welcome to asia, where the heart of the continent beats.

the main holiday of russian graduates is universally recognized as one of the most beautiful holidays in the world. ale sails is taking place at these minutes in st. petersburg. there is a big concert on the palace square, and in about half an hour or so, all attention will be focused on the water area of ​​the neva, where that same legendary brick should appear. you can watch the full broadcast on the russia channel. our correspondents alexandra are also watching what is happening in real time perfilyeva and dmitry akimov. they are joining our broadcast now. hello, dim, first of all, how is the concert going, what emotions do the audience have? yes, colleagues, hello, well, i think that about the emotions that this concert brings, it’s better to ask the direct participants, the graduates, guys, how are you, let’s talk to someone, hello, what’s your name, hello, i sofa, sofa, how’s the mood, how’s this concert, what do you think, damn, this is off the charts, this is just amazing,

well, this is what we have now the situation we have now on the admiralty embankment is incredible, because in literally 36 minutes the most long-awaited event in st. petersburg will take place, of course. the same

sailing ship with such a native name russia will enter the equatorium of the neva and will pass a new route this year, it will pass under two bridges, first it is under troitsky , according to... tradition, so to speak, then under the palace one, we will tell you now, guys, now we will also show you, under the palace bridge, it will turn around there and walk along the peter and paul fortress, all this to the sound of volleys, fireworks, this is the largest technical show in europe, and of course, let's talk to the graduates, guys, unexpectedly for them, guys, but we were waiting, yes, of course, what emotions, very exciting.

akimov, told how the graduates’ holiday is going on in st. petersburg. and we’ll come back and watch the culmination of allah sails live at 0:40. now let’s look at

the advertisement, immediately after the question of science, about artificial intelligence. the sber-thanks loyalty program has been updated. pay with any sber card and get double more bonuses than before. every month, choose up to five top categories with cashback up to 10%. supermarkets, gas stations, cafes, restaurants and others. and even a category for everything. transfer money to sberbank from other banks for free and receive cashback of up to 70% from partners. more profitable! with a subscription from berp, one of my friends went to the wrong area, boy, there is a call, but i myself count zero, you have a megaphone, mega strength communication at zero is already included in the tariff for free, our subscribers will not be left without communication, word megaphone, alfabank

- the best bank for business, send all payments and transfers without commission, open an account for business in alfabank, for the first time on... and discounts for people from people, only until july 7 khvatamba on so many things, from cabinets to auto goods on avito. thank you sber loyalty program, updated. pay with any sber card and get twice as many bonuses as before. every month, choose up to five top categories with cashback up to 10%. supermarkets, gas stations. cafes, restaurants and others, and even a category for everything. transfer money free to sberbank from other banks and receive cashback up to 70% from partners, more profitable with a subscription with berprime. this is not the opening of the swimming season, this is a promotional code for

100% cashback from estrobka. book hotels and apartments and participate in the draw. investing with sber is easy, get three sber shares as a gift when you open your first brokerage account, deposit the best percentage in the sberbank online application. touch the rock star burgering water on fire tastes better with yota your money won’t fly away

, there are still big minutes left, we return rubles for them i wanted to open a deposit at 18%, it turns out you can at 20. you can open a deposit in one bank, it turns out you can open a deposit in different ones at once, for money you have your own marketplace, it turns out, open a deposit with a rate of 20%. financial services, financial marketplace from the moscow exchange. thanks to shopping at the megamarket, someone will reach the finish line first, someone will create comfort in their apartment, someone will go on a long-awaited vacation, sellers at the megamarket are changing the world, megamarket - just grow with us, business needs acceleration, for sure, open a checkout account on wtb.ru and transfer money to any bank instantly. vtb, we help business.

too powerful to be controlled. strong artificial intelligence from the pages of science fiction books is again breaking through to the front pages of major news publications. another open letter from experts in the field of artificial intelligence has been published. a group of specialists, mostly represented. projects in the field of artificial intelligence are expensive and require the attraction of enormous funds, so the company seems to need either funds from the state or funds private investors, because of this, any hype, any statements about revolutionary nature, will attract much more capital than general arguments about how great it is to live with... artificial intelligence, we need to create a common field, that we are better, that is, we have developed such a powerful artificial

intelligence, and even we ourselves are scared by it. industry leaders are constantly accused of playing with fire, and some of these concerns are understandable: artificial intelligence has made incredible progress in recent years. now he imitates people so well that he can deceive even specialists, and at the same time has gigantic knowledge and capabilities. what will happen if technology turns against its creators. here, for example, is the latest report from the work analytical center. for the us government. they believe that artificial intelligence systems - quote, can threaten humanity with extinction. the discrepancy between the tasks we set for artificial intelligence and what it actually does is the rule, not the exception. as intelligent models become more powerful and developing their ability to communicate with the world, the threats associated with this discrepancy are growing. what is strong adept intelligence and should we be afraid of it? how close are we to creating a fully functional digital mind? question of science, i’m alexander gasnikov, rector of enopolis university, and today our guest is alexander

krainov, director for the development of artificial intelligence technologies at yandex, alexander georgievich, hello, hello, alexander, well, the first question is, what is artificial intelligence anyway? oh, this one there is an endless amount of debate on this topic, it seems to me that we are only now agreeing on this definition. classic artificial intelligence is the ability of a machine to demonstrate skills. comparable to a person, that is , to demonstrate something that was previously unique to humans, but i would formulate it this way: artificial intelligence is programs, services, algorithms, whatever, created using some kind of automatic algorithm for selecting parameters and values, here is ordinary programming, ordinary composition programs, when the developer himself decides at every moment that if this is the case, then we will do it.

situations when one person is superior

to another, yes, but not in everything, that is, there are some very, very narrow areas where he is superior, because how long have we been superior to human artificial intelligence, what has been happening lately, recently time, two things happen: firstly, there are more and more areas where artificial intelligence surpasses humans, and secondly, unexpectedly for us thus, these areas turned out to be those that we had previously considered. as our absolute advantage, here again we remembered about chess, chess is a very complex game, it is the prerogative of man or some other games, such as the game of go, which are even more complex, check implies,

there is such a moment that, let’s say, before this moment they somehow didn’t think about it, and then everyone became, unambiguously, unambiguously, there is just such an absolutely clear moment when this transition occurred, such a quantum leap, it happened in the seventeenth year, when a computer just beat a person in the game of go, the company deep, which, yes, but they won at the beginning, this was the first one from fan huye, he is chinese by birth, but a european champion living in france. change, before this the computer played go , just why was it so serious , very weak, well, that is, the state has its own levels of skill, but roughly speaking, at the level of the discharger and above it did not rise in comparison with the leading people, there was an abyss, here at once such a leap took and beat, and it was thanks to the fact that they approached very well

convolutional neural networks, which were then very popular for tasks, say , image processing, they... were very good for playing go, and well , huge progress was really made in this area, and the deepmd company did a very good pr job, worked out this whole story and apparently this is how i imagine now that when they were thinking about how to put it all out there, imagine that machine learning will win against a human in go, well somehow it’s not very interesting, or even the neural network is already more interesting, but still some pepper is missing. here this could not have been more suitable: the term artificial intelligence, until that time it lived in computer games, when a person played with a bot, he said, there is strong artificial intelligence or not very strong artificial intelligence, determining by this the computer plays strongly or not, artificial intelligence sounded in a gaming environment, it fits right here, but a person

wins a game against a computer, that means the person wins the artificial intelligence, and if the artificial intelligence wins against the person into the game, this means the artificial intelligence won. anything, but somehow artificial intelligence is not used very much, but in the entire press, in all media, yes, now we have artificial intelligence everywhere, well, the example that you said, as far as i understand, there is still a lot of an important element was, well, let’s say, some self-learning, that is, that the program played by itself, in general, in your opinion - this track is

self-learning, it plays an important role in artificial intelligence, it plays an important role and - here’s what’s interesting, we slightly misunderstand what self-learning is, well, we are people, when we talk about artificial intelligence, we imagine some kind of program in ourselves that lives and learns, with it somehow it becomes better and smarter, well , actually this is not true, here we are people at the same time learning and doing something, if we play some kind of game, we are smarter with every move, well, presumably smarter, at least with every move we play somehow... something different. we have some kind of process action program that we have and learning occurs simultaneously, this is not the case with computers, a machine can perform the same action a thousand times, 10 thousand times, any number of times and does not learn anything at this moment, that is, we have some - a written program, this is very important to understand, it does not change in any way during execution, nothing

happens inside, a robot that i don’t know, will there ever be a robot that will walk around... there is a separate absolutely the process when it occurs, not just under control of developers, without developers it will not even happen, with great efforts of developers, when a certain mathematical problem is formulated for the algorithm, and the algorithm internally goes through the parameters until... they become as close as possible to solving this problem, it usually doesn’t work out completely, but it brings it as close as possible, but learning is a process of sorting through these parameters, what then is self-learning, if? we are talking about a game, suppose we have a task, as it was in the game, to guess the next move in the game, we need to take some set of real

games, preferably the strongest ones, because then we will guess the strongest moves, and guess the opponent’s next move , meaning any move at all, that is, we have some kind of game, in in this game in this position the next move was made, the task of the program is to guess what the next move was made in the game played , see. then her training is to guess what the next move was made, in this regard, yes, this is self-training, but i repeat, i repeat, a completely separate controlled process,

let's say, we have some kind of autonomous delivery robots, yandex ones, that deliver food, while the robot is driving down the street, it does not get smarter at all and will not get smarter as long as it just drives like that, a taxi here's a yandex taxi, come on, or an unmanned taxi , yes, but the car records some things. events happening around, not just some, everything, then on this experience, on these events, the learning process is launched separately somewhere on the servers, after which it is thoroughly tested, if it turns out that yes, indeed, it is better than the previous one, then the updating process occurs, that is , now new versions are being uploaded to all these robots, now they have become a little smarter, but on its own, if it is an autonomous robot or drives a car, then it does not get smarter in any way , well, we have now raised a very important question, this is indeed the so-called training sample, that if it is not very large, then there may be problems with the quality of work, now many experts say that

modern language models, they are already use almost the entire internet, or at least half, well, in this regard, we can expect that the rapid development of artificial intelligence will slow down in the coming years precisely because there is not enough data, this is the question that we raised, this is the question: are they related to self-study, that’s how you look at it, that is, to what extent, well , let’s put it this way, not self-study, but some kind of search for additional information, communication of non-native networks with each other can help solve this problem, i think that based on the data which are now on the internet, at using this data you can be much, much smarter, but we still don’t know how, we people haven’t yet learned how to present data correctly, to understand which piece of data is in the beginning and which is later. how much data should there be, because there certainly shouldn’t be perfectly good data, that is, the model should, in theory , see everything, well, in the end, in order to distinguish good from bad, but how

to teach, in what proportion, in what quantity, we don’t know, and we can’t get this experience from anywhere, we have some experience, understandable, how they learn people, but people still have some completely different hardware device, and it’s not entirely suitable here, very expensive, expensive experiments, that is, training one big one.

neural network, when it writes a text, it has some kind of model of the text, it imagines the relationship between the words, but it does not imagine the relationship between some phenomena that these words describe, watching what is happening on the video, it’s probably possible and learn this, and of course there is also audio information, but i’ll combine it here with video, because video can be full-fledged, audio and video, it’s not that it’s not happening now, it’s happening, this is a very interesting area. many people do this, but it’s incredibly difficult, because of course, text is a much more compact compressed data stream compared to video, well, the next wonderful stage is,

which is... we balance this bottle, we understand how the weight changes, when we tilt it, how the weight changes, when the amount of water in the bottle changes, and we do this somehow imperceptibly in background, we have this information, we take the next bottle, we are already trained, this information is not stored anywhere in any form, not digitized, it is too

simple and natural for us to digitize. we haven’t agreed on what is considered strong . we don’t yet fully understand, it seems to me, artificial intelligence, that is, what is happening, here they say that intelligence will be strong when it beats humans. ok, so he beats him in some tasks, we’re like, well, no, he doesn’t do anything at all not very strong, because he can do something in this task, but i don’t know, there is a program that beats a person in a goal, and a program that writes poetry or answers questions intelligently may not be able to add 2 + 2, and what this is intelligence, and then the idea appears that, well, strong artificial intelligence is a kind of general algorithm that is capable of solving various problems and being comparable to beating or not beating, but

being comparable to a person. in different tasks and act in conditions of uncertainty, but he was not taught some specific problem, and he solves it, demonstrates this solution, to some extent this is what text-generative networks show us, because the way they learn, they learn to continue the text in a logical way, they insert every time the next word so that it would be naturally appropriate here, i still can’t get used to it, i still can’t get my head around the fact that they sometimes do it this way... they even solve math problems, why do i underline the word sometimes because in most cases they issue complete heresy, but the fact that sometimes these cubes are put together like this, this random next word is inserted, that this turns out to be the correct solution to the problem, this is of course amazing, unexpected, to some extent yes, this is a step in that direction, but still, it seems to me that we are still very, very, very far away from strong artificial intelligence, that is, i

don’t share it at all. strong artificial intelligence has appeared, we must have such a program, such an algorithm, such artificial intelligence that can effectively solve a problem that he was not taught to solve, and what is missing for this strong artificial intelligence? and there is not enough computing power, and there is not enough data on which to learn data from different modalities, not just text, well, in the sense

that it’s not that the data itself is not enough, to understand that this appearance of each language model is super expensive, very few companies can afford it, and it’s not paying off right now, such investments in the future, in development, in everything else, but it’s very expensive, right now and any significant increase will lead to an even greater rise in price, that is, i think that we will also be stubborn here, we will slow down in this place, and i also have some kind of

consideration that... i’m trying to analyze it.