There are companies very interested in connecting AI directly to the brain. I cannot imagine a future in which we will be the dominant side in the long run. How would this happen? Well, if we believe that this model is starting to build a representation of another player, then there is a good chance that it is building a representation of itself. We call it self-awareness. Self-awareness is different from consciousness. And that's what self-awareness is all about. There are these distinctions. We are now able to turn consciousness on and off with increasing precision using chemicals that we don't fully understand how they work. Neural network programming is like a nozzle for growing a plant. We will certainly change and we will change in unimaginable ways, in a way that we may become unrecognizable. A well-defined technical problem is whether we can record an experience from one brain and transmit it to another brain. Human evolutionary development is slow. We cannot speed it up at will. Where is it really a matter of 6 years for models to be able to solve problems lasting 100 years?
You started beautifully. I absolutely agree with that, that what we are observing now is the beginning, the beginning, the beginning. It turns there, and what is the end? The end, the end, the end.
The conversation you are about to see was created within the walls of the Jagiellonian University as part of the Be Humans in the Age of AI event organized jointly with our This is it channel. The channel partner is Zencom, a technology that improves shopping and payments for informed users who want more and for companies to grow faster. Isolution is a Polish software house that produces software for identifying, anonymizing and retaining personal data. Wizards.io.
Wojtek, I have the first question for you. I'm going to read something to you and at the end I'm going to ask you if you think it's true?
I thought I was supposed to say whether it was a charge pity or a human.
No, I will say right away that what I am about to read was written by a human being. It is not the neural networks that are programmed, but the training algorithms that modify these networks, often without human supervision. And only the training algorithms are understandable to us, not the product of training. We understand the evolutionary mechanisms that led to the development of the frog, but we have almost no understanding of how the frog works. Frog does not have its own designer in the same sense that GPT chat does not have its own programmer. So I would say that programming neural networks is closer to growing a plant. It's like this: we put in a seed, tear it up, maybe add a little fertilizer, and see how it grows. And we even know, we do some research. We add a little water here, more there, and on average here we see how it affects growth and then we know how much water to give, how much fertilizer to give. We adjust the light from which side it should come so that it grows most beautifully. So this is how we grow these nets.
Hello, this is Maciej Kawecki, the host of this channel. I know I'm interrupting your viewing of this debate, but I wanted to share some important news with you. I wrote my first book. For many years I have been having conversations here with the greatest minds on Earth, the titans. I have collected previously unpublished interviews into a book that will be released in October. You can pre-order it via the link in the description of this debate or via the QR code you see, and receive the book with my signature as well. Order it, support what we do and delve with us into the secrets of reality.
You mean we hardly understand how neural networks work. This is a quote from Andrzej Dragan, who is here next to us.
I could say so. It's even a little bit similar to raising children. Yes. We have some psychology about how to raise children. We have a way of teaching. There are weekly magazines that suggest how to raise children, but the parent doesn't really know why the child did a certain thing. No, he doesn't know exactly, he doesn't understand the elements of his brain.
Why did you share this tweet, Andrzej?
I will start such a discussion on this topic because I am a physicist. My business is understanding. I'm trying to understand, and the standards of understanding for physicists are a little different than the standards of understanding everywhere else. For example, to understand the theory of relativity or quantum theory, it is not enough to recite a fragment of a textbook, you need to be able to look at this fragment of knowledge from several different directions, understand the angle examples, know where it comes from, be able to derive it, understand where it comes from, which stages of understanding we do not understand, what falls from the sky, what is a postulate, what is in this understanding. So, in general, physicists' need to understand is much deeper. And it's interesting that neural networks, as Wojtek said, actually design a very detailed training mechanism, and some training mechanisms worked better than others, for example, recurrent networks converged slowly to results, but transformers accelerated this convergence process, and even though we have these approximation theorems that say that in principle there is even a single-layer neural network that approximates any function, finding that function, well, it's unclear how there are no rules on how to look for them, so we experiment a bit. Transformers actually worked better than recursive CCDs. Well, great. If they didn't work, then they wouldn't work, obviously. We keep looking. So we really do have a great understanding of the training algorithms we design. Just as we understand evolution very well. Well, we understand evolutionary mechanisms. We know that a frog has its information about the frog's structure encoded in a string of amino acids. We know how this sequence multiplies. We know how the process of natural selection and sexual selection takes place at this meta-level. Well, we roughly understand it and know that this iterative process, after a billion years, led to the creation of the frog. But we have a very vague idea of how a frog works. We know how, that he has an eye here, he has a leg here, he has a lung here. We can cut ourselves apart, look inside, and we start to understand something. But I don't know how a frog cell works, I don't know how a frog brain works, I don't know how a frog's tail works. Our level of understanding of the frog is very murky, to say the least. And it's the same with language models. Well, we can do post-facto tests after such a model has been installed, to see what these models do. It is extremely interesting that we live in times when a program programmed by a programmer or his product of this programming is later studied by psychologists. This in itself is deeply shocking. But what's interesting is that on the one hand, our level of understanding of a frog is similar to the level of understanding of a persistent neural network, especially for these large neural networks. There are some stories about small numerical networks, for example, convolutional networks. They say there are some fairy tales about the first layer detecting small elements, the next layer detecting structures between structures, and these subsequent convolutional layers are some kind of process of perceiving more and more details, which is probably true for small convolutional networks, but probably not for larger ones. But we say things like that, or do people who lecture computer science students to create the impression that we understand something, so that, but I just wanted to say one thing: The fact that we don't understand networks well is a fact, and that's also true. Besides being interesting, it's also why networks are generally interesting, because training networks in this way allows us to solve certain problems that we wouldn't be able to solve with classical programming techniques. For example, I can barely play chess, but if I were to set up a neural network that would train itself, like the alpha zero you mentioned, through reinforcement learning, through self-play, playing with itself many times, then in this way I would be able to train a network that would beat the world chess champion. It will not lose at chess at all. So in that sense, I have no way of knowing how this network works, since I don't know how, I don't know how it will make this next move, I don't know why, I don't know what's behind it. And this is also something that worries me as a physicist, because I want to understand and I know that I can't, but on the other hand it opens up new possibilities for me, because my understanding is no longer the upper limit, it is not the ceiling for what I can achieve.
This is also Wojtek, your misunderstanding. You agree with what Andrzej said.
I agree and I think that if we look at the world, they are reality. Physics is about understanding reality at the most fundamental level. And at this level we are able to understand, or want to understand, every aspect of reality. As Andrzej said, you want to be in a situation where you look at the equation from different sides and you understand its consequences precisely. And even though we understand every aspect, we can understand every aspect of the molecules, how they interact with each other. If we have large enough systems, for example the human system, even though we have equations that describe how the molecules interact with each other, we are unable to use the same equations to describe human psychology. It's just another level of abstraction. And it seems to me that different topics, different fields require different levels of abstraction and will have explanations that are very precise using equations or less precise, like psychology.
You want to say that this is due to the number of parameters, the complexity of reality that we have created inside neural networks.
As I sometimes hear, and unfortunately I hear this unawareness even of people who train neural networks, and this is unfortunately quite common at a certain level, this thesis. Whereas if we could basically trace it neuron by neuron, we would basically understand everything, and after all, it's just a function, right? Because the neuron is just a function. Anyway, when I hear something like this, I say that for a human being it is only coal and water. But it's not just about the large number of parameters. Well, the stone also has a large number of parameters, but generally speaking we don't feel that we somehow don't understand how the stone works. Well, at the level of atomic physics I understand how one atom works. I understand how 10 atoms, 1000 atoms, a million atoms, and so on work. And the stone doesn't have many new structures emerging there, even though there are a lot of atoms there. There are about as many atoms in the brain as in a stone. There the order of magnitude may vary. However, this complexity is not just a multitude of parameters. It's not just large scale and that's it. There are a lot of mechanisms there, probably very simple mechanisms, that we don't understand. As a student in Amsterdam, I attended a lecture by a physicist inside a cell. A cell also has many, many atoms. Basically, what's complicated here is to understand, because there are a lot of atoms, but not when you look at the mechanisms of how individual cell elements work on organelles, often these are simple tools, simple, even mechanical devices that resemble simple machines, levers, gears, there are simple mechanisms for opening and closing some membranes, chemical mechanisms, mechanisms, so there are a lot of solutions of this type, which are not only, it's not just a matter of a large number of parameters, but many ingenious solutions that we don't understand. We begin to understand the simplest things, but we don't understand the complex ones. He doesn't understand how a cell phone works. No Demis Hasabis swears that within 10 years he will be able to map an entire single cell using a neural network. This is a very ambitious plan. But it's the same with the brain. Well, it's not just that there are a lot of parameters, that it's such a large function, there are various clever solutions there, because the brain, this function, is very clever. But there are many clever solutions, even if only to compress knowledge, so there are many discoveries that we haven't discovered yet, so it's not like it's enough to scale it up and I'm saying that I'm saying that just like we don't understand what's going on in the brain and there are many ingenious solutions that will teach us many things one day. Likewise, there are a lot of clever solutions in language models that allow us, if we understood them better, they would allow us to understand a lot more about the mechanisms of text production by humans, for example. I point out that people often speak very quickly, faster than they think. This often happens to me, I speak fast and think slower. This is one such system, as Kaneman defined it, and it seems that the language models they have are a system on steroids, at least for reasoning models.
I remember my interview with John Hopfield, this year's Nobel laureate, one of the fathers of artificial intelligence, Hopfield networks, and when I asked him if he had any regrets, he said that I regretted one thing: that from the moment the concept of Hopfield networks was created, the development of neural networks did not take place within the framework of biological sciences, but within the framework of computer sciences. Don't you have the feeling, Wojtek, that if the development of neural networks had taken place together with biologists and not computer scientists, we wouldn't have these question marks we're talking about today.
I mean, historically, a lot of people have tried to find inspiration from the brain in training neural networks. Even if you look at Deep Mind, they hired a significant number of neuroscientists, brain scientists. And the main idea that comes from biology, you know, it might sound trivial at this point, but it's this idea that a neural network is a digital brain. And I can say that this perspective actually had a significant impact on the development of artificial intelligence, and in general, on the realization that it would be possible to create a network that would be as smart as a human or even smarter. However, so far the results from these biological sciences have been limited. This does not mean that they will always be limited. It's possible that now that many more people from different fields are starting to be interested in neural networks, there may be new ideas coming from there, so there's a chance that this will change. However, so far, progress in neural networks has been closer to the field of so-called optimization.
Well, yes, gentlemen, but this digital brain defines the world through mathematical relationships between data, and the human brain defines the world based on experiences.
I can tell you whether bits of information are coming in from the brain's perspective and to what extent those bits of information are converted into weights in the brain. The human brain is also capable of performing tasks in the world. Yes. Now if we look at AI, it is also the case that bits of information come in. There is an algorithm that turns these bits of information into action in the world. Now there are some differences. Correct. There are some differences. It is so that, for example, in the case of how we currently train neural networks, in the first phase they learn on their entire Internet. Where in the case of a human being, a human being mainly learns from, as you said, his own experience and it has a certain influence. It is so that initially neural networks learn to mimic people on the Internet. In the case of humans it is a bit different. I'll try. I want to try to say such an interesting point. For example, if you train a neural network in a computer game, let's assume it's a three-dimensional computer game, at the very beginning of training the model learns to distinguish between the sky and the ground. Then he can start to notice where the apples are at the points. What happens is that it starts to build a 3D model of reality internally. For example, if you have a network where you can ask the model to tell you what it will see if it rotates a few degrees, it can somewhat predict what it will see. If you ask him to tell you what he will see when he turns 360 degrees. At the beginning of training, it thinks it will be something different than what it sees now, but if you train a little longer, you predict that you will see exactly the same thing, which means you are starting to understand 3D sufficiently. If there are other players in the game, then obviously this neural network has to learn to be able to represent the other players. He must be able to know whether if he goes to fight this other player, whether he will win or not, whether the amount of energy he has and the items he has are enough to kill this other player or not. If we believe that this model is starting to build a representation of another player, then there is a good chance that it is building a representation of itself. So you could say that we call it self-awareness or self-knowledge, and it could be a consequence of training models in environments where there are many players and as training in that environment, because you start out, if you want to score points, you have to understand other players and you have to start representing yourself. This may be the result of such training. At this point, just as we mainly train models like chat GPT, this first stage involves imitating a large number of people. The thing is that this model doesn't do much action, at least not in the first stage of training.
Can this Andrzej, this imitation, not imitation but consciousness, self-awareness, awareness of oneself, be a source, a seed for the subjectivity of decisions, models?
I think I don't know and I think we don't know and I just wanted to point out that there are so many opinions in this discussion. You talk about subjectivity, about whether there is consciousness in national networks or not, whether it is not there. Well, Geoffrey Hinton, who is the creator of the first simple language models, claims with great conviction that consciousness is there. Yann LeCun laughs at this. He says it's just stochastic parrot and matrix multiplication or things like that. Well, we have two outstanding specialists who know how it works inside and out and they completely disagree. Which means that the nonsense content in this discussion is at least 50%. And at least 50%, because it's one for one, one is wrong, it doesn't mean the other is right, so at least 50%. It doesn't matter which side I'm on, it doesn't matter what my opinion is, but I would like to draw attention to this, because opinions are irrelevant in science. I don't really want to add another opinion to this discussion. There are too many of them, not too few in my opinion. However, I would like to draw attention to one thing. You're thinking a little bit here and it sounds like this in this talk about consciousness, thinking as information about this or that system or not, then self-awareness is something different than consciousness. Well, well, what does this self-awareness mean, self-awareness and consciousness, there are these distinctions, for example, if you have a toothache, I had such a discussion with my friend Tomek Czajka recently. You have a toothache, and you feel this toothache and this is information that something is wrong with your tooth and the brain knows this information and interprets it as pain. You can tell me all about it, describe the pain precisely, I will basically have the same complete set of information, but it is you that is hurting this tooth, not me. And basically we know the same amount, and we have the same information, but you, your set of feelings is completely different. So these are the sensations that are in our brain, that we feel, that seem to go far beyond just the complete set of information about this or that subsystem, or about myself. And my standards of understanding make me think that we completely do not understand what this feeling is. Another is another matter. People, critics of neural networks, rightly emphasize that networks, every neural network is just a function. This is just a function. This is the word, only it is here, only you can unbutton it, because in the end everything is just a function. Well, physics is just physics, physics is just multiplication of matrices, because the entire quantum theory is linear algebra, which of course only and exclusively determines the matrix elements of the probabilities of some process. So everything is just mathematics, to put it a little more precisely, so this word is just completely unjustified. I can't quite hear myself like when someone says that it's just a function, because everything is just a function. As Wojtek says, we disagree. However, there is an interesting question whether this complete reading that we have in our brain also loses some function for it. Because at this fundamental level, basically everything is a function in some sense. However, I had such an interesting conversation with someone in the room today. Is it quite good between an artificial neural network and the brain network, which is based on wet biology and some other kind of interaction that we model with neural networks? Aren't there any structures other than just function? We don't know that. Well, Penrose, for example, is convinced that there must be something non-obligate there and he has been repeating this for many years without any arguments that he considers reasonable, but he repeats it with great conviction. Well, nevertheless, it is really difficult to imagine that, if we could imitate the behavior of a system like a human, which responds to some signal, a set of input signals, with some signal, a set of reactions, which can be digitalized and replaced with a function. So you can probably imitate a human with a function. The question is whether we can actually recreate everything we feel inside this type of simple network burrow. I don't think we know. I mean, yes, we don't know, we don't know. At this point, we do not see, and we are probably not aware of, any structures in the brain that go beyond just functions. But some people, especially those who want to know the uniqueness of man, seem to be waiting for it very much. The interesting thing is that this question is there are scientific steps to even answer it. There are a few startups at the moment. The most famous one is Neuralink, but there are many more companies that deal with connecting computers to the brain. And so from specific or what, from specific technical problems, which is a well-defined technical problem, it is the problem of whether we are able to record an experience from one brain and transmit it to another brain. The whole experience. This is a specifically defined technical problem. And I think in the next decade people will make progress on this topic. There are also companies that are very interested in connecting AI directly to the brain. At the same time, if it is possible to connect a computer to the brain, it will also be possible to connect brains to each other. I wanted to say one thing, because I absolutely do not believe in predicting the future. If there is any lesson in physics that we have learned after 400 years of practicing modern physics, it is that when people predict the future or share their opinions, they are talking nonsense. That's why no one believes in quantum theory, no one believes in the theory of relativity. We have been forced to accept that this is how the world works. Despite our beliefs that this obviously cannot be the case. Deterministic worlds, of course time is something simple and so on, because here it turns out time is very complicated, it can slow down, speed up. Reality is not true, all this strangeness of quantum mechanics has been forced upon us despite our beliefs, and that is why physicists are often so cautious when predicting the future. Well, the most future will share the opinion of this Dzdor and even here this discussion Hinton vs LeCun shows it. Yeah.
Well, yes. Hinton now talks about it, he says that saying that we will not decide with discussion and a wise face what consciousness is and whether it is there or not. I suppose we will achieve this, for example, through anaesthesiological research, where we are increasingly able to turn consciousness on and off with chemicals that we don't fully understand how they work. They work. Not more than consciousness, because we don't know what it is, I'm interested and I would like to come back to that. Subjectivism, because the subjectivity of decisions is a process that accompanies humans. And I still don't understand how we can compare the way a human thinks and a synthetic neural network, when the synthetic neural network sees the world through vectors, and a human sees the world through associations, through connections.
