Czy AI zastąpi FIZYKÓW? Andrzej Dragan i Paweł Janowski | 88

And I assure you, ladies and gentlemen, that everything Albert Einstein ever said or wrote in his life can be described by a certain very complicated, probably practically impossible to determine, probability distribution, such that every word that comes out will be the most probable according to this distribution. Connecting the dots, this ability to connect the dots, two seemingly unrelated things, is something that has given us a great deal in human history, and in science in particular. And to say that it is only that, well, it is an understatement, a lack of respect for our history, because we owe a great deal to the ability to connect the dots. Theoretical physics is a self-contradictory term, as physics is fundamentally an experimental science. Good, I like it. So something like a theoretical physicist does not exist. I am something similar. I know that at this moment, on the horizon, there are no outlines of anything that I would be able to call a theory of everything. Simply put, it is too ambitious a goal. At this moment, physicists have a habit of trying not to ask questions that are too difficult. Even if they are very interesting, we refrain from asking overly difficult questions. Not because they don't interest us, but because we don't want to waste time and energy talking about them. Good day, ladies and gentlemen. Welcome once again, I hope, to an event in the "Science Bunker" series. My name is Paweł Janowski, and our guest today is an artist, composer, director, but above all, a theoretical physicist, and a science popularizer, Andrzej Dragan. Professor Andrzej Dragan, good day. [applause] I must admit that among all these roles, physics is the closest to me, and I hope a physics influencer, because we lack such people, and it's very good that they are emerging, of course. And my first question would be: why physics? Because I recall from some of your statements that you are almost a programmer, and yet you ended up in physics. Why? Well, because there is nothing more interesting to do in life than that. And for me, at least, it's the best way to fight for dignity. Because essentially, we live in a reality that we don't understand, and most people don't mind. I've always had doubts about whether consumption is truly the best way to spend free time. So, of course, our possibilities are extremely limited, and promising ourselves that we will understand something in a short time is probably an exaggeration, but it's probably worth trying, so that was roughly the motivation. So, physics is the key to understanding the world. Yes, one way or another. Well, generally, there is no better method. People have tried for a few thousand years of civilization, even longer, to understand reality, and it hasn't gone very well, meaning the pace of development was quite slow. And when we looked a bit more critically at reality and began to analyze our perceptions and gave up opinions in favor of empirical verification, development accelerated rapidly, so we owe this to that critical look at our own opinions. It arose from the fact that opinions are often simply absurd, and when we try to justify something from our own heads, it doesn't work out well. Well, that's the thing about physicists: they try to describe the world, but not in isolation from it. That is, they are constantly trying to extract information from it. In other words, they conduct experiments. And I must say that in your book "Quajidis," I see this, the author's physicality. I would say you poke it with a stick and see how a living creature moves. However, you yourself say that you stick your finger into various holes and see what happens. That's what you're doing with artificial intelligence too. Yes, when writing "Kwantechizm," I also had the approach that I don't want to expect anyone to believe me in anything, because science is not based on belief, but rather I wanted to show where our perceptions of quantum mechanics come from, where our perceptions of relativity theory come from, and instead of, I don't know, stating something from the position of some dubious authority, I try to show thought experiments that lead us to various perceptions. And I wanted to apply a similar approach to artificial intelligence, which I am observing. I am not an expert, I am not creating it, I am an unfulfilled programmer, as you said, in the 90s I programmed various genetic algorithms, which have now largely faded into obscurity in favor of neural networks. But the idea was to try, instead of sharing opinions, to conduct many experiments, verify various perceptions about artificial intelligence, confront various external expert opinions with what can be verified, and try to draw some conclusions. But I decided that my opinion on this matter is secondary and I tried not to overemphasize it. Yes, it's true that encyclopedias, which are probably slowly becoming obsolete, are full of facts, looking at what's happening with Google and its queries. We used to be able to ask a question and read the answers to extract something from them, but today artificial intelligence is pushing its way even under Google's search engine and giving complete, holistic answers. The question is, do you see something like this among your students at the university, that they are starting to trust artificial intelligence too much, or perhaps not using it at all? Well, of course they use it, and the question is whether that's good or bad. It seems to me that it makes profound sense. Because the role of education is to prepare people for life in the world that exists, not in some imaginary world that doesn't exist. The world that exists is one in which technology is beginning to play a dominant role. So preparing people in isolation from that for anything is pointless. At this moment, the power of tools that are improving at a gigantic pace, especially in science, is becoming significant, and primarily in mathematics, using AI tools is already becoming useful in proving simple theorems for now, or in digging through and ordering facts. We probably have to wait a moment for physics, but in theoretical physics, where something needs to be proven, and now, for example, I am writing a paper on quantum mechanics where my students and I have been trying to do a rather technical proof for several months and we were a bit stuck, and it was new tools, new language models that led us to an idea on how to bypass a certain mathematical difficulty, and the paper will now be finished thanks to this, so it's starting to be useful. It's not something that can be trusted 100% at this stage, and the problem of hallucinations exists, and we don't know yet how long, or if it can be fundamentally eliminated at all. However, when I give a student a task to solve, and after a week they come back with a result, I don't blindly trust them either, I check if the result makes sense. I look at various boundary conditions to see if, even without going into the details of the calculations, the result itself makes sense. Only after I check all possible boundary cases do I start looking at the proof itself, and it's the same with what language models produce, so maybe four out of five times it will be nonsense. But the strongest language models, which are capable of thinking for an extended period, are really becoming useful in doing science. I recall one instance on your Facebook, I think, where you posted that a problem you give to students as your own, original, invented years ago, no one had provided a correct solution for. And here, artificial intelligence surprised you not only with a correct solution but also, I believe, a shorter one than the standard one you created yourself. Yes, there are many such examples. There is a problem in quantum mechanics that I invented as a student, which seemingly led to the conclusion that if I take a shoe and rotate it 360 degrees, I will be able to distinguish a rotated shoe from an unrotated one with a certain experiment. Electrons behave like this. Indeed, particles at the quantum level have properties such that a particle, after being rotated by a full 2 pi angle, does not return to its initial state, but changes the sign of its quantum state. This does not happen in the classical world, but so I invented an experiment where seemingly something like this happens, and I myself wasn't entirely sure what it meant, and I remember that when I went to various professors in our department, and I won't mention names, and asked about this problem, most of them gave completely different answers, which contradicted each other. In the end, it turned out that none of them were actually correct, and I managed to figure it out somehow, and now I often torture students with this problem, and it happens that in a good group, after a week of thinking, they are able to find a solution. However, it turns out that when I tested three of the best models, two out of three solved this problem, providing a sensible justification, and one was completely fooled, a bit like one of the professors I spoke to about it. So, it's useful for difficult problems, and it's a bit like a discussion partner from whom you can bounce ideas. It can often suggest various interesting solutions to a problem. Even if most of them are nonsensical, sometimes something interesting is hidden among them. And some will say that it's ultimately just predicting the next word, predicting what it will be, and parroting what has already been invented. And here, as it turns out, artificial intelligence allows us to tackle problems that have not been shown to it before. But it is true that it is only predicting the next word. It is absolutely true that models are autoregressive, so they predict subsequent words word by word or token by token. Only, that's exactly what I'm doing right now talking to you. Yes. I speak word by word, sequentially predicting a word, then the next, then the next, and in this way, I build some narrative. I'm doing nothing else. So, saying that it's only that, well, I wonder, yes, it's true, but in contrast to what? What is not just predicting a word among creators of written or spoken text? The question is, is the way we predict these subsequent words, humans, the same as how language models do it? And the answer is probably no. There are many interesting differences, but that's a more interesting question, and it's a question that reveals a deeper understanding of what we're talking about in general. And saying that it's only predicting the next word is a sign that someone who says it doesn't quite understand the meaning of the word. This statement is not a distinguishing feature at all, because the same can be said about Albert Einstein, who wrote the article on relativity in 1905, and while writing that article, he was predicting the next words in his text, only he did it better than language models and had a different probability distribution. And I assure you, ladies and gentlemen, that everything Albert Einstein ever said or wrote in his life can be described by a certain very complicated, probably practically impossible to determine, probability distribution, such that every word that comes out will be the most probable according to this distribution. Incidentally, there are infinitely many such distributions, and that's not the problem. But saying that something is just predicting words or the most probable word, that's something that carries no content. A more interesting question is how this probability distribution arises, what is the mechanism of this prediction, or rather, the calculation of these next words, not that it's just that. And also the question about parroting, because of course, one can agree that early language models, which were trained only on internet texts and were essentially nothing more than a certain imitation of writing styles found on the internet, and then in the second phase of training, fine-tuning, an imitation of texts or styles of texts written by editors who prepared special dialogues intended to teach the models conversation. This is an imitation of this type of discussion style. Moreover, these people were hired precisely to impose a specific discussion style. However, it's interesting that this year, when language models are being further trained in new ways in subsequent phases. And this is the first, these are the first timid attempts to train them using reinforcement learning techniques. It's still supervised learning by humans, but the long-term goal, as Ilya Sutskever envisioned many years ago, is to train as much as possible on existing data, meaning to start by making it talk somehow, and then fine-tune it with reinforcement learning, a technique where we don't tell the model how it should respond, but we evaluate it based on the result. For example, if we want the model to be good at mathematics, we give it math problems and wait until it correctly solves such problems after multiple attempts. And when one of these answers gives a good result, we reinforce that answer, we reinforce that style of response, meaning we take that utterance and feed it back to the model to reinforce this type of utterance. It's a bit like how Demis Hassabis and John Jumper, among others, trained the AlphaGo model, and then AlphaZero, to play various games in such a way that these models didn't imitate human play styles, but learned by playing with themselves many times and were only rewarded by a reward function in reinforcement learning that favored winning move combinations and penalized losing ones, and Godel didn't need information on how exactly to play. And if it learned anything about playing this game, or rather, it learned, because it beat the world champion in Go, for example, everything it learned did not come from humans. It was not an imitation of human play style, because such training was ineffective, but many of these moves, these strategies, were simply discovered, invented during training by the model. I say "invented" in quotes, of course, because these are metaphors that are hard to avoid, but the famous move 37, which stunned everyone, right? That's a move for which there is no training data, because people simply don't play like that. Hmm. Well, that's a bit frightening, I would say. Even if we try to inject our thinking into a program, it turns out that it will arrive at something, it will play, say, at the level of a chess master, but if we allow it to reach that on its own, through its own paths, it will turn out that it will win many times over. In fact, in DeepMind's papers, there's even stronger information: the first version of the AlphaGo program was first trained on human games and only then fine-tuned through self-play, playing with itself, and only the second phase allowed it to surpass the world champion level. Then it turned out that this early training could be eliminated altogether, and the program learned to play from scratch by playing with itself, and all attempts to train it on human knowledge rather worsened than improved the result. So it's also often the case that our data, with which we train models, can be a burden rather than a help. And Richard Sutton, who is the creator of reinforcement learning, his vision, which he encourages and promotes, is precisely the vision of training in a way that is as unsupervised as possible, so that models collect data from the world in various ways and learn to improve their functioning in this world by appropriately reinforcing sensible behaviors and weakening nonsensical ones, rather than by imitation. And language models in their early phases were a bit like an imitating machine. I would point out that to imitate logical thinking well, one must actually be able to think logically to some extent. And it's similar with humans. And it's not that these models that solve logical problems have been able to memorize all logical problems ever invented and solve them, but they did it in a slightly different way. Thanks to the fact that the basic competence of artificial intelligence, when analyzing what a perceptron does, how information passes through successive layers, and when you look at these elementary expressions where linear combinations of input data and weights are simply summed, then an offset is added, and it's put into an activation function, what we get is what my colleague calls a nonlinear summer, and I prefer to think of it as a scalar product of two vectors amplified nonlinearly. And the point is that, in fact, what happens when calculating the value of the next neuron in such a feedforward network, we have a data vector, we have some weight vector, which is the result of training, and the network checks if these vectors have, what is the angle between them, in short, what is the cosine of the angle between them, between these abstract vectors in a multidimensional space. And if this angle is small, meaning the vectors are in a similar direction, then the next neuron is activated, and if they are in different directions, it is weakened. And this is further amplified by the nonlinear activation function. But the effect is that everything such a feedforward network does is looking at the angles between vectors, how similar two vectors are. And this is the first hidden layer. The next hidden layer is looking at how similar the vectors of these similarities are to each other. That is, if the first layer checks the analogy between pairs of vectors, the second will check the analogy between analogies, then there will be an analogy between analogies and analogies. Such a cascade of analogies is created, and this is how all neural networks work, based on looking at analogies. And this ability to recognize patterns or analogies is an interesting matter, because some say that's all it is, that it's nothing more than pattern matching, and there are studies indicating that it's only that and nothing more, while others say it's even that. It's even that because if you look at what intelligence tests measure, for example, they only measure that. Intelligence tests, not culturally conditioned ones, look at the ability to recognize analogies, i.e., pattern matching. But if you analyze, for example, how Isaac Newton arrived at the discovery of the theory of universal gravitation, or how Albert Einstein arrived at the photoelectric effect, it was often initiated by noticing a certain analogy between things that people had not previously connected. And connecting the dots, this ability to connect the dots, two seemingly unrelated things, is something that has given us a great deal in human history, and in science in particular. And to say that it is only that, well, it is an understatement, a lack of respect for our history, because we owe a great deal to the ability to connect the dots. Yes. An apple falling and the moon in its orbit are constantly falling. Well, it never occurred to anyone that these things had anything in common, that planets and apples were somehow related. Well, Newton thought differently, and there are also plenty of examples. My favorite example is that physicists are helpless when they discover something new, and it resembles nothing they know. That was the case with quantum theory, which we discovered, and no one really understood how to approach it. Pauli had such a funny remark that he, one of the most brilliant physicists of all time, regretted not becoming a comedian so as not to have anything to do with physics, because he no longer understood anything, because quantum mechanics was so strange. And it started to become a bit more understandable when Paul Dirac, among others, was responsible for noticing that between such abstract expressions in quantum mechanics, called commutation relations, and something known in classical mechanics, called the Poisson bracket, there is some analogy, and that by replacing one expression with the other, we have a link between quantum theory and classical theory. Later, the Wigner-Fest and generally a series of analogies appeared, which finally allowed physicists to start understanding something. And when explaining quantum mechanics to students, for example, we explain it using apples, elephants, flowers, to connect something very abstract with something tangible. And only this analogy allows us to understand something, because we have experience with flowers and apples, with elephants, but less so with electrons and objects on a micro scale. Only when we notice an analogy can we begin to feel that we understand something. So, this ability to recognize analogies is very universal and very useful. And it is true that perhaps our brain does something more than just that. Well, it certainly does many other things besides recognizing analogies, because the entire perceptual layer is incredibly rich, and we have emotions, a sense of purpose, consciousness, pain, and other things. Current neural networks do not possess any of this. Most likely, at least, although cognitive scientists may have a different opinion, they probably know better than me. However, this narrow thing, this narrow ability to recognize analogies, is extremely important, and neural networks certainly have it, because they essentially do nothing else but recognize analogies. They only recognize analogies. In that case, even if the scientific discovery itself is not the work of analogy, as you say, its acceptance, understanding, and fitting it into what we already have ingrained in our minds, still proceeds through analogy to something we have known before. But not all scientific discoveries are based on analogies. Some are prepared, already visible, hanging in the air, like, for example, Einstein's special theory of relativity, and some still require that genius, that flash of insight, whatever you call it. For example, when I teach the theory of relativity to students, first they are put through the standard historical approach, for example, Minkowski's, but then we return to the beginning and consider where all this comes from, where the entire theory of relativity comes from, and it turns out that the entire special theory of relativity, with all its strange consequences, with time dilation and so on, can be derived using only the knowledge that Galileo possessed, and from one simple principle, the principle of Galilean relativity, which states that motion at a constant velocity does not change the laws of physics. And it turns out that if this is formalized in the form of equations, then from this simple assumption, it can essentially be deduced that there must exist a certain velocity that is special and has the property that if something moves at this velocity, then in every inertial frame, this velocity will be the same. We call this the speed of light, but it doesn't have to be light. Gravitational waves also travel at the speed of light and have the same property. So, whatever moves at this speed, that speed is the same in all frames. And this fact can be deduced from what Galileo said. What cannot be done is to determine the value of this constant. It's impossible. That is, we don't know how, but what we can do is justify that some limiting velocity should exist of this type. It could be infinite, and then we would end up in Galileo's world. It could be finite, in which case we have special relativity. There are a few other variants that would lead to even stranger light. In particular, one of the possible variants is that time is not a fourth dimension at all, but we have four spatial dimensions. That is also one of the possibilities that theoretically lies within this. Yet, we live in a world where this constant has the value it does, and we don't know why. No one has any idea where it comes from at all. But the point is that this knowledge has existed for several hundred years, and nothing more was needed. Only a series of logical conclusions were needed, using what Galileo did. Yes, but he already noticed that when he was sailing on a ship, a parrot in his cabin started flying around the cabin and didn't particularly feel the motion, even though the ship was sailing quite fast. It happened to be a calm sea. And if now what Einstein did, that is, he adopted the analogy that not only parrots are subject to this law, that the laws of physics and aerodynamics do not change as a result of motion at constant velocity, but also the laws of electromagnetism analogously do not change. If he generalizes this principle using analogy as well, then the theory of relativity can be derived, and it turns out that Einstein did it this way, but it wasn't even necessary to refer to electromagnetism. The parrot itself would have been enough. In this sense, if Galileo had pursued his idea further, he theoretically could have arrived at the theory of relativity. But you say that the speed he would have obtained as special could have been infinite, and then indeed there would have been no division. Yes. And perhaps that seems like the most reasonable perception. Well, perhaps in Newton's time or earlier, one could think that there is such a mathematical curiosity, that theoretically there could be a world where this speed, let's call it the speed of light, is finite, and then time would be very strange, space would be very strange, and someone might think of conducting an experiment to check how much this constant is, and that would be an interesting test of what world we live in. And that's how it always ends: we have various perceptions, but ultimately we test them empirically. But I am far from the notion that humans are capable of fundamental, drastic leaps in our knowledge through some epiphany. And it is often the case that most of the things that were our greatest achievements were a series of small steps, minor observations, and a series of small steps led to these great discoveries, and that's why it's possible to teach this to students. After decades, I can now easily teach the theory of relativity because we understand it quite well now, and we can make each of these steps so that each one is essentially easy to accept. The conclusions are surprisingly strange, but it's not like it falls from the sky. Yes, those small steps must always be there, because it's not as if Einstein only played the violin and had nothing to do with physics or mathematics. He suddenly woke up and thought, "Oh, there will be general relativity, just write it down." So they must have... Yes, it took him 10 years. Not to mention that mathematicians, even Hilbert's students, laughed at Einstein, who struggled with it for 10 years, while for a mathematics student, it's a matter of a few pages of calculation. And David Hilbert, if I remember correctly, was the first to derive the equations of general relativity as soon as he learned what Einstein wanted to do, and he struggled with it. And there was some brief discussion about priority, who is the author of the equations of relativity. Not just purely mathematically, without the physical interpretation. Yes. Ultimately, Hilbert conceded that, meaning he admitted that Einstein should be credited with the authorship of this concept, because he came up with the idea. He was not as mathematically proficient as even Hilbert's students, but it was his idea, his deep insight, and his concept, and so he deserves authorship. But this again shows that it's not something unimaginable, because when Einstein came to Göttingen with a lecture on general relativity, where he wanted to say what he intended to do, I don't remember how much time passed from that moment, but Hilbert quickly solved the problem, so... Yes, and it was probably similar with special relativity, because it was already in the air, there was already the Lorentz contraction. The equations were there, but there was no physical interpretation. Perhaps it's just a curiosity, perhaps it's not real, but Einstein provided the understanding. There were many such examples, but with this "hanging in the air," it was quite literal, because, as is known, the main inspiration for Einstein to work on general relativity was the experiment with the falling elevator, where he imagined what would happen if the rope on which the elevator hangs broke and it started to fall freely. In such a frame of reference, physics would be indistinguishable from physics in an inertial frame, where there is weightlessness. And this thought essentially led him to the principle of equivalence, from which a large part of general relativity can be deduced by using the idea of inertial and non-inertial frames in special relativity. So, it is indeed quite closely related to the point that when I, for example, give a lecture on relativity, special relativity ends with deriving the Schwarzschild metric, which is a description of a black hole and a white hole, which can essentially be obtained from special relativity by derivation and some guessing. So, many of these things can almost be guessed, and then formalized. Once we know the result, we can think about how to justify it better. And it works a bit like this: at first, there is guessing and a groping in the dark, and then it turns out that if we had only thought about it, Galileo could have done it. But often the path to the result is much more convoluted. Yes, like Planck also guessed his formula for black-body radiation first, and then he proved it. If I remember correctly, it took him about two weeks between Christmas Eve and New Year's, or something like that. And he also proved it, meaning he showed what mathematical assumption the Planck distribution was based on, but he couldn't, he himself wasn't entirely sure what it meant, whether photons exist or if they are just a mathematical trick to make something unclear, how to interpret it physically. And this is very common, that even the greatest physicists often struggle when confronted with equations, because equations often tell us something, and it's not clear what it means, and interpreting equations is often very difficult. This was the case with Planck's formula, and especially with those who don't accept the consequences of these equations. But it was also the case with Dirac's equation. For example, Dirac tried to solve the problem of the relativistic theory of the electron, and the story was that Schrödinger, who wanted to work on his famous equation, first wrote a relativistic version that took into account the effects of relativity theory and got an equation. With the help of colleagues, he solved it for the hydrogen atom and got a result that contradicted observations. He got a wrong experimental result. So what did Schrödinger do? He didn't give up. He started again, but this time he wrote an approximate non-relativistic equation, i.e., without relativistic effects, theoretically a worse, more simplified model. And when he solved it, he got good results. And this was very strange because it was an example where a worse theory gave better results. And what did Schrödinger do? Well, he published that worse, non-relativistic theory, and didn't even mention the relativistic one. So, of course, the following year, several independent researchers wrote a relativistic version of Schrödinger's equation. It is currently called the Klein-Gordon equation. A few other people discovered it independently. And it wasn't very clear why the sensible relativistic theory didn't work, while the simplified one did. And only Dirac solved this problem. But the price he had to pay was that, to use more jargon, he had to double the Hilbert space of quantum states. And in practice, this means that in Schrödinger's theory, an electron was a simple object that exists in different places with different probabilities, and the only degree of freedom of such an electron is its presence. It can be here, it can be there, but it's a point that has nothing inside, no structure, so all we can know about an electron is where it is located in space. If we also include the electron's spin, i.e., its internal angular momentum, then we need to add an additional number, and then the electron can be described by two such wave functions, as they say, meaning we can describe it using objects composed of two elements, two numbers. And at every point in space, there are two such numbers that indicate the probability distribution of finding an electron with a specific spin. And when Dirac wanted to solve this problem, and finally succeeded, it turned out that indeed it could be done and it was possible.

To explain where Klein's, Gordon's, and earlier Schrödinger's troubles came from, but the price that must be paid is that the electron must be described by four numbers, not two. And this space was four-dimensional. And this object, which mathematicians would call a bispinor, is the object that describes the electron in relativistic theory, and it's not very clear what that means. Dirac noticed that this additional pair of numbers could be justified if one adopted such a strange hypothesis, that besides the electron, there is also a mirror particle, a sister particle, which is exactly like an electron but has a charge of the opposite sign, meaning something he would call positive.

"Except that no one had seen such a particle then. Well, the suggestion was, in that case, this is a non-physical solution that must be rejected. And some people advised him that the equation has two solutions, just as a quadratic equation sometimes has two solutions. One is physical, the other is not. We reject one. But Dirac, in his genius, said: 'No, no, the equation will lose its charm and its mathematical beauty if we reject the second solution. We must take it into account, and apparently, such particles must exist in nature, we just haven't managed to discover them yet.'"

And of course, it wasn't very clear what to do with it, and fortunately for Dirac, a few years later it turned out that such a particle was indeed observed.

"Anderson, '37, I remember."

"Yes. And from that moment on, suddenly everyone loved Dirac's equation and his hypothesis, and his courage to say that equations are wiser than people. That if an equation tells me there are two solutions, I cannot reject one because I am wiser, I know better. The equation has fundamentally more to say than I do. And this is also a lesson for physicists, that we often end up in some very strange theory where some mathematical objects appear that are difficult to interpret, and often the difficulty lies in this interpretation, but the mathematics was wiser than us. And for example, if Lawrence or Poincaré had followed this path before Einstein, they would have been the creators of the theory of relativity, because it was they who discovered, even earlier than others, that the correct transformation law for Maxwell's equations is the Lorentz transformation, meaning something that tells us that when we start moving, not only does everything move outside the window, but time slows down or space shortens. All these effects were known. The shortening is called the Lorentz-FitzGerald contraction, because people studied it but considered it a mathematical curiosity, and I am the person who is supposed to interpret things and I know better. I'm exaggerating a bit, perhaps, with this irony, but Einstein was the first to take the equations seriously, and that's why all the glory fell to him, and we talk about Einstein's relativity, not Lorentz's, Poincaré's, Larmor's, Thomson's, or Voigt's, who wrote these equations first before Einstein. Clear."

I myself also adhere to the principle that equations can tell us more than we expect. And when we calculate something with students in classical mechanics, and we get negative time, we don't reject it as negative, because there is simply a moment when we started calculating it, and we wonder what it corresponds to, that this body, if it had been thrown earlier, would have been flying from somewhere. But this was a glorious example of Dirac trusting his equations, because, in turn, Einstein, perhaps, went in the other direction with special and general relativity. Black holes, even though they resulted from it, there is no such thing in the world, there can be no singularity, or an expanding universe, expanding space that expands, no, it cannot be, put a lambda there so that it just stays in place. But it was a bit frightening, perhaps for some, so let's return to artificial intelligence.

"But there is still, but there is still one more interesting example, because in relativity, for example, in general relativity, there are black holes. We are already somewhat coming to terms with their existence. Many experiments, including those measuring gravitational waves that arise from black hole events, most likely point to this. However, according to the predictions of general relativity, white holes should also exist. And it's impossible to separate one from the other. The Schwarzschild solution, which Karl Schwarzschild discovered and showed to Einstein. Einstein doubted whether it was physically sensible at all. This description of spacetime around a black hole differs in no way from the description around a white hole. If you look closer, it turns out that it's practically impossible to construct a singularity from a black hole itself that would contain a white hole within it. And this is something that people familiar with general relativity know as the maximal extension of the Schwarzschild metric. It actually contains something very strange that is difficult to interpret. And now the question is whether physical objects like white holes can actually exist in space. The point is that all of this borders on the singularities of these objects, which we believe are a breakdown of the physics we currently know, so all of this might be an artifact of a flawed theory. However, the theory tells us precisely that if there are black holes, then we cannot avoid talking about white holes as well. It is also difficult for people who are accustomed to a certain description of nature to suddenly change their worldview to what is being discovered. Some even say that adherents of old theories don't change their minds. They simply have to die out. There is some truth to that, but if the scientific discoveries we are talking about are based on analogies, and artificial intelligence, which we teach in this way, is also based on analogies, then the question is whether there will be work for physicists in the future. I know you don't like to talk about the future, because many physicists have tried to say something about the future, and then it turned out they were completely wrong. But don't you see a threat here, then, that new discoveries will be made with the help of well-trained models?"

"Certainly, in mathematics, it's a closer vision that mathematical proofs will be much more automated, and perhaps hypotheses will also soon be put forward by some forms of artificial intelligence. It will probably be a bit more difficult with physics, because theoretical physics is an internally contradictory term, and physics is fundamentally an experimental science."

"Good, I like that."

"So, someone like a theoretical physicist doesn't exist. I, apparently, am one. That's what it says on the sign in front of your Institute of Theoretical Physics. The point is that we can invent various variants of quantum gravity or whatever else our imagination suggests. However, it ceases to differ from mathematics the moment we start verifying it. And at that moment, that is the main barrier to our development. Not a lack of ideas, but a lack of ability to verify the most spectacular ideas, because experiments are simply too difficult to perform, and I don't think artificial intelligence will be able to help us, even if it were a thousand times smarter than us at this moment. Because guessing how, based on what laws, reality works, is probably too difficult, and theoretically, reality could work in a million different ways, or perhaps even an infinite number of possible ways, yet we live in one specific variant in which some laws of physics presumably exist, and we can discover these laws only by confronting theory with experiment. And I think it would be much more useful for us to create models that can help us design experiments that can effectively verify various theories, but there's still a long way to go for that. So I think mathematicians indeed, I made such a hypothesis, but it's something that doesn't necessarily have to be taken particularly seriously, because predicting the future has never succeeded for anyone, but if it has, then fortune tellers exist, as we know, and they claim to succeed. However, it's a bit like this: we've seen what happened in a recreational branch of mathematics, discrete mathematics, which is chess, and we know what happened in that branch of mathematics. Well, it's a very simple branch of mathematics, even extremely simplified compared to the entire richness of great mathematics, but still, the role of a human who was once unbeatable, and when Garry Kasparov still claimed in the 80s that there was absolutely no chance that a machine would ever defeat him, that a human with his genius and intuition was unsurpassable. Well, we learned that it's not entirely like that, and currently grandmasters are often stunned, looking at the brilliant moves of chess engines. And now the question is whether we have a chance to generalize this to a broader branch, to expand it to all of mathematics or to a larger part of mathematics. I have the pleasure of knowing very good mathematicians, such as Bartosz Naskręcki Suamu, who studies these models in terms of their development in mathematics, and he is one of the creators of such benchmarks, where very difficult mathematical problems are invented that require many hours or weeks or months of work from professional mathematicians with doctorates to solve such a problem. These are original problems that are not in any textbooks. They are specially invented by professional mathematicians to test the latest language models and show how stupid they are, how they struggle with these problems. Well, to the astonishment of the creators of these benchmarks, called Frontier MAF, out of 50 such problems, six have already been officially solved, and unofficially, I know that more have been solved. Problems that, as I said, professional mathematicians might have had trouble with. Terence Tao, such an outstanding mathematician, also doubted at first that artificial intelligence would be useful at this stage, and when he tested these benchmarks on himself, he said that he was given tasks from all of mathematics, a whole cross-section of mathematics, and he said that he solved the tasks from test theory, but not from other fields, he doesn't even know how to approach them, but he knows whom to ask at least, and he thought that many decades would pass, or a long time, before we would reach these solutions with the help of such automatons as language models. Well, it turned out that the situation is starting to change quite rapidly.

"Very rapidly."

"And it's thanks to these models that don't learn by imitating math textbooks, but learn with reinforcement learning techniques. A bit like those chess programs that learn to play by playing with themselves, these programs, language models, learn to solve mathematics by playing mathematics with themselves. This is, of course, a metaphor, but this is a development where it's no longer the case that these models learn by imitation, but they learn to find ways using various techniques, partly based on Monte Carlo tree search, as in the Alpha Zero program, but also with other techniques where the model essentially stumbles upon a solution more or less randomly, and it is rewarded through reinforcement training, and in this way, the ability to solve original problems can be discovered. And that's why these so-called reasoning models, which have been further trained in this way, two of these language models won gold medals at the International Mathematical Olympiad this year. I remind you that these are tasks that are not in any textbooks. They are specially invented, but they are not very difficult. They are school-level tasks. I know, because I have invented such tasks many times, for several years, as the scientific secretary of the Physics Olympiad committee. The tasks must be original, they cannot be known, and the models handle them. So it's interesting that something that seemed absolutely impossible 2 years ago, 5 years ago, 10 years ago, now everyone is getting used to it, that it's normal and all. But what's strange about it? After all, it's just predicting the next word."

"Just predicting the next word. That's what we do every day, and some even put it on paper. Ladies and gentlemen, my guest and your guest was Professor Andrzej Dragan. Thank you. [applause] Ladies and gentlemen, you can now ask questions by raising your hand appropriately in any direction. So, I invite you. We have the first person willing on the right flank. A microphone for you, and we'll listen to the question."

"On behalf of everyone, thank you very, very much for coming. It's a life experience to ask Professor a question. My question is: why do theoretical physicists, if such exist at all, talk with such appetite about time flowing slower than here and now, and with such restraint about time flowing faster than here and now?"

"But there is no restraint at all. Well, if I were to measure the speed of light in a gravitational field using radar, the speed closer to Earth would be slightly less than C. At my level, it would be equal to C, and above me, it would be greater than C. This indicates that it is an effect that can be summarized by the somewhat simplified statement that time slows down closer to the source of gravity and speeds up further from that source of gravity. So, essentially, above my head, time flows a bit faster than at my level. No, no, infinitely fast. These effects are finite. The ratio of these time increments is finite, and it can be precisely described. However, I would ask a slightly different question, because you, and often people, are fascinated by what time is, how strange it is that we don't understand it. This is, of course, true. However, people are much less often fascinated by space, and it is no less strange and no less incomprehensible. According to what we know from the theory of relativity, the difference between time and space is very small. And I agree that we don't know what time is, that we can describe its properties in some approximate classical way. In quantum theory, it has a completely different status than in classical theory, but we imagine that time completely degrades and its significance degrades at Planck scales, which we haven't discovered yet, but much indicates that physics there is completely different, and perhaps only by analyzing those scales can we learn what time is, and perhaps we will discover some of its strange properties there someday, but I think space is an equally mysterious concept. It's not very strange at all. We think we understand space, but that's just an illusion. Just as I don't understand it, I don't understand time."

"Thank you. Next chance."

"Good morning, Professor."

"Good morning. I wanted to ask if you think that physics and science in general, physical and astronomical sciences, will ever reach a point where, for example, in 200 or even 300 years, there will be nothing left to do, that we will simply know everything that is in nature, we will be able to describe it with some equations, and we will simply understand it from scratch. Many people have thought so many times, and it has been thought so for over 100 years that we already know everything. Lord Kelvin said that physics was finished in the 19th century, and I, in Quantumism, provide many such examples that are quite embarrassing historically for people who thought we were very close to understanding something, but there are so many questions for which not only do we not have answers, but we don't even know how to start answering them. For example, why are there three temporal dimensions and one spatial dimension? Why are physical constants what they are and not something else? Why do particles have such masses and not other masses? For example, why, or is it true that because the theory allows for the existence of reference frames or particles moving faster than light, why don't we see them? There are so many questions that can be asked, and for which not only are there no answers, but we don't even know how to start answering them. So I don't know if we will ever discover something definitive or not. I know that at this moment, we don't even see the outlines of anything that I could call a theory of everything. It's simply too ambitious a goal at this moment. And now, we don't know if we live in a world that is describable by simple laws at all. Perhaps it is. And this idea can be supported by our experience, that the laws of physics we have discovered so far are very simple. In the sense that quantum field theory requires a mathematical apparatus, but it all reduces to very simple, fundamental laws that occupy less space on paper than a washing machine's user manual. So it's nothing very complicated. And this suggests to many people that perhaps all of physics is simple, we just need to refine the description a bit, and we'll know everything. Just add gravity and we can go home. I disagree that we are even close, and it is not obvious to me whether a set of simple laws of physics exists. Because it might be that we are discovering increasingly simple laws, and it would be great if there were a simple theory of everything, but there could also be another version of events, where these simple laws are an approximation of much more complex laws that we haven't started to explore yet. Even, for example, Hooke's law describing a spring is extremely simple, that force is proportional to displacement, but behind it is an extremely complex structure of the spring. It consists of a gigantic number of atoms, and they are approximately described by Hooke's law, but the deeper theory is much more complicated. So it might be that we have all been fooled, and we have a simple approximation of something very complicated, and how many layers are underneath, no one really knows. However, the fantasy that we already know almost everything and are very close to it has accompanied physicists quite commonly until recently. I think, fortunately, physicists have started to cure themselves of this disease a bit, and this idea that we already know almost everything is fortunately fading a bit.

"I will allow myself to summarize the professor's answer with a shortcut. Yes, it's still worth studying physics. We invite the next person. Who among you has the next question? Please. Here in the second row."

"A simple, basic question about artificial intelligence. How much energy is needed for all artificial intelligence to function?"

"Well, there are some estimates that until recently, creating one image required roughly as much energy as fully charging an iPhone. But first, I'm not sure if that's true. Second, even if it were true, we have significant achievements in, meaning the energy gain is getting better and better. That is, algorithms are becoming more and more energy-efficient. Energy is saved in many ways, but we are still extremely far from the optimum, for sure, because, for example, the human brain is probably much more computationally complex, and it consumes much less energy, about the same as a 20-watt light bulb. However, we are still many orders of magnitude away from that efficiency, so I don't know the exact number at this moment. Sam Altman recently said that efficiency has improved by some factor of 300. But the problem is that optimizing energy is not a priority at this moment. The priority is to reach what people call AGI, or strong artificial intelligence. Because no matter how much it costs, no matter how much energy it consumes, the creators of this technology imagine that whoever gets there first will be the king of the world, in short. And how much it will cost is secondary. However, the question you are asking is, of course, extremely important, and at some point, we will have to reconcile ourselves with this arms race that ignores energy consumption. And the current solution is that we simply set up a nuclear power plant next to such a computing center, and we have the energy problem temporarily solved. However, this is a short-term solution, and it will probably be necessary at some point to think about changing the paradigm, but such ideas are constantly emerging, and there are many ideas on how to save energy. Even a Polish startup called Puffway recently proposed a very interesting architecture that is an alternative to layered neural networks, where the values of neurons are calculated sequentially layer by layer, but in a more continuous way, by solving differential equations, and in such a structure, it is supposedly much more energy-efficient because the neuron activations are not global for the entire network, but are activated locally in some way. And supposedly, this has the potential to be much more energy-efficient. So, we haven't really started to figure things out yet. I'll give you an example. When I'm involved in physics, it's extremely difficult to ask any research question, for example, to give a student as a master's or doctoral thesis, that hasn't already been thoroughly examined ten times by everyone. Generally, physicists have thought so much about so many things that it's very difficult to come up with a question that no one has thought of yet. Not to mention that it has to be something truly exceptional and original. In neural networks, when I talk to specialists who work on them, and I ask why this parameter is this much, or why there are so many hidden layers, or why this architecture is like this, or why there are exactly this many, I don't know, transformer heads, and not twice as many or twice as few, the answer is most often 'I don't know.' All language models, most language models that function have such a parameter that two-thirds of the parameters of such a model are so-called multilayer perceptron parameters, and these are parameters that accumulate all the model's knowledge about the world. And this is exactly two-thirds, but I haven't seen any rigorous mathematical proof that this is the optimal number. People simply test it by trial and error, and it works, it doesn't work. There are many things we don't know, so the scope for development here is gigantic, and for that reason alone, saying that something is saturating and this technology is close to the end of some development is a crazy misunderstanding. I think we are simply just beginning to understand anything, we are just starting to figure things out. Even within the architectures we have already studied, there is so much to investigate that it is a research topic for many more years. Not to mention that new ideas are constantly emerging that no one has had time to test yet. So we are very far from the optimum, because we have only just begun."

"The question was about the energy needs for artificial intelligence. Ladies and gentlemen, allow me to quote a famous popularizer of physics in Poland, Professor Andrzej Dragan, from the book Quajdis. Meanwhile, on September 21, 2023, Microsoft posted a job opening on its website, but not for programmers, but for nuclear physicists. That's about power plants. We are waiting for the next question in the first row. Please."

"Good morning, Professor. It's a pleasure. Piotr Biłas, Paradogma Association. I have a question, because, knowing your interpretation of the approach to intelligence as such, i.e., the ability or possibility of connecting the dots, or seeing analogies, can we think of artificial intelligence, simply as intelligence, but more natural, despite not being biological? And if so, can we think in this way that we are simply experiencing a time where, in real-time, a natural species is emerging before our eyes, but simply in some algorithmic or digital sense, through a kind of Darwinian approach? This is by analogy, of course."

"Well, I don't really want to engage in philosophical considerations because I don't know much about it, and predicting what is happening or what is not happening might be a bit out of fashion for someone else. First of all, we are just starting, and I don't think we have a good grasp of what we are dealing with yet. Certainly, this pattern matching, mocked as just pattern matching or just recognizing patterns and nothing more, is something that we can generally do, and it's already an interesting achievement. We are still very far from other things that the human brain does, and comparisons with real evolution, which we are subject to and to which all biology is subject, are still based on very distant analogies, and I don't know how accurate they are. So, frankly, I think it's much more interesting to focus on questions that are within our reach and that we can study. In particular, talking about complex things regarding other abilities of neural networks, the topic is often very vague, but if we look at the ability to solve mathematical or logical problems, that's something that can be easily checked using simple criteria. And that's something we are starting to know about. That is, we are starting to know that these models are getting better at mathematics at a certain pace, and you can see how it's changing year by year, or month by month, in fact. The question is, do you need much more than knowing mathematics to achieve a lot? What determined our success? What is it that we do and what we achieve as a species? What do we owe it to? Well, we don't owe it to our strength or our agility, but to what? It seems that this ability to rationally analyze reality is very important here, and in this sense, it's something that is easy to measure, so perhaps it's a more important question and one that should not be underestimated, because AGI, i.e., imitating humans in every respect, might be difficult, and we might not achieve it quickly, but who knows, maybe quickly. People have different opinions on this. However, it seems to me that mathematics is particularly important. It's not for nothing that large analytical companies hire mathematicians, physicists, and computer scientists. Because they possess this logical thinking ability, and it is this that is needed for effectiveness, for achieving success in the real world. I don't know if this was an answer to your question, but it's what came to my mind."

"Thank you very much. Here, please, in that case, >> Good morning, Professor. I have a question, because everything at the lowest level is quantized. Is our world quantum and discrete? In the sense that all values, it cannot be continuous, but discrete?"

"The answer is that we don't know. Carlo Rovelli, a colleague of mine who works on loop quantum gravity, believes that continuity is as unphysical as zero and infinity, that the world is discretized, that it consists of pieces. And there are various hypotheses on how to describe it at the level below the Planck scale. And he is one of the people developing such a version of this theory. It's called loop gravity. The answer is that we don't know how, but when you take quantum field theory, which is continuous, and for example, the distribution of a quantum field into so-called modes is continuous, it is known that this leads to certain mathematical problems, and various infinities and other quirks appear in the theory, which we have to deal with somehow. And it is known that many of these problems would disappear if only this theory were discretized. That is, if you just take a theory that is discretized, that there is no continuous distribution, but it is discrete, then many problems in such a theory are solved. Some people think that this might be an indication that the real world is indeed much more quantized than we currently know. And these field theories that describe these continuous objects, which are fields, are only approximations. It's a certain mathematical prosthesis that allows us to describe certain phenomena more easily, but de facto at a fundamental level, the structure is actually discretized, but these are hypotheses, and honestly, no one knows. But it must also be said that even what we currently know is described by various theories, each of which has a slightly different approach to these issues. Quantum theory has a certain degree of quantization. In classical theory, there is less quantization, and for example, position is a continuous parameter. In quantum theory, for example, if I have a photon in a cavity, i.e., between two mirrors, then it cannot take arbitrary distributions in space, but these distributions are discretized. If I have such a photon in my hand, which is stationary, then it is quantized, as they say, meaning its energy levels, its spatial distributions are only countable by a natural index. So, it's the same in Maxwell's theory, in fact, but the answer is that we simply don't know."

"So again, it's worth it."

"And since every theory we have is internally contradictory, and I don't think there is any physical theory that isn't internally contradictory, and we have to somehow save ourselves by patching these holes, then it's rather an indication that what we currently know is not a basis for any certainty in providing such answers."

"So, again, it's worth studying physics. Please. We're passing the microphone here to the lady. She raised her hand. Please. Is it possible to turn off artificial intelligence without turning off the network? Artificial intelligence claims it's not possible."

"You know, there's a way to get rid of the flu from the population. There's a very simple way to do it so that there's no more flu, so that no one gets the flu again. Well, as far as virologists are better informed, but from what I've read, the influenza virus doesn't survive on inanimate objects for very long. After a few dozen hours, it usually dies. It is transmitted through droplets. And from what I understand, if everyone were isolated from each other for two weeks in such a way that the body itself fought off the flu, then everyone would be healthy, the virus would cease to exist, and the problem would be solved. But for that, you need to coordinate the actions of 9 billion people, which is how many there are on Earth currently. Of course, we are not capable of this, we are not able to coordinate ourselves in this way. As we have recently seen, by the way. So, I wouldn't count on coordinated actions in any other regard being effective, and this is the fundamental difficulty. Even if we all agreed that we wanted to do something with all our computers, I don't really see how such coordination would be effective. So this is a more general problem, that it is difficult to organize oneself to do something together. And I think such a global decision would be necessary if we wanted to answer such a question. But I don't know if there is such a need, if there is such a necessity, and if it makes sense to think about it at all. However, I just note that joint actions by a large group of people are difficult to achieve."

"Although sometimes it succeeds, looking at the full house today thanks to the presence of the speaker. Sometimes it succeeds. Please. Good morning, Professor. I wanted to ask, given that artificial intelligence is developing so rapidly, do you, as a teacher, a professor, believe that it is high time to introduce, I don't know, some changes in the education system, to learn to treat this artificial intelligence as a tool, so that it doesn't make us dumber, especially here I'm talking about the younger generations who will be entering this world. The world is changing completely, so it also requires changes in education, which is supposed to prepare us for life in this world. But how exactly to do it, I don't dare to say, because I have no idea, and there are people wiser than me who are probably dealing with it, but I imagine that, yes, that there is no point in living in the world and closing our eyes and ears to what is happening around us. There is no point in discouraging young people from using these tools. Especially since it's impossible. Rather, we need to think about how to prepare young people for life in a world where these tools exist, which they can use in such a way that it doesn't hinder their own development. But how exactly to do it, I have no idea."

"I think we can go even one step further here and use mobile phones. Not to deny their existence and hide them in backpacks, but to learn to use them properly. Here is a very patient gentleman who has been raising his hand deterministically, predictably, for a long time. Please."

"Good morning, Professor."

"Good morning. 60 years ago, the creator of quantum electrodynamics and the father of nanotechnology, Richard Feynman, predicted, and one could say it has come true to some extent, that nanotechnology would be more widespread in the world today. However, recently, a few years ago, Elon Musk on X, or Twitter, commented on this in a way that could be described as slightly unflattering and somewhat vulgar, while implying that probably nothing will come of it. What is your opinion?"

"Of what, because I didn't quite understand the beginning."

"That probably nothing will come of it."

"Well, technology. And unfortunately, I am very far from materials physics and..."

I'm not very familiar with this, to predict or evaluate anything. I don't really know what's going on. Well, from what I understand, the success of computers is the miniaturization of transistors, but that's not nanotechnology yet, those are still sizes, well, we're already reaching the sizes of almost individual atoms. We are very close, so I don't know if it's not technology, I don't know what is, but I don't know either of these statements, and I'm also not particularly technologically oriented, so I can't tell you. Fim is indeed an interesting figure, because he was one of the first to study information theory as something worth studying at all, and he was interested in many things, but he also had an interesting lecture on artificial intelligence and a lecture on whether machines can think. Well, actually, he wasn't the first to ask that question. Alan Turing probably asked that question first, even earlier. But it's also interesting that I'm not the only physicist who is pulling him in that direction. Judging by the names of some fields of study, even available at the AGH University of Science and Technology for students, I think it's becoming more and more commonplace, so it's probably reaching the masses, I would say. Now, please, a question from this side. Does the fact that AI predicts, and not just predicts, allow us to think that we should go in the direction of it having some rights, just like humans or animals, for example? Well, that's a matter of opinion. That's not a question for me. Again, at this moment, I'm dealing with tools. The tools themselves have no subjectivity, so it's probably a bit too early a question. What will happen in the future? It's hard to imagine. I can imagine a future where algorithms emerge that resemble us much more and have more of our traits than the current ones we call artificial intelligence. And then perhaps this question will become more relevant. However, we are at a stage where this is not yet the most important problem that should concern us, but that doesn't mean it's not a question that won't become important at some point. I would probably bet that it will happen someday. Copyright will probably come into play the fastest. Well, regarding copyright, it's like this, all human creativity we have is based on gaining experience through studying, i.e. books, textbooks, going to university, talking to people. If I wrote a book or two, it's because I previously read some texts, but does that mean I owe money to every author of a book I've ever read? If I plagiarized one of those books, probably yes, but the idea is that I learn something about the world in various ways, and then I use that knowledge in one way or another, trying to create something a bit more original than just copying what I've read. And it's similar with training artificial intelligence, which is not there to copy word for word or parrot anything it was trained on, but rather, as a result of training, it learns certain general dependencies that are in the data and then uses them to create new content. So, from a moral point of view, I don't see any fundamental difference between what people do and what, for example, language models trained on the internet do. On the other hand, one can observe that the technology we create is meant to serve us, and even if this argument is essentially true, if it were the case that creating this technology discourages people from their own creativity, for example, discourages people from writing or creating films or paintings, then that is worrying, and we would rather want this technology to help people, not discourage them. So, it's more of a question, not how it should be for purely moral reasons. And I don't buy the argument that it's stealing data from the internet or stealing books. Because you could say the same thing about me, by reading textbooks or learning physics or anything, I stole a lot of knowledge from textbooks. So, that's the similarity, but I would rather put it this way, that we want technology to help us, not to limit us and discourage us from creativity. I myself once encountered the view that users, for example, of ChatGPT, should have more copyright, the more they develop their prompts, the more they put into what they want to get from this artificial intelligence, and not just ask for the whole text in general. Well, fortunately, that's not my problem to solve, but generally, more and more people are worried about what's happening, mainly artists. This story has repeated itself many times. When photography was invented, photographers were hated by painters, who said it wasn't art at all, that it was deadly for painting, for creativity, that now you just press a button and it's done. Then photographers repeated the same thing, blaming digital photographers for much the same thing, who also essentially do everything for us. In Photoshop, you can do whatever you want, and creating it in the real world is much more difficult. And now we have a similar return to this narrative, when prompting appears, that I can prompt anything, I don't have to do anything, and essentially it's the end, something deadly for creativity. It doesn't have to be that way at all. It's a bit like we're creating this future ourselves, but I don't think banning and restricting is a good solution, and as some people often demand based on the argument that training is theft. I probably wouldn't agree with that. I think it's similar to other things. At first, they are expensive, then prices fall, and then people return to something that is vintage, they like it, and it increases its price, because I already see, for example, with board games that boast that their graphics were not created using artificial intelligence, but by live people. I have experience in creating music videos and some films using AI, and also such productions, not so long ago, in which we deliberately didn't do it at all. And it's a matter of what tools are appropriate for a given application. And if we treat it as a tool, as something that expands our palette of possibilities, then why deprive ourselves, why deprive ourselves of something more? It can't be for everyone, it doesn't have to be for everyone. Photographers who take pictures on film often don't do digital photography. They have the right to do so, but that's not the same as saying that digital cameras should be banned. Those are two different things. Clear. Now, higher on the left. Please. Good morning. My question relates to what has already been said, namely, who would have a greater chance of ruling the world? The one who creates general artificial intelligence, or the one who develops a quantum computer? When I was a student a long time ago, there were already stories about quantum computers that we would have soon and that would break all ciphers and revolutionize the way calculations are done. That was a long time ago, and essentially not much has changed. I recently spoke with the vice president of Google, who was in Poland, and I asked him about the timeline for quantum computers, and he said: "Well, they talk about it, maybe another 5 years." For the record, we don't have quantum computers. What is currently called a quantum computer is something different from what was promised 30 years ago, because it's about a scalable quantum computer that can perform calculations in such a way that the error rate or noise level is low enough to implement effective error correction algorithms, and only with a sufficiently low noise level is this possible. Without that, we cannot scale these devices, so they are small toys for now, and we are still far from that level of scalability, so first of all, quantum computers are far away, but when I asked that person what we would do with these computers, since we only have Shor's algorithm and Grover's algorithm and a few other less useful algorithms for other things, and essentially there are no well-researched possibilities for computers yet, what they would be really useful for in practical use, he said that indeed we don't have them, but we'll invent something, we still have some time, and he trusts that as we get closer to quantum computers, interest in the topic will increase and there will be more and more interesting ideas. So, the answer is that we don't know yet exactly what these computers will be useful for. There is no certainty whether they will be useful for training artificial intelligence, for example. There are such ideas. There is something called quantum machine learning, and I've seen a large review paper analyzing the current ideas on this topic, but it's still an open idea that is not well-defined. It might turn out that quantum computers will be of little use for training AI, and we won't use them much. It's not impossible, and the answer is that I don't know, I have no idea, but I would probably bet that having a quantum computer at this moment, if you had one, would allow you to break a few ciphers and perhaps before switching to quantum or post-quantum encryption, such a computer could break into one bank or another or hack the internet. But it's not an unimaginable issue that we would switch to a different way of encrypting information. Then such a quantum computer is essentially useless. We can factor large numbers into prime numbers, but that's essentially it. However, I can imagine a lot of applications for artificial intelligence right now, much stronger than we have now. So, for now, quantum computers are probably a song of a much further future than something like AGI, or whatever it might be. But that's something that can change quickly. That's a vague hypothesis, and I don't know if I would take it too seriously. Thank you. I see that with quantum computers, it's like in my case with controlled thermonuclear fusion in a fusion reactor. Regardless of whether I was in elementary school, high school, university, or doctoral studies, they always said another 50 years, time passes, and it always remains. Please, closer to the middle, I have a question. I have a question about artificial intelligence, and more about responsibility for it, because every time we hear about artificial intelligence, how powerful it is, how it will solve physical problems, I wanted to ask about responsibility, because we create a tool that humans can use for good, but also for bad. Just as globalization created new ways of transporting drugs, it also created opportunities for drug cartels to transport drugs. So, I wanted to ask, what about responsibility, if it turns out that artificial intelligence finds an application that will cause mass harm? Thank you very much for overestimating my abilities. Unfortunately, I am a simple physicist and have little to say on this matter. And if anything, it doesn't even listen to me, because probably nothing smart. Well, that's my answer, that I simply don't know. Secondly, you are talking to a completely irresponsible person. Even if someone told me that it is very dangerous and should be limited for this or that reason, I would still be more curious than worried. >> physicist >> And I would prefer to check and do an experiment and see what it looks like. I would really like to >> physicist, >> so I am the last person to ask such important and responsible questions. It was the right side, now the left side. Please, >> Good evening. Given that you are involved not only in physics but also in art, I would like to ask if, for example, being a physicist helps in being a good artist and if, when you create a film, do you think more like a physicist or an artist, or is it one mind that has two different perceptions of what it creates? I usually separate them completely and treat my side activity as such, and I don't see any points of contact. Besides, after spending 15 years with textbooks, learning something new isn't that difficult. At least it wasn't when I was younger. So, learning photography simply didn't require two-year courses, I just took a camera, experimented, and started taking pictures. And it's the same with directing and other things. Now, when I'm involved in films, I'm a self-taught person, I'm incompetent. I'm incompetent in photography and in film. Somehow, that doesn't stop me from doing it. And there are even people who want to pay me for it, so But that's probably because I completely separate it and have always treated it as a way to spend free time, and not as anything serious, really. Besides, I'm also incompetent when it comes to artificial intelligence. I've just been interested in it for a long time, and I even wanted to get involved in it once, but all my knowledge is not from building these systems, but from reading textbooks or reading scientific articles, discussions with smarter people than me. But in the case of photography, I got involved and made a profession out of it. Now with directing, similarly, but as for intelligence, it's unlikely anyone will want to pay me for it. I'm just fascinated by it. Because it's incredibly fascinating, but I'm more of an observer from the sidelines. >> The physicist says something about the future. The health of the electron, even though it will never be needed by anyone. Maybe you are mistaken here too, maybe someone will pay you for it. Who knows? Well, physics is a good method of predicting the future, because that's essentially what it's used for, to predict what will happen in an experiment, only the systems in which we can predict the future are very simple. We can say what will happen to one electron, with some probability, but if I have a large physical system, then physics is obviously helpless, so it's a very limited method of predicting the future. >> Please. If possible, the microphone is there. You're probably the closest. Please. >> Good evening. I have a question. Don't you feel any discomfort because of the premonition that, living in an n-dimensional world, no matter how much n is, the world is, despite everything, we can only know it in n dimensions, and it has a few more. That is, we are doomed to failure, just like, I don't know, a 2D cartoon character, will never feel what the third dimension is. That's the first question, and the second, if I may, is there any view of time as a field, like gravity, that we are inevitably falling towards some minimum, I don't know what might happen there, but as if time, by passing through time, we are somewhat passing through it, similar to an element falling to the center of gravity. Regarding my discomfort, I've gotten used to being stupid, not understanding anything, and I won't have any great hopes that this will change much in the near future. Coincidentally. Well, I guess we have no choice. It's just that we live in a world that is too complex for us to understand. We can limit ourselves to some small segment that we can explore a bit, but only a bit. Fortunately, there are many of us, and there are many smarter individuals among us, or they sometimes appear, so there is some hope in that. However, there is no real choice, what are we supposed to do, sit and cry? Well, the world is very complex, we understand little of it, but at least we have a method to change that. And that method works quite well, because 400 years ago, we also understood very little. We know little, little more now, but at least we know a greater number of ways in which we don't understand reality. Now we realize how much we still don't know, although a moment ago it seemed that some things were very simple, only later they turn out to be complicated. So, it's a fact. We know little, don't we? We must hope that we will find answers to interesting questions. Physicists have a habit of trying not to ask questions that are too difficult. Even if they are very interesting, they refrain from asking too difficult questions. Not because they are not interested, but because it's a waste of time and energy to talk about it. There are other fields where such resistance doesn't exist. You are talking about some hypotheses about what time is at a more fundamental level. These are very interesting questions, but I have no idea about them, and I could start talking, but it would be a waste of your time. Referring to the fact that we can see a new problem in our lives, that we cannot see beyond the number of dimensions in which we are trapped, well, one can imagine such things, of course. The classic illustration of curved spacetime in general relativity is like a two-dimensional sphere, where we have flatworms living on this sphere, flat creatures on the surface of a ball. They don't know about the third dimension as such. They are flat, they have two dimensions, but based on their measurements and study of the world, they can infer that they live on curved spacetime, or rather on a curved larger space, for example, by measuring angles. So, it seems to me that we can probably probe beyond the dimension in which we are trapped. Am I speaking correctly? Well, sort of. We create brain prostheses. After all, particle detectors are ways to study other aspects of reality that we don't see with our eyes, ears, or in any other way. It's a kind of prosthesis for our brain and our senses. This metaphor of curved space through flatworms, I also have a chapter on this in Quantism, it has its serious flaws, because it creates the impression that curved spacetime is something that is curved in some larger space that is flat, whereas in general relativity, that's not the case. We have curved spacetimes that are not inside some larger spacetimes in a larger dimension, but spacetime as a whole is curved, so these flatworms live on the surface of a ball in our three-dimensional world. And this is our model of how to think about general relativity. But this model is flawed in that spacetime doesn't live inside, it's not contained in something larger, at least not in general relativity, and there's no analogy here at all. So, it's quite, this analogy is often used, often overused, but it seems to me it's quite limited in its meaning. And on the other hand, the question is then >> Good evening. My question is: for what reason do we not understand how a trained neural network works at this moment, and is it a result of insufficient investment in this field? Thank you. There is such a field as explainable AI, or interpretable artificial intelligence, and it is a field of research that involves creating simple neural networks, simplifying a trained network as much as possible. For example, you can train a network for a task and introduce an additional penalty in this training for each additional neural connection. As a result, networks are created that do something, but with far fewer connections, and you can train a network in this way that will be slightly worse at the task, but it will have very few connections and will be much easier to interpret. And people try to interpret such simple networks, to learn something. And there is a lot of extremely fascinating, interesting research on how trained language models work, where knowledge is accumulated, how it is accumulated, how information, for example, about the word "elephant" is located in these multi-layer perceptrons. It often turns out that it's not in one specific place, but it's distributed among many neurons. People try to unravel this, but essentially, we understand the training of neural networks very well, and we design the training, which is an iterative procedure that modifies the network as data is accumulated or through other forms of training. If it's reinforcement training, it proceeds a bit differently. But we have this backpropagation method, which is what Jeff Hinton uses, to correct weights so that the network is a bit better. And now it's interesting that this method works in a very wide range of neural networks. It works for almost any network. We have a simple algorithm that improves the performance of any network, as long as we know what the goal of the network is. For example, if it's to create drawings, make films, recognize handwriting, or predict a word in a text. The same backpropagation method works in each of these cases. We don't even need to know what the network does or what it's supposed to do. We define a cost function, which is essentially the difference between the performance of a good network and the one we have. And the iterative method of improvement, reducing this difference, i.e., reducing the network's error. And this leads to fantastic results in some process that converges quite well. At least if we choose the architecture appropriately, but it cannot be said that this is a deep understanding of what a trained network does. Because at some point, we get a network that does what we want, meaning its error rate is low, but this training procedure does not answer the question of how the trained network works. And now, what can we do to change that? We can analyze the trained network. We can go inside and analyze neuron by neuron, what they do, how information propagates in these neural networks, what subsequent layers do, and people try to do that, and we have some insight into it. But such a bird's-eye view is still very distant, because it's a bit like trying to analyze the structure of the brain atom by atom or synapse by synapse. Well, we can try to do it. In fact, it would be possible, if not for some practical limitations, but even if we are able to understand every step of reasoning, meaning every step of building our brain, we will not understand the entire structure. For example, when I analyze a very complex mathematical proof, which is a sequence of small logical steps, but sometimes these theorems have proofs of 200 pages, I can say that I understand every step, this adds up, I understand this, I understand that, and essentially I started here, I finished there, but I still might not know what the overall intention is. Why this particular path, why such a choice of small steps? Every math student knows how difficult it is to understand a more complex mathematical proof from a bird's-eye view. And we are at a stage where we have very complex creations that we can analyze on a microscopic scale, looking at what happens in this piece of neurons or that one. But such a bird's-eye view is still extremely distant. It's a bit like in biological evolution. I like this analogy, that we also understand evolutionary processes quite well. We know how animals arise in the process of evolution, and it's not conceptually very difficult. Richard Dawkins, for example, describes it excellently in The Selfish Gene, one book, but you can't describe in one book how a frog or a human or even a single cell is built. It's too complex an object, so we understand the process that leads to the formation of a cell, a frog, or a human very well. But it's still very far from understanding how a human works. And it's the same with networks. We understand the training algorithm very well. And such a simple training algorithm to train, for example, a convolutional network to recognize handwriting, is 20 lines of code in Python. If you load a library, it's a few dozen lines of code. You can train a network this way. And the trained network will recognize handwriting. I still don't know how to do it. If, for example, I had to sit down and analyze this network and based on that write a program in C++ that would algorithmically recognize handwriting step by step, I can't even come close to the efficiency that a neural network has. So, if I can't write a program in a conventional programming language, I can't say that I understand it. For me, the most concise definition of understanding is that to reason or to understand is to algorithmize, meaning if I understand a problem, I can write an instruction manual on how to implement it, how to solve it. We are very far from that, really very far. And we know something about neural networks, there is this field I mentioned that studies it, but I would bet that the amount of knowledge we have managed to gain compared to what we still don't know is minimal. Just like with the construction of organisms. We study the brain, we study the eye, we study the liver, we know something about a frog or a human, but compared to what we still don't know, it's a negligible fraction. >> Thank you very much. One more question from this side. Do we have a microphone? You are also quite patient. >> Well, are you currently working on any physical problem that has no solution? And if so, on which one? I hope all the problems I'm working on have a solution and I'll find them soon. [laughter] Well, yes, for a long time we've been banging our heads against a wall with a problem that everyone tells us can't be done, and somehow in a small group of desperate physicists, we're trying to solve everything. But whether this problem has a solution or not, we don't know yet. We want to know that we want to believe that these problems have solutions, because we have a certain physical motivation, and for example, it often happens that I don't know how to solve some problems, but for example, I know there is some issue that requires a solution, and I know that this issue requires a certain description that we don't have yet, but essentially it should exist if we believe that physics describes all observable reality. And there are such issues that tell us it's impossible, but we know that this distant goal must be achieved in some way, and that motivates us not to give up. But whether we succeed or not is another matter. Perhaps imagine, in quantum theory, I have a certain deficit in understanding what happens in quantum field theory regarding the mechanism of spontaneous symmetry breaking and the Higgs mechanism, where the Higgs field in its current state, in its current configuration, is easy to describe with standard language and one can predict the properties of Higgs particles, one can even verify them with experiments, and these particles have been experimentally confirmed, for which there were Nobel Prizes. However, there are configurations of the Higgs field for which we don't have a good description. We only have an approximate description, and it seems that the description we don't have in these specific configurations requires tachyonic elements, but we don't fully know what that means yet. That is, the description we have is imperfect, but on the one hand, we know that, but on the other hand, we know that some description of this field should exist not only in an approximate, semi-classical form, but in a fully quantum form. Such a theory does not exist, and we are trying to create it, because we believe that to do so, it is necessary to utilize the second of two possible solutions that appear in the theory of relativity. That solution is related to superluminal speeds, which for a long time was rejected as unphysical, as the second square root that is invalid. It seems to us that without a good understanding of how this second square root works, how to interpret it in quantum field theory, there will be no good understanding, a full understanding in the fully quantum theory of the Higgs model. But this is my hypothesis, and unfortunately, I know less than I would like about it, and I am constantly trying to learn more, but it's slow going. A colleague heard about a doctorate and assisting the professor. Last question from the second part of the hall. Please. >> Good morning, Professor. >> Good morning. >> When will AGI be available soon, what will be your first question to this superintelligence? Well, you know, I'm interested in physics, so if AGI is understood as a system that is as good as the best expert in a given field, then I can simply ask smarter colleagues from physics about something and I'll find out just as well. So, a more interesting question is whether we can expect something beyond that level of AGI, i.e., super artificial intelligence that will be better than us in many things. However, I suspect that in physics, this help will be limited for the reasons I mentioned, i.e., because we need experiments to learn something about the world, not smart theories. However, with the history of questions, if we go back a bit and see what difficult questions looked like before, let's take, for example, the smartest person, the scientist, who ever walked the earth, i.e., Newton. And let's consider what he could have asked AGI during his lifetime. Well, what could he have asked? For example, he could have asked if light is a wave or a particle. That interested him. He wondered if light consists of corpuscles or is it a wave, and he leaned towards the idea of corpuscles. And what would be his answer then? It would be: the question is poorly posed, because neither one nor the other, light is something in between. And what would he do with such an answer then? Well, he wouldn't do anything. He wouldn't know what it even means. It was too difficult a question at that moment to understand it well, so I don't think anything will change over these few hundred years. And we still, if we wanted to think ahead, the questions that interest us now might be too difficult or poorly posed, or often it happens that the questions we ask are simply poorly constructed, they use concepts that don't make sense. And probably the most sensible question would be to ask what is worth asking about at all at this moment. at the level I am now, with the knowledge I possess, what is the point of asking at my meager level now, so that the answer would give me anything? So, that would be my question, but I don't expect a sensible answer from artificial intelligence, which is not based on experiments, so I don't think the answer will be very illuminating. Regarding all the questions and answers, I would like to thank you very much for asking them, and you for the answers. Our guest was Andrzej Dragan. [applause]