Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
#34 John MacCormick: Can Machines Think Like Us?
40 Plays3 days ago
When an AI tells a joke, writes an essay, or solves a complex programming problem, is it just performing a statistical magic trick—or is there something deeper happening under the hood?
On this episode of The AITEC Philosophy Podcast, Roberto and Sam sit down with computer scientist John McCormick, author of Thinking AI: How Artificial Intelligence Emulates Human Understanding (Princeton University Press).
For over 70 years, we've been asking Alan Turing's classic question: Can a machine think?. John reframes this debate by taking us past the polished, fluent surface of modern Large Language Models to look directly at the underlying code. We discuss whether AI merely mirrors human output or if its internal structures have begun to replicate pockets of human cognition. Despite the existential dread surrounding superintelligence, John shares a surprisingly optimistic perspective: even if we build entities that exceed human intellect, they will never replace the deeply rooted, biological fulfillment of our shared humanity.
Uncover the machinery behind the curtain and join the conversation at ethicscircle.org.
Recommended
Transcript
Debate on Machine Thinking: Philosophers vs. AI
00:00:16
Speaker
For more than 70 years, philosophers and computer scientists, among others, have been circling one deceptively simple question. Can a machine think?
00:00:28
Speaker
It sounds like the kind of question that should have a yes or no answer. Either the machine thinks or it doesn't. Either there's a mind in there or there isn't.
00:00:39
Speaker
But the closer you get to modern artificial intelligence, the harder that question becomes. Because today's ai systems can write essays, translate languages, explain jokes, produce code, summarize books, and carry on conversations that feel, at moments at least,
00:01:00
Speaker
uncannily human.
AI: Imitation or Genuine Thought?
00:01:02
Speaker
So maybe we start with the question of whether or not AI really thinks, but maybe the next question is, what exactly is it doing when it appears to think?
00:01:16
Speaker
And once we ask that, the next question becomes, is the resemblance only on the surface Is it just fluent output, polished imitation, a statistical magic trick?
00:01:32
Speaker
Or if we look under the hood, do we find something stranger?
John McCormick's 'Thinking AI': Book Discussion
00:01:39
Speaker
Little pockets of machinery that begin to resemble the way human minds work.
00:01:46
Speaker
Today on the A-Tech Philosophy Podcast, we're joined by computer scientist John McCormick. John McCormick is the author of Thinking AI, How Artificial Intelligence Emulates Human Understanding, recently published by Princeton University Press.
00:02:06
Speaker
And our question is simple enough to ask, but very hard to answer. Can AI think like us?
00:02:20
Speaker
Let's begin here by talking about maybe like ah an overview of of Thinking AI, your latest book. So tell us what what is your main goal in writing this book? Who is it for? And and tell us maybe about some inspiration there from Alan Turing.
00:02:38
Speaker
Sure. Well, it's a general audience book, so it should be fun to read and understandable by anyone. That is certainly my goal. And it addresses a fundamental question that is inspired by the great mathematician, and I guess now we would call him a computer scientist, even though computer science didn't really exist um for most of his lifetime.
00:03:01
Speaker
ah In 1950, Alan Turing published a paper in a philosophy journal, um the Philosophy Journal Mind, and the title of his paper was Computing Machinery and Intelligence.
00:03:15
Speaker
And at the very start of that paper, he states that he proposes to investigate the question, can a machine think? Now, he used the word machine. That was in early 1950s. Today, we would use the word computer to mean the same thing.
00:03:32
Speaker
He was talking about computing machines and abbreviated it to machine. So really what he was saying was, can a computer or can a computer program think? And that is a problem that has continued to trouble computer scientists, philosophers, cognitive scientists, psychologists, and many other people ever since the birth of digital computers in the
Understanding AI: Neural Networks and Learning
00:03:54
Speaker
1940s.
00:03:54
Speaker
So clearly there's been a lot of work done on this. And the really nice thing is that in the last 15 to 20 years, we've seen amazing new developments in artificial intelligence,
00:04:07
Speaker
that really help us to gain empirical evidence and understanding about this question of whether a computer program can think.
00:04:18
Speaker
So the book really has two sub-goals. The first sub-goal is to explain how these modern artificial intelligence programs work. And they're based on two pillars of technical understanding.
00:04:31
Speaker
One of them is called deep neural networks, and the other is called reinforcement learning. So the book first explains those ideas in a way that anyone can understand.
00:04:42
Speaker
And then equipped with that technical understanding, we are ready to address that classical philosophical question in artificial intelligence of whether a computer program can think and hopefully give the reader a chance to make an informed decision about how they feel about that question based on how modern artificial intelligence systems work.
Beyond Binary: A Nuanced View on AI Thinking
00:05:04
Speaker
Awesome. Yeah, I would really recommend to to those interested in you know yeah getting a grip on like certain key developments. Your book is really good at laying out clearly yeah what is ah deep what are deep neural nets and what is reinforcement learning. So yeah, I just wanted to echo that. And ah so you know turing to go back to the Turing point, I guess, he posed the question, like you said, can computers think? um Now, you kind of prefer a slightly different formulation within the book of in what ways can ah computer programs appear to be thinking?
00:05:42
Speaker
um So, yeah, i maybe could you kind of walk us through the differences between those two questions and you know why you prefer to ask you know in what ways can computer programs appear to be thinking?
00:05:54
Speaker
Sure. Well, there's one obvious difference, which is that second question about whether a computer program can appear to think like a human. That is a lot more nuanced than the question, can a computer program think?
00:06:07
Speaker
That immediately gets us mired down into a binary question, a yes or a no for something which is actually a very complex question and deserves to be approached in layers with with nuance.
00:06:20
Speaker
So we definitely want to remove the binary and think of in what ways and to what extent can a computer program think. And then again, take another step backwards and say, well, there's a lot of intuitive baggage associated with believing that a lump of silicon can think.
00:06:42
Speaker
We have living brains. A lump of silicon is a
Human vs. AI: Limitations and Resemblances
00:06:46
Speaker
dead piece of inorganic matter. And it's just difficult for us to bridge that gap intuitively.
00:06:55
Speaker
So I like to just remove that whole question and just say, don't worry about whether it can truly think, whatever that means. Let's just say, can it appear to think like a human?
00:07:06
Speaker
And if so, in what ways, to what extent? Can we find some resemblances between the way this computer program works and certain brain processes that we know are going on in our own minds? And that's why I like to address this more nuanced version of Turing's original question.
00:07:24
Speaker
Can I just ah kind of double click on that and make sure that that listeners understand here? Because this is a basically key to understanding the book and basically the rest of our conversation.
00:07:35
Speaker
ah let me Let me try to say it in in my own way and tell me how how accurate I am, John. ah So it's basically, it seems like two questions. Okay, so um can can it... um Can someone be convinced that they're speaking to something that is thinking? That's that's sort of one question, right? can Can they appear to be thinking? Sure.
00:07:56
Speaker
ah But there's also this more and maybe more nuanced question, which is when you look under the hood, do those processes that enable this appearance of thinking actually resemble human brain processes? Did I get that right?
00:08:11
Speaker
Yes, you absolutely did. And my response is we should be interested in all of the above. So we can take a purely what a psychologist would call behaviorist approach to this question and look only at the outputs. So if you are interacting with a computer program and all of the outputs are indistinguishable from a human, then ah the behaviorist approach might say, then yes, this thinks like a human, no question about it.
00:08:40
Speaker
On the other hand, many other people would say that behaviorist definition is not good enough. We do need to look under the hood and see if it's coming to its ah answers in the same way that a human would do it and that the process is also important.
00:08:55
Speaker
So in my book, we do both, mostly the latter. We primarily ah reject the behaviorist approach, but sometimes it's good to think at that level and especially because we are willing to take that step back, remember the gap between the lifeless silicon and the living brain, and just say, does this
Ethical Considerations: AI Consciousness Debate
00:09:16
Speaker
appear to think? Well, that might be from the even the appearance, even having taken that step back, the appearance could be from a behaviorist point of view,
00:09:23
Speaker
um Or it could be, when I do look under the hood, does that appear to be thinking? It's still happening in lifeless silicon, so I might not expect it to be, quote, truly thinking, whatever that means. But when I look under the hood and look what the program is doing, does that appear to be ah resembling some kind of human brain process?
00:09:43
Speaker
So yes, you are absolutely right to just make that distinction. And I think we need to address all of those levels when we are making our personal decision of how we react to programs that have outputs that might b appear to be creative, intuitive, understanding, and so on.
00:10:08
Speaker
At this point, the conversation could have naturally gone in kind of an easy direction. right Just yes or no questions are crude. Nuance is better. End of story. But Sam presses on something important.
00:10:22
Speaker
Sometimes we really do need a yes or no answer. If an AI system were conscious or if it had moral status, then our obligations to it might change.
00:10:34
Speaker
Turning it off, deleting it, copying it, experimenting on it, those might not be morally neutral acts. So even if John's gradual approach helps us avoid a simplistic answer, it doesn't make the sharper philosophical questions disappear.
00:10:54
Speaker
and So real quick on the whole binary question issue. So Yeah. So, so, you know, the binary question would be, you know, can computers think and they, and the binary answer would be, you know, either yes or no.
00:11:10
Speaker
Um, and you know, you might think that in so far, like, like if you connect thinking to consciousness, you might think that the question would be a yes or no. Um,
00:11:26
Speaker
I don't know, another one one might be like moral status, right? Like if something has a moral status, you might think it's like a yes or no question. Like does this thing have, you know, a right to life, for example, you know, yeah and that might seem to be a yes or no question. So it seems like sometimes, sometimes a sharp yes or no question is appropriate because there is a sharp like yes or no fact connected to it. Yes. And especially if there's going to be some kind ah of action predicated on that answer, as in,
00:11:55
Speaker
you know Is it considered cruel and unusual punishment to terminate this program by turning it off? And the questions like that. So ah actually, you, Sam and Roberto, are the experts in philosophy who have been studying that field for many decades. I have done my best to master the aspects of philosophy that are relevant to my book and my field, but I don't pretend to have mastered all of that literature and have good answers to these questions about moral status and consciousness.
00:12:31
Speaker
So the book doesn't address moral status at all. um that That question doesn't even come up. Consciousness, I mentioned a few times, and at one point, I basically say,
00:12:42
Speaker
okay, Rita, I know that you're almost certainly dying to know whether computer programs can be conscious. So I'm going to give you what you want. I'm going to tell you what I personally think and what I think we can infer from our understanding of these modern AI systems.
00:12:58
Speaker
But personally, i think that is a rather separate question and at least for the purposes of working with AI, not necessarily a useful one.
00:13:11
Speaker
ah For me, I don't really see a difference between a working with a computer program that appears to think and understand like a human, but is not conscious because someone has figured out how to define that and test that.
00:13:31
Speaker
um just I'm talking hypothetically now. We don't we don't have a good um testable definition right now. But ah hypothetically, in a future world, suppose I'm working with a very advanced AI, more advanced than we have now, and it is able to give the appearance of Thinking, understanding, being as creative and intuitive as any human in my interactions with it, hypothetically. And suppose hypothetically, someone's also figured out an experimental way to determine whether um that entity is conscious.
00:14:02
Speaker
um if ah If the answer is yes, I'm not sure what that tells me. If the answer is no, I'm not sure what that tells me. For the questions that I'm interested in, it's not so important. However, i don't want to just dismiss all of those other aspects of philosophy for which it would be important. It's just that I haven't anded addressed those and I'm not an expert on them.
00:14:22
Speaker
Right. Yeah. And and just, ah you know, maybe to think about a few reasons why the more gradual approach seems fitting, um you know, thinking, it doesn't seem to be like one single thing, you know, it seems like it includes like language, memory, you know,
AI's Spectrum of Capabilities
00:14:41
Speaker
reasoning.
00:14:42
Speaker
creativity, perception, planning. So it seems to include a lot of things. And so it's really interesting to think about, you know, AI maybe can emulate some of these better than others. So that's like one thing. um And then, yeah.
00:15:02
Speaker
yeah And then another thing to think about too, is that like human thinking seems to kind of come in degrees in some ways, you know, in the sense of like, you know, children, animals, experts, people with impairments, like there's like kind of varieties of thinking there too. So that might made make you think that a kind of gradual approaches his approach is appropriate.
00:15:22
Speaker
and that Anyway, I'm just kind of giving giving some thoughts about why you know more gradual approach. Yes, I mean, you're absolutely right that there's a spectrum of abilities in the realm of thinking. And we have seen in the last few decades good evidence in the psychology literature that non-human animals have some kind of reasoning ability and some kind of what many people would define as a mind or consciousness.
00:15:54
Speaker
So that's indeed a level, ah on a point on the spectrum that is not at the the level of human thinking and understanding.
00:16:06
Speaker
So ah as you know, it's kind of interesting that AI programs tend to just come in at a completely different point on that map because they're so incredibly good at some things and then very limited at others. So it would be very hard to position an AI program on, ah say, you know a modern chatbot like Claude.
00:16:29
Speaker
it would be hard to position Claude on that spectrum. Like, is it more intelligent than a dog, but less intelligent than a human? I don't know, because it can do a lot of things way better than a human and way faster. But ah clearly, it has other limitations, as as we all know.
00:16:43
Speaker
So, yes, ah maybe even spectrum is the wrong word, because that seems to imply a one-dimensional tunable parameter of either... better thinking or less capable thinking. Whereas in fact, there may be a ah whole complex space of different abilities and thinking. And again, that is another reason to address a more nuanced version of this question inspired by Alan Turing in 1950.
00:17:09
Speaker
ah So that more nuanced version being in what ways and to what extent can computer programs appear to think like a human?
00:17:21
Speaker
Yeah, this is actually a a great ah segue into kind of setting up um the the Turing test and John Searle's counterargument because these are exactly the the kinds of questions that they were trying to, well, that arise from the their and interaction between these two, I almost want to call them traditions because there's been so much ink spilled about these ah particular articles. So I think a lot of people already know what the Turing test is, but just to make sure everyone's on
The Turing Test: Modern Chatbots and Human Mimicry
00:17:51
Speaker
board. ah Tell us, John, about ah the Turing test, the idea of the Turing test, ah maybe the original version, and also what you mean when you're going to talk about, you know emphasizing more so the the concept of emulation.
00:18:04
Speaker
Sure. The Turing test was part of this inspirational 1950 paper by Turing called Computing Machinery and Intelligence.
00:18:15
Speaker
And in that paper, he proposed an experiment that he called the imitation game. So many of the listeners may have heard of or seen the movie, Biopack, about Alan Turing called The Imitation Game.
00:18:28
Speaker
And the the thing that he called The Imitation Game is what today we simply call the Turing test. And a modern reformulation of this Turing test would be if you were having a conversation via text message,
00:18:44
Speaker
with some unknown entity. And that entity may be a chatbot, some kind of computer program, or it may be a real human. You do not know. And your objective is to somehow determine what kind of entity you are having a text conversation with.
00:19:03
Speaker
And what Turing proposed was that if you can't reliably distinguish between humans and computer programs over a large number of tests,
00:19:13
Speaker
um and conversations ranging over a wide variety of topics. If you're unable to distinguish between humans and computer programs, then we should simply admit that the computer program can think.
00:19:28
Speaker
So this is a 100% behaviorist test. Remember, behaviorist means just looking at the outputs and not at all being concerned with how the outputs are produced.
00:19:40
Speaker
That is Turing's imitation game, or as we now call it, the Turing test. It's interesting to note that modern chatbots, I think it's widely agreed, can pass the Turing test.
00:19:55
Speaker
ah The way they're set up today, you can't just open up Claude or ChatGPT and do ah do a Turing test with it, because you can just ask it, are you a human? And it will honestly reply, no.
00:20:08
Speaker
But um I think experts agree it would be pretty easy to reprogram ChatGPT, to pretend that it's human and give convincing answers. It certainly is able to respond to an immense variety of queries on an immense variety of topics in a way that produces realistic prose in natural English.
00:20:31
Speaker
So let's at least, even even if your listeners don't completely agree that current chatbots can pass the Turing test, I think They're so so close that it makes no difference and certainly we can expect that within a few years these things will be even better and really if the objective was to program them to them to have Exactly human-like conversational abilities. I think that could be achieved very easily. Just on that note, um I was going to say that I was talking to some people about this and their objection was that it's actually the in the other direction. They say it's it's too good.
00:21:05
Speaker
I know it's not a human because it can make a poem about like, don't know. ah photosynthesis that where every word starts with ah the alphabet, right? A, B, C, D, right? And obviously humans can't do that.
00:21:18
Speaker
But what you're saying is that we can take this technology, constrain it in a way, actually make it less smart, I suppose, so that it can't do these human-like, ah sorry, above human-like capacities. And so that's what you mean by compatrioturing test.
00:21:33
Speaker
um Yes, you're exactly right. So um I think there are still areas where you can catch them out. And ah if you if you sort of know the tricks of how to trip up these things, you can get them to do a response that is very unlike a human.
00:21:51
Speaker
However, that is becoming more and more rare. And then there's also a question, by the way, of how exactly are you going to judge the outcome of a Turing test? Are the participants going to be experts who are trained in AI and know all the ways to trip up the system, or is it just going to be a sort of average person brought in off the street and said, here, have a text message conversation and tell me if you think you're talking with a bot or a human.
00:22:18
Speaker
So ah there's there's many different nuances there. But to come back to your original point, yes, um Another problem is that these chatbots are too powerful. They can very quickly answer complex mathematical questions, come up with um elaborate poems that may not be ah thematically and stylistically as creative as a true genius human poet, but still it can quickly produce um very powerful
00:22:49
Speaker
um realistic-looking poetry that is better than the vast majority of humans could produce. So, yes, ah what I'm saying, I guess, is that you could train it not to do that, to deliberately be slower, worse, and to even make mistakes. And ah Turing was well aware of this, by the way.
00:23:12
Speaker
he um There's a ah really hilarious... line in his 1950 paper where he's talking about, well, yes, you could you could just ask it to do computations. And you know he gives an example of ah a long, complex addition.
00:23:26
Speaker
And then um he doesn't comment on it. But if you check his math in the original paper, the solution to that addition actually has a mistake in it. So he actually put a little Easter egg in his own paper there where people who could be bothered to check every detail would discover that he had introduced this sophisticated point right at the start that maybe these things could make mistakes and maybe that would actually be part of the process of proving that you're human um or, you know, lying. Essentially, the job of the computer program is to lie about itself and imitate a human. That's why he calls it the imitation game.
00:24:06
Speaker
Now we go under the hood. This is the most technical stretch of the episode, but the basic idea is less mysterious than it first sounds.
Mechanics of Large Language Models
00:24:16
Speaker
A large language model is not a little person sitting inside a machine.
00:24:20
Speaker
At the most basic level, it is taking inputs, representing them as numbers, passing those numbers through a huge network of artificial neurons, and producing probabilities about what should come next.
00:24:36
Speaker
That can sound deflating, just numbers, just multiplication and addition. But one of the themes of this episode is that scale changes everything.
00:24:49
Speaker
When you have billions or trillions of tiny operations interacting, simple parts can produce surprisingly sophisticated behavior. So listen for three ideas.
00:25:01
Speaker
artificial neurons, weights or parameters, and attention. Those are the pieces John uses to explain how the machine moves from arithmetic to fluent language.
00:25:21
Speaker
um Many of your listeners will have heard versions of this before, but for concreteness, let me just go over the basics. So all of these large language models that we're using right now are the mathematical model that underlies the AI assistant that's layered on top of it.
00:25:43
Speaker
So the actual application that you download or interact with on the internet, which may be called Claude or ChatGPT or Gemini, for example, that um is taking whatever your inputs are and sending it to an underlying mathematical model that is called a deep neural network.
00:26:04
Speaker
Now, a deep neural network is made up of typically billions or trillions of
00:26:14
Speaker
artificial mathematical models of, it's called an artificial neuron, but don't think of it as modeling a cell a biological cell. This is a very simple thing that can take a bunch of numbers that come in, multiply those numbers by certain other numbers, add them all up, and if the result is um bigger than some fixed threshold,
00:26:38
Speaker
then that neuron is going to send a message on to all of the other neurons, the upstream, sorry, the downstream neurons that receive messages from this one. So you've got trillions of these things, each connected to, let's say, somewhere between one and a hundred other neurons upstream from it.
00:26:57
Speaker
And the whole system is going to receive some data as inputs, which could be the text that you typed as a prompt. And then that's going to flow through the network.
00:27:09
Speaker
um Trillions of these things are going to decide whether to send a message on to their downstream neuron or not. All of these numbers are being multiplied and added. um It's just simple arithmetic.
00:27:20
Speaker
And then out pops a number at the end, which is a probability for every word that the system knows about.
00:27:31
Speaker
um Typically, they have a vocabulary of, let's say, 100,000 words. ah for every one of those words, there's a probability of what ah should come next.
00:27:42
Speaker
And the large language model is typically going to not always pick the one with the highest probability, but it'll pick one of the words with highest probability. So for instance, if your prompt, and this is ah an example similar to one that appears in the book, if your prompt was, please give me three words...
00:28:01
Speaker
that are related to flowers and rhyme with each other. Suppose that's your prompt. Then, and it goes to the neural network, all of these trillions of calculations take place, which are all simple multiplies and adds, nothing else.
00:28:16
Speaker
And now we have a probability distribution, which might say um the word ah mountain has a low probability. The word book has a low probability, but the word bloom has a high probability. The word rose has a high probability.
00:28:35
Speaker
And it's going to pick one of those high probability words and output that. So now we have a new text prompt, which includes your original input. Please give me three words that are related to flowers and rhyme.
00:28:49
Speaker
Bloom. It's got the first word of its response. We take that. put that new phrase, including the word bloom, into the same neural network. Away it goes. It does its trillions of calculations.
00:29:01
Speaker
But this time, the output probabilities will be different because it's going to be trying to find a word that is related to flowers and rhymes with bloom. So how is it going to do that?
00:29:12
Speaker
Well, There's lots of clever mechanisms in the recent versions of these neural networks. is By the way, the particular architecture that most of these systems use is called a transformer neural network, and that's the architecture that's explained in the book.
00:29:27
Speaker
and And these transformers, like I say, there are many important aspects to them, but perhaps the key one for us is a mechanism called attention.
00:29:39
Speaker
This is a technical term in the neural network transformer architecture. It's not exactly the same thing as a human giving their attention to something, but it reminds us of that. So what is attention? Attention is at each, let me see, yes, at each neuron in the network,
00:29:58
Speaker
um that neuron has the capability to pay attention to any other neuron at the same level in the network.
00:30:10
Speaker
So the reason these networks are called deep is that the neurons are arranged in layers. The input ah arrives at the first most shallow layer and then progresses through the trillions of neurons, which are all arranged in layers, until it comes to the deepest layer, which will produce those output probabilities for the for the different words.
00:30:28
Speaker
And at any one layer in there, there could be like somewhere between a dozen and hundreds of layers, depending on the architecture. But at each one of those layers, the neurons can pay attention to the outputs of the other neurons in its own layer.
00:30:45
Speaker
And this gives it a way of expressing connections between concepts. And as one simple example, going back to that specific concrete example I gave earlier, when the input prompt is...
00:30:58
Speaker
please give me three words related to flowers that rhyme bloom. Now, um at certain points in the network that a word bloom will be paying strong attention to the word rhyme.
00:31:12
Speaker
And this is likely to trigger other neurons within that network that in the huge bulk of training data that this, um,
00:31:22
Speaker
architecture was exposed to when it was trained, which is basically the entire internet, every book that was ever written, and a bunch of other stuff that these companies have been able to get their hands on. um Based on everything it's seen, there will be a neuron somewhere in there that fires when there's combined attention on the concepts of bloom and um rhyme and also flower um but especially bloom and rhyme and then that is going to naturally there's no sort of homunculus in there saying oh this means rhyme rhymes with bloom therefore we need now I'm struggling to think of an actual actual word
00:32:08
Speaker
I think in the book you said gloom next, right? Yes. Okay. I was trying to avoid gloom because of of the confusing fact that it does not ah immediately make you think of flowers, but let's go with that and I'll explain afterwards why gloom actually is is often associated with gloom. Yeah. So um ah gloom does at least rhyme with bloom and ah that that will get a high probability. And if it's sufficiently high, that could be chosen as as the next output.
00:32:38
Speaker
ah Yes. And so, parenthetically, in case listeners are wondering why on earth, so this is a real example. I mean, I did it on it on an earlier, simpler example of ah the the family of GPT models.
00:32:49
Speaker
um I think it was GPT 3.5, where this exact experiment was done that appears in the book. um It turns out, if you go and look on the web, the words bloom and gloom do actually have fairly close associations because there are lots of articles and advertisements that say things like, oh, remove the gloom of your day with some blooms.
00:33:10
Speaker
And therefore, those two concepts do get an association in the real world. And that is something that the model has picked up on. I wanted to mention something about ah the feeling of reading your book, which is kind of ah an interesting idea. But um i I remember I'm, so you explain ah wonderfully all these processes. You explain the difference between an a biological neuron and an artificial neuron. And then obviously the, the, biological networks and the artificial networks.
Training AI: Backpropagation and Error Correction
00:33:40
Speaker
And as you're learning the mathematics of the, I mean, just so that everyone knows the mathematics are very simple. It's like this multiplication and addition, basically. um What is, um what gives the, these, ah the software its surprising power is the scale, right? I mean, we just no no human, I guess no human consciously would want to, you know, multiply and add all these all these numbers together. so there's roughly, so if there's like a trillion neurons, there's roughly a trillion connections, right? You know, I guess. No, there's actually something like, every neuron has, in our human brain,
00:34:20
Speaker
um the ah we have about 10,000 connections for every neuron. In the, um ah The computer models have not got to that level yet. That is just a vast number of connections that are not supported even by the most fancy GPUs and hardware that have been produced so far.
00:34:38
Speaker
But ah depending what kind of task it's doing, i would say you could see maybe hundreds of connections going into a single neuron in the transformer neural network architecture.
00:34:52
Speaker
Yes. So in terms of what are the kind of settings that makes this all work, it's what we call parameters in computer science. You can imagine that every one of those connections between neurons has a number associated with it.
00:35:05
Speaker
And that number, we call it the weight, but it's really how much importance does this connection have? And that's the thing that's going to be multiplied. So when the connection fires, a signal arrives from the upstream neuron, it's going to be multiplied by the weight that's sitting there on that connection, which could be, i don't know, plus 10.
00:35:23
Speaker
That means the signal gets multiplied by a factor of 10, or it could be minus one, which means it gets negated and it's actually going to reduce the chance of the downstream neuron firing.
00:35:34
Speaker
But yes, so we have all of these trillions of neurons. There could be hundreds of trillions of connections. If each neuron has 100 connections, we're just throwing out numbers here.
00:35:44
Speaker
And yes, there are therefore, in that particular network, there will be 100 trillion parameters that are affecting the output. I just wanted to mention one more thing here about the technical stuff and that's the back propagation. I feel like that, that there's, we need some good visuals around this because it just goes through a bunch of computations and then based on the output, it'll get either flagged as that's a good one or it's a bad one. And and if it's a good one, it'll go back.
00:36:15
Speaker
That's the where the back back propagation comes in. Go back into network and kind of strengthen those connections that led to the correct output. Did I get that right? Yes, you did. Okay. So, yeah, this is kind of interesting. So one of the things you said earlier was the mathematics of how these transformer networks um produce their output is amazingly simple.
00:36:37
Speaker
In the way it's explained in the book, a few of the bells and whistles are eliminated. So it's not literally this simple, but the most important features are just multiplication and addition.
00:36:49
Speaker
So yes, literally the numbers come in, they're represented as numbers. The yeah inputs come in, they're represented as numbers. So you can think of every word in English, like the hundred thousand words in English, each having a number that codes it.
00:37:04
Speaker
And those are the things that actually come into the network because everything is dealt with as as a number. And then from that point onwards, all we're doing is multiplying, um,
00:37:14
Speaker
signals, which are numbers, by the weight sitting on the connection and added those all up. And if it's bigger than a certain value, send the result on to the downstream neurons. So you are absolutely correct that this is a very simple process.
00:37:32
Speaker
The process of training the neural network in the first place uses more complex mathematics, but not irretrievably complex. And in fact, in the book,
00:37:44
Speaker
So I hope I was able to give a reasonably good flavor of how this works. And so let me have a go at explaining that now, even in the audio only format. I think we can get some idea of what's going on here.
00:38:00
Speaker
So. the um when you train a neural network the object objective is to set all of these parameters remember each one of the the connections has a number sitting on it that we call a weight or a parameter and that's going to determine how much importance signals going through that connection are given so what we can do it is let's for example say we are trying to um train an object recognizer. So this is going to be a visual recognition system. We present a picture um at the lowest layer, ah sorry, the yeah shallowest layer of the deep network.
00:38:39
Speaker
And the picture has information about all the colors at each pixel, which are translated into numbers. And so we have these numerical inputs coming n And then...
00:38:52
Speaker
That flows through. Initially, this is going to sound horrifically unscientific, but this is the state-of-the-art way that a professional will initialize the parameters of a neural network. They'll choose a bunch of random numbers.
00:39:04
Speaker
So initially, all of those parameters have random values. um We send the first picture through, and it's going to produce a completely nonsense result, like reindeer, when in fact, um it was a picture of a suspension bridge.
00:39:18
Speaker
But what we can do now is we can say, oh, this is what we call labeled data. So I'm going to tell the computer system what it should have done. I know this is a suspension bridge and not a reindeer.
00:39:29
Speaker
Therefore, please go to the probability um that you just output for suspension bridge, which may have been 0.0001, and adjust all the parameters in the system in such a way that that gets increased a little bit.
00:39:43
Speaker
You might wonder, why don't we just change them all so that the suspension bridge goes up to probability one immediately. There's technical reasons not to do that. We have to adjust the system gradually. But what we do is we make small adjustments to the parameters so that instead of being 0.0001, that probability for the suspension bridge is now maybe 0.001.
00:40:02
Speaker
point zero zero one And then each time the system sees that picture of the suspension bridge, it might see it many times while it's training, even though there are millions of other images also. um each time it sees that, the probability of the correct output is going to increase a little bit because you can think of these parameters as little dials that get adjusted.
00:40:21
Speaker
And we just make small adjustments to gradually move the network parameters to values that produce good, accurate answers on as high of a percentage of the training set as possible.
00:40:37
Speaker
Now, why is it called backpropagation, this process of twiddling the dials on the parameters so that we can slightly increase the probability of the correct answer.
00:40:48
Speaker
The reason is that when you do it mathematically, you need to start from the deepest part of the network and work backwards. So what we say is, oh, I want the probability for this for the suspension bridge to be a little bit higher.
00:41:00
Speaker
Therefore, this parameter should have been a little bit smaller. This one should have been a lot bigger. And this one should have been a little bit bigger. and so on. And then we say, oh, well, for those parameters to have those changes, we look in the previous layer and say, that means these parameters should have been adjusted in this way, and so on. And we work backwards, computing the changes that are needed at each stage.
00:41:21
Speaker
And this technique of backpropagation has been reinvented and rediscovered several times in the artificial and intelligence literature. The person that most often gets the credit for discovering it, and he did independently invent it, um it's just that he may not have been the first person um to do so, is Jeff Hinton, who not only won a Turing Award, the highest award in computer science, for his contributions to artificial intelligence and neural networks, but also and they threw in a Nobel Prize in physics. yeah The physicists were so impressed at his work in neural networks that they also gave him a Nobel Prize.
00:42:08
Speaker
At this point, John has walked us through the basic machinery. Artificial neurons, weights, probabilities, and the process called back propagation, where a system adjusts itself after getting an answer wrong.
00:42:21
Speaker
That can make AI sound almost disappointingly mechanical. A picture goes in, numbers move through the system, the output is wrong, the weights get adjusted, try again.
00:42:33
Speaker
But now comes the deeper question. Human beings also produce fluent language. We also recognize faces. We also connect concepts. We also learn from correction.
00:42:45
Speaker
But from the inside, we don't feel like probability machines. We don't experience ourselves as adjusting thousands or millions of little dials. So is a similarity between AI and human thinking only at the level of output or are there deeper analogies hiding inside the system?
00:43:07
Speaker
That's where Sam picks the conversation back up.
00:43:11
Speaker
yeah. So, okay. That's really helpful. So, um, so the system, it's not just kind of producing probabilities, you know, during training, it's also, ah being corrected through like back propagation, you know, um,
00:43:26
Speaker
starting from the error and the output and then you know kind of working backward through the network and making like tiny little adjustments, like you said. um So i guess that raises the question with with kind of human thinking. So you know from the inside, it doesn't seem like we're calculating probability distributions over thousands of possible networks, next words or whatever, or adjusting all these parameters, right?
00:43:52
Speaker
So you might think that the similarity to human thinking is merely at the level of output. you know Both us and ChatJPT produce fluent language output.
00:44:05
Speaker
So I guess my question is, you know do you actually think that there might be a deeper analogy, not just the output analogy, but maybe there's something deeper going on in terms of the underlying process?
00:44:18
Speaker
or So what's your position on that? My position is, yes, I think we can point to similarities within these deep networks and the human brain. um it's going to be at the level of little pockets of similarities. We can think of them as circuits. I'm not using that in ah in a technical, in a rigorously defined technical sense. But when I say circuit, I just mean a little part of a neural network or a little part of the human brain.
00:44:46
Speaker
And for large language models, I think ah we still don't have, or at least at the time the book was written,
00:44:57
Speaker
we still We didn't have great hyper-convincing examples of similarities. So in the book, what I have is the example of these attention values that are calculated between pairs of neurons at the same level in the network.
00:45:17
Speaker
And those, at a high level of analogy, i think it's reasonable to say this is a kind ah of way that that network is able to connect abstract concepts to each other, which may resemble the way that we do it in the human brain.
00:45:38
Speaker
I'm not saying that the human brain uses attention scores. I am saying that somehow the brain must have an ability to connect concepts and produce new associated concepts that relate to the original ones.
00:45:54
Speaker
And that the attention scores in modern transformer models, in some sense, capture that same ability. So this actually links into a concept called emergence that we may come on to later, but I'll leave that there for now, because the thing I do want to mention instead is to go back to the visual recognition problem.
AI in Visual Recognition: Parallel to the Human Brain
00:46:18
Speaker
Because there, because we already have more than 10 years of experience with those models. so the The visual ones were the first ones to make a real breakthrough in this new revolution of AI that involved deep neural networks and reinforcement learning.
00:46:33
Speaker
um Those networks were the were the ones that first really cracked an AI problem that had previously been thought to be essentially insoluble. um For example, in my own career during the late 1990s, I was working in grad school on certain aspects of computer vision, getting computers to analyze images. And I didn't directly work on object recognition, but I knew people who did. And that was just considered a problem that would take many decades to solve and was extremely intractable.
00:47:05
Speaker
And we were making chipping away, like making these tiny little progress, progress you know getting the success rate up from 32% 32.5% some
00:47:15
Speaker
on some pretty small database of objects. Whereas when Hinton and his folks came along in the 2010s, they were able to train up on this image recognition and challenge called ImageNet that contained a thousand different categories of very specific things like um the suspension bridge I mentioned earlier, but also duck-billed platypus, wheelbarrow, nurse, cowboy hat, lots of very specific and specific breeds of dog, um you know golden retriever,
00:47:46
Speaker
um some specific type of spaniel, I can't i remember what, that kind of thing. So 1,000 categories, um they got to train on about a million images, so plenty of examples of each category, and then test on 100,000 images that the system had never seen before.
00:48:03
Speaker
So brand new images that it has no advantage on. And I don't remember the exact results, but they were able to achieve very high accuracy, significantly higher than any previous person.
00:48:15
Speaker
system by using a a deep ne deep neural network. So let's talk about some potential similarities in that kind of visual recognition system with the human brain.
00:48:28
Speaker
What we find is if we look into one of these systems um trained by Jeff Hinton and his collaborators, And we start looking for neurons, not at the very deepest layer, because that layer is the thing that outputs the probability for suspension bridge or golden retriever.
00:48:47
Speaker
um We look somewhere in the middle. And ah if we look at a few neurons, eventually we'll find ones that respond to certain types of inputs much more strongly than others.
00:49:00
Speaker
And for example, as you can see in the book, um There are actual outputs reproduced there of the image patches that produce strong responses at one particular neuron about halfway down the layers.
00:49:13
Speaker
And what this neuron responds to is regions of an image that contain a face, a human face. So this is really amazing because nowhere in the training data is the concept of face presented.
00:49:29
Speaker
There is no object category in those 1,000 object categories. There is no category called face. All of the objects are much more specific and contain multiple parts, whereas a face, in this context anyway, um is always part of a larger object like nurse.
00:49:47
Speaker
So therefore, what has happened here? What has happened is the network has spontaneously Somehow, and I'm going to use some anthropomorphic language here. I don't want to push it too hard.
00:50:00
Speaker
It's not like I believe there is some um human-like figure in there guiding it. But I'm going to use the and ah the the word realized. The network has somehow realized that it's useful when trying to recognize cowboy hat, for example.
00:50:18
Speaker
It's useful to be able to recognize a face because very often a picture of a cowboy hat um it's being worn by someone and there's a face immediately below it. So that's one of the clues that the thing you're looking at is a hat, if you can see a face below it.
00:50:31
Speaker
And you can imagine dozens of other examples where it might be useful to be able to recognize a face because it gives you clues about the other bigger objects um that are present in the image.
00:50:44
Speaker
So we found um a face neuron there, and you can see the the actual outputs in the book to convince you of that. We found other neurons that respond strongly to, for example, um the heads of animals with floppy ears.
00:51:01
Speaker
So this is a very bizarrely kind of specific thing that you might not think you have a neuron in your own brain that responds to heads of animals with floppy ears.
00:51:11
Speaker
um But I would actually be willing to bet you do ah somewhere those trillions of neurons of yours. um Anyway, definitely in the network ah that we were examining, there's a neuron that responds just to those heads. And of course, again, there's no training data that has the concept of ear or floppy ear or head with floppy ear.
00:51:35
Speaker
um This is something that the network has somehow, quote, realized it is useful to have for recognizing various breeds of dogs that also have floppy ears, like spaniels.
00:51:47
Speaker
So we found a neuron that for the floppy-eared heads. We also found another neuron for pointy-eared heads. And of course, that neuron is useful for recognizing things like foxes and cats.
00:52:00
Speaker
So this is really quite remarkable. And I think this is a very inspiring connection between the way that deep neural networks are able to produce results on AI problems and the way that human brain processes function.
00:52:19
Speaker
um I won't go on about it now, but there are other experiment experiments in the neuroscience literature demonstrating that humans do have these same kinds of neurons, neurons that can can recognize specific objects and parts of objects and other low-level concepts that can be combined to produce a high-level inference.
00:52:46
Speaker
Notice the move John makes here. Instead of claiming that the whole AI system thinks like a whole human brain, he makes a more modest claim. look for pockets of resemblance.
00:53:00
Speaker
Look for small circuits, small patterns, small functional similarities. And to make that case, he turns away from language for a moment and toward vision. That might seem like a detour, but it gives us one of the clearest examples in the episode of something genuinely surprising.
00:53:18
Speaker
An AI system developing internal features it was never explicitly taught. Not just memorizing labels, not just matching surfaces, but discovering useful structure.
00:53:37
Speaker
so Just to echo real quick what you what you were just saying, John, like
Emergent AI Properties: Understanding and Creativity
00:53:41
Speaker
so it's really important that the network is not just like memorizing labels or like matching surface and images. It's it's discovering kind of useful features somehow that were not given in the training data. and and i mean you You mentioned emergence earlier. i mean the The concept of emergent AI is really one of the key concepts in your book. would you want to maybe unpack
00:54:13
Speaker
like This would be an example of emergent AI for you, right? and maybe So maybe you could define that. and Yep. So the concept of emergence itself is just the fact that we observe in the world systems made up of a large number of small, simple, interacting components.
00:54:34
Speaker
Systems like that can often produce qualitatively new, interesting, definable, observable behaviors and properties.
00:54:45
Speaker
And um ah the simplest example that's often used is that of a copper wire. It's made of many trillions of identical copper atoms. um This wire has a macroscopic property that it can conduct electricity. It has conductance. It has flexibility. You can bend it.
00:55:04
Speaker
But if you look at one atom, um that can't conduct electricity. That doesn't have any flexibility. It doesn't even make sense to consider those properties. So emergence is a very obvious, natural thing in certain situations like that copper wire.
00:55:19
Speaker
But we can also apply it to things excuse me such as um the human mind. So many people would agree that the human mind is has reality.
00:55:33
Speaker
it's It's a phenomenon that we can investigate, discuss. It has properties. um it's ah It's a real thing. But where does it come from? Well, um I and many others would say it's an emergent property.
00:55:46
Speaker
So there's no ah mysterious mechanism operating operating inside your head. All of the neurons, we have a pretty good understanding from neuroscientists of how those neurons individually work.
00:55:58
Speaker
They're all connected to up to 10,000 other neurons sometimes. And by interacting, this large number of relatively simple components somehow produce the emergent property of a mind.
00:56:13
Speaker
So that's emergence um as as it occurs in biology. What about in a computer program? Well, the terminology that I use in the book is to apply this word emergent to human-like behaviors that we can see in these networks.
00:56:33
Speaker
So I would say that the neurons that spontaneously became able to recognize faces and pointy ears without ever being specifically trained to do that are an example of emergent recognition.
00:56:48
Speaker
Somehow the training process with its millions of iterations and the artificial neurons in that network with their millions of interactions have produced an emergent property, which is the ability to discover a new feature, that of the pointy ear with a human face and to successfully recognize it.
00:57:09
Speaker
So that's emergent recognition. The book also has examples of emergent understanding, emergent thinking, emergent creativity, and so on.
00:57:21
Speaker
So far, the conversation has moved from the outside of AI to the inside.
AI Understanding: True Comprehension or Symbol Manipulation?
00:57:26
Speaker
The outside question is Turing's question. Can the machine produce behavior that looks intelligent?
00:57:33
Speaker
The inside question is John's question. When we inspect the machinery, do we find processes that resemble parts of human thinking? But now we arrive at one of the most famous objections in the philosophy of artificial intelligence.
00:57:49
Speaker
Even if a machine produces fluent language, does it understand anything? Or is it merely manipulating symbols? This is a challenge raised by philosopher John Searle in the Chinese Room argument.
00:58:05
Speaker
And in a conversation about modern AI, the Chinese room becomes newly strange. Because the question is no longer just whether a rulebook can imitate understanding.
00:58:17
Speaker
The question is what happens when the rulebook is replaced by something like ChatGPT.
00:58:29
Speaker
to see Let me try to summarize for you John Searle's Chinese room argument, and I hope to highlight all the things that you want to talk about. What he does is provides a thought experiment where he gives ah um a way to demonstrate that it doesn't seem like computer programs really have any real understanding of what's going on. And so he imagines someone in a room, he calls it ah the Chinese room,
00:58:57
Speaker
And this person is speaking with a ah Chinese, ah someone who is competent in the Chinese language, maybe a native speaker, and they're getting notes through, let's just say, you know, ah under the door, right? And so the person inside the room is getting these notes ah in Mandarin, and they have a rule book or a ledger or whatever, so some um some book there that kind of instructs them, if you see these symbols in Mandarin, ah even though the the person in the room doesn't know Mandarin, they are able to, through the rulebook, figure out the appropriate response to slide back under the door.
00:59:39
Speaker
And so John Searle's ah point here is that obviously the person inside the Chinese room has no idea what's really going on. He doesn't know what the conversation is about in Mandarin. But to the person outside the room, they they feel like they're having a fluent conversation in Mandarin.
00:59:57
Speaker
And so for this reason, the same thing is happening with the computers or computer software. It might seem like it understands, but really it's just engaging in formal ah processes and it doesn't really know what's going on. Did I capture all the elements that are important to your argument, John?
01:00:16
Speaker
Yes, absolutely. So the the place where the book would pick that up is to say, h It is obvious that I call um that person in the room the human subject.
01:00:31
Speaker
And I think of the people outside the room as the experimenters. They're trying to do an experiment to see if whatever is behind that door ken can understand Chinese. That's what they want to know.
01:00:41
Speaker
And we all agree that the human subject cannot speak Chinese, does not understand it any Chinese. um All they can do is follow instructions written in English and laboriously copy Chinese characters that are in the instruction manual without understanding what they mean.
01:00:57
Speaker
So, yes, I mean, Searle's basic point is there's no part of the system that understands Chinese. Clearly an inanimate Instruction manual can't understand Chinese.
01:01:08
Speaker
We all agree that the human subject doesn't understand Chinese. um So no, there's nothing that understands Chinese here. Even if it but even if the behaviorist approach would say, the system's producing perfect Chinese, we must declare that it it understands Chinese.
01:01:25
Speaker
So what if we take that instruction manual and replace it with a computer and in modern times, I mean, Cyril didn't have access to this, but in modern times, we could just give it an internet connection and let it use ChatGPT or Gemini to um produce response ah appropriate responses from the Chinese inputs.
01:01:49
Speaker
So now the human subject is reduced to doing almost nothing. I mean, if if if they're receiving pieces of paper slid under the door, I guess they have to pick up the paper and maybe hold it up to the camera um on the computer.
01:02:05
Speaker
because optical character recognition is also very good now in these AI models. So it can probably just read it by holding up the piece of paper in front of the camera, produce an appropriate output, print that out, and the human subject just is literally shuffling pieces of paper and doing nothing.
01:02:21
Speaker
And so now our intuitions are going to respond quite differently to this whole system. Again, everyone agrees that the human subject doesn't understand Chinese. But what about the computer or what about the the um AI system that's producing these outputs?
01:02:37
Speaker
It's a lot less clear. And depending on how much we look under the hood and how much we're willing to take a behaviorist versus um more internal-based approach here, we might agree that there's some kind of phenomenon here that we could label as emergent understanding.
01:02:59
Speaker
So rather than just saying it's true understanding, because I find that actually a difficult concept to even define, um let's just use a different label. We could use artificial understanding or emergent understanding, which emphasizes the fact that this understanding has emerged from the complex interactions of those many artificial neurons inside the large language model or the or the transformer model.
01:03:23
Speaker
And yes, so I think this is An interesting new challenge to Searle's thought experiment because now that we're armed with programs that can do such good machine translation and natural prose responses to questions, I think that really starts to push our intuitions closer to thinking that there is an aspect of this system that changed.
01:03:57
Speaker
emergent understanding. It may not be the same type of understanding that humans have, but it is a phenomenon worthy of study and it is a form of understanding. So we can push it even further, but that's that's one point at which we ah you know readers and listeners can decide like what what level of comfort do you have with this approach? Are you willing to grant that there is a phenomenon going on here that is some kind of understanding?
01:04:23
Speaker
And would you be comfortable calling it emergent understanding?
Future AI: Hybrid Models and New Possibilities
01:04:28
Speaker
This might be the part where i i put my my foot in my mouth a little bit speaking to a computer scientist here. But um ah Sam and I have been interviewing many people about large language models and their capacities and their limitations. And so let me you know kind of run this ah idea by you and kind of get your response. um one might feel that, you know, ah large language models, maybe because they can't engage in causal reasoning, according to some people, or maybe because ah they're still so prone to hallucinations or they can't respond to anything they haven't been trained on or for, you know, reasons X, Y, z um There is ah some limits there that might make it so that large language models are not really, um you know,
01:05:16
Speaker
ah sufficient to to get at um real emulation of human thinking, since human thinking does have causal reasoning, for example. um But do you think that um even this line of argument will eventually be... um you know, kind of ah rebutted by new hybrid models where we take a large language model and, and you know, I don't know how to connect it with some Bayesian, you know, networks and some, a world model and whatever, right? Will there be new hybrid versions of of AI models where it would be increasingly difficult to sort of make these arguments? Yeah.
01:05:57
Speaker
I would actually approach that issue another way. I mean, certainly i think there are going to be new architectures and who knows what the next big thing is going to be. Maybe it will be a new architecture of neural network model. Maybe there'll be some completely different thing that comes along and is even better. I'm actually crossing my fingers that doesn't happen anytime soon because I don't want my book to go out of date like you know within two months of of it being published.
01:06:19
Speaker
I've made pretty strenuous efforts to future-proof this book in the sense that all of the examples I give are evidence of emergent thinking, emergent understanding, emergent recognition.
01:06:32
Speaker
um It may be that next month someone comes out with even better examples of those phenomena, but this book still ah sort of... ah puts a stake in the ground at a certain point in this AI revolution and shows, look, even now we can see these glimmers of circuits in the deep neural networks that have similarities to circuits in human brains.
01:06:58
Speaker
And that's where I would stop my strong claims at that point. So I'm not trying to say that LLMs as they exist now understand English or understand natural language perfectly.
01:07:13
Speaker
ah That's absolutely not the case. And specialists know how to how to catch them out and trip them up, cause them to hallucinate, cause them to and make false statements. So ah that's not the case. But what I hope the book does is point out some examples of some aspects of these models that have resemblances to human brain processes.
01:07:35
Speaker
And we can expect that as these models get bigger and more sophisticated or maybe completely new things come along, um that the match between how many of the circuits in the AI program have similarities to those in the humans, it may increase to be increasingly similar, or maybe it will diverge, and we'll find that In fact, it's much easier to write AI programs that work in this completely different and clearly non-human way, but but is still useful ah for humans.
01:08:07
Speaker
Of course, I can't predict that. But there's another angle of the book that we haven't talked about at all yet, which is fairly early on, I go through an argument about brain simulation computer programs. So computer programs that literally try to emulate the inputs and outputs of every neuron in a human brain.
01:08:27
Speaker
That's far beyond our capacity to scale ah right now. But um there has been a lot of progress in brain simulation. And ah for example, the entire brain of fruit fly was mapped as of a couple of years ago.
01:08:44
Speaker
And we're also to keep making progress from there. So um i I believe, and I think the book explains that it's very plausible and that I hope most people would agree um once reading the arguments, that it's possible in principle to simulate a human brain. Just take all the connections between all the neurons and calculate what would the outputs be given these inputs, and hey presto, here we have a computer program that is producing the same outputs as a human brain given the same inputs.
01:09:22
Speaker
And so this, you know, imagine you get that program and put it into the computer in the Chinese room. Now, um are you going to credit that with some kind of emergent understanding?
01:09:33
Speaker
It's literally emulating a specific human brain that's been scanned and we know all the ways that the neurons work and we're simulating them with enough precision that it produces the same outputs.
01:09:44
Speaker
um It's not going to convince everyone, but I think at this point, you know, many people would be willing to at least adopt some terminology of saying, yes, it's emulating understanding. It's producing the exact same outputs as what a human brain would in that situation.
01:10:02
Speaker
And therefore, it's emulating the understanding of Chinese that the real brain but clearly produces. So that's one angle of the book. But I don't like to lean on that too heavily because this whole brain simulation thing, number one, it's pretty impractical. there's a It can be done in principle, but to actually do it in practice I don't know if that's ever going to work, um and it certainly would be a long way in the future.
01:10:26
Speaker
I think it's more productive and interesting to look at real AI programs that have been developed and try to understand which aspects of them resemble some human processes and to what extent.
01:10:38
Speaker
And that helps us answer that original question in the book of in what ways can a computer program think like a human?
Conclusion: AI's Place in Human Experience
01:10:50
Speaker
There is an anxiety underneath all of this. If machines can produce language, solve problems, recognize patterns, imitate creativity, and perhaps even develop forms of emergent understanding, then what is left for us?
01:11:07
Speaker
That question can sound threatening, but John's final answer is not nihilistic. It is, in a quiet way, humanistic. Even if intelligence is not ours alone, human life is not exhausted by intellectual performance.
01:11:24
Speaker
We still have relationships, humor, grief, ritual, love, the strange texture of living a life from the inside. So the episode ends not with a declaration that machines are human or that humans are machines, but with a more generous possibility.
01:11:42
Speaker
Maybe intelligence was never the only thing that made us worth caring about.
01:11:48
Speaker
John, is there anything we haven't asked you yet that you want to um you know add to the conversation? Well, I guess we didn't really talk about the implications of this for humanity. If if if we accept the hypothesis that computer programs can exist,
01:12:07
Speaker
equal or exceed humans in all of these intellectual arenas, such as creativity, intuition, thinking, understanding. um What's left for us humans to do? Some people might get a kind of nihilist feeling um if if if we're forced to accept that hypothesis. And ah so towards the end of the book, I try to put a positive spin on that. i um I certainly am very optimistic and positive about this.
01:12:40
Speaker
i'm I'm quite happy to accept that human brains aren't the only thing in the in the universe that can do sophisticated computations and understand things like natural language and have possessed creativity. I'm quite happy for other entities out there in the unit universe to do that. And I'm happy for some of those entities to be AI programs that that we've built on this planet.
01:13:01
Speaker
um that That doesn't bother me at all because I'm going to continue to have relationships with other human beings and to enjoy things like jokes. I'm going to laugh. i mean I'm going to cry.
01:13:13
Speaker
um for those who ah receive fulfillment from religious practice, they can continue to practice religion and receive fulfillment from that.
01:13:25
Speaker
I don't see any of that as being in conflict with the fact that we also have these entities in the world and and some hypothetical future world that exceed us in all intellectual realms.
01:13:37
Speaker
So I'm positive about that. I think um while there are dangers and potential problems with AI, I think i'm I'm optimistic that we can avoid catastrophic outcomes and that overall, those entities will tend to improve our lives rather than harming them and that we can continue to and enjoy our own humanity.
01:14:06
Speaker
That was computer scientist John McCormick, author of Thinking AI, How Artificial Intelligence Emulates Human Understanding.
01:14:19
Speaker
Today's conversation began with Turing's old question, can a machine think? But maybe the better question is the one John leaves us with. In what ways can a computer program appear to think like a human?
01:14:32
Speaker
And what do we discover when we look under the hood? For the A-Tech Philosophy Podcast, I'm Roberto Garcia. Thanks for listening.