Episode Transcript
[00:00:00] Speaker A: Welcome to ID the Future. I'm Andrew McDermott. Today's episode comes to us from our
[00:00:05] Speaker B: sister podcast, Mind Matters News, a production
[00:00:08] Speaker A: of the Discovery Institute's Walter Bradley center for Natural and Artificial Intelligence.
You can learn more about the show and access other episodes at mindmatters.
[00:00:19] Speaker C: AI.
[00:00:24] Speaker A: Hello, everybody. Welcome back to the Mind Matters Podcast. I'm your host today, Pat Flynn. I am joined once again by Dr. Robert Marks. And special guest today is Dr. Angus Manouj, a returning guest. He has a new paper out. The paper is called Powers of the Soul beyond AI, published in the journal Religions. A fascinating paper. It's going to make for a fascinating conversation, I'm sure of it. So, gentlemen, thank you so much for being here. How are the two of you?
[00:00:53] Speaker B: Very well.
[00:00:54] Speaker C: Just right.
[00:00:54] Speaker A: Just right. That's just good enough for me. So let's get a little background here. So, Dr. Minouj, it's been a minute since we've had the opportunity to speak together. If you wouldn't mind, Just reintroduce yourself to the audience, who you are, what you do, and give the background, the relevant background to this paper, if you don't mind.
[00:01:11] Speaker B: Yeah. So I'm a professor of philosophy. I've been at Concordia University, Wisconsin for 35 years. And my area of specialization in philosophy is the philosophy of mind and agency in particular.
And early on, my research looked at comparing the mind to machines, the thesis of mechanism.
This recent paper is an attempt to take stock of some of the most recent developments in artificial intelligence and to see if they provide any reason to think that AI can actually develop a mind of its own. And my view in the paper is, is that there is good reason to think that even the most advanced AI systems still lack critically important powers of the mind, or as I prefer to say, of the soul, because I think there's more going on here than just the mind.
[00:02:12] Speaker A: Great. Now, Angus, if I may. In the paper, you do articulate your particularly preferred version of the soul and the human person, but you also note that nothing about your paper or the arguments depends upon that particular conception. Can you tell us a little bit more about that?
[00:02:32] Speaker B: Yeah. So I'm attracted to the view of the soul that was developed by Augustine, according to which the whole soul is present everywhere in the body, where there is sensation, rather analogous to the way that God is omnipresent in the world.
I think that helps to explain the unity of consciousness and things like free will and so on. But I also think that you need the soul to explain our being, one organism, both at a time and over time.
And that is because otherwise we would just reduce to a bundle of parallel processes, and then we wouldn't be one person.
So that's the view that I have become most attracted to. But I recognize that many other defenders of the soul have different views. And the argument of my paper is really only to show that there is something immaterial. There's something beyond the material mechanism of the brain. And so it's really compatible with any of these views of the soul. So it's a good principle in philosophy that you don't want to fight about everything all the time.
And since this is not an issue that divides these different views of the soul, I decide to put that on the back burner and just argue, no, there is something in the soul that is clearly beyond even our most sophisticated machines.
[00:04:03] Speaker A: Yeah, yeah, great. Or even if you do want to fight about everything all the time, maybe you just shouldn't for the sake of a publication.
All right, so, yeah, awesome. Thank you for that background. So at the beginning of the paper, you introduced this idea of the anthropic mechanism. So can you help explain what that is?
Like, how broad your target is in the paper? Are you. Are you, like, just specifically arguing against the computational theory of mind, or are you trying to rule out physicalism across the board? Help us, help us understand your particular aims here.
[00:04:36] Speaker B: Yeah, I really am taking aim at a very general thesis that is sometimes just being called mechanism. Anthropic mechanism just means that human beings reduced to organic machines. It was suggested very early on by Thomas Hobbes, who says that material objects interact with our senses, causing motion and the senses causing motion in our brain. So therefore, our thinking is just motion in the brain. And he even suggests at one point that reasoning is only reckoning, which is an old word for computation that was developed by Julien Onfre de la Metrie, a French physician and philosopher who basically argued the soul is just the mainspring of the whole machine.
And of course, that view initially seems very naive because the machines of the time did one particular thing, like a sewing machine is just good for clothes. So the thesis was really rather implausible until the development of the idea of the Turing machine, Alan Turing's idea of an ideal digital computer. Because what's special about that is that now you have a multipurpose problem solver, a device that can be reprogrammed to do just about anything, that can be reduced to a set of instructions.
And of course, that then makes you think, well, could it be that A human being is essentially an organic realization of a Turing machine. Right. The Turing machine seems versatile enough. Maybe that's really all there is to us.
And in the background, the big target though is just any version of materialism that wants to say that we're nothing but an arrangement of physical parts.
[00:06:30] Speaker A: Great. Bob, anything you want to add at this point?
[00:06:33] Speaker C: I totally agree with Angus in his paper. It's kind of interesting because in engineering and computer science we use similar terms but different terms for the same thing. I think, for example, if, you know, with the idea of insight, which is one of Angus's five different things, I think it's kind of asking whether the computer can understand a number of other things. Like part of his. He goes through five different criteria and I'm sure you're going to get to that, Pat, but five different criteria of why machines are different than human beings. And I love it all because I learned a lot of new vocabulary, at least that I can put in my thesaurus for synonyms.
Really, really good stuff, Angus. Thank you.
[00:07:18] Speaker A: Good. Well, yeah, I think maybe we should move in that direction. Obviously, you know, the big and interesting development that many people think lends support to some sort of physicalist understanding of the mind is artificial intelligence, the LLMs.
So yeah, Angus, do you want to say anything about that in particular?
And then I want you to outline, at least in broad scope and get into more detail, the five powers that you think are beyond any physicalist understanding of mind.
[00:07:46] Speaker B: Yeah, I mean, the development that is huge at the moment is the transition from weak AI. That said that these artificial intelligence systems would just do a good job of simulating intelligent human activities. They could produce the same kind of output to the this claim of strong AI that now wants to say that the machine actually has a mind or that it can be credited with understanding or reasoning or creativity.
And the artificial general intelligence. All that really is saying is instead of being domain specific, like theorem provers and game players and expert systems that are dedicated to one problem domain, we now with large language models have systems that can find rules that they were not initially programmed with, such as ChatGPT, finding rules of grammar or alpha fold, to solving this problem of how you go from a one dimensional genetic code to the actual protein. Okay, so they're amazingly more powerful and they seem in some sense to be able to come up with rules that were not explicitly programmed with.
Right, and that's why you see a lot of the language about, you know, machine learning.
Now my problem as a philosopher is just that I'm seeing knowledge based systems. Right. And learning algorithms. And the words that are being used here, knowledge, learning, create, aren't used with the same meaning that they have been used for thousands of years by philosophers and other scientists. And so there's a serious worry here that we are going to fall for an equivocation where we're using the same word, but we mean very different things by it. And so we need to really laser in on what we mean by these terms, what these powers of the mind really are, and then to see if we can literally attribute those to the system or if in fact they're doing something else.
[00:10:07] Speaker A: Yeah, yeah, great. Okay. So that sets us up perfectly for where I want to go next. If you wouldn't mind. Just give us the five powers, Angus, first just in, you know, in broad overview, and then let's take a deeper look at each one.
[00:10:21] Speaker B: Yeah. So the five powers are insight, or as Bob likes to say, understanding. There's a difference between finding the truth and understanding why it is true.
The second one is matter insight, and that is we are aware of where instructions are going, what they're going to do. So you move up a level. Or if I do this, then what's that going to do? Right.
So it's a higher level form of insight.
Then there is free will.
Can we, with the same data, make different choices? Or we're forced to just do one thing given that data? Then there's our ability to see that some things are necessarily true, like the laws of logic or statements of mathematics, that it's not just that they have been true in the past, but they're necessarily true.
And then finally, perhaps the most difficult one, because it's become a bit of a weasel word, creativity.
And the problem is that people mean different things by that word, creativity. And so we need to be careful to distinguish ways in which someone might loosely say a machine is creative to the highest forms of creativity that we're aware of. And, you know, is there a good reason to think machines will ever be creative in that stronger sense?
[00:11:48] Speaker A: Right, yeah, you're great. Okay, so we have the powers now. And the claim is not. Is not just that machines do not have these powers, but they will not ever have them.
So. Right, great. Let's go to the first one, insight. Tell us about insight, Dr. Manuj. What do we mean by this? What is understanding?
[00:12:08] Speaker B: Yeah, so there's a difference between stumbling over the truth and recognizing the truth. If I'm in my basement with the lights off, I might stumble over gold bullion. Right. But it doesn't mean I recognize I found treasure. And likewise, ChatGPT or another large language model might discover the grammatical rule that a declarative sentence always requires a subject and a predicate.
But does it understand why that's true?
And if you think about it, the reason that it's true is otherwise you don't have a completed thought.
If I say most birds can fly, that's a declarative sentence and it's true or false in this case, true. If I just say most birds, you've got a subject but no predicate. If I say can fly, I've got a predicate, no subject, right? And so it's the same thing with other rules, like the rules of logic. And I give the example of the simplest rule of logic, modus ponens. If A, then B, A, therefore B.
It's one thing to find that rule, it's another thing to understand why that rule is true in this case, what logicians call truth preserving.
And I don't think there's any good reason to think that these AI systems actually have insight in that sense.
One of the reasons is when you discover how they build these rules, they do it on the basis of uninterpreted tokens. They don't need to know what the tokens mean, their sensor reference.
Okay? So if it finds the cats on the mat, it doesn't need to know anything about cats or mats. And likewise, if it finds this rule of logic, it doesn't really have to know what a conditional is or what its first part, the antecedent, is, or the consequent, or any of those things.
So that's one reason. But the other reason, going back to Plato, is that he pointed out that when you really know something, when you have insight into why it's true, you're not going to be budged by any amount of contrary claims. Once you see that A equals A, or that two plus two is four, no one is going to budge you from that.
But if you look at these systems because they are inductive learners, based on their training history, even if they discover a rule is true, all you have to do is retrain them on bad data and they will deviate from and abandon that same rule very, very quickly. And that seems to me to prove pretty conclusively that they never understood the rule in the first place.
[00:15:01] Speaker A: So the idea would be something like if once I understand why affirming the consequent is fallacious, right?
Doesn't matter how much you try to train me to the contrary, Right. I'm going to be deeply entrenched, right? Because I understand it. I've had that aha moment, right? That's the way we think about insight. We have that aha moment, right, because there's lots of instances where we might work with rules and math, I think is a good example. We're kind of working with something and we're getting at truth, but we don't really understand it. And then you sometimes have this flash of insight, this aha moment where you do understand it. What you're suggesting is that at that point it becomes deeply entrenched, that, you know, if you, I guess, entered some sort of propagandistic universe where they tried to train you otherwise, right. You're not going to budge. But you're saying this isn't the case with AI, Right, yeah. Sorry, Bob, go ahead.
[00:15:50] Speaker C: No, no, no. That flash of genius you're talking about is.
It used to be a requirement, by the way, for a US patent. Used to be US patent law required that you had a flash of genius, that you had this insight. And I, you know, I agree with Angus, of course, on computers. I think the old argument is the computers can add the number 5 and 7, but fundamentally they don't understand what the number 5 and the number 7 is, and they don't understand what add is, or simply following, following an algorithm.
And Angus in his paper goes through Searle's Chinese Room, which I think is the big smackdown in terms of computers understanding. Now, it's interesting, I think it was both Dennett and Ray Kurzweil that talked about Searle's Chinese Room. And they said, well, you know, if you have enough people in the Chinese room, you know, a thousand of them, maybe all of a sudden you would have consciousness or sentience or understanding or something like that. And they maintained that that's exactly that, the way the brain works. But this is under the assumption of emergentism, that if you have enough stuff, then boom, all of a sudden something magic is going to happen, you're going to have understanding. So, yeah, I think that the idea of insight or understanding is something that computers definitely do not have.
[00:17:07] Speaker A: So let me ask you this, Angus.
What do you think would be the case? Or what do you think we would see if machines did have insight, if
[00:17:19] Speaker B: machines did had insight, then when they found a truth, particularly if they found a necessary truth, then they're not going to abandon it.
I had someone at one of my talks recently ask, why is it none of these recent AI systems can follow Isaac Asimov's rules for robotics, which are about not harming human beings. And I said, well, it's very clear that they're not able to really understand any of these principles. And even if you program them in because they are able to shift their rules on the basis of their training, there is no guarantee that they can hold on to them.
So if you wanted to prove me wrong, you need to show that there's a system that can by itself, find a rule, establish to itself that it's true, and then it doesn't budge. It doesn't leave it behind. But so long as all I have to do is retrain the system and it loses the rule, then it is, at best living in the world of opinion, not knowledge.
[00:18:23] Speaker C: Yeah, you know, Asimov's rules for robots, I think, are kind of stupid. I think the second law is that a robot can allow no harm to come to humans. So this says if a policeman is chasing a criminal and the policeman takes out his taser gun, then that robot's going to take out the policeman.
You know, when you put together laws, you have to think of all the detailed consequences of those laws. And I don't think Asimov drilled down into that.
[00:18:47] Speaker B: Yeah, that's totally fair. But we also don't want to have systems like some of the recent AI systems, where in solving a problem, they've concluded that the best path is to eliminate a lot of human beings. So I just, you know, give that. I give that as an example. I don't think Asimov's rules are well worked out either.
[00:19:09] Speaker C: Yeah, well, I think those. Those AIs are trained on Mein Kampf. Right, okay.
[00:19:16] Speaker A: All right, so let's, let's move now to the next level. The next level up, literally, meta insight. So, yeah, reintroduce this notion to us, Angus, and then explain why you think it is something that we do not see happening with machines.
[00:19:31] Speaker B: So the idea of meta insight is that you essentially know the implications of an idea or where you're going in if you follow it. And so I thought that the simplest way to dramatize that is with the example of an infinite loop, because it seems like a fair test. We know that computers get stuck in infinite loops, and human beings can also follow instructions that constitute an infinite loop. So then the question is, do computers and human beings handle infinite loops in the same way?
And I argue that the computer doesn't give any evidence of being aware of where it's going once it gets into an infinite loop. And that's why it requires some kind of external agent to terminate the loop. Right. So maybe it's a user interrupt, maybe it's the operating system, maybe it's one of these fancy agents they talk about in AI today. But it has to be external.
The evidence that we know that we're in an infinite loop, at least in some cases we don't always. Right. But at least in some cases we can tell we're in an infinite loop, is that then we simply internally abort that loop. It's an internal power that we have.
And I don't think any computer has the internal power to break out of an infinite loop based on its understanding that it is an infinite loop. It's just following those instructions, and it might be forced out of those instructions, but not on the basis of seeing that it's in an infinite loop and not on the basis of an internal power to break out of the loop.
[00:21:20] Speaker C: I think that fundamentally the idea of meta insight goes back to Godel when he proved his incompleteness theorem, where he actually had the mathematics refer to itself, and he came up with a contradiction that couldn't be done. Turing took this, and with the halting problem, he proved the halting problem with a meta statement which was applying the halting problem, the hypothesized halting problem program to itself. And he came up with a contradiction that was a meta analysis of the halting problem. And he used that in order to establish that the halting problem was simply not possible. Just genius derivation in mathematics.
[00:22:00] Speaker B: Yeah, that's that. Bob's absolutely right. The standard proof of the unsolvability of the halting problem, that is, can you have an algorithm that will always tell you whether a machine will halt or not?
Depends on the idea that no machine is able to solve the self halting problem. And that's because if you say, here, oh, here's my halting module, what Turing does is if your halting module says halt, he'll put it into an infinite loop. And if he says it doesn't halt, then it will halt. And because that's possible, there cannot be a solution to the self halting problem.
And yet human beings halt themselves all the time. It doesn't have to be an infinite loop, it can just be a finite loop. That's getting to be tedious.
Right?
We just have this internal power to say, this is pointless. I don't want to do this anymore.
[00:22:54] Speaker A: So the idea isn't that the halting problem is essential to the meta insight argument because, I mean, because we can't solve it either. Right? It's more, it's the idea that from the first person fact that the machine clearly cannot become aware of and evaluate its own activity that's the issue.
[00:23:12] Speaker B: Yeah, I don't want to put all the emphasis on the halting problem simply because there are plenty of cases where a human being doesn't know for sure if he or she isn't infinite in an infinite loop because it just could be too complicated.
The critical point is the way we handle loops is different. We can certainly understand that we're in an infinite loop in at least some cases, and we can certainly break out without external intervention.
And John Lucas's great point was to note that with a machine, if you add any code, then formally it's a different system and therefore a different machine.
So adding code to get the first machine out of trouble doesn't prove that the first machine could get itself out of trouble. It just gets you a different machine. And what's worse is because of Turing's result, then that new machine has got the same problem. It won't be able to get itself out of trouble.
[00:24:08] Speaker A: Great, great.
I like this, this question about proving Angus wrong, or what would prove you wrong. I think it's good to stick our neck out there. So, same thing that we considered with respect to insight.
What would you have to see, Dr. Minouj, to give this point or argument up?
[00:24:30] Speaker B: Yeah, it would have to. The machine would have to autonomously provide evidence that it had recognized it was in the loop and by itself break out of the loop without the addition of any more code. Because as I say formally, if you add code, you technically now have a different machine. So that doesn't prove the original machine could do this. So you need a machine that can do it by itself.
[00:24:57] Speaker A: Great, great. Okay. The next one, if I remember rightly, is free will.
A very difficult and controversial and controverted subject all its own. So take us down this avenue now, Angus, if you don't mind.
[00:25:10] Speaker B: Yeah, so what I think is that we've got very good evidence that none of our machines, not even our most sophisticated AI systems, have free will for a very simple reason.
That is, if you hold fixed all the instructions they have and the data that they have, there's really just one thing that they're going to do next. At least a standard machine is deterministic. And if you bring in random causes like the decay of a nucleus, well, then you know, that will determine the chances of what happens next. But in both cases, the key point is the next state of the machine is purely passive.
It can't do otherwise than what is dictated by the priority state of the machine. But we're human beings. I think we've got loads of evidence that we're not like that.
And this goes back to a brilliant insight of William James, who was a philosopher and a psychologist who studied the phenomenon of selective attention.
And he pointed out that human beings have this amazing ability to process the same data differently.
And the classic example is the cocktail party. You go to a cocktail party, everybody's talking. So as a result, you can't make out any voice because everybody. Every signal suppresses every other signal.
However, if you want, you can choose to tune in to just one of the voices. And we've all experienced doing this. And so now I'm processing the exact same data, the same data that's entering my senses, but I'm doing it differently.
And there are a whole bunch of what are called cognitive therapists in neuropsychiatry that work in the same way. When someone has a phobia like arachnophobia, or if they're suffering from a certain kind of depression, or if they have obsessive compulsive disorder, they find themselves doing something over and over again. All right? In the case of obsessive compulsive disorder, it's interesting because the individual knows it's futile. They've just locked the door, they feel the urge to check it again.
And it causes this phenomenon in the brain called brain lock, where there is a neurological cycle that causes this repeated sense of unease.
But what they found with cognitive therapy is you can, using selective attention, consciousness, you can call the patient's attention to this fact, and you can encourage them to use their willpower to do something else, such as reading or gardening or something like that. And this has been used for many, many neurological problems. And it ties into neuroplasticity, the fact that the brain can be remapped.
And one of the ways it can be remapped. I mean, obviously it can be remapped through conditioning or drugs. But one of the ways it can be remapped is through the conscious effort of the patient. This is also used in stroke rehabilitation.
And you have to make yourself try to do something that's very difficult. If you do it for long enough, it will rewire your brain and you're able to do a task which previously you couldn't do before.
So that, to me, is good evidence that we have active power.
By active power, I mean that we can do something which we're not simply made to do by prior Causes.
If we have active power, then we can't just be a big set of switches or a set of what in metaphysics is called passive liabilities.
Like salt has a passive liability to dissolve in water. Right.
So it's a purely passive ability to respond given certain input. But we've got evidence that we've got active power.
That active power seems to belong to a substance, to a thing in its own right. And what's more, since it's guided by reason, it suggests that it is an immaterial power. It's not a physical power, it's an immaterial power because reasoning that we want to get well, we use our ability of conscious attention to focus on a more productive way of acting. And over time, we can sometimes break free of a neurological problem that way.
[00:29:54] Speaker C: Yeah, let me elaborate on that a little bit. And totally Angus is spot on. But neuroscience, Michael Egnor, talks about a process in the brain where there is a signal in the brain that tells you what to do before you know you want to do it. And that certainly seems contrary to free will. And this was studied by.
Yes, okay.
Anyway, there's a signal and you are told to do it. I used to smoke cigarettes and I was told, bob, you want a cigarette? So I had a cigarette. Then there was a signal that says, bob, you want a cigarette, have a cigarette. But then my wife says, no kids, as long as you smoke. So I had this signal that I want a cigarette, and I had to say, no, I don't want to. And Liebet called it free won't. Not free will, but free won't. And as Angus says, the beauty of that, if you exercise free won't long enough, you get this neuroplasticity going on in the body brain, and it totally rewires. So today I have no desire for cigarettes at all.
It's just a wonderful way that God created us. So I think that that's parallel to what Angus was saying. And the interesting thing, according to Egnor, they have discovered no signal in the brain that correlates with free won't.
So if you negate an impulse to do something, an exercise free won't, they have not discovered yet a signal in the brain that corresponds to that exercising free won't, which I think is just fascinating.
[00:31:28] Speaker A: One thing that I liked about everything you said there, Angus, was.
Well, you brought up the old randomness objection to free will, right? I mean, this is a classic objection against free will, that either something is determined or random, but in either case, neither is free.
And the traditional, I think, convincing response to that is it ignores the unique class of action that is reasons based action, which you really articulated quite well that we just had this power to consider the space of reasons. And I agree, I think that this power indicates an immaterial aspect about us and to deliberate. Right. We can, we can concentrate, we can focus on reasons for one action, reasons for another, to. To smoke, to not to smoke. Right. Whatever the pros and cons are, and then the will traditionally is understood, is, is really just, as you put it, the active power to end deliberation, to close down on some motive for action that is by itself incapable of completely determining us. Right. To give it that final I choose you or I don't choose you, the free won't. It's an interesting aspect of it too. And so it gives us this sort of unique category of action, this sort of, we might want to call it like a constrained indeterminism. Right. Or reasons based action.
It's not random because it's not for no rhyme or reason. We can cite reasons, yet it's not determined either. And it does seem like something that is not done to us, but we become ourselves agents of action. We can direct the course of history. It's not just the course of history that directs us.
And if I understand what you're saying, it's very clear that machines do not have that power. They do not have that category of action. Is that right?
[00:33:15] Speaker C: Well, there is a degree of probability in large language models. Of course, large language models are word completion upon steroids, but at the end of the process you get a probabilistic ranking of the words that should go next. So there is a probabilistic thing, but I think the point is still applicable because instead of just having a single response, you have a small arena of responses. And so definitely you're limited.
[00:33:40] Speaker A: Yeah, I think, you know, Angus anticipated that, is that we could, we could have chanciness and probabilities, but that itself is not adequate for a robust notion of libertarian freedom. Right.
[00:33:52] Speaker B: Yeah, because it would still fix the chances is the phrase that people use of what you do next. But when you act for a reason, you can alter what those chances predict. Just as much as you're not forced to do something deterministically, you can also alter the probability of something.
And I think that C.S. lewis put this very well years ago in his book Miracles in Chapter three, where he says that for reasoning to explain our actions, it has to break free of what he called the causal nexus. It doesn't matter if the causal nexus is strictly deterministic, as in Newtonian physics, or if it's probabilistic, as in. As in quantum physics. But you have to be able to look up to and be guided by reasons themselves.
And in the end, it is your desire to comply with reason that explains your action, not just a passive result of either chance or necessity, you know, occurring beforehand.
[00:35:01] Speaker A: You mentioned something else as well. I think it's important to spend a moment on it that, that we have this power because we are a certain sort of thing. We're a substance, right? We're. We're a rational substance. And our power, one of the powers that you talk about, freedom of will, kind of flows from the types of thing we are, types of thing that we are.
But machines, these at least large language models, are very plausibly not substance. They're an aggregate. So explain these notions to us, I guess, if you don't mind.
[00:35:32] Speaker B: Yeah. So this goes back really to Aristotle. He makes a distinction between a substance and an artifact. If you have a substance, then you have something that develops according to an internal principle. And it remains the same kind of thing over time. So going from a chicken embryo to a chicken, all right, for example.
And that's what's typical of all organisms. And it's at a higher level with human beings. You can also talk about rational beings persisting over time as the very same substance, despite the fact that they change the particular things that they are thinking about or doing.
But in case of artifact, it's not internally directed in its development.
It is externally composed, built by people from the outside.
And there is no internal principle that is the same anytime because it's parallel distributed system.
All that persists over time, all you see as flux. So in the case of these transformers, there's just these endless adjustments of weights that belong to these artificial neurons.
You can't see anything that persists as the same thing over time.
But at least most of us believe that there is such a thing as personal identity, right? That you can persist over time as the very same person that you were a few years ago and will be the same person tomorrow. Why is that? It must be that you belong to a different metaphysical category. Because if you were just an aggregate of shifting parts, then you would keep changing every time those parts change.
And of course, our bodies do change, and our recurrent thoughts, the particular things we're thinking about, keep changing.
But we don't think that that's a reason to think that we as persons change. Why not? I think because we are a substance. And once you have a substance, you have a being that can have its own powers. Whereas I think that the fact of the matter is, is that all computers are at the end of the day, just an enormous big bag of switches or passive liabilities that can be configured different ways by external causes, but they don't have any basis for active power. There's no self there that persists as the same thing over time as there is with a human being.
[00:38:04] Speaker A: So is the idea there that when it comes to an aggregate, there is a reduction that can be had right, to its. Its component parts, whereas with a substance there is not, at least for the type of substance that we are, there's not a reduction that can be had to at least the component physical parts. Is that right?
[00:38:23] Speaker B: Yeah, that's right. Metaphysicians will tell you that an aggregate is simply defined by the particular part. So as soon as a part changes, you have a different aggregate. In the case of a substance, this is a bit difficult to understand, but the whole precedes the parts. In other words, the kind of being you are precedes the parts that you have. And this is why you can change all kinds of parts in your body every so many years. You've got completely different matter in your body and so on, and yet all of those parts are continuously conjoined with the same substance. And you don't just become a different human being or a different person because of those changes.
[00:39:11] Speaker A: Yeah, great. So the idea here is that my cells have the nature and activity they do because they belong to me. Yeah, right. Yeah. Good. And that seems. Right, that seems obvious. I know not a lot of people accept it, but it seems obvious to me. So I'm satisfied with that. All right, let's move to the next one.
Access to necessary truths or necessary conceptual relations. So what are we up to in this section of the paper, Dr. Minouche?
[00:39:38] Speaker B: I think this is the biggest challenge to AI. The. The most impressive forms of AI are the ones that they claim learn, you know, new rules and new truths, and let's not quibble about the word learn. The problem is, even if they did learn those, they do it on the basis of past data and past interactions with the world. Let's suppose you supplement the AI system with all kinds of sensors so it can get real world input, like with some of our robots. You can do this.
The trouble is, no matter how sophisticated it is, it's trapped in a contingent. It can learn what has happened to be true or what has happened to Work and you can never logically get from has been the case to will be the case or must be the case. Let's suppose, for example, I toss a coin a billion times and to my astonishment, it's heads a billion times.
It doesn't prove that it'll be heads the next time or that it must be heads.
Well, the problem here is simply that human beings can, by an act of insight, in the Latin this is often called intellectus. We can directly apprehend that certain truths are necessarily true.
So Lewis loved the example. You know, A equals B, B equals C, therefore A equals C.
He says, I don't have to base this on the fact that I've never caught these things doing otherwise.
I see already that it must be true.
So I'm grasping something that must be true in the future and in every possible world.
But so long as physical systems, I call them adaptive physical systems are confined to their past interactions, all they can ever show is what has been true in their past.
They can't show it must be true in the future, and they certainly can't show that it's true in other possible worlds than the world that they've interacted with.
[00:41:55] Speaker C: So, Angus, I have a question for you. I read this and it reminded me of axioms like Euclid's axioms, where you have a statement of truths that are almost so. Well, they're transparently obvious. And many mathematical systems have been generated on these sort of ideas of axioms.
Is there a relationship with what you're talking about now here to axioms, things that are obviously true?
[00:42:22] Speaker B: Yeah, there is, although we have to be careful because we know that in the history of thought about this, even some things that were thought to be self evident, like Euclid's eczema parallels, turned out later to not be necessarily true.
But gosh, I can't think of a circumstance where you're going to doubt A equals A or some more very basic axioms like that. And so this goes back to Aristotle, who famously pointed out, you can't ever succeed in arguing for anything unless some of your premises are immediately given.
Some things better be self evident, because if they're not, you have an infinite regress of arguments and then you never really establish anything.
So the notice, by the way, that the AI researchers have assumed that themselves because they're trying to show something.
If you're trying to show something, you better have some starting points which are rock solid, because if not, you can't show anything.
So I think it's very important to See that everybody requires as there to be some truths that are foundational. Even though we can find particular cases where we turned out to be mistaken, there better be some fundamental absolute truths that we can see unnecessarily. True.
[00:43:49] Speaker C: Okay, but wouldn't you agree that Euclid, that his axioms were true if he limited himself to the universe of planar geometry?
[00:43:58] Speaker B: Yeah, that's, that's fair. That's like the point that people will make about Newton is as long as you clarify the assumptions, you don't have to say that Newtonian physics is false. It's just. It's true within a particular domain. And I think it's fair to say that Euclidean geometry is completely valid for a certain kind of space.
He just didn't think that there might be other kinds of spaces.
[00:44:21] Speaker A: So my question here for clarification is, is the idea here that a physical system is confined to induction?
Because clearly that, you know, AI, it can. It can implement deductive procedures, or is it the stronger claim that it cannot perform deduction, it just mimics it? Because to my mind when I'm thinking about this is like, no, to perform deduction, you have to understand deduction.
Otherwise you're just. You get an approximation, you get a simulation. And to most people, that simulation might seem like the real McCoy, but you're not really performing deduction. Yeah. Help clarify some of this for us. What your position is on this. Right?
[00:45:01] Speaker B: Yeah, yeah. So my position on this is, of course, the machine can act in accordance with reason. If it's got a properly engineered arithmetic and logic unit, it is going to obey in its computations basic arithmetic and logic. That doesn't show that it understands arithmetic and logic or grasp the fact that these are necessary truths. And the reason.
And here I agree with Selma Brinschort. The reason I don't think it grasps these necessary truths is that they are about abstract objects, numbers, logical relations. Logical rules are not found in space and time.
Any physical system is confined in what it can ever learn to its interactions with other physical systems. It cannot interact with abstract objects.
Now, what's interesting about the human mind, you might say, well, we don't interact with abstract objects either. That's the view that many philosophers hold. But we don't need to somehow. We simply perceive them.
All right? And you can perceive. Perceive that there are certain relationships.
This is what allows you to follow the soundless proofs of logic. You can see, yeah, these rules are sound and these other rules are unsound, right? And once you follow that proof you know that this is always the case, even though you've only done so many proofs in your life, a finite number, you know that those correct rules are valid for an infinite number of proofs, including proofs that no one will ever do.
[00:46:45] Speaker A: Yeah, good. No, that's. That's. That's very helpful. Excellent. Fascinating. Okay, the next one I'm really interested to hear more about. And this is the idea not just of creativity, but transformational creativity.
Now, before we began recording, we were. We were talking about music. Bob and I were talking about music. We had to halt that conversation before we got carried away. And I can think of many instances in the history of music where it seems like we've had transformational creativity and of course, so many other areas as well. So maybe we could bring out some of those examples at some point. But talk to us first, what you mean about transformational creativity and why you think machines are incapable of it.
[00:47:27] Speaker B: Yeah. So the basic contrast is between combinatorial creativity. AI is doing a good job of that. If you mean it can produce a poem or a piece of art, which in a sense is new, but it's just recombining works that were already created by human beings.
Transformational creativity is utterly different because you have to question the assumptions of the original conceptual space to get there.
So as long as people thought that space was planar, as Newton did, there's no way that light could bend.
When Einstein suggested that space time is curved because gravity is a distortion in space time, it predicted that light could bend. And then it was confirmed by Arthur Eddington. So he had to change the rules of the whole problem space and create a new conceptual space.
Well, Kurt Godel had the audacity to challenge the idea of that there's a difference. You know, you think there's an obvious difference between a statement of arithmetic, two plus two is four, and a statement about arithmetic, like arithmetic, is consistent.
He challenges that and shows that arithmetic can actually talk about itself. And as a result, you can come up with this sentence that says this sentence is unprovable, which then, of course, you know, you realize that that sentence has to be true and therefore it is unprovable, Right.
If it's a consistent system, it better not be able to prove it, and therefore it has to be true.
In both of those cases, you had to change the original problem space. That meant you had to thoroughly understand what were the underlying assumptions of the problem space and then dare to question them. Or when Einstein said, maybe time isn't constant, but can slow down this is just amazing stuff. Right?
Because you're doing something that you have never experienced. The evidence for that came much later when they took synchronized. They used synchronized clocks right and left one on Earth and put one in orbit and they come back and put them together and they tell different time. But the point is that Einstein came up with that idea before there was any evidence whatsoever.
And the same thing with Godel trying to show that David Hilbert was wrong. You can't reduce mathematical truth to what can be axiomatized or what can be computed.
He came up with that idea before there was any proof and then he had to prove it.
That was the great insight. Or Turing comes up with the idea of the modern digital computer before there was a single one of them. In the sense that he defined lots of calculators. Yes, but not something where you could make it do basically any task that can be reduced to a set of steps by changing the program.
So I think that's the kind of creativity which so far at least is certainly unique to human beings. I see AI doing remarkable work, finding patterns in data that it's given.
And sometimes it can just by brute force explore a great big space.
There's a good example in England there's an AI system called Homes2, which is an AI assisted data analytics system system that has helped investigators solve major crimes involving serial killers.
But it's done it by noticing that there are patterns that overlap on particular individuals. And until you suddenly realize, oh, that guy is always very close to where these crimes occur.
Right. And so when you boil it down, it's still just doing a very fancy form of extrapolation. Very helpful.
But it is not changing the problem space.
Or think about people like Tolkien coming up with whole different worlds.
See, now you can conceive of a world that operates according to totally different assumptions. Maybe, maybe magic is possible. Maybe you can do all kinds of things in that world that you can't do here.
[00:52:08] Speaker C: I like your term non combatorial creativity. And I mentioned, I think in the beginning that I was learning a thesaurus of different words for different things. I have heard this referred to as horizontal innovation.
In other words, it was putting together concepts which were already known. Let me give you an example of a recent one that was touted by OpenAI. They said that OpenAI had solved an open problem in, in mathematics called Erdos conjecture number 90. Now Paul Erdos was kind of a weird guy. He had over a thousand conjectures that are listed. And what they did, it wasn't creativity. What had happened is that the people that had attempted to solve erdos problem number 90 were concentrating on a single area of mathematics. What OpenAI did, ChatGPT, if you will, is is they looked at different areas of mathematics, they brought in an area of mathematics which hadn't been considered before, and boom, the problem was already solved. Now, Terence Tao, who is a Fields Medal winner, maybe one of the greatest mathematicians living today, is very invested in looking at AI and the solution of math problems. And it turns out a lot of these Erdos problems were actually published elsewhere and they had been considered open, but other people had published them and nobody had determined this. I think that currently there's 300 math papers published every day and there's no way mathematicians can keep up with it, but AI can.
So we see this horizontal innovation, and yet there's other problems in mathematics that have been around for a long time.
Goldbach's conjecture, the Riemann hypothesis, the Collatz conjecture, these have been around for a long time, which have been with well studied by mathematicians. And when AI can solve one of these problems, that will be not horizontal innovation, but vertical integration, which is creativity.
[00:54:09] Speaker B: Now that's very helpful. I mean, I think that's a very fair way of putting it. AI is impressive because it can now bring together multiple domains. It's not tied to a particular genre or a cognitive domain, but it can bring together multiple ones. And so sometimes by laying two things side by side, which no one had ever done before, you say, oh gosh, that's really important. I don't think the AI system is recognizing the truth, but it can help us recognize that it's the truth.
But nonetheless, what that is is just exploring lots and lots of combinations, more than human beings would. And that's, that's where AI is great. And I think it will help us make some great, great discoveries.
But it, I think it's fundamentally different than being able to, as it were, make that leap where you see your way beyond something. And sometimes creativity is just plain mysterious.
You know, when you think, for example of, in Bob's book on the discovery of the structure of the benzene molecule, right, you talk about that.
And I mean, that was based on a dream, on a vision, right? The vision of a snake that's basically eating its own tail. And then, well, maybe the structure of benzene is a ring molecule.
[00:55:40] Speaker C: Yes. And again we get back to this idea of flash of genius, which is used by Roger Penrose in his book the Emperor's New Mind. So he talks about a flash of genius. You know, there's another case. There was recently headlines that AI had beat the got 100% score on a Math Olympiad. And my response is, of course it did, for the same reason it's going to get 100% on SAT test and GRE test and PSAT test for law, because it has seen the corpus of all of the material basically that has been published in the world and has access to it and we know that it has the answers. Why do we know that it has an answer? The answers, because somebody made up those tests and they know what the answers are and they were in this corpus of material. So I'm not going to be surprised when AI does all of this incredible stuff and passes all of these tests.
That's horizontal, if you will, innovation.
[00:56:33] Speaker B: Yeah. You need an information audit. And there's great work there's Bob has done and Bill Dembsky has done where you. You need to follow the data and say, where did it actually get this result from? It can seem very impressive if you don't know where it got the answer from, and very unsurprising if you actually trace back where the data came from.
And it seems to me that, no, we're not primary creators like God, but Tolkien is right that we are sub creators.
And that's a different kind of creation because it depends on your grasping universals. I don't think philosophically that machines ever grasp universals because they are purely physical systems. All they can do is interact with lots and lots of particular examples, but they don't have that ability to abstract to universal.
Once you have a universal, you can, as Tolkien recognized, imagine it being instantiated somewhere where it's never been instantiated before.
So I abstract green from grass and now I think of green blood. Or, you know, I abstract fun from a party and now I imagine having fun at the dmv. Right. Even though I've never experienced it.
[00:57:57] Speaker A: Nobody has.
[00:57:58] Speaker C: That's an ox. That's an oxymoron.
[00:58:02] Speaker A: It's like government intelligence.
So what's the, what's the, what's the link here then? I think you're hinting at it not just, you know, not just the inadequacy of reductive physicalism and computational theories of mine, but for the immateriality of the soul? Is it for you, Is it the ability to play with, if you will, universals, to entertain these conceptual rather than just perceptual ideas?
[00:58:33] Speaker B: Yeah. I think the best evidence philosophically that we are immaterial beings is that we can reach beyond the spatio temporal world. If we're purely physical, things are always spatio temporal and they are limited in what they can do to physical properties and particular physical causes and effects.
But it seems very clear that we can access a different realm. We can access a realm of things which are necessarily true. And as I said before, there's no way that any physical system can interact with the abstract truths of logic and math. And yet we've proved all kinds of things are necessarily true. Now, of course it is the case that the large language model or some other AI system can sometimes find these truths in the sense that it finds the pattern. And we say aha, yes, that's true. But it doesn't mean that it could ever prove that it is necessarily true because it simply found that because it was common in the particular items that interacted in the past.
So it may be discovering something that is a necessary truth, but it doesn't understand that it is a necessary truth. That's the difference.
So I think the fact that we can see that there are all kinds of things which are not contradicted, contingent, when everything that is physical appears to be contingent, is really the argument. Even the entire universe, right, doesn't have to be here, it is here.
So it's contingent. And yet we were able to think about things which must be true and therefore go beyond the contingency of space and time.
[01:00:28] Speaker A: Yeah, great.
Dr. Marks, any final questions for Frang is here.
[01:00:33] Speaker C: No, we've covered some, some great things, of course. Creativity I like to go back to Summer brings your Lovelace test that the AI is generating something which is beyond the intent or explanation of the programmer. Unfortunately, the programs that we generate today are maybe too complex to do that. So I think that setting up some goals, like I mentioned, proving some of these, these outstanding will well explored open problems in mathematics. If AI can solve those without the interaction of a human being.
It's been said, and it's true, Terence Tao, for example, mentioned it, that AI is a great research assistant. It can really help you go through a lot of papers and find good stuff. And I think if it proves something and a human being is involved, that credit should probably go to the human being which has guided it through through the solution. But if it can solve one of these problems without prompt, without human intervention, then I will reassess my evaluation of AI, but it isn't going to happen, so I don't have to worry about it.
[01:01:39] Speaker B: The other huge thing I'm worried about in my paper in Turkey and another one that I just finished is on this is the issue of morality.
I would be astonished and impressed if AI could actually show that it understood moral rules and that it understood the value and dignity of human beings.
I don't think that it does. I think that it could be programmed with certain rules that would express human value to us. And I think it could be trained on scenario like they're doing with the autonomous vehicles and the, of those trolley, trolley problems. Right. That are designed in a sense to get it to make hopefully moral decisions.
But I don't think that it can actually access that realm too. And the reason is it's the same issue.
These moral relations are not contingent physical relations.
I have a, a retired colleague at our other campus who completed a book recently about the nature of moral necessity and oughts.
Why is it that we think that there is some moral necessity that you shouldn't torture babies for fun or whatever is your favorite example of something that's just off limits?
How is that?
I think that you have to be able to access a different sphere and notice that we never see good and evil, right or wrong, with our senses, nor do we physically interact with them in any way.
And yet that's another area where we can grasp very important truths. And what bothers me is that we are increasingly handing over decision making to AI systems. Neil Postman back in 1994, three in his book Technopoly said we're, we're already in serious danger of what he called agentic shift. An agentic shift is where you hand over the prerogatives of rational agents to a being that can't handle that. I, I don't think our AI systems are moral beings. I don't think they have a conscience. I don't think they understand the moral law.
It's wrong for us to ask these systems to make really important moral decisions for us. We're just trying to pass the buck and we're trying to engineer a situation where no one is to blame. But ultimately we are to blame. If we hand over these weighty decisions to a being that is not equipped to make them.
[01:04:27] Speaker A: Yeah, that's an important warning because even if all the arguments in your paper are sound, and I think they are, and that we don't have to worry about this self aware terminator situation, it doesn't mean that AI doesn't come with real dangers.
Right. And that there are still some very concerning things that we should be aware of and some areas we should be quite hesitant to go into. Is that right?
[01:04:58] Speaker B: Yeah. And the fact is we've got a Paradox here. The early AI systems were terribly rigid in their following of rules. And then the complaint was, well, that doesn't show the ability to move from one domain to another. So we need to have systems that can go from one domain to another. Right. And learn new rules. Well, now they're doing that.
The problem is that comes with a tremendous cost that there's absolutely nothing that you can pin down for certain.
And so given their huge complexity, many AI researchers are saying that in practical scenarios, these systems are essentially black boxes. They don't meet this criterion of explainability that we can explain how they made a decision that they did.
And this is rather frightening. It was frightening because, for example, two years ago there were a couple of sets of parents who testified before Congress that AI systems had aided and embedded in their teenage son's suicide.
And there are other cases like this.
So the worry is that we might, because these systems act like human beings in various ways and they engage in conversations, sensation, which is human, like, in various ways, that we start to attribute agency there and compassion and other qualities where there's just no justification for that. And of course, they might give good advice sometimes. And so then we're lulled into this false sense of security. But if they don't really understand moral concepts and they don't really care and they don't really know what a human being is or why a human being is valuable, sooner or later they may betray us.
That's my biggest concern.
So I think that at a practical level, we at least need to be doing a lot more work on.
I'm a big advocate of the human in the loop standard hitl, that if it's something really important that lives depend on. Yes, you know, use AI to do what it's good at. Look, looking at lots of data, but you don't ever allow it to make the final decision if it's a morally weighty decision.
And yeah, I know that this is going to happen because people are so infatuated with its power that they then sort of say, well, you know, the AI did this. And notice as well that it then gives us an out. If doctor uses AI in medicine for, for, you know, brain exploration and surgery, and it goes wrong. Well, you blame the company that designed the AI system. It's like, hang on, did the doctors spend time to find out if the results of this system were explainable?
Because if he didn't, then it's not clear that he avoids all responsibility for what happens next, even if he didn't know.
[01:08:01] Speaker A: No, very important. And Good warnings there, Dr. Minouj. I appreciate that. On a more positive note, I want to congratulate you on the publication and encourage readers, listeners to go and read it. It is over at the Journal Religions. I believe it's open access, so everybody can get that, get that article and download it. It's very well written, quite accessible. So if you've enjoyed this podcast, please do head over there and get the article.
Gentlemen, final thoughts as we wrap up here.
[01:08:31] Speaker B: I do have one more thing to say, and that is what should we learn from these AI systems about education?
I tell my students that it's very unwise for them to take classes which are training them in sets of procedures because they are precisely the kinds of things that can be automated. Good education should be focusing more and more on creativity, innovation, wise moral stewardship, things that are uniquely human.
That way we will have a reason to still be doing these things, which are things that we should be doing. I mean, theologically, God called human beings to do these things. He didn't call machines to do them.
But it also means that it's less likely that these poor students will end up finding their jobs have gone away because they automated.
[01:09:26] Speaker C: I think the important takeaway is that we have to remember that AI simply simulates human beings and it doesn't duplicate human beings.
There was an example recently of the Japanese who came up with artificial skin. And this artificial skin could actually detect and feel different sensations.
It can detect pain.
And imagine somebody getting waxed on the chest, that really hurts. If you get waxed on the chest and you're a human being, you feel pain. The AI only detects pain. And there's a difference between the feeling and the detection. And I think that the human condition is one that needs to be taken into account in all interpretations of artificial intelligence. And the idea is simply simulating but not duplicating the human being.
[01:10:19] Speaker B: That's really good. And I just would would add this. Our culture is already way too pragmatic, utilitarian.
It loves AI because of what it can do. And that's impressive. But much more important is what a thing is. It is the intrinsic nature and powers of a human being that are so important. That's why we're irreplaceable. But our machines are replaceable because they're just defined by what they do. No matter how expensive they are, you can always get another one. As long as it does the same thing, it's equivalent to the original.
Each human being is utterly unique and irreplaceable.
And it really bothers me when I see business leaders who think that the only thing that their workers have value for particular things that they do to contribute to the bottom line.
So they're looking at that output. Well then of course it's easy to think I can get a machine that can do the same or better.
Yeah. What about what it means to be a human? So if you've got this system that's got artificial emotional intelligence, is it really good for me to relate to a non human being? Of course, some people say they like it more, but even if they do, it doesn't mean it's good for them. Another real human being is going to have that otherness and that pushback. You're going to have to learn to negotiate and compromise and you will develop as a human being in the company of other human beings. If you have systems that are engineered to be compliant, well, then you, you end up actually becoming a pretty horrible, selfish person.
And there's been a whole bunch of studies that have shown this, that to the extent that young men have got these sort of female chatbots, it makes them more abusive because there's no consequence if you abuse this system. Whereas if it were a real woman, of course, there would be consequences rather quickly.
And so I think we need to hang on to what matters inside of us, the intrinsic nature of a human being and what's irreplaceable and not just focus on the. The outside.
[01:12:35] Speaker A: Yeah. Really, really good conversation, gentlemen. I appreciate the insights. We talked a lot about insight. A lot of insights in this conversation. The philosophical and also the practical and existential. Very valuable.
I want to thank everyone for tuning in. If you enjoyed this podcast, please, please do subscribe, share it around, leave a comment where it is relevant. We'll link the resources in the show notes and we'll talk to you all next time. Thank you, gentlemen.
[01:13:00] Speaker B: Thank you.
[01:13:10] Speaker A: This has been Mind Matters News.
[01:13:14] Speaker C: Explore more@mind matters.com that's MindMatters AI.
[01:13:22] Speaker A: Mind Matters News is directed and edited by Austin Egbert. The opinions expressed on this program are
[01:13:29] Speaker C: solely those of the speakers.
[01:13:31] Speaker A: Mind Matters News is produced and copyrighted by the Walter Bradley center for Natural and Artificial Intelligence at Discovery Institute.
[01:13:42] Speaker B: Sam.
