Transcript for:
AI Overview and Foundations

So, he saved me from the pond. His we'll leave it anyway. And what do you have? is happening? Yes, in the MTS family group maybe 10 people, but there are 11 of us. Here Let's find out who's extra. Maybe you're the odd one out? And the bald ones speak for themselves: "We're deleting it." Think about who add. Family group from MTS, after all each participant will receive a discount of up to 30% Get in touch. The moon is the perfect starting point platform. If you use it wisely, then it's much more. You wanted explain how you will bring so much there in total. There is no need to drive there, it needs to be made there. No no, No. To manufacture a plant means to manufacture a plant needs to be brought. Well, well, just, well, First we make small Christmas trees and sticks, yes, additive technologies, automatically For all this you need brains engineering. For us, our brains for you and me This is clearly missing. Intelligence not enough. What do you need? You tell me more artificial intelligence. Artificial We need intelligence with us, by the way, our fellow traveler. We have a great specialist on artificial intelligence Sergey Markov. Great and terrible. They do happen such coincidences. It's amazing. A you're talking about how to best apply artificial intelligence and what What is this anyway? I don't understand. What is this such? You don't understand. No, I'm honest. I honestly don't understand. Listen, you have you ever programmed in any languages did you write? Well, and at a high level in basically. Well, not counting some attempts there are no simple ones there at all. At a high level language. Did you write the codes too? Yes. No, well. No. And I didn't write the inputs, but I, for example, I can write programs in six languages. Yes. But I don't understand what it is artificial intelligence. Sergey, I write the program for a specific task. So I have a task, I wrote algorithm on how to solve it. And artificial intelligence is something else. Now I'll tell you. In fact, it means, well, if we talk about the term itself, it has quite a lot of interesting story. In general, his first the tracks are from 1955 when John McCarthy prepares his proposals for organization of the Dartmad seminar, in the future of the famous seminar. This mathematics, which took place in Dartman College. Well, that's it. computer science programs there is no such word was. Ah, well, the term computer science as it is traditionally translated as computer science, yes, in Russian. Here the term computer science didn't exist, but as a specialist in the field of coater science, a, and in the fifty-sixth year, uh, so, for this very freebie a large group gathered for the seminar specialists. Some of them were famous already then, someone became famous later. AND that's exactly at the Dartman seminar this has already sounded, so to speak, openly the term itself is artificial the thesis artificial was voiced intelligence. It seems to me, Sergey, correct me on this point, what The Russian translation suffers from some bias values. Artificial intelligence. Intelligence is the ability to some kind of reasoning. Right here intelligence. This is, you know, the problem here. that and with the definition of intelligence there are many disputes, so it is a great sin There is not. Here. Nuno, it just happened that way, scientific traditions to translate, as artificial intelligence. And if there was artificial intelligence, would you agree? It was intelligence in general. I would do the same. But here, look, yes, the system is like they say that there are a lot of disputes going on there long, yes, the topic of defining intelligence, reason and so on, but disputes definitions, probably not very much for us are interesting. Here. I wonder how what do we mean by this anyway? the industry itself, right? That is, when we we talk about artificial intelligence today, that, uh, we usually understand such an industry science and technology that deals with automation of intelligent solutions tasks. That is, there are some tasks, which people traditionally decide when with the help of your own mind. Here I am I put my fingers in, I already have objection. I'm holding back for now, I I'm not interrupting you, I'm bending my fingers. It's okay, it's okay. So that's it in that sense, yes, that means it is an area science and technology, which deals with automation, intelligent solutions tasks. Intellectual tasks, well, that's it. quite neat too, yes, we need to do this thermom. Game of tic tac toe, for example, and that's all. This too, this too artificial intelligence. Well, it's been like this for a long time the programs are written, no one called them artificial intelligence. No, they called. Well, they called it, of course, Certainly. Moreover, didn't you know? Everyone knows. Yes, you see what the thing is? Here, uh, yeah, even mental addition, yeah, that's it also intelligence. Here. But there is no need for this to be afraid, you know? Here. The child is there he's learning the multiplication table and he's doing it mathematics at the same time. Well, we say, Well, it's formal, right? That is arithmetic is a part of mathematics, right? But we still call him a mathematician, Yes? That is, when we talk about it there specialists in the field of mathematics, we we are talking about people who work for at the forefront of science with tasks that complex, they create, complex for the current level of development. But here in this case the task of memorization is not fundamental, it is also intellectual, but we also have it brain help vs. Let's give it to him We'll seal it today. Look, that's it completely mechanical. Yes, plants. they remember, cats remember. And the brain it's also a mechanical thing, physical object. In it the physical object is understood, but we are getting ahead of ourselves now. Here look. Well, is this magic or something? Well, this is the functioning of the physical systems in the physical world. Here. Uh, here it is really important to understand Well again, yes, we are there now Let's get ahead of ourselves. the connection of all this with physics, with entropy, with e always, when we are talking about calculations, yes, or about in intellectual operations, it is always underlying physical processes, but subjects in the sense of underline, below lying in the conditions. Here. And that means there is such a general concept as effect artificial intelligence. Sometimes too the effect of macordoc is called honor broom macordoc. And it is connected with the fact that as just some intellectual task is solved with the help of automated systems, people stop recognizing the intelligence of this task. Yes. That there is, for example, in the fifties, if would you stop a man on the street there and asked him: "And here is the program that she will beat the world champion in chess, will be a real artificial "intellects"? Here. And when in ninety-seventh year, So, in the second match Kasparov lost Blo, what did people start saying? Well, yours the car, it's, uh, actually stupid. There is no intelligence there. She's just very fast. She's picking at it, she's rough. forces of 200 million positions per second I was looking through it. Man the truth, but not the whole truth Truth. That's the whole point. That, sorry, there was a moment when there was a move that Kasparov suspected human intervention because supposedly it's gentle to guess. Later, when won at go, it turned out that This thing makes intuitive moves. No, I In No these are two different systems based on different, quite different principles. As for intuition, this is generally now the road will take us into the byways. Here. Means, but people at the same time, as if, when they became use the brute force argument, they as if they forget a little that it is very many processes that occur in in the human brain, they are outside conscious control occurs. A the brain is a very complex device electrochemical, right? So, 86 billion neurons to a quadrillion synapses at peak development. One synapse to model today, we need there as at least a thousand binary elements, right? That is, if you imagine the scale, the computational scale of this scheme, if look at the brain as a machine, yes, We don't have anything like that yet. If calculate the approximate number of conditional ones there there are binary operations that the brain in unit of time is carried out, while brute force is still on the side of the brain, not on the side of the electrode computer inferior to the brain's calculation speed. He is he he speaks not according to speed, he speaks not according to speed, but by the number of elements. Well That's exactly what I mean, yes. But, you see, consciously, uh, here we are looking at picture and we understand what is on it is the cat drawn or not drawn cat. What is behind this? There is a reason behind this huge cascades of electrochemical signals. We ourselves cannot explain, why did we uh realize now that this a cat is a person we know, and this is something else, yes. But the underlying physical processes that ensure pattern recognition, for example, yes, they, if we recalculate them in the operation cars, yes, well, they're huge scales of calculations of conditional, which the brain performs in order to solve these problems. And when a person looks at a chessboard board and understands, for example, that there are some such such reasonable people here continuations, and here I will be consider in your analysis, yes, on in fact, the operations to be performed are such that this, what is behind it, why do we we understand that these moves make sense analyze, but these are not even worth it watch, right? There is a huge reason behind this the work of this powerful one electrochemical device, rough the computing power of which is very great. A person doesn't go over a million options intuitive modern modern programs they are also no longer are sorting through. Here. But forgive me, please, but the number of calculations, which they perform, unit of time, it's very A forgive me, enlighten me. those trained nets, if left aside, trained, but on Cat entertainment, for example, is not there cats, dogs and, say, hamsters and so on further. Putting aside hell training, let's leave that aside, she it does this quickly and not for very long a large number of computational steps, perhaps much less than in the brain. Correct me. Now, now I will answer this question. I'll finish talking about the effect. Makorda, we kind of got interrupted like that, Yes. That's what people started talking about brute force to say, yes, that there is chess there - it's not an intellectual task, it can be solved This is a rough estimate, but there is a wise one the eastern game of go, which is uh cannot be solved by enumeration. That's when you show us the car, which means, will be able to compare with a person here, then we lost Lisdol out of five. Yes, when Fox Sedol is the strongest there, well or there is one of the several strongest in in the world of players Goa lost to alphago, that what did those people start talking about? Well, none at all. there is no intelligence there. It's simple matrix multiplication. That's when you tell us show the car that will be painting draw, yeah, uh, whatever you want order, then we will be with you let's talk, yes, that is, in fact in the case of Dadada. Well, they did in this case, Yes, expensive models are similar, but it is important understand that, of course, people spoiled by technological progress, yes, this is how it is with Luisikei, when a person flies on a plane and suddenly he has Wi-Fi stopped working. What the hell is this? I was offered some nonsense here. Wi-Fi is not working. Does it matter that there are 3? a year ago there was nothing at all Wi-Fi. I'm sitting on a chair at an altitude of 10,000 m and I'm flying at a speed of 1,000 km/h. Nothing, Yes. At a distance of several thousand kilometers, yes. And my ancestors are several generations ago, if they wanted to move the same distance, then a group of people who would go to the way would be different from that group, which would be the conclusion. Conclusion. Artificial intelligence in every era implies your own. [music] Imagine you work in an international company with a stable salary, you move up the corporate ladder stairs and suddenly you quit your job, oh which millions dream of, and you go study to be an astronaut. This is history Konstantin Borisov. You probably have it you know. A couple of years ago he went to ISS on the Dragon spacecraft and spent in space for almost 200 days. But before that he had to prepare for many years and study. You can achieve it at any time new heights. And there is no need to make a U-turn by 180°. Most of us don't want to change your life completely and become better at what you do. And of course, earn more. Very interesting the master's program is now available in Central University. This really in-demand professions, in which past experience will come in handy. You don't you will start from scratch, but you can advance in your career or move into related specialty. You have a choice several programs: grocery analytics, product management, backend development or machine learning. Lectures are held in the evenings and on weekends. 80% of students combine their Master's degree with work. Also in Central the university is launching experimental two-year program. by design human experience. It is for those who already has education or experience in design industry. During training you gain management competencies and deepen your skills in product and communication design. You will study in Moscow with international experts, practitioners of large companies and teachers from Moscow State University and IN InfoTI. On Upon exiting you will receive a state diploma. And all the skills to be competitive in the market with They will help you with an internship and employment. We wanted to see the campus ourselves. modern university. So ours the space train made a stop at Russia, in Moscow. People of the Central the university is strong students, and as well as their experienced mentors and coaches. This atmosphere inspires you to grow and develop. Dear friends, here we are finally in Moscow in Central University. Elizabeth, hello. Hello. You are in graduate school Central University. I'm done Baumanka. I am an engineer, I worked in Roscosmos, built rockets. And in some At that moment I realized that I was missing something some managerial managerial skills. And for this I came to the Central Control Center product management, and to with its engineering fundamentals education, but to make a superstructure modern, but also managerial, learn how to be engineers and manage IT specialists. You will hear engineers thanks to your engineering background, and here you will get useful management skills. Thanks to a large number of engineers will be work more efficiently, yes, and do cool, useful products. Participate in the grant competition. This support program for applicants, which covers up to 75% tuition fees. The grant is issued for the entire period of study. To accept participate, take the test and interview. All details in description. Go for it. [music] I think you just have to understand that artificial intelligence is a big discipline. There are a number of trivial ones in it tasks that we have today, well, interesting only from a historical point of view vision. We have mechanized them, yes, but there are the frontier of this research. And of course, when the average person talks about artificial intelligence, it expects that, Well, they will show him either some tasks that have not yet been solved, or which were recently resolved, or which we, well, sort of, aren't solving yet Very good. Frontier is a borderland, yes, yes. The border of science. Here she is always excited about technology, yes. So of course it's normal that People treat it that way, yes. But if we will look at the history of artificial intelligence, then do the first one there system for playing Tic Tac Toe not it was a joke. It was very difficult at the time. time. There is Sergey, Sergey is still there. things are algorithmic, but now we want talk first of all about the big ones machine learning, big language models on one side and on the other sides, forgive me, for God's sake, but it's true when they say it now artificial intelligence, e mass the consumer of this term is small interested in what programs there were earlier. He now believes that he is will replace scientists, decision-making, judges. And the main thing is that what we can’t do, what is the common opinion, in case you don't know in the know, what we can't do think of it. We have an opportunity here. limited but artificial intelligence will decide everything for us. Let's do this We will comment on this thesis in due course, when when in the mid forties Rosshbi did a device he called homeostat, which can be considered as one of the first prototypes artificial neural networks, like this one made in the form of a device. then the first the same article that was dedicated to this the device was called the clicking brain smarter than a human. Here. And there are journalists there, naturally, they wrote that soon, who knows, this The machines won't need us anymore. They themselves he clicked. Why? The relay is there, of course. So, that means, there the weights were selected very interestingly this artificial neural network. There there was a drum finder that substituted random resistances into chain. Yes. And the resistance capacitance was selected using random tables numbers. Here. So, it's very interesting, actually, look at history development of these technologies. This is fine in a sense. Well, how is it normal? Normal in the sense that we, I guess they got used to it, that ordinary people very often judge many scientific concepts, based not on anything reading books by scientists, yes, relying on Hollywood film productions there. Listen, here's what's bothering me right now. Our brain volume is 1.5 liters. Yes, I think so. Well, here is the head, here it is, the head is 2 liters. Oh, and Kasparov also has something like that. head. And he is on equal terms, well, in general, solved chess problems with almost equal this BlueBlu. Who is he? Now, wait, wait. And the size of this of this computer on which it is implemented there was a program, it was just like this hospital or was it also standing on the table and could it happen? Several cabinets then it was, but today it's a mobile phone copes well with the game on the super supergrass. A mobile phone can It's easy to beat Kasparov with a huge breakaway, right? Yes, I want it disappoint. Yes. Here. But this is this look in nature in living nature there is creatures whose nervous system is strongly smaller than human, but separate tasks due to super-specialization they can decide on superhuman. Try to catch a fly, right? Or this one a bee, for example, will find it in the hive the optimal route is faster and better than a person will do it. Although she has it there hundreds of thousands of neurons, yes, we have 86 billion. Oh, and in general all the successes of artificial intelligence in many ways is about that, that scientists were able to carve out a large seven Shapovtsy, scientists, engineers. That is it systems that, due to their very large size, specializations reached superhuman level of decision making individual tasks. Who's talking about hats? I don't Understood. Well, it's a cartoon, you know. that's when the merchant demanded from there The furrier, so that he I made six hats for him. This is a task for optimida. As an optimal pattern, yes, that's it let me, well, well, yes, that is, as in in essence, to solve an impossible problem, yes, and having such res Well, look, electronics are still very much in some things are inferior to human brain. Here is the human brain 20 watts energy consumes only, yes, and successfully will be able to cope with the enormous the number of intellectual tasks, which machines are not yet capable of. Yes, about 20 watts. Here. And the cars they consume, well, there's a lot of it, yes, and They weigh a lot, yes, and they give off a lot of heat. We have a supercomputer at Moscow State University. You pass by this building where he is there is a hot stream of air from there, with such huge air conditioners. I I thought that I was this cool one, but I think yes. And this kind of hints to us the fact that there is a lot of common ground here works for physicists and computer scientists, yes, and specialists to sew to make miniatures, find new new physical substrates for calculations. And a quantum computer in general this is a realizable quantum quantum. And how can its implementation be be very different from a physical point of view of vision? That is, quantum is essentially things are, well, kind of an abstraction More mathematical, right? And how about us? realize this very quantum? What what separate elements? That's before classical electronics appeared there in the thirties, uh, John Vincent Atanasov, the man who created the first electronic computer car, I did everything. For example, so he used paraffin cubes as the underlying system for calculations. Soap films, laplaciometer. But here are the soap films are described by the Laplace equation. That's why it's like what's possible to set in a certain way surface of the soap film some pattern, yes, and it's a soap film, in it physical processes will occur, due to which we will be able to result count answer weigh. Here you go hydrointegrators heap. We have a lot of different physical processes were tested for the role of candidates for the creation of computing devices. All these memories were on different principles. Yes. Yes, yes. Yes. Here. And, uh, that means, electronics is classic, it approaches certain limits as times of a physical nature. And on the way general development of computing systems there are several barriers of varying degrees irresistible, right? So, but there is fundamental barriers. In the fifties years, for example, they talked about the limit there Bremerman, so called. So, the idea in what? That we have information faster The speed of light cannot be transmitted. It is forbidden. The machine element is smaller we can't make it to the Planck scale because of uncertainty, right? So, u a machine of mass M will always be able to produce no more than some limited number of calculations in unit of time. Whatever it may be arranged. Well, the Planck scale is for us still far away. This is such a limit, not very scary still, yes. It has been generalized to quantum systems, called Margolus theorems Levitin. Here. which simply says, that within the framework of a certain energy you can't have more than a certain limit number of calculations to perform, right? This need to. So you're striving for something, right? Is there a need for this or is it a game? It's just a game, beads. Here. Well, we have there are a huge number of different ones tasks from real life, for which a lot of calculations are needed. How did we do them before without a computer? did you decide? No way. Badly. Well, that is to say by approximate methods. Well, as always, we want to improve algorithms the incredibleness of alcohol, yes, better, for example, the same engineering tasks creating optimal designs for some tasks. Eh, yeah, yeah, whatever, In fact. No, well we have a number of tasks is expanding. For example, we are now we are interested in protein coagulation, which we only 50 years ago, for example, yes, generative biochemistry, yes, create new molecules that will have given properties, that is, in advance calculate their structures. Yes. Yes. That is uh system, which and many of these tasks, they still, unfortunately, relate to difficulty categories expй complete, then there is as the problem dimension grows exponential way computation time is growing. That's why cars, of course, are for us are needed more and more if possible fast. Sergey, I'm sorry, please, I insist, I'm sorry, For God's sake, I insist. Electronics is electronics is developing, but at the modern electronics, it seems to me, correct me, we have been lately this jump happened when we from the written programs that are performed more or less deterministically in such a way, and it is approximately clear to us how we moved to machine learning. System she learned, she does something. And then, wait, wait, I want to ask the most interesting thing, otherwise I won't have time. A then, then it turns out, and you are here correct me, that because she is so she studied well, and it arose in her something that we don't even really understand taught. She can do something else. This in general the thesis has the right to existence? it does, yes, but with a caveat that a learning machine is a well-deserved discipline, the origins of which there they start 70 years ago. First self-learning primitive system, we they were able to create even more there than half a century ago, right? A simple example. So, what is called a matchstick automatic, right? uh, you can create a system from matches from matchboxes for that so that she can learn to play crosses zeros, yes, shifting to a certain like a match between boxes. Here using a simple set of rules, yes, having played a number of games with of this system, it learns as if the rules necessary for, and therefore, correct game of tic-tac-toe. A, and in principle, in the most primitive way training, which is, well, the method is just there are no errors, yes, let's just do it a system that will generate random decision yes we will reject the wrong. We as people or we with the help of the same machine? Well, that's it. it can be automated. There what? the problem was solved there at the time BCTL park cryptographers, right? So, here you go they had messages sent during with the help of an encrypted key, right? Means, there are operations with keys or key removal decryption decryption decryption Machines in the ready world war. Here, so, and, accordingly, uh uh cryptographers suspected that, therefore, Well, according to the rules cryptography, after the message encrypted with some key, cipher I have to turn the handle, yes, so that next cipher combination formed, yes, if you send two messages seem to be the same ciphers, it is not safe. Why? Because that the codes are superimposed as if on the text messages using the operation exclusive or if we take two adjacent messages and through exclusive or combine them with each other, the key will be deleted and will be just text first message xor text of the second messages. And here are the semi-combinations not a cipher, but the texts of two messages, which makes it much easier to understand. And this is it it is already possible using classical methods cryptanalysis back to decrypt. But we need to try to catch this situation, right? That is, we need go through all the message pairs there, yes, and look at how they stack up against each other will it lead to such a combination, which will have certain statistical anomalies, right? Here. A These statistical anomalies can be explained with the help of a simple mechanical device determine, yes? What we found here the very couple of messages, encrypted one and the same to the face. They will carry some features of the English language. Well then approximately inherit there will often be frequencies of different elements be very different from the situation, when we messed around with the key. That is purely mechanically, yes, this is the task, you see, it was in the early forties make a machine that is simple goes through many, many options and stops at the moment when a variant with the given characteristics was found parameters. We learned to do this even back then. By the way, every like and comment and subscription brings the release of new videos closer Universe Plus channel. So if you I like it, you know what can be done right now. And if we talk about the homeostat of Rosa Ashby, what is this, huh? That is, you there are such blocks, this means they consist of electrolyte bath. So, it is connected to one side contact. There's a float floating there, its surface, therefore, the arrow, to which the next contact is connected to, Yes? That is, depending on how the arrow deflects, the distance changes between one contact and another, and the resistance changes. Yes. Why? Here. A now it doesn't really matter why exactly this little block, but in fact they were used in sighting mechanisms bombers at that time, yes. Here. But what's important is that it's like potential controlled resistance, right? That is, we can, uh, apply voltage and using this change the resistance, yes, uh, another chain. Here. Uh, right now it is clear that they are much more miniature make devices, yes. Yes. Here. A and What do we do next? We connect -e These blocks all have all the contacts. we pass current, yes, and in each one eh in each conductor that connects two blocks, we install this an interesting device, then, which, if our arrow has deviated more than 15° from the central position, the stepper is triggered, it starts the drum spins and is inserted into the chain random uh resistors. This is not encryption, Right? I understand? No, no, it's simple This is such a car. homeostat. So, in What was the idea behind Roszhbio? Now, maybe, no matter how difficult it is for people, having a cybernetic, so, background, to understand this idea. But What's the idea? So, we, uh, gave current to the circuit, Yes? eh some of the arrows started deviate more than 15°, these drums started spinning, yes, and in at some point such a thing was chosen combination of resistances in which all arrows were inside 15°. More do not deviate. The system came quite well I am happy with my condition. System I am happy with my condition. She came came to a stable stand. Now we enter some kind of indignation. We cut one wiring, for example, yes. The system again comes into a chaotic state again the stepper switches start to turn again a combination of parameters is selected, at which the system again comes into that there is a system capable of saving homeostasis, like, yes, like biological systems, Rosshbi was a doctor, yes, and he was interested in how biological systems work in this regard systems, how they adapt, yes, to environment. And he essentially created such a a primitive model of adaptive systems, right? That is, which we can artificially introduce various disturbances, but we can be different, we can damage something in this system, we can change its power voltage that we apply to the inputs this car, right? Well, well and so on. Here. And this is if we talk about physical side of the matter. And if we talk about the mathematical side of things, then, of course we have what is there? We have a theory of optimization. Here is the theory of optimization, it is again about What? About what it means, uh, we can, uh, imagine some procedures, uh, which will be the parameters of the mathematical models should be selected in such a way that minimize the value of some target functions. Well, for example, here's what we have neural network, right? Here is a neural network, we are moving towards bigger things language models, right? So, uh, Well, big language models are large neural networks that are used to pass the language exam modeling. But neural networks, yes, here is the first artificial one neural network, it was described in 1943 year in Mac's article. Listen, I lost it. thread, guys, sorry, but let's after all, it is an analog system, which strives to calm down. Yes. What the hell? What problem does it solve? I have a cup and There is a ball in it. I knock it out of the bottom of the cup, He rode around and came back there again. Same analog computer. The parameter is larger, just a cup is complicated. Now here it is look, we, for example, want to create such a system, into which, if we push a photo with a cat on it, the light would come on, right? And if there is no cat, then the light bulb wouldn't work it was lit up, right? So, in this system which is like a disturbance, right? If the system malfunctioned, yes, that is, if it is for a photo with the absence the cat lit a light bulb, and for a photo with on which, on the contrary, has a cat, there is no light bulb lit up, yes, an error, yes, of this system. And we can use this error as feedback, yes, in order to adjust the parameters of this system. So it turns out that this error decreased. No, no, no. We create automatic system. Here's how there is a certain reason behind this mathematics. What is a neural network? Let's take a look. Let me explain now. Now we'll quickly get to what we need. result. A neural network is something like this big formula. It's such a big multi-level formula, which is actually superposition, a function that is superposition of simpler functions. Here. What kind of functions could these be? Well, actually, they are different. Ah, well, for simplicity, for example, we can, it can to have such a function means that it exists, uh, several parameters, uh, that means She multiplies each parameter by constant, adds all this between itself, and, and substitutes it into the so-called threshold function. Ah, yeah, well, uh, for example, sigmo. Well, generally threshold functions may be different. May be piecewise linear simple function, for example, relu is a function like this. So it's her, right? She goes from zero to zero zero, and then y = x. Yes. What is the task? Now I want to keep the task in mind. Here Recognizing cats in pictures. Let's take this problem. And we we are building Here we write a program that is capable of doing this. Yes. Yes and no. We don't want to ask explicitly the rules by which this system will be act. We want her to somehow myself based on the set examples that we will give her. Well, here it is Well, finally, yes, she has it standard photographs. We have, yes, we put some guy in jail who I took 10,000 photos from there and laid them out two stacks. In the first stack there are photographs with with cats, the second one without cats. Now we want to create a system, which will actually guess, this is it. So, now what is it? photo, right? Well, we can do it turn into a set of numbers. And the set of numbers how? Well, for example, e means mesh breaking, mesh covering, right? So, each furry is an image, or something, It's not just a matter of laying down a grid. Each cell means uh consider that inside this cell the color is the same. Well, well, average the color, yes, I swore Curse every pixel. Here we have it get those same pixels, right? And everyone pixel, well, if it's a black and white picture, then one number can describe zero, black, one white. Intermediate The values are some shades of gray. If it's a color picture, you can use three represent each one with numbers. Yes. Well, for we have simplicity there, for example, 16 by 16 such with a small resolution picture, right? So, we have 256 numbers. it turns out from zero to one, right? And u we have a kind of formula in which we We substitute these 256 numbers, yes, we calculate them her and the answer: we want it to be equal one, if there is a cat, and zero, if there is no cat. Here. What should we do now? Here we have it There are two stacks of, uh, pictures. We are them digitized. That is, we have, in fact in fact, there are 10,000 sets of numbers, right? 200 by 256 numbers. Here. So, for each of these sets of numbers class label value. Well, the answer is correct, yes, which is like this marker, teacher, one who laid out in these two stacks of photos, he A formula-then one for all or one for all. And how? is this possible? Here. How is this possible? A this formula, it's, uh, complicated, yeah, it, uh, has a lot of parameters. That is, we, look, uh, so it's a superposition these are the functions, yes, that I am talking about said. Means, the element itself is separate, artificial neuron, yes, as they call it, this is it such a function. So, f from x0, x1, x2 there and so on. Xn = x0 on W 0 + X1* W1 + X2 to W2 and so on. Xn on W. The following is said. This is linear function. You enter 10 parameters there. You give each parameter a specific one. weight, multiply it by this weight. And this you fold it and put it under the threshold a function that either triggers on it is or it isn't. And then his task is pick up all this crazy stuff number of scales. Yes, yes. Here. That is, ah, in this formula there, for example, there 1,000 parameters, right? We need select the value of 1,000 parameters so that the total error recognition was minimal. What is this a recognition error? Well, that's it. difference between predicted value true meaning, yes. Accordingly, what we what do we do next? So there are some methods that are called method backpropagation of error. He is for specific a specific type of these formulas, superposition works. It allows us calculate the gradient of the direction of change these very parameters that reduce the error. Yes. And we move And we are moving in the direction of the gradient space in multidimensional, I don't know, in the ten-thousandth parameter space. ten thousand black space. We are in this we go through ten thousand dimensional space with the help of optimization methods that are still pick up the pace, pick up the direction, this very gradient is recalculated. U they still have inertia moment, so called to local optima don't fall, right? Here. And in this in a sense, from a mathematical point of view, as if the task of machine learning is What? This is an optimization problem, this is the one in which we must decide, Well, it's like two tasks, right? That is, First, we must choose such a form functions, yes, which would provide us with such a maximally smooth surface objective function in this space, so that later gradient methods optimization, yes, they would have found a solution for us as close to global as possible optimal. Listen, we have such a difficult time. conversation that I think is not for everyone, Well, it's pretty obvious. I want illustrate the gradient method search for their own incident, when My friend Sergei Lomzin and I went to pre-graduation practice in astronomy Observatory. And the Observatory in Georgia, Georgian Astrophysical Observatory Bastumanskaya, it's on the mountain. Telescope stands, naturally, on the top of the mountain. AND We once went for a walk surrounding area, went down into the valley. Do you want me to guess? We got caught in the fog. Into the fog To be precise. And we got caught in the fog. Dark fast. This is the south. Everything became dark. And where should we go? go? We are a wild place. I say, "Listen, We had a method for mathematical physics gradient search a maxima some function." She says, "Well, so what? what?" He says: "What do you mean what? Let's "feel with your feet." The coolest direction. Now you feel where the hardest part is walking. Feel. Here you go let's close our eyes and just move our feet wherever we want steeper, much steeper, much steeper. We made their way through the bushes, but did not lose this very task. And to look for the most cool direction. And boom in the dark came across a telescope tower. Note, the condition that Sergei had stated was fulfilled said: "The surface should have been smooth. If there was a breakup, if there wasn't this is either a gap or an overhanging one the rock is sharp, you wouldn't notice anything it worked. That, yes, yes. That is, it is needed. and he said, you need a smooth surface for good gradient descent. But yes, but yes, we need it, then, to come up with either such a format and it is important come up with a kind of function form, which will give such a reparametrization tasks, yes, what is the objective function in this the space will be the smoothest. This a person does, and the search for architectures, and A person can do this, but there is something in that including automated methods now search for architecture. Here. And then there is this too also optimization algorithms. Here. A so, and then, well, naturally, as would uh around this whole story uh very many interesting problems and questions It's growing, yes. starting from that, then, really, how to choose the right architecture neural networks, yes, that's which ones to choose threshold functions, yes, what, then, well here again is the human brain, it too, Well, like this whole model, it was inspired by human brain. We have neurons, which exchange with each other electrical impulses. Here is each one the neuron essentially performs certain operations on their own inputs to produce some washing it is a cell that has entrances and there are exit. And he receives signals and then on he gives a way out, he does something with them, does something, somehow puts together some resources. The universe whispers: "Subscribe." Is a neuron a basic cell or is it a small computer that is difficult process? It's more like a small one computer, yes. That's because that's all, of course, it's not very easy, because uh let's say, eh, how does it change, well, that is, it was clear already to Freud at the end of the 19th century. Freud, by the way, before he started to do his mischief, he was quite a decent physiologist, worked under the command of Gerloch. Here. And he wrote in In 1892, such a work without a title, which is now called an essay on scientific psychology, in which, therefore, he reasoned about the fact that the brain consists of neurons, which means, there are some differences between them contact barriers as it was, yes, it is those were just the years of neural development doctrines. That's just when Romon Kajal wrote his magnum opus, he He referred, among other things, to Freud. The point is that it was then, at the end of the 19th century, learned how to paint nerves correctly tissue under a microscope, yes, sonny appeared. Here. And so Freud reasoned, that there must be some kind of process, which changes the permeability of these contact barriers. He called it facilitation of contact barriers. And here it is Today we know what is behind this process there are six molecular mechanisms there known neurotransmitters are released, yes, but not only. That is, change ion channel populations, e, change patency of neural ion channels channels, uh, conformational change dendritic necks, about growing dendritic spines, a, i, and two more, which I can't remember right now, but I'm in the newsletter, if anything. And about what? book speech? And the book is called "Hunting" on the electric vehicle". The Big Book artificial intelligence. So let's give it to her so we got it at the last station brought. This is the very thing. For us too The conductor brought her. What is this so heavy? We even thought about how he dragged it. There are two awesome volumes here. electrician. This is what is called the Big One. book of artificial intelligence. Until we read everything, not today let's go our separate ways. Oh, this work is hard. IN in the literal sense. I lifted her up with difficulty. Listen, you can immediately turn in another direction direction. Artificial Intelligence Program - it is physically a computer and a program, which was written for him. Nothing more No. Well, sometimes they can be specialized machines, right? Analog, right? Analogue. Well, yes. That is, this is already a direct connection with Here it is analog computer. Yes, yes, yes. Well here, uh, the thing is that We have this with digital machines some problem. What's the problem? consists of? It's called bottled vonneumon's neck. The thing is that in modern cars, in most cases we have, uh, memory cells separately exist, and the computing core, computing devices separately. And in In the brain this is not the case. In the brain there is a neuron is also a storage place information and its processing. And here is a machine like this, Phoneymon's a car in which storage is divided data and their processing, we need data transfer from memory cells to the location where they are miscalculated. Results calculations return, yes, it is called tire. And here she is narrow neck, yes, slena. Yes, yes. That is What's the problem? Yes, there are memory cells There are a lot of cars, yes, but the registry the processor is not enough. That's why we need constantly transfer these volumes of data. And here, although the frequencies are very high, electronics, yes, it turns out that here this is a bottleneck, that's why it's for us, when we start simulating these things like biological neural networks, and we get a big fine, yes, because of that our electronics have to here is the data to run through This is approximately reasons why video cards are good cope with the tasks. For what? video cards are essentially matrix calculators that can operations with matrices of numbers at once to produce here even multidimensional They say tensors, but that's not quite it mathematicians scold us, yes, but in the field of machine learning calls this tensors usually, that's why they say tensor calculators, tensor processors. Here. Well, uh, in everyday terms video cards, yes, actually today specialized ones are being created there video cards that are actually for neural ones are originally intended, not for graphic tasks. Here. So, but they solve the problem only partially, because which is still the case even in a modern GPU graphic processor, graphical processing unit. So, that means, anyway, it has, uh, like, a quantity memory, yes, more than that, that's why there is something more than e than a number tensor computing cores. Here so, uh, well, someone has to wait, what should the queue be like? I don't understand, what are you talking about. There is a chemical here the speed of information transfer, and there the speed of light. But there is evolution here has been optimized. But we have a very a large diagram and a three-dimensional one, and there is a diagram flat. And, well, we have problems there, We have physical problems there. Here We need other electronics. Of course, yes, we come across physical limit. I was talking about Brumerman's limit and Morgolos's theorem Levitin, but there are similar things there. Like Landauer principle. He brings us one trouble. We do highlight it, yes. In case of loss one bit in the system is allocated some a small amount of heat. If we we want our car not to melt and it doesn’t evaporate, we cool it. Well, well, it doesn't matter anyway we cool it as heat removal organize fairly quickly. If everything company and everything in one place, try it take away thermodynamically reject three-dimensional computational matrix, in which can be immediately built in heat dissipation elements, yes, with three-dimensional connections. We don't have it yet good technology for such a thing. We we can make flat diagrams with very high frequencies, but they are by number elements are still far inferior to the brain. Yeah, yeah, yeah, because, well, that's it. look, I don't have billions there transistors. No, look, we have it, the problem is that the brain is 86 billion neuron, but a neuron is not just one binary element is not one e transistor. For us, only one synapse in neuron model, needed thousands binary elements. much more, an order of magnitude more than than neurons. Here. Well, it's rude, yes. But so the size of the scheme in electronics, and when do we start try to make big schemes, we have marriage occurs. Here is a modern one electronics, she ran into what? We already, uh, often we assemble an integrated circuit from chiplets, Yes? Because if we do everything pancake, that is how it is arranged, there is Damn, this is the backing, yes, on which it is printed there, yes, in fact, transistor structure. But, uh, if use the most modern point technologies, where is the size of the individual elements, this is a unit of nanometers, there the percentage of defects increases monstrously when production of such a system. That's why They are coming up with all sorts of tricks there now. Let's put it together from the pieces. chiplets, so called. Well, there We'll reject one chip, okay, we'll throw it out. Here you go, yes. Or we'll print it on the whole pancake the scheme is big, and we'll check all the blocks, which of which it consists, yes, and if some block is faulty, we are there Let's weld the entrance there, right? Well, that is to say let's make a scheme that will be bypass faulty elements. Here. But that's all equally uh, it's clear that we uh got closer to the possibilities of this one classical electronics. There is a huge direction which is called neuromorphic systems. Here are the ones are exploring a variety of alternatives different existing phoneymonian architectures. And they are also being studied other physical substrates that could replace the classic electronics. You are not too much you demand from the computer. There are tasks today, which a person solves better, than the coolest computer. Here you are can you solve this problem correctly without fail. In my opinion, there are no such tasks. within a reasonable time. There are, of course, Of course there is. There are such tasks that A person can do better than a computer, Well, I decide whether to punch him in the face or not. I think he decides better. In full seriously this is a task, you know, this is it, Of course, the task is interesting. Here. Certainly Well, people are still solving many problems better than cars. Well, for example, Well, for example, he writes poetry better, better paints pictures, right? Well, the best people better than the best cars. Well, about pictures of poems, this is not optimal I can not coefficient qualities, so to speak, to attribute to them. One likes it. there are no computers yet fat James, Joyce, uh, Kavka and and so on. And we tried, we tried. On in fact, in fact, it is very bad, as long as the car copes with it problems of the real physical world. That there is, for example, to make a robot, who can solve an arbitrary problem in physics world. But it's not just good that's needed recognize images, yes, and give control signals to the system, but it is also necessary do it very, very quickly. Well, that is, for example, when I know that this thing is going to start falling now, yeah, and I know approximately, but not exactly, mind you, I no, but I know roughly how it will fall. Why do I need to pull my leg away? Although I I don't calculate the exact trajectory. Computers are, as I understand correctly, for modern neural networks this is quite a difficult task? Here you go look, we have an automatic driver not yet. We can make cars, and we we make machines that look like this they drive along a specific route, it's good there maneuver and so on, but colliding with tasks as if out of sample, yes, with a stranger comes out onto the road, who has a stop sign on his shirt drawn, yes, the car stops in front of him, yes. Well, that's right. Human in front of the car. No, well, he's not like that. road, went to the side of the road, yes, on on the side of the road. And he has a stop sign on his shirt. Yes, Sergey. But still, look, but nevertheless even before that very much it's terribly interesting about neuromorphic, maybe this is the future we don't have yet we know, really, that it will work out come up with. Even in what we emulate on von Neumann architecture. We learned to make big language models. We train them. And look, after everything you told, from matrix multiplication they talk with a human voice, they write poetry, and they make intonations, laugh, joke, understand our jokes, draw videos on our order, pictures and so on. They, wait, I can't help but ask, I I just have to ask and you just will please insist that you answered. Are they conscious, intelligent or No. Although I know, I was told that there the matrix is multiplied only very a lot in seconds. Ah, well, that's work, so to speak. The brain is a physical object. We can too write with a set of mathematical operations, in theory to describe the movement there particles, yes, that make it up. If we are not Sir Roger, Sir Roger thinks, that consciousness is uncomputable. Oh, spinrose. Everything has been forgotten. We don't remember Penrose. Everythingeverythingeverything. I take my words back. Half Penrose. Who is Penrose? This English mathematician. Roger Penrose. Mathematics, Nobel laureate, time in physics for black holes. There is, well, for black holes, what does the brain have to do with it? Yes, that's it, we forgot, we forgot. Here. Means, about consciousness. Eh, here of course, the question is more about the theory of consciousness in the first place queue, yes. What do we know about it? in consciousness, then? What is your opinion? They consciousness, but in general, then no, Of course not yet, not yet. I'll explain, in general, why. And here is the evolutionary one the theory of consciousness, it is in general in general converges to what at the moment time? That's where the development is in history social beings appeared in the biosphere, yes, that is, those who live in the environment, a significant part of which is representatives of the same species. And for successful adaptation arose the need to predict behavior other representatives of your own kind. And so a mental model emerged another one, right? That is, at some point She found herself closed in on herself. That is I got this idea of myself as if about something else, right? I'm like someone else. Here. AND here is consciousness, it is the result this short circuit, unexpected closures within themselves. Well, yes. Not, no, well, I guess it's like, well, evolution did not plan anything in advance, but that's how it turned out, yes, that's what it turned out to be, that the model itself appeared. Here. Ah, but, uh, so that's what we see here, right? Here it turns out that in order for this to happen kind of system that has this consciousness, arose, yes, we need the process, eh, optimization, yes, it required this system uh well that's basically it forecasting the behavior of the same systems, the same kind as herself, right? And emotions here have some kind of meaning biological, yes, they have their participation too emotions, yes. No, no, the machine can do it recognize emotions machine. And she herself can show emotions. That's why I say it, I have this Alice on my desk, and here it is my granddaughter sometimes tries to offend her, says different bastards. Well, here's a child, yes, conducts an experiment, researcher. Alice is not offended, she is trying answer correctly. Good educational samples were not studied in such cases. A there are those who are capable, listen, I don’t want to I grew up talking, I'm sad. For what are you trolling me? Yes, of course there is. Yes, I, frankly speaking, am not that big emotional domain meanings and for some reason they like to talk about it, that that's how Sergey is, the emotional area - this is the very survival among yourself similar. I read, I read evil in his face and decide whether to leave or fight. Emotions, look, we need to not, well, it is important to separate the emotional coloring there are voices, facial expressions, there is whatever, but another matter is subjectively experienced feeling, huh? But subjectively the sensations experienced, they can arise only when it arose reflection when that very thing arose consciousness, right? Well, that's what I need, damn it, I'm sad today. This morning is bad mood, sad in the morning, right? That is I kind of reflect myself, right? That is me as if I perceive it with the help of the psyche the state of this very psyche. Yes. At your place physiology inside, there is no cooking stomach, the computer does not have this. He is not will be sad. Here is the problem in that to consider to be sad. That is, as always, yes, here we are uh we're counting, the machine made a mistake, yes, we we give some kind of feedback signal, to adjust its weight. Can say that it is the same as pain. Wow. And what? Can. Well, up to a certain point comparable things to some degree. You jerk your hand away from the fire, right? Feedback works. Yes. Here is the feedback. Well, no, well. you know, when you do this, I It's amazing, I realize that I am myself I contradict because I am completely I agree with you, but when you say such, I say: "Well, no, the iron doesn't It might hurt." I understand that the living creature is in pain, but there is no iron. A decide what is painful? Here I am I understand the contradictory nature of my position, please. Therefore, Sergey, let me clarify. Not, I just in this sense, well, as it were, there is the problem is uh quality, yes, that is in general the problem of how it relates physical processes in the human in the body, in the brain, in whatever, with the subjective world, yes, with ours subjective worldview. And this, well, there is a very difficult question there, yes, because all we have is this correlates, right? That is, what are the correlates? what do you mean? Here we are observing some changes in the physical system, and behold, Well, the guy is reporting there, saying something about your internal state. We, uh, are there. we perform surgery on an open brain, here here we put the electrodes, yes, and The man says: "And I am there now, happiness I feel it, and I'm there, uh, I don't know, now I feel pain, and now my ear is there itched." Yes. Here. Eh, that means, well, Somehow it turns out there is a connection, right? That there is between the physical object itself, yes, and the fact that we are in a subjective world. Here. But it is clear that, uh, it means, uh, that is this such a consciousness? Try to find him in a separate element, right? That's how it is there no, right? An individual neuron does not possess consciousness, right? That is, just a system, yes, she has this emergent thing there property, yes, there, if you like. Well, that's all. they say emergent, but still and they say that when we trained a large language model, something arose in it, I will repeat my main concern. They say that, well, experts say that it arises in it something she was not taught. And this, damn take it, it's amazing. Well, no, no. Well, listen, what's the problem with creating, but here's a dice, well, you're any the system is capable of generating random pseudo-random state, it creates something new, yes. Why wasn't she taught? Here you take, eh, 100 playing cards cubes, just like that, suck them out of baskets on the ground, yes, and you will write down the numbers that fell on them. This the combination of numbers has never been seen before met. It's brand new, right? You you can match them with the letters of the alphabet, write the text and you will be confident, what will it be it will be a nightmare. Here Now the question. What's the difference there? texts that are good and useful to us, Yes? But just random ones. Ah, and here is my namesake at one time Andrey Andreevich Markov, yes, he took it. Evgeniya Negina, yes, these are the Markov Chains, right? Yes, yes. Here is a model for you too statistical machine learning. And he started counting frequencies, right? Means, frequencies of what? Letter frequencies in depending on the previous letter. Here you go us after the letter A with what probability the letters A, B, C go. This is a feature language. Yes. Yes, yes. Here. And that is language, a statement in natural language, it represents some a pattern, right? That is, we have the following letters, yes, depend on previous ones, yes, in the ones that are interesting to us statements, yes, that's what people they say, they write in books, they publish in the Internet and so on. If I subject statistical analysis all these texts written by people, yes, we we will find that they contain statistics dependencies. Only they are next context from all approximately approximately approximately from all the previous analysis of that, something that has already been created by someone. Is it possible to based on this analysis to create, here it is and speaks happens with the language on on the basis of this it is compatible with the game cube. Here we are with the help of neural networks predicted the probabilities of the next letters in the text. Now let's throw it away playing cube. Let's simulate this distribution. Yes. Here we have it continuation. Neuronka predicts that there is a 99% chance that the next letter is there or there's okay, with a probability of 50% the next letter A, with a probability of 10% the next letter is O, there, yes, with 1% probability soft sign. Now we are proportional We throw a dice at these probabilities cube, right? Well, that is, which one the probability of getting a face is equal to this this probability, which predicts neural network. And now we take what fell out. We get the next letter. And so she speaks to you in a human voice voice. You say to her: "How are you? today?" She says: "I'm doing fine, and where were you just assigned probabilities with the help of our this the formula is big, right? At the exit of it probability distribution". Now we from this probability distribution we choose proportional ones. I was sure, that she has the phrases written down. No, she generates approximately starting from the first letter. Every time, every time, does the next one, remembering the whole previous conversation. Well, modern ones in about half an hour, yes. So when you talk to them for an hour, they they start to get stupid. Well, the length of the context this is a separate interesting story, yes, but just a block of self-attention, the so-called building block of modern ee non-growing transformer models, he has a quadratic computational complexity from the length of the context. That is it has the number of calculations that it is necessary to calculate it, it grows as square of the context length. Well, glory be to you God, it's only a square. Here. Yes, but there, as they say, there are always all sorts of things computing tricks to make after all, this square there is a little smaller. And there are varieties of specially sparse ones attention blocks that, well, sort of they don't look at everything, you know. OK. Well, Apparently people do the same. We don't We remember everything. We remember that this is some reference points. We have the same thing the mechanism of attention, yes, it is no coincidence that here is an abstraction, which attention, yes, that when we translate a text, for example, yes, from one language to another and when we write the next word in translation, at this time we are looking at some word from the original text, right? That is, our attention is not to everything follow, look, but they are because of this, because there is variation, they hallucinate, they get into that area, into that area and that They say. And wait, and yours too colleagues say that this is what distinguishes them from living people, because they start to hallucinate. They got into the area where they generate deliberately not as if grammatically correct text, but they do obviously impossible things The statements are incorrect. And then the dock, when you say: "Why are you like this "You think, yes, I trained." They prove why this is rightly disgusting. Try the student on the exam to pin someone against the wall. well with the question of the answer to which he doesn't know. He will be yours later too justify why his answer is correct. So far we have only talked about language programs, and artificial intelligence - this is an automatic driver, yes, and solver with language. Look, a trick. Which. Well, with language, language is the formal system is complete, yes. In fact in fact, any intellectual task can be solved present with the help of natural language and conditions. Ah, well, for example, let's imagine. And on in fact, any task at all intellectual can be represented as the task of continuing some text. Class. For example, for example, look here, you have this text: apple dash Apple, table dash table. Yeah. Now if you could continue such text, then you can translate words from one language on the other. You can also have this text introduce. White's position. And the king E1, Rook E2, Black King H8. White's best move is a colon. And if our model is smart enough and it can continue and such texts sequences, yes, that means it is can play chess. Play in chess, yes. Here. Can I have a picture? present also describing the color of each pixels, for example, yes. deification car turning, steering right, to the left, speed, and so on. So, Certainly. That is, we have reactants actions are the way out, right? That is, here it is should i press the brake now? You you are saying something amazing. You say: "The whole world" from this point of view is, damn text. It can be imagined like text, right? One can imagine to rule the world as an extension, you said an important thing, as a continuation text. From a human point of view the whole world is electricity for the brain. That there are all our sensory systems, they turn everything into electrical signals. And feedback - these are, well, also electrical signals from brain that control contractions muscles. Here. And this is what they taught, no, I want, I want to finish it, sorry, for God's sake, the fact that she doesn't surprise you, that she was taught there to just read everything texts, and then she starts doing from there some conclusions about how things are there prove the Pythago theorem, well, no definitely theorem, something more difficult, some medium difficult starts solving mathematical problems, although she never, these tasks never I haven't seen it. Where does this come from? That's right what if I trained her on this one sample, it would seem that this would be the average in the hospital, any, and the result of the training on everything - this is very good mediocrity because it is being taught somewhere in between. And sometimes it seems that these weirdos, these neural networks, large language models, can do something to produce something that didn't exist before. Here. A you are here some very some very did you take a stingy position? This is very just produce something that is not was. It makes sense, yes. Well, look, we, uh, what is a large language model modern, right? So, we hit, we take lots and lots of texts, here we have collected all of them the internet, everything that lies in the open access, it worked five some. then they are still in the public domain read and learn from this and then copyright holders complain to the holders, who go around the internet, they, well, you know they don't know, well, if the page opened, opened, we downloaded it, right? That is, Well, in that sense, unfortunately, we don't we can, even if we really wanted to, yes, review everything that was downloaded there. Well, OK. So, we have collected the entire Internet, deleted duplicates, uh, so deleted some kind of trash, yes, with the help of there some set of algorithms, it doesn't matter, received a corpus of texts, e- all all about everything, right? So, here is its volume a few there, maybe dozens a terabyte now, right? So, uh What do we do next? We can't, of course, as in the case of cats, yes, to plant a person, yes, who will say: "Here is a good answer, What is the answer to the problem here? What is the task? So, this is the problem we formulate. So, let's learn how to predict. the next, well, for just say, word according to the previous text. Text again. Fine. Just so you understand. She will be speak to you in human language, with intonations, chuckle, joke, say you have a green shirt in white border, that you are sitting against the background of this myself. And if someone's hand sticks out you are so worn out at this time about she's discussing something, I'll ask her later, what happened during this time, she will say: "I stuck my head in, the conductor came out, and entered, patted the one sitting on Let's do it for now For now, let's talk about the texts. In fact this is also true for other modalities interesting, but in the case of texts, yes, so, in the case of texts we teach in At first the model simply predicted the next word is based on the previous one." Simple task, right? So, here he was making a stern noise Bryansk. Who means with what probability, right? Maybe a dog, maybe, She can write a book. Here. And I even I think it's her. Yes. Look who! wrote this big book. Here written by Sergey Markov, and written below book of natural intelligence. So whose? She, Markova or intelligence? And here it is People are not much different from machines in that sense. We are also born, as it were. How [music] Here. But the cars helped me search spelling errors, Here is our neural network that was used. Again, let's distinguish from the texts, Guys. Let's just finish the story now I, yes, to finish talking about how where does the ability to do something come from? do. new new great create to create science to do that means uh and us what we really want to do is we want to get a formula that will predict the next word in all texts created by mankind according to the previous words but the number of parameters in this system it is much less than the number of these texts than the number these words, that is, in essence, in the process optimizations within this system will be to be produced such to be found such patterns that will allow minimize the forecast error. This means that they will be created there high-level abstractions within them, Yes? That is, it means, and exactly the same as when we distinguish cats from non-cats in the pictures, in fact, yes, that means, Our retinal cells do react just there for the lighting, yes, then deeper cells already react to slightly more complex patterns that are a combination of simpler ones, yes, gleonary, yes. Here. Then, then, the visual cortex is connected, yes, through optic nerve. It's the same there. How the deeper we go, the more abstract signs will be there to be formed. Yes, this is Pavlovi too said that the cortex is an analyzer. Yes, and a large language model, in fact in fact, solving this problem, how to cram unfitting, yes, that is, like uh and here this number, the parameters that are unique yes, to describe all the diversity, yes, it is actually produces such internal abstractions, which then prove useful for solving new problems tasks. Amazing. This is similar to that, how do we teach a child at school, what is there the bunny and the squirrel went into the forest, right? Bunny I found three mushrooms, the squirrel found five. We are the same we understand that in adulthood at this age it is unlikely that he will ever will meet a bunny and a squirrel, who sums up the mushrooms, yes, and here they are. But that, what did he learn in the process of solving solutions to different problems, what is it inductive operation of addition, it it turns out that it is now applicable out of context. I have great doubts. If you will learn on a huge body of texts this program find the next word after the previous phrase there, then taking the phrase: "And could you play a notturno on "flute?" And what should be the next word? water-eastern? From where? Well, no, well. this is full of Mayakovsky in And she just I would have found it. No, no, no, wait, wait, wait, wait. No, Sergey can. to do when we get home, when we get there, when we get to Mars, Sergei can do the next task. Take computer, feed it to her, take the net, a large language model, feed it only Pushkin. And it will be better Pushkin. She will write all the unwritten remember the poems. That's not what I'm in right now, It seems that I understood Alexey's question better. Let me explain. But we can set it up like this generation parameters, which will not always be the case the most probable answer is chosen, right? U and that is, as I already said, yes, we are we throw a dice to get the result from this sampling to choose, and the probability words uh pipes there small. small, she's rather big just because there are many times Mayakovsky ones are here already was yes it's not 100% anyway, it's not our task search and the task of creativity creativity she you no you contradict yourself you you want to repeat what is known and speak what is this creativity no no I meant what if she didn't know poetry She would never have finished Mayakovsky why did you add this line, it's brilliant this and no I think that's exactly it She could well have figured out what was going on here problem. Uh, I'm saying it again, uh. A dice can create something new. The question is how to create such new, which will be for people there I like it, it's high quality, yes, yes. How manage this generation? Indeed, there is a lot of garbage on the Internet, yes, and a lot of bad texts, yes, and a lot good ones. So there are a lot of tricks for this. For example, what is the difference? good texts from bad ones? Yes, we can. write Author: Alexander Pushkin. Yes, we can, we can in the left context provide some clues for models that will control it generation. Actually it's called promt eng engineering, of course. Here. That is we thus, as it were, attract the attention of the model we draw attention to some of the essence a subset of the representations of its internal, yes, which she has there associated with such texts. Here. AND due to this we can, as it were, write instructions on the essence of what is being done. But in fact, at the end here I'll tell you more about this model now. At first we taught them to just continue texts. We still have some there at the end the standard case of such dialog, yes, where right here there are questions and there are good answers, complete, detailed and and so on. And we are at the very end after of how we pre-trained there trillions of these texts, just a little bit more we are finishing our studies on this target building, which is needed to slightly fine-tune the model at the end, train it her, so that she leads, well, the way we want need to. Here she is, when she is learning more, she relies on everything the abstractions and ideas she learned, and that's why she doesn't do it clumsily, not just remembering what's in this set examples, yes, but as if understanding it. Look, so far we've been talking about harmless tasks. Playing chess, writing poetry. But that doesn't help much our life depends greatly. But after all artificial intelligence, that is, this complex programs now, but they will be they won't, but they are already entrusting driving cars, and the other day some train I started running around Moscow with art, well, with the driver automatically, well, medical diagnoses to begin with diagnoses. And soon she will begin to control nuclear weapons, because a person thinks for a long time and does not makes the optimal decision. And this is not it's dangerous, is this somehow legally allowed? Look, I'm right here, of course, you can need to. And for this, well, there must be technical regulations. In any there dangerous systems are created risk management associated with their exploitation. But here's what I want pay attention. Isn't it dangerous to leave people with acceptance? some solutions? Look, this is ours. technological power is growing at a tremendous pace. Uh, if we 100 years ago would have wanted to inflict irreparable damage to the planet's biosphere or in its appearance, yes, we didn't have such tools, right? Yes. Nuclear power has appeared weapons, the stakes have risen very sharply, Yes? Today, it is not only nuclear weapons that We have and uh biochemistry, right? So we do too we are progressing very quickly in this direction, right? Potentially our technologies are here are becoming more and more dangerous, Yes, dangerous. And ours is like from the movie about Spider-man, we know that with a big Greater force must come responsibility, right? That is, skill It is wise to use this power. Here. Ah, but our brain is biological, it’s, well, not progresses as fast as ours technologies progress He, yes, is changing. much slower. Here. And, uh, that means, here, uh, I'm saying, like, the apocalypse is getting cheaper, right? Ah, uh, so, uh, a club in the hands of a monkey it becomes more and more difficult, yes, but how do we solve this problem, right? So, in this sense, too, technologies artificial intelligence is not it is the same as any other the tools we create to compensate their own biological imperfections. We don't have sharp teeth, claws, etc. further. When necessary, we take peaks there, knives, everything we need is there. We don't have any thick skin with thick fur. When It's cold, we put on clothes. Yes, no We have enough natural intelligence. We we take the tools of human enhancement intelligence. And artificial intelligence - it is nothing more than technology, which humanity uses in order to to expand the possibilities of your intelligence. Look, this person has How so to speak, our biological of reason, there are several fundamental e barriers that are difficult for us to overcome. Well, for example, speed, right? Why in in the brain, the length of the impulses that neurons exchange - this is a unit milliseconds. That's quite a lot by electronic standards. Why is that? Because that the system is electrochemical. There is ion channels through which it is necessary push ions. To make the ions quickly slipped through, a big difference is needed potentials. But if there is a difference in potential more than 1.27 Vl will stop, the water starts decompose into hydrogen and oxygen, and the brain can go boom, right? This is thinking It doesn't really help with such a small amount potential. Yes. Yes. Here. Yes. 1.27 V and It is interesting that approximately 1 V is also for semiconductor elements too working voltage. Yes, it's interesting. So, the speed is because of this, well, something is wrong with the human brain. Here. Next, uh, then, uh, People are quite fragile creatures, yes, we are we exist in a limited range temperatures, accelerations, concentrations oxygen and so on. The vast majority of places in the Universe will kill us. Yes, we need it very often solve some problems in these aggressive means. Yes, already on Mars The Mars Rover is running around with quite a cool strong intellect, cool, I would said. Here are the sensory limits. Yes, we can't, we have a pass The capacity of the sensory cortex is limited. We can't have seven films at the same time. look. Although it would be cool, Probably seven times more fun it would be possible to obtain, yes. Here. Means, further problems. Uh, so some we don't want to solve problems, yes, that's the problem motivations, yes, there are some intellectual tasks that people They just don't want to decide, yes. There is a scaling issue. Well, for example, I don't know, do you need it tomorrow a call center of 10,000 operators to call, I don't know, all the people and tell them there, I don't know, about approaching. You persuade us that this is a new tool that requires intelligence manage, and then it will be safe, got used to it. I'll say more if we are now let's say, like, that's it, we're imposing a moratorium on the development of these technologies, We will develop everything else, but We won't do that here. And this is exactly what is very dangerous. Because it's dangerous. This dangerous, yes. It's just because, well, you know, a person also presses the red button press maybe. And the man who, for example, will know the consequences of his possible actions better thanks artificial intelligence tools, which will be used there comprehensive analytics and so on, uh, he will already act more safely, yes, in this situation. Today many systems security, they are based precisely on automated systems that A robot uprising - that's it fantasy? Fantasy. Well, uh, I think it's fantastic. You know, I like this one the most comic artist Nick Aragua. Uh, so he has, there is such a caricature, there, then, such a servant there to the master: "Master, robots rebelled, and there outside his window There are already robots like this with villas there. Here. And so on." He says, "So what, "What do they want?" He says they want, so that we don't use them as such anymore metaphors of our class and national conflicts. Here. So, and, Certainly, the story of how people were created rebel against themselves. Talas, Pandora, golem, then, does not exist examples of all this. Where does all this come from? did it come about? Uh, well, I understand, because for very often technology is a human a conductor not of his will, but of the will of another a person, yes, who has other interests, Yes. Well, in this sense, well, uh Here's what the killer machines say. Here. And what about? We are now people in cars, let's say, kill other people? Yes, we have been for a long time they said that an anti-personnel mine is a machine that is allowed to man kill. Here. Well, this is this car. primitive artificial intelligence. Please. Here. So, this is, of course, uh, na na I look at this problem this way. That is it powerful tools that we we create. Natural enhancer intelligence. Yes. Yes, yes. And we, naturally, we can, among other things, use them in harm, and all our shortcomings will simply be are reinforced by this system. Yes. we will. Apparently, we must hope that in first of all and for this reason It is worth trying to strengthen ours advantages, and disadvantages, on the contrary, compensate, smooth out or to be toned down, if it works out. Nobody said it was Maybe. I said, I would say, hope There are two things here. Well, first of all, that these technologies can not only attacks, but also for defense be used, yes? That is, in this too I mean, that means, [music] uh that's good news. Yes, maybe perhaps even better news is that if the system is smart enough that it's called, yes, that's it when taking steps that mutually assured destruction they will bring, for example, she will not be, yes, if it's programmed, well, if She knows how to calculate the consequences well, yes, the actions of those or others, then, uh, in this I mean people, they, unfortunately, sometimes, ah, they wrote in the newspaper that we We can bomb them and they can bomb us they swallowed them. Well, a person has fear. own death. He often doesn't is taking some aggressive action actions, fearing to lose their lives. And at The computer does not have this fear. Big problem from my point of view. Well look, we create these systems and We formulate goals for them. That's why is it possible to create a machine that just will destroy everyone around? Well of course Can. Well, you can. Well, for that reason, you know, artificial intelligence isn't really needed, right? I mean, well, he took the bomb, like They say they set it on fire, that's all. No, no, no. I meant something else. Not always an optimal solution in some sense it is for the benefit of others. There is tension there the network weakened a little, and the program thinks: "So, what needs to be turned off so that there is no way to lose power supply nuclear power plant?" Oh, I'll turn off the hospital, please there. Well, that's the most the little that I can. Well, these are our goals must, yes, formulate, yes, naturally, no need, yes, and here, how is it It's not paradoxical, but in the era of e such a fast technological progress uh especially big requirements, they are presented specifically to human qualities, to humanity, to our responsibility, ability to use this power for good, and not to the detriment. Here. And now is the era, which is very humane is being tested. Please tell me, maybe one person, a cool programmer, write such programs or is it always remedial work and you don't even know it how many people took part in it, because that you assemble programs in blocks It's impossible to do it alone, right? Well, of course, This is not done from scratch, right? I mean, well, software, yes, no one has been around for a long time doesn't write everything from scratch. Well, that is, for with some very rare exceptions, there, with low-level kinevers and so on then, of course, modern neural networks are developed on the basis of ready-made languages programming, ready-made frameworks. By the way, what languages? Well, Python. popular now, yes. Here. But of course, It is important to understand that everything high performance computing, they, of course, it doesn't happen inside python, but in A that is, the translator works and then at the code level, yes, all of this is here. Yes. Here. But overall, well, there are popular ones there. frameworks for neural network development networks, there are Porch, Tenerflow, JКС. Here in general, of course, when people are there are afraid of some technological failures that may lead to troubles, but it's quite there justified fears. I want to say that very often people are there under under impression of Hollywood film productions, like Terminator, they start worrying about some rather unlikely scenarios. And here it is at the same time, the real risks associated with the development of these technologies, they are very often end up out of the spotlight. Vodka. Well, there is, for example, this the so-called problem of digital secrecy court. So, our society is in it algorithms are generally like this automated systems are becoming more and more are gaining more importance. You come to the bank for a loan, solution accepts algorithms, yes, looking at your personal data, history and so on further. You come to apply there. work for some mass speciality. Your resume too the machine analyzes, yes, which used by HR specialists. And different accounting systems are used, production management. The solution to many of these systems is life people can be influenced very strongly. Well, for example, I don’t know, they didn’t give it to you there credit, yes, you have, I don’t know, so, well, and and so on, right? That is, if it were a person and you would tearfully ask him, maybe he would have mercy, right? And the car Clear, all criteria. The power of influence, as it were decisions of machines on our lives, it sometimes comparable to the court's vertical. Well, yes. Here. But in court, for example, yes, the court is in fact, the whole system of jurisprudence - this is also a set of algorithms, yes, if so figure it out, that's exactly how it works, without emotions. Yes, yes, yes. And the goal is precisely such, yes, but in court we have the right to the adversarial process, to access, for example, to all those gathered against us evidence. Well, yeah, uh, so, on a lot of things, yes, on a lawyer, on qualification of legal assistance. A when we are dealing with an algorithm like this of the kind that belongs to the corporation, which is its trade secret, proprietary algorithm, who knows, maybe to be, uh, by mistake a person who knocked down my age, yes, instead of 43 years printed 443, yes, and the machine model I decided after studying that I wouldn't live long already, well, at such an age, yes, and I'm unlikely to be able to repay the loan, so I don't need to be given a loan, yes, but how can I I can check this, yes, there is none there is no mechanism institute, yes, which would protect ours rights in such situations. Here, uh, some time ago, I think it was about a year ago a long-term trial has ended, so which, uh, so it was a trial between the British Post Office and its employees. Yes. So these employees at different times of their work, so, you received notifications about shortages in the units entrusted to them, yes, about what it means to lack money in at the checkout, yes, that means some amount people lost their jobs as a result, some number of people fell under court, some number of people were marriages are broken. And what was the reason? So, the reason was in the accounting defects systems, in the algorithms that are in it were. Here. But since the software the provision was opaque, opaque, yes, that means, unopened, the company, well, has been claiming for years, what, well, it's a machine, a machine can't to make mistakes, yes. Well, I'm exaggerating there, of course, argumentation, but in the end So, it was possible to prove that and find these mistakes that were behind this. But human destinies were destroyed, and there was even a suicide, yes, which, therefore, presumably connected with this whole story. Or, for example, there in the eighties the sensational story with this medical emitter 25, which is several gave lethal doses to patients irradiation. Yes, you remember now individual unique cases. And how much? cases where people drove others to people to suicide? That's exactly what I wanted to say, yes, that what's really more important to us always not so much, so that a person or a machine could take it there decision, how much is what kind of profile there security, yes, what statistics. Machines have the potential to be even more high level of security in a number of processes to ensure. At the expense of what? For due to the fact that we, for example, such We can test the system millions of times. Yes, we can't, we get into a taxi to to the driver. We can't before we we get into his taxi, force him drive a million kilometers. Is it true, Guys, another example from my recent days. And I'm riding in a taxi and I see that I'm tired the driver, that here he is wiping away the kiss: "Well, I'm going, I'm going, what should I do?" And suddenly some traffic light when everyone is standing at red, friend green lit up, that's it: "Let's go, we're standing." Well, what's the matter? I looked, he was sleeping. Only a person could fall asleep like this, Yes? The computer wouldn't fall asleep. On the other hand sides when I am driving and There's a car with an automatic transmission in front of me driver, I just admire how she fits perfectly. All the rules, all the nuances She complies. I am, of course, much I feel more relaxed about the car computer than to a machine with a live human. This is already a fact today. Just a fact, yes. Well, here we go again some borderline cases for now, yes, There are always such cases. And just in on in the relationships of people we We encounter them all the time. But for us it's pretty clear who's to blame, uh whose fault is it, who missed something. AND we feel better somehow because we know, well, relatively speaking, who will be punished, from whom they will take a fine, who and, perhaps, excuse the expression, they'll put you in jail unpleasant cases and so on. But here at We have a problem with the transfer responsibility. And we are social, and we we started with the fact that we are social creatures. We have a problem with the transfer responsibility and its distribution among their own kind. Here she has to distribute responsibility somewhere else put. Who wrote the program? Well that's it, let's get everyone from all of you. Well, of course, this is a complex question and not by chance there are questions related to transport safety legislators they've been doing it for so long. Yes. Here. Well It's clear. Uh, there's a general approach to it, I can look at it differently, yes, but there is some kind of, uh, complex instrument, Yes, it is a complex system, as a result exploitation of which arose human sacrifices. As we have a similar elevator system, for example, here The man entered the elevator. The elevator is also automatic driver. He just drives up and down, yes, on its own route small. Here. And so the man entered elevator, the elevator fell, right? That is, who will be guilty, right? figured out how to do it, yes, because there are certain systems elevator certification certification, yes, there is regular maintenance of them, there and etc. etc. etc. Here, accordingly, in the emergence of such a situation as it turns out, well, they go and figure it out, were Have all necessary measures been taken? all necessary measures to prevent occurrence this situation. If someone didn't accept appropriate measures that he should take were required, then, well, that means, what for the work of artificial intelligence should answer naturally. Naturally, and what I hear that cars and help from artificial intelligence is a given from which we cannot escape. And what we need to do is, like it is known if something is not allowed resist, it is better to lead it. On in fact, it is best to understand this understand a little bit about him, in that including thanks to the fact that we are today learned in order to, if possible, consciously and skillfully use and so that it is possible it was for the good. Look, we're already approaching. to the station, yes, we probably need to talk will stop, but I think that it will be possible to delve into this problem, taking this book in my hands. She is like that It's heavy, it's written artificially intelligence. But the next thing, Sergey, following. Well, admit that you are no longer you will write it yourself, and you will issue it issue a book written by an artificial person intelligence, for his own. We are not for anyone we'll tell you if you confess now, after we get off at the station. Thanks a lot. Thank you. Huge Thank you. Dear friends, please, so whether they are electric sheep or not, put it Like for if you want to put like for the electric sheep, especially not Don't forget to subscribe, turn on reminders and so on and so forth further. Take care of yourself, be careful to your own intellect, that of others and artificial intelligence around. Total good, goodbye. Our Universe and plus you are a great force in general. Knowledge is power. All the best, Friends. Bye. Goodbye. [music]