in the universe? Meaning every evolutionary process that we could observe, whether it be genes or memes or ideas in a single mind, do they all involve the same basic principles, yes or no? And what I’m hearing is that Campbell would argue no, unless we kind of twist his arm a little bit and still manage to accommodate the notion of replication, albeit not the notion of replicators necessarily, but the notion of replication into his theory.

Ella: Yeah. So I think that Campbell would say, and I would say, that there are sort of fundamental aspects that you’ll see in all processes of evolution. And those aspects specifically are blind variation and selective retention. It’s just that replicators aren’t one of those fundamental aspects to all processes of evolution. And I might be misremembering what Dawkins argued, but I thought his argument was just that wherever life occurs in the universe, whenever there is complex life that looks vaguely like what we have on Earth, that must have arised through neo-Darwinian evolution, through chemical replicators that eventually lead to more complex organisms. And that I agree with.

I don’t think that there’s any complex life out there in the universe that evolved by something other than neo-Darwinism. I’m just saying that more broadly, the notion of evolution as, you know, the creation of adaptivity, I don’t think that all of those processes have to involve neo-Darwinian evolution, which is to say, you know, evolution involving pools of replicators. So I’m not actually sure if what I’m saying disagrees with Dawkins, though we might just have different recollections of his argument.

But I suppose, I’m not sure that I agree, as you said, that including replicators in the theory, has a lot of explanatory power. Or actually, before we talk about that, I guess, the reason why I think that there is, that the process of evolution that happens in the human mind and that will happen in AGI minds, is different from the, the process of evolution that happens in the biosphere or in the memetic world, memosphere, is that, as Popper argued, contradiction must play a really important role here.

So whenever two ideas are, you know, contradictory, that is a criticism, that is a problem, you know, in my way of thinking about things. And so there isn’t really this notion of a contradiction in, you know, genetic or even really memetic evolution. And so I think that the fact that we have to account for contradiction at kind of the root level of how the mind must work, I think that that sort of hints that there’s some different type of thing going on here.

And it’s more about individual ideas contradicting other individual ideas, rather than, you know, a bunch of copies of ideas competing for dominance in some, you know, fitness landscape.

Dennis: Yeah, maybe, it definitely could be the case, although I think this could be accommodated relatively easily within a neo-Darwinian model of the mind where we instead promote the idea of contradiction from a lower-level concept to a higher-level concept, so where you argue that a contradiction, or detecting whether or not something is a contradiction, that is itself an explanatory theory. So it is quite sophisticated, and so it has the appearance of design, and so it must have itself arisen out of an evolutionary process. It is not built in at the bottom, so to speak.

So within a mind that involves replicators, at the bottom, so to speak, where those replicators compete, some cooperate, some compete, and so forth, there wouldn’t be any direct notion of contradiction. There might be ideas, and sooner or later this seems to happen in every mind, because it’s useful for criticism maybe, that sooner or later an idea evolves that happens to incorporate an explanatory theory of what it means to contradict something. And so then the mind can then use that to criticize other ideas. Whenever an idea arises in that idea pool, that is a very useful survival technique because it gets itself invoked a lot by the mind to criticize other ideas. So I don’t know if the notion of contradiction, though I agree that it is of fundamental importance for creativity, like Popper said, I don’t know if it has to exist on a low level.

Ella: Okay, interesting. Yeah, I guess just, if, I’m not sure that it would really, you know, make sense to call it contradiction, if it’s just some, you know, something that some evolved idea, you know, considers two ideas to be contradictory. I don’t know. It just, I think that, I think that the notion of contradiction needs to be more fundamental than that, not just something that might evolve and might be useful once it evolves. I think that, I mean, my guess, following from Popper, is that that is, you know, one of the most essential characteristics of how the mind works, is detecting ideas or detecting contradictions between ideas.

And so I should say, I think that you’re, you know, the vaguely, the sort of neo-Darwinian model of AGI that you’ve proposed, I think that a model like that could exist and could be interesting and could have genuine evolution going on there. I just don’t think that it would be much like a, you know, a conscious mind. I don’t think that it would really be an AGI. It would just be some form of kind of simulated Darwinism, which would still be, you know, an interesting achievement. But I just, I’m not seeing, I’m not seeing how exactly, we can be, how exactly, what exactly is making you think that this is actually going to be a mind in the same way that we have minds.

Dennis: Right. I mean, there’s several reasons for that. Our listeners may not necessarily be familiar with, with my neo-Darwinian theory of the mind, unless they’ve read my book, and many of them won’t have. So we sort of have to decide how much we want and how much detail we want to go there. I don’t disagree with you that contradictions are of fundamental importance. I agree with you and Popper there.

The problem I see, and one of the problems that I think need to be solved, I like to think I solved it, if you want to count that, but one of the problems I see with Popper’s philosophy is that he describes this interplay of conjecture and refutation as an evolutionary phenomenon, which I think is correct. But the problem that remains is that conjectures themselves always, I think, have the appearance of design themselves. And the criticisms themselves do too.

So it’s not just that the, that I think arguing that, and you didn’t argue this, but Popper did, I think arguing that, conjecture and criticism sort of egg each other on, so to speak, and so improve each other, I think that isn’t quite sufficient to say, to explain the appearance of design that both conjecture and criticism have by themselves. Every conjecture, I’m not just— Go ahead. Sorry.

Ella: I guess I was just going to say, what exactly do you mean by conjecture and criticism having an appearance of design? I’m not really sure that I get what you mean there.

Dennis: Well, so an example of problem solving that I’ve given in the past is, I think, I thought I put my keys on my kitchen table, and now I go look for them and I can’t find them. That would constitute a problem in a very practical sense. And so now any solution I think of, I mean, so one thing I could say, oh, maybe I put them on my desk, or maybe they’re still in my pocket, or I left them at my friend’s house, whatever. All of these have the appearance of design in the sense that they all make sense. They’re well-formed, both in terms of language, in terms of logic. They don’t have any internal contradictions. You know, they all already serve the purpose of solving this problem.

Now, we may then find through criticism that most of them are false and only one of them is going to be true. But there’s something going on underneath the hood here very rapidly that generates this conjecture, maybe the keys are already in my pocket, so then I go check my pocket. But presumably, the vast majority of ideas that are involved in this process never even come to the surface of this conscious exploration.

Ella: Right, definitely.

Dennis: And so that is what I mean, is that the conjectures that we become aware of, those have the appearance of design because, like I said, they’re well-formed, they don’t have internal contradictions, and so forth. So that is why I’m guessing that there is a much larger underlying evolutionary process going on. And I think that, I would imagine that this process involves replicators.

Ella: Right. Okay. Well, so I’m with you that, you know, there must be a lot of, the mind must be coming up with far more conjectures than actually rise to the level of consciousness, right? And I think Popper would agree there, probably, you know. But it just, I’m not seeing how introducing the idea of replicators really solves that problem at all. I think the solution to that problem, right, the problem of the appearance of design, is simply there is some kind of evolutionary process going on, and we’re only seeing, you know, a fraction of the results of that process, right? We’re only seeing the most well-adapted ideas.

But, you know, it could just be that the mind is not, the evolution there, the evolution going on there isn’t the neo-Darwinian kind, isn’t the kind that involves replicators, it’s kind of more the kind that I’m advocating for, where it’s just, you know, there’s only ever a single instance of each idea, and new ideas are created as sort of blind variations and recombinations of existing ideas. And so I just, I get the problem you’re pointing at, but I just don’t see how the notion of replicators adds anything to the solution, rather than just thinking of the solution as being, you know, some kind of evolutionary process, maybe involving replicators, maybe not.

Dennis: Sure. Well, I think the notion of replicators adds several beneficial things to a theory of the mind, which I can get into if you like. But even in this particular instance, if I recall correctly what you just said, you did grant that there does seem to be a vast number of ideas involved in coming up with a good conjecture, say.

Ella: Yeah, definitely.

Dennis: So what you get when you have a pool of replicators is, when a replicator replicates, the original sticks around, right? So the number of those, and again, modulo competition and any self-regulating forces that exist within that pool of replicators, you have several copies of that replicator floating around in that pool at any given time. It seems to me that in something like CTP theory, or in general, in something that doesn’t employ replication, only variation and selective retention, you always only keep a single instance of an idea. Is that right?

Ella: Yeah, that’s right.

Dennis: So it seems to me, just in terms of tractability here, you would lose the vast majority of possibilities within that idea space. You would have to generate, you would have to employ selective retention, or you’d have to employ variation at every stage. But so you’re basically, if you picture it as a tree, you’re going down a specific branch of that tree, but losing potentially all of the other branches that you could go down. And so, the, most of the tree remains hidden, it seems to me.

Ella: I don’t see what you’re losing if you only have one copy of each idea, right? You still have all the information there, right? You still have, you know, you still have the, all the ideas, at least one copy of them, to use as kind of the basis of new conjectures. So I don’t really see what you’re losing by not having many copies of each idea.

Dennis: So in Campbell’s model of blind variation and selective retention, does he, once you make a variation to an idea and it is better than the original, does he keep the original? Does it stick around or does he discard it?

Ella: Well, so Campbell didn’t have a particular theory of AGI or of how the mind works.

Dennis: Sure, in CTP theory, how you picture it.

Ella: Sorry, so go ahead and ask the question again, just so I make sure I answer it correctly.

Dennis: When you vary an idea blindly, and by some criterion, say explanatory power or whatever it might be, you then decide that the variant is better than the original. Does the original get to live on, or is it discarded?

Ella: So, conjectures, generally speaking, are motivated by problem solving, right? So, the mind is going to be creating conjectures in an attempt to solve a problem. So if there is a problem with some idea, you know, an idea A, and then you create a variant idea, an idea B, right, that, you know, is based on A, so it’s a variant of A, and it turns out that B doesn’t have the problem that A had and maintains all the explanatory power in terms of, it satisfies all the relevant requirements, then A will be discarded, right? Because it’s causing a contradiction, and because it doesn’t need to resolve the requirements anymore because B is now doing that.

So in that case, A would be discarded and B would be kept around. If A actually isn’t problematic and there’s no reason to get rid of it, then yeah, you could have an idea and a variation of that idea both in the mind. There’s no reason you couldn’t have that if they aren’t causing any kind of contradiction.

Okay. Does that make sense? Does that answer your question?

Dennis: Yeah, I’m still processing it. So basically what you’re saying is it depends. Sometimes the variant and the original both stick around if there’s no contradiction between them.

Ella: Yeah, so the mind, broadly speaking, the mind only gets rid of ideas when one of those ideas is involved in a contradiction. And so getting rid of it solves that contradiction, solves the problem. And so, you know, if you have an idea and a variant of that idea, and neither of them is involved in any problems, then yeah, you can stick, you can keep both of them around. But, you know, if one of them is involved in a contradiction, then maybe you get rid of them if that helps solve the problem.

Dennis: Yeah, basically that brings us back to the notion of contradiction as the fundamental regulating principle.

Ella: Yeah, that might be our biggest disagreement. It seems to me that contradiction doesn’t just need to be present. It kind of needs to be present on a low level. And that kind of has to be one of the fundamental driving forces for the mind. Whereas if I’m understanding your argument correctly, you think contradiction might be important, but it’s something that can sort of evolve later. And it’s just something that the mind can kind of discover for itself. Is that right?

Dennis: For sure, yeah. Theories involving contradiction themselves have the appearance of design, so I think they must have arisen through an evolutionary process themselves. And I mean, we see this all the time, right, where people have, to others, clearly contradicting theories in their minds, but don’t discard either one, they’re happy to just accommodate both. Now, I guess you could argue within the CTP framework, you could say, well, there can still be higher level theories about contradictions. The contradictions you’re speaking of, those are only lower level.

Ella: Well, there’s that. And then there’s also that, you know, if I think two theories in your head are contradictory, but you don’t see that, that might be because, you know, they’re only contradictory in the context of some third theory, which I have and that you don’t have. So again, if I just return to the calico cat example, if you tell me my cat is calico and then you tell me my cat is male, right, and you don’t seem to notice any contradiction between that, but I do, maybe that’s because I know all calico cats are female and you don’t know that. Does that make sense?

Dennis: Right, definitely. But there are cases where people are just sort of in denial. They have the knowledge that in principle tells them that their ideas are contradicting. But they can be ideas that they hold very dear. And so they just lie to themselves. I think people do that a lot, actually.

Ella: Yeah, that’s definitely something that can happen.

Dennis: I don’t want to paint a cynical picture of people. I think people in general are great. I’m just saying, yeah, definitely a lot of people are happy to accommodate cognitive dissonance. And so the question then, even though the knowledge for how to resolve that dissonance is in principle accessible to them, they have generated that knowledge. So in that instance, it strikes me what’s going on, I think, is, to bring it back into a neo-Darwinian picture, is for some reason, the idea that would resolve the conflict finds itself in heavy competition with the ideas that are happy to facilitate the conflict and to keep it around. And those ideas that facilitate the conflict have an easier time spreading through that pool of ideas, and therefore are more numerous and stronger, and therefore the mind keeps the conflict rather than resolving it.

Ella: Why would they be more likely to spread throughout the mind? I don’t see why those would necessarily be stronger. Are you just saying that they could be, and in that case you would get that kind of, you know, irrationality, hang-up?

Dennis: Yeah, the details are always going to, they’re always going to vary depending on the ideas involved. The answer would be the same as, as to why, why do some species spread through the biosphere better than others?

Ella: Yeah. Okay. So it’s just kind of a case-by-case thing.

Dennis: Yeah. It’s a complex answer. There’s, there’s no simple answer.

Ella: Okay. So just, like the, the problem of, you know, obviously people and AGIs must be able to, you know, have irrationalities and hang-ups, right? That, that’s something that humans can do. So, so surely that must be something that AGIs are at least capable of, even though, you know, obviously we don’t want them to. But that is something that their minds must be capable of housing.

And I haven’t really considered that in the framework of CTP theory before. But just kind of off the top of my head, the way that I would model something like that is that, so there are two contradictory ideas, like, you know, my cat is male and my cat is calico. And, you know, there’s an obvious solution to this problem, and maybe the person is even aware of a way to resolve that contradiction, but there is some desire or some requirement in the mind, right, that getting rid of either of those ideas would, would stop, right, and, or would, you know, leave unfulfilled. And so, for that reason, that they don’t want to get rid of either of them.

And so, you know, it’s kind of odd in the, in the example of, you know, my cat being calico and male, but, you know, imagine that somebody is really, really attached to the idea that their cat is male and the idea that their cat is calico, right? And even though they know those ideas are contradictory, you know, the fact that they’re so attached, which in lower level terms means that they have, you know, some sort of requirements that they need fulfilled, that those ideas are fulfilling and that no other ideas are fulfilling. In that case, you know, they would see the contradiction, right, but they would just, you know, withstand the cognitive dissonance, because it’s going to leave, you know, because otherwise they’d have to, you know, think about those desires and find some other way to fulfill them, or to, you know, remove the desires by getting rid of whatever caused them in the first place.

Dennis: Yeah. I mean, in general, what you just pointed out at the beginning there is, I think, is super important, that most models of AGI miss entirely, is that an AGI is capable of having irrational, irrationalities, hang-ups, bad ideas, you know, all these kinds of things. I think most, most models of AGI just forget about that completely. They just want some super, superhuman or superintelligence that just solves problems reliably, which, which both you and I agree on can’t exist.

In CTP theory, when you, and remind me if I’m, or tell me if I’m remembering this correctly, please, because we just touched on it five minutes ago, you said a theory is only discarded if its variant is considered better and is in direct contradiction with it. Is that correct?

Ella: Well, it’s not just if the variant is in direct contradiction with it. So you have two ideas, A and B, and A and B are in conflict. And then you create a variant of A, we’ll call it C, right? And C satisfies all the requirements that A satisfies, so it maintains all the explanatory power. And then so you might get rid of A because it contradicts B, right? Not because it contradicts C. So it doesn’t have to contradict its own variant. It’s just that if it’s involved in some contradiction with some idea in the mind, and its variant is a satisfying replacement for it, in the sense of having all the explanatory power and filling all the requirements, then that’s the case in which you would get rid of an idea.

Dennis: Got it. So at that point, that idea is gone. The mind, like, it’s either deleted explicitly or garbage collected at some point. The mind doesn’t have any remaining reference to it.

Ella: So, I mean, the simplest answer is yes. But I mean, it might be worth keeping around at least like a copy of old discarded ideas, just to kind of maybe be the basis of new conjectures in the future. And so, like, it wouldn’t be an active idea in the mind, you know, producing implications and interacting with other ideas. But if it turns out to be, you know, efficient to keep around a copy as the basis of future conjectures, then, you know, surely you could do that. And that wouldn’t be in violation of CTP theory at all.

Dennis: Yeah, okay. This is one of the, it brings us to this notion of memory, that, and I don’t mean memory in the sense of, you know, storage space on a computer. I mean memory as in remembering things. Basically what I was trying to see is if CTP theory can accommodate the, what we know about how memory works or seems to work. I think many of the neuroscientific theories there are completely wrong. I’m guessing you may at least partially agree there. We’ll get into that later.

But yeah, one of the, I mean, if you can accommodate that within CTP theory, because I think memory, as in remembering things, is just one of the emergent properties that a creative mind has. I don’t think there’s anything special or extra going on. You know, there’s, I don’t think there’s a separate committing-something-to-memory algorithm. I think it’s part of the creative process. But it sounds like you could accommodate this in CTP theory, but you would need a little bit of extra logic there. Like you just said, you would, in addition to the ideas that are still alive, so to speak, that are still accepted by them, or that are still in sort of this main, how should I put it, in the main idea pool, I don’t know if you use that term, actually, but that set of ideas, you just introduced this idea that you could have a separate set of ideas, of things, of ideas that you’ve discarded in the past.

Ella: Right. So actually, I, when I recently, when I was writing the new article that I put out about CTP theory, I initially used the word idea pool, but I didn’t want to, you know, steal that from you and get our two theories confused. So I just called it, like, the idea set or something. So for the sake of not being confusing, I tend not to use the word idea pool, because I don’t want to imply that there’s anything neo-Darwinian going on.

But so I see what you’re saying. And this is something I’ve thought about before, because being able to remember things is really important. And I actually don’t think that sort of the normal process of just memory requires that specific sort of additional system that I was talking about. That’s just sort of maybe an optimization thing for the process of conjecture. But I think that actually normal CTP theory without that addition can still account for memory, though it requires a little bit of explanation to see exactly how it would do that.

Dennis: Go ahead.

Ella: Yeah, sure. So if, again, I’ll use the example of the perihelion of Mercury and replacing Newton’s theory with general relativity. Say you’re a scientist, right, and you read, you know, Einstein’s new paper about this theory called general relativity that supposedly has all the explanatory power of Newton’s theory but doesn’t have the contradictions that Newton’s theory had. So if you read that theory, or if you read that paper and you’re convinced, you don’t just forget what Newton’s theory says, right? You don’t just, right?

In a certain sense, general relativity replaced Newton’s theory in your mind. And it replaced it as, you know, your model of the correct theory of how objects move under gravity, but it doesn’t replace it in the sense of you just can’t even remember what Newton’s theory is anymore. And how this would work on a low level is, imagine that you have an idea that says Newton’s theory is governed, or Newton’s theory says motion is governed by these equations, or gravity is governed by these equations, right? And then you have another idea which says Newton’s theory is true, right?

So if, if you have that first idea, right, which is, the, you know, in quote, “Newton’s theory says X”, end quote, right? And then you have another theory, quote, “Newton’s theory is true”, end quote, then if you get rid of that second idea, then you still have what Newton’s theory says in your mind. You just don’t any longer think that it’s true. Does that make sense?

So, so you could still have, you know, like, so I could have an, I, you know, I know that Newton’s theory is false, but in my mind, I have, you know, an idea presumably that says something akin to Newton’s theory says, you know, F equals MA and, you know, the gravity equation, et cetera. And the fact that I have, the fact that I have that description of what Newton’s theory says doesn’t mean that I believe that Newton’s theory is true. And so, you know, Newton’s theory says X is still a fact about the world, right? That’s, it’s not, it’s not false that Newton’s theory says this just because Newton’s theory is false.

And so that’s kind of how I imagine that CTP theory would be able to, you know, accommodate even ideas that it’s decided aren’t true anymore. It would just, you know, you would have some of the remnants of the old theory, you just wouldn’t have the idea that those remnants are actually true. You would just have, you’d have the ability to say, oh, well, Newton’s prediction, or Newton’s theory would predict this. You just don’t say this is true, right? It’s just, that’s what Newton’s theory would predict. But of course, Newton’s theory is wrong. So that’s not actually true.

Dennis: Right. Okay. I understand. So then let’s invert it. How do you explain the phenomenon of forgetting in CTP theory?

Ella: Yeah. So that’s an interesting question. If you just sort of imagine how CTP theory would work in a basic sense, I imagine that the mind would just sort of, you know, keep growing, right? New ideas would just keep getting created as, you know, part of the problem solving process, as part of conjecture. And the set of ideas in the mind might just grow without end. Now, that’s problematic if we’re actually intending to implement this on a real computer, because we have memory constraints.

And so, you know, in a real implementation of CTP theory, unless you’re just planning on, you know, continuously buying more hard drives to have more memory space, you’re going to need to have some mechanism for saying, okay, once there are 10 million ideas in the mind, we have to cut back a little bit, right? Whenever the 10 million and one idea is created, we need to get rid of one just so that we have enough memory space. And, you know, exactly how you would choose which ideas to discard is kind of one of these, another sort of small open problem with CTP theory.

I don’t mean small as in it’s not important. It’s just that I don’t feel the theory is sort of developed to the point where we can really make, where we can really make, you know, concrete answers to that. I think that once we have sort of the explanatory power problem solved, and we’re actually, you know, thinking about implementing the AGI fully, then those ideas will be kind of what we need to think about. And so I guess just for now, forgetting will happen by some mechanism that exists to conserve storage space. And it’s just, I don’t know exactly how that mechanism will work exactly.

Dennis: Sure. Okay. I bring it up because, I mean, for one, because I was just interested in how CTP theory might solve this, but also because I think that is something that the neo-Darwinian theory of the mind has going for it, which is that phenomena like memory, but also the inverse, forgetting, are simply solved just by virtue of describing the mind as this arena of replicators, because you’ve basically forgotten an idea at the point that all the replicators that encoded it die out. So there’s nothing really left to explain there.

But let’s go back to the difference between Dawkins and Campbell then, which, the point that Dawkins brings home repeatedly in The Selfish Gene is that evolution is this process that, it favors three things. And I think only those three things. That is copying fidelity, longevity, and fecundity. So how accurate are the copies: evolution will favor replicators that are very good at copying themselves, that don’t make a lot of mistakes. It will favor replicators that have longevity, meaning those that manage to stick around longer than their competitors will be favored by evolution. And fecundity, in the sense that those that create more copies of themselves will be favored by evolution.

So if these are the driving forces of evolution, then again, it’s, I wonder if, what, it seems to me that what Campbell is describing is sort of a version of evolution that, I’m using the term very loosely here, but that denotationally looks the same, but structurally is entirely different. And in fact, all these driving factors, these three driving factors, they all depend on replication. And so it seems to me, and correct me if I’m wrong, that Campbell’s model of evolution misses those three. Am I wrong about that?

Ella: So I see what you’re saying. And I guess I think the way to resolve this problem here is just that Dawkins is using the term evolution in a more sort of, in a sort of smaller scope way than Campbell or I am using it. And what I mean is that, so as I said before, I consider evolution just to be sort of any process which leads to adaptivity. And I think that any process which does consistently lead to adaptivity is going to have to involve blind variation and selective retention. And that’s what Campbell argued.

But when Dawkins uses the word evolution, right, in The Selfish Gene, say, he’s referring to biological evolution, which is understandable, given that he is a, you know, an evolutionary biologist, or, well, I should say, he’s referring to, yeah, that’s what I was about to clarify. He’s referring to what I would just call neo-Darwinian evolution. So that includes, you know, genetics and memetics and some other stuff as well. And so when Dawkins says the word evolution, he means something more specific than what I or Campbell mean when we talk about evolution.

And actually, I might not even be representing Campbell here. I should probably just speak about what I think, because I’m not really sure exactly if he uses evolution in exactly the same way that I do. But the way that I think about it is that evolution is just any process that creates adaptivity through blind variation and selective retention. And neo-Darwinian evolution is a kind of evolution that involves the, you know, the pools of replicators that Dawkins describes.

And so I think that Dawkins is correct in saying that, you know, those three characteristics are what leads to the success of replicators in neo-Darwinian evolution. But I think it’s a mistake to try to apply that to evolution in the broader sense, because there just simply isn’t replication in all forms of evolution, such as sort of simulated annealing, as I talked about earlier. There’s not really, you know, the neo-Darwinian notion of replication there. And so talking about fecundity just doesn’t really make sense. So I think that Dawkins, you know, those three characteristics, those are a good description of how neo-Darwinian evolution works, but they don’t apply outside of the scope of neo-Darwinian evolution.

Dennis: Yeah, I wonder if, I mean, Dawkins certainly seems to speak in universal terms in The Selfish Gene. It’s just, when I read the book, it certainly seems, it seemed to me that what he was describing was a universal phenomenon. I don’t recall, I might certainly be wrong about it, whether he left room for any other models of evolution. I guess it would be really interesting to find if Dawkins ever addressed Campbell’s suggestion directly.

Ella: Yeah, it would be interesting because, yeah, like, I think that there is something very universal and deep about neo-Darwinian evolution. You know, I agree with Dawkins, as I said before, that anywhere that life arises in the universe, it will be through neo-Darwinian processes. It’s just that, you know, a slightly deeper understanding of what is going on on a logical and epistemological level, reveals that, well, neo-Darwinian evolution is, you know, it’s common, you know, it’s happened a lot in our universe, so it’s easy to see, not all types of evolution, which is to say, you know, adaptive processes, not all have to, have replicators. It’s possible to get, you know, an adaptive, it’s possible to get adaptivity without replicators at all.

And so I think that, I don’t know what Dawkins, thinks about this. I, you know, I might be misremembering, you know, exactly. You might be right. Maybe he does think that it’s completely universal. And in that case, I would disagree with Dawkins. I think he’s slightly wrong. It’s, it’s quite universal. It’s quite important, but he’s making it out to be a little bit more universal than it actually is.

Dennis: Yeah, it does. It does bring up the question of why Dawkins then thought of memes as being replicators when he could have, it’s, it’s a little more difficult to say in hindsight, because once you understand meme theory, it’s harder to imagine how that could have worked without replication. But when he was coming up with it, assuming Campbell’s work was already done, he could have been aware of it, maybe not. But it seems to me that he could have conceived of memes as arising through something like simulated annealing, and replication just sort of being a, you know, the simulated annealing could happen within a mind, and then the replication could just be sort of incidental, that it spreads from my mind might not be really that important in meme theory. But in any case, that, of course, is not a conclusive criticism at all of what Campbell is suggesting. It’s just a curiosity of history, so to speak, that Dawkins didn’t phrase it in those terms.

Ella: So I do think that, I don’t think that meme theory needs to be, you know, rephrased as not working in terms of replicators. I mean, I do think that memetics is a neo-Darwinian process, right? So ideas do replicate when they pass from, you know, one individual to another. It’s just that I don’t think that ideas replicate within a mind. I don’t think that they ever create exact copies of themselves within a mind. Or if they do in the human mind, I think that’s just kind of, you know, a peculiarity of the way that humans evolved. And it’s not essential to, you know, the epistemological logic of what’s going on. And so I think that meme theory being phrased in neo-Darwinian terms is perfectly fine, because once you’re talking about the spread of ideas from one person to another, it makes sense to talk about them in terms of, in terms of neo-Darwinian replicators.

I’m not actually sure if Dawkins himself specifically claimed that it’s sort of a neo-Darwinian process within a mind.

Dennis: You can kind of just treat the mind as a black box in memetic theory.

Ella: Absolutely. And yeah, so whether or not the mind itself uses replicators to create adaptivity, it doesn’t really affect whether or not memes spreading between minds should be thought of in a neo-Darwinian fashion.

Dennis: Yes, I absolutely agree.

Ella: I don’t think Dawkins, at least in The Selfish Gene, talks about whether or not ideas replicate within a mind.

Dennis: Yeah, I don’t recall that either. I think you’re right. He only speaks about ideas replicating across minds. So yes, within meme theory, it can remain perfectly neo-Darwinian without introducing the notion of replication within a mind, and the mind can just simply be a black box that either works due to neo-Darwinian mechanisms or CTP theory mechanisms. But in any case, okay, although we disagree on the role of replication within a mind, we do, we should clarify, I think we actually agree on the vast majority of the implications of Popperian epistemology on AGI research.

Ella: Definitely. Yeah. Even though we have spent most of this, you know, talking about our disagreements, you and I are probably closer than almost any two people in the world in terms of, you know, how we think AGI will work, compared to, you know, other AGI researchers. We are, you know, kindred spirits.

Dennis: I agree. You mentioned early on in this conversation that perhaps, or maybe I’m misremembering, that perhaps you used to think, before you read The Beginning of Infinity, that you were much closer to maybe more conventional, to approaches to AGI? Were you interested in building AGI before reading The Beginning of Infinity?

Ella: Yeah, so I was definitely interested in AI. And, you know, I think anybody who has an interest in AI, you know, has somewhat of an interest in AGI, though that might not be, you know, their immediate goal. And so I was interested in kind of getting into the field of AI, you know, and this was when I was in high school and college, when I was, you know, reading The Beginning of Infinity, I think. And so I kind of, you know, I liked AI and machine learning and stuff. And I had, you know, some thoughts about AGI, but it wasn’t really until reading The Beginning of Infinity that I really started focusing on AGI rather than just AI and machine learning.

Dennis: Right. Okay. So what, what do you think, like, how has Popperian epistemology informed your opinions on narrow AI and machine learning and so forth?

Ella: Yeah, so I think that you and I mostly agree here, and that, well, the epistemological framework behind, you know, a lot of sort of machine learning, and this is often left implicit, you know, people don’t explicitly talk about this, but the epistemological framework that’s being assumed is kind of an empiricist, inductivist framework. So it’s just assumed that you can derive the truth about the world from sense data. You don’t need to blindly conjecture. You just need to sort of apply the right inductive process and do the right computation on the input data, and then you get the right answer.

And so I think that the fact that that is sort of underlying most machine learning algorithms means that those algorithms are never going to lead to true AGI unless, you know, the machine learning community comes to a much better understanding of epistemology and realizes that they need to, you know, be utilizing radically different methods.

Dennis: Yeah, I agree. What is your opinion of neuroscience?

Ella: So neuroscience, I know you’ve talked a lot about how you see sort of neuroscience as a sort of a dead end. And I agree with you mostly. I think that neuroscience isn’t likely to lead to AGI anywhere in the near future because of where the field is at right now. But I do think that if you imagine a sufficiently advanced field of neuroscience, maybe to the extent where we could just scan a brain on like a neuron-by-neuron level while a person was still alive and, you know, read the internal state of every neuron and, you know, all the connections between them, and, you know, then simulate that on a computer, if we had a sufficiently good understanding of how neurons work internally.

If you did that scan and then simulated it on a computer and played it forward, you would have what could be called an AGI. You would have, you know, a human mind that could do anything that any other human mind could do, just running on a computer. And so I think that a sufficiently advanced neuroscience could lead us to AGI, but I think that that’s very, very far away. And I’m hoping that, you know, kind of the epistemological approach will get us there much sooner than, you know, waiting for incredibly advanced scanning technology.

Dennis: Yeah, let me clarify that I think neuroscience is a dead end when it comes to AGI research. Of course, I think neuroscience, you know, when it comes to studying brain disease or stuff like that, is the way to go, because brain diseases are a hardware issue, whereas diseases of the mind, I think neuroscience is completely misplaced there, because that’s a software issue. I mean, there may be some things where, you know, just like when you spill a cup of water on your laptop and it messes with the hardware, then it will mess with the software in a certain sense, too, right? But there can arise software issues, even though the hardware is working perfectly correctly, in which case neuroscience couldn’t possibly help us diagnosing or, more importantly, curing those ailments.

But yes, I agree with you that applied to AGI, it’s mostly a dead end. And I also agree with you, though, that the field of, I think it’s called whole brain emulation, where you simply, you just scan, if you have a sufficiently small resolution, that buys you something, namely you don’t require an understanding anymore of what it is that happens in that brain.

Ella: Exactly.

Dennis: So it’s sort of a convenient way to do it. Not convenient in the sense that it would be easy, but yes, you don’t require an understanding. You can simply run that forward and it would just sort of work. But yes, you would still be very reliant on the underlying hardware structure of the brain. And of course, what we’re really after is we want to explain how AGI works, because then we can, well, for one, because it’s just fascinating, but also because then we could actually implement it from scratch as software without needing any understanding of the brain.

Ella: Yeah, exactly.

Dennis: Yeah, I do think that potentially whole brain emulation will happen first, though. If the, and that we should focus efforts on it. I do go back and forth on it a little bit, but, sometimes I do wonder if we should focus efforts on whole brain emulation just for the purpose of getting people off the planet. Right? Because once you run on a computer that’s not the brain, you’re then untethered, and you can, you know, you can, it would be trivial to send people to other, other planets.

Ella: Yeah, absolutely. Creating the first AGI, whether it’s through epistemology or through whole brain emulation, that’ll be a huge step forward in mitigating existential risks for the future of humanity. Once we can become a multi-planet civilization, it would be much harder for a stray asteroid or gamma ray burst to end our civilization. So yeah, I agree that, I think it’s definitely worth putting resources towards whole brain emulation and whatever fields of neuroscience are, you know, involved, since obviously we’re far away from actually being able to do that.

I don’t, I mean, following kind of, you know, Deutsch’s arguments about the impossibility of predicting future knowledge, I guess, you know, we can’t really say whether a whole brain emulation or a deep epistemological understanding of what’s necessary for AGI, we can’t say which of those will come first. So it’s definitely worth focusing on both.

Dennis: I agree there. Yeah. I guess there’s one remaining thing that we haven’t really touched on that we could touch on if you like, which is consciousness. I don’t think we’ve talked about consciousness much, you and I. What is your—

Ella: No, I don’t think we have.

Dennis: What is your take on it? Any, any thoughts on it? Have you thought about it at all?

Ella: Yeah. So it is something I’ve thought about, and, you know, I think that the hard problem is a real problem. I wish that I had, you know, an answer to the problem of consciousness, but I really don’t. And I think that David Deutsch has conjectured that a solution to AGI will be a solution to the problem of consciousness and vice versa, right? We’ll have an understanding of why we’re conscious once we have, you know, a theory of AGI. And I would hope that that’s the case too. You know, I hope that it’s true that, you know, we can solve both problems at once and that they’d turn out to be connected like that. And that certainly seems like a reasonable conjecture. You know, we have these two open, big problems about the mind. It wouldn’t be surprising if it turned out that they were related.

Unfortunately, CTP theory, as I’ve developed it so far, hasn’t given me any insight whatsoever as to how consciousness works. So that could be an indication, you know, if Deutsch’s conjecture is correct, that CTP theory isn’t going to work, right? Or that it’s, you know, I’ll need to have a better understanding of consciousness before I can make it work. Or it might just be an indication that Deutsch’s conjecture is wrong, right? You know, that the hard problem of consciousness is actually just a completely separate issue from AGI, and we can create AGI without a solution to it. So I wish that I had more to say on the subject, but all I really have to say is, you know, it’s baffling. You know, I don’t have a solution to it, though I wish I did.

Dennis: Right. What do you think of Popper’s conjecture that it has to do with disappointed expectations?

Ella: Yeah, so I definitely think that that is part of it. And to kind of phrase that in CTP theory terms, you know, any disappointed expectation would be a, you know, contradiction between the predictions of some theory and, you know, another theory, which, you know, interprets sense data to say, you know, this, that didn’t happen, that prediction didn’t happen. And so if that’s true, then, you know, perhaps that could be integrated into CTP theory. And that seems like a reasonable conjecture. I think that there’s something there, but, I don’t really have much more to say on it than that. I don’t, it, well, it’s certainly not, you know, a complete theory. That’s, you know, that’s more of the beginning, or, you know, a suggestion for a real theory rather than a theory itself. So, yeah, I wish I had more to say on the subject, but I don’t think I do.

Dennis: Okay. No, this is great. Well, then it was great having you, Ella. This was fascinating.

Ella: Yeah, thank you for having me. This was a great conversation.

Dennis: Likewise. Where can people find you?

Ella: Yeah, so I am a blogger for fourstrands.com. And so you can go there. I recently published an article on CTP theory, kind of giving more details than I was able to get into in this conversation. I also have my own website, ellahoeppner.com. Hoeppner is spelled H-O-E-P-P-N-E-R. There’s a silent O there. And I also post all of my writings there. So either of those two places would be a good place to sort of read what I have to say. And I’m also on Twitter at Ella underscore Hoeppner if you want to follow me there. I’m trying to be more active on Twitter than I have been in the past. But in the past, I haven’t been super active there. But if you want to follow me, then maybe that’ll motivate me to be more active. So yeah, I think that’s all the places that people can read my stuff.

Dennis: Excellent. Okay, well, I encourage everyone to check out what you’ve written about CTP theory. And yeah, it was great having you.

Ella: Yeah, thank you so much.

References

There are 2 references to this episode in:

• Post ‘The Neo-Darwinian Theory of the Mind’
• Episode ‘16 – The Neo-Darwinian Theory of The Mind’ of this podcast