the wave seemed to roll in and the germans retreated to the so-called siegfried line, and this is a line of long-term fortifications that was created back in the thirty-sixth, in my opinion, the forties continued to develop, yes, it continued to develop naturally, on this line, so to say, the allies were stopped, the wave, so to speak, of the offensive stopped, and then

they said either to wipe out all of gaga from the ground, because they are mobile and you can’t exactly catch them there, or it means carrying out an offensive bypassing the siegfried line, come in, so to speak, from another country and move to occupy a number of dutch ports for supplies, because this was one of the big problems for the allies, they believed that they could reach them with a throw.

when - he gave the task to his paratroopers, he promised there: you will land, take the bridges, you need to hold out for two days, then the main forces will arrive, we did not succeed, in general this plan failed, heavy oncoming battles began, with very large losses, in in the end the americans stopped and all of them... plan to meet

christmas in berlin turned out to be unfulfilled; we already took berlin in 1945. we continue our podcast, today we are talking about the allied landing in normandy in 1944. it must be said that there were ideas, in particular from churchel , to try, so to speak, to win the race for berlin. including in the forty-fifth year already, but it must be said that the commanders and montgomery and senhauer, realistically assessing the resistance of the germans, in general, considered this plan, these ideas in general mythological, that is, there is also in some soviet historiography has the idea that on the western front the germans surrendered in entire units, and that accordingly the walk through germany was so easy for...

responsibility for what they did on soviet territory, but that will come later. in addition, germany was able to launch a counterattack, this is the famous offensive operation in the orders, and the order in december of 1944. esenhaur later said that the german command made a serious mistake by launching the varden offensive; it was an offensive of the desperate. this is partly true, but the americans lost 30,000 prisoners, only prisoners, the city of boston was besieged there, accordingly, and the americans a couple of times...

they bombed their own troops with missiles, but this is also a common situation in military history, fire on their own, but that however, this wardenach operation showed that the german army was still combat-ready and capable of taking action. counterattacks, and the counterattack ended with what? when the flying weather began, the breakthrough of the german tanks was again stopped by air strikes, well prepared positions of the americans, as a result, the german troops rolled back to their original positions. i must say that hitler really hoped for this offensive, he hoped that he would be able to push back the americans and the british quite strongly, then transfer the resistance to the eastern front, and... then, prolonging the war, still wait for a conflict between the allies. hitler already believed then, well , in addition to mystical hopes for superweapons, there were hopes that the allies would quarrel among themselves, that at the final stage of the war the united states and great britain on the one hand and the soviet union, respectively,

will come into conflict with each other. well, as a result, we can say that, uh, this is the normandy landing, it became like this for the west -

100 km advanced to a depth of 600 km, the scale, yes, naturally, this did not make it possible for berlin to transfer any additional allied forces to france. in general, an example of a combination of operation bagration and overlord is an example, a clear example of cooperation between the allies, which ultimately led to the decisive defeat of the nazis understanding the fact that the war is lost. many germans, even generals , understood that the war was lost after stalingrad, after kursk this conviction grew, a conspiracy against hitler began to take shape, which was carried out on july 20 of 1944, an attempt on hitler’s life, but that’s what finally pushed him towards it, and

that pincers came from both sides, on the one hand, the operation, which smashed the army group center, by july 20 it became clear that all that the front line had been broken through for a huge length was to be resisted. the only choice was to remove hitler and try to make peace on less shameful terms than otherwise, but the plot failed, history, unfortunately, did not follow this path, which maybe would have saved hundreds of thousands, and maybe millions of lives, but unfortunately...

with which, given the most varied ideological positions, nothing can be done, and of course , without the soviet victories near moscow, especially of course at stalingrad in the battle of kursk , the exhibition in normandy would have been simply impossible in the form in which it happened, but for the year forty-four it was really one of the decisive battles of the forty-four year, which made a great contribution to the victory of the anti-hitler coalition, and here is such a tug of war... of the rope, yes, attempts to prove that this is all complete nonsense, on the one hand, or on the other hand, that this is the main event of the second world war or 1944, both of which, of course, do not fit into historical objectivity in any way. in our history with the west there were not only periods of confrontation, but also periods of cooperation. stalin himself spoke at the tiger conference, i will quote. i want to tell you what the presidents of the united states did from the russian point of view to win the war. the most important things in this war are cars. i'm missing a fragment of the quote. russia can.

recognized and emphasized, you have already quoted his famous phrase about the significance of the normandy landings, when he said that this was undoubtedly a brilliant success for our allies, history will mark this matter as an achievement of the highest order. the fact that in the west they prefer to remember only the landing in normandy, what can you say, well, we shouldn’t follow this example, because a story with torn out pages is a distorted story. history always distorts the soul. this was a historical

podcast russia-west on the swing of history. today we talked about the normandy landings and the role it played in world war ii. pyotr romanov and sergei solovyov were with you. study history with us. you can watch all episodes of the podcast russia and the west on the swing of history on the channel one website. dear friends, the creative industry podcast is on the air, there are still living, real copies of the presenters with you, this is still elena kiper, producer and video director, and i won’t touch it yet, you need to make sure, roman karmanov. general director of the presidential fund for cultural initiatives, our guest, denis dimitrov,

our guest today, managing director for data research of the largest bank, i’m very glad to be here, hello, thank you for coming just like that, it started off so flimsy, because really artificial intelligence is a topic that everyone hears about, but very few people understand what it is, so everyone is already afraid, afraid, well, in general, of something, well... if we knew what it is, but we know what it is, we know a little, we use it, what artificial intelligence is, let’s first at least figure out what this thing is, in fact, artificial intelligence, and there is absolutely no need to be afraid of it, since we started with this, this a tool that helps us do different things, that is, it automates some part of our intellectual work, mostly the routine part, artificial intelligence - this is actually usually... they mean some kind of neural network, what is a neural network? a neural network is something you know, you can imagine it as a black box

that takes something as an input and gives something as an output, that’s exactly how we imagine it, like a black box, what’s inside the black box is a kind of mathematical function. the function that now, for example, lies behind, for example, generating a picture from text, has - well, there are billions of parameters, so they are so complex. what should a simple person who is now sitting... in front of our digital tv about artificial intelligence, have to pass, within the framework of this everyday understanding, artificial intelligence is artificial intelligence - this is such a black box, you can imagine it this way, which everyone simply imagined a black box, what to do with it, which learns to make some decisions, these decisions can be a huge number of problems, in fact, which it solves, that is, you take a certain volume data, well, let’s say, a certain amount of knowledge, well, relatively speaking, you take everything. the house of the tolstoy lion, you load it into this black box, and it can give you, well, conditionally, another volume,

as it were, of the tolstoy lion, for example, well, in a sense, yes, that is, this function, it learns in this way, that is, the parameters are configured in this way, these billions, based on, as it were, so that it is configured, this means, this means that this is it, that is, the learning process occurs, if we are talking about language models, such as, for example , well everyone famous chargept. gigachat, for example, is a function that takes a set of words as input and tries to predict the next one, and thus, if we predict a word many times.

various very, very cool things, in my opinion, that is, she can process and, for example, make some objects in the style of some artist, here you can argue whether this is creativity or not, yes, these debates are already going on, quite fierce, but the very fact that this can be done beautifully, especially lately, is an indisputable fact in my opinion, but in fact, i can say this, that indeed the field of artificial intelligence, it is now being democratized, such a term, maybe it is not there... a little complicated, but i will try to explain what this means, that artificial intelligence can now be dealt with already by schoolchildren, say, from the ninth grade, sometimes someone from the seventh grade, in general they study at a good level, which was impossible to imagine there even 10 years ago, that is, the level of mathematics is really there, the level of interest of schoolchildren is sometimes even

the quality of training programs, which are actually made on artificial intelligence. they allow you to enter into this science, and this is really science, it is complex, this is mathematics, this is the junction of mathematics and the actual calculations of machine learning, to enter already in the seventh grade, and we hold a huge number of competitions, hackathons there, and we see that children in general they cope with complex tasks no worse than people who graduated from college, but what they do, that’s what they do, well, actually the task, the task of a person who has artificial intelligence deals with, well, let's say neural networks, although in fact there are a huge variety of algorithms and types of algorithms... these are not only neural networks, just neural networks, they are very flexible and therefore they allow you to solve a huge number of problems, even creative ones, that is, a human task, who wants to create an artificial intelligence model or artificial intelligence, how to select an algorithm that suits the task, although recently there are universal algorithms that solve

a huge number of problems, and how to train it, what the most important thing, you know, what surprisingly i discovered is that...

in fact, a skill that is especially relevant lately is industrial engineering, that is, the creation of these very texts that are loaded into this black box in particular, yes , although in fact this is the most common, say, task, although a neural network can create pictures using other types of data, for example, you can draw a sketch and ask to make it realistic, this is also a task for a neural network as a whole, such as kandinsky, for example,

the neural network is named after the great russian artist, who came up with the idea that it would be named after? and you know, we’ve actually been working on this problem for a long time, about 3 years already, and our first model was called this, it’s rudali, she came up with it herself, in short, yes, and we, we sort of thought how to animate it, that let's say, to make her more recognizable, and we decided, at that moment, 3 years ago, she created abstract objects, and we kind of thought, what kind of russian artists are there who painted in the style of abstractionism?

art means that you need to do and come up with some new details, new features, make slight changes to existing architectures, thereby making... the algorithm the best in terms of the quality of understanding the text of this industrial product, the visual quality of the image, in fact , when we did kandinsky and are doing it, we set the first scientific and technical task, firstly to create something new, secondly to improve the current architecture, even western, there is huge scope for improvements there, and kandinsky’s line 2, here is kandinsky 202122 - three models, they have a unique architecture, we even have a set of articles.

if it is not ours, then what we could lay down if it were ours, roughly speaking, well, you know, yes, that the model learns from data, and we can, for example, make a model that better generates russian objects, russian entities , it is clear that western networks, neural networks, they certainly saw, there is the kremlin, and some iconic russian objects, and a character, it is still unknown which kremlin, well, that is what kind of kremlin will be there, exactly like that, yes, that is, they do not specifically set the task of making it generate a model, their task is simply everything in general. generate, but we can set the task of showing the models a lot of our data of such a domestic cultural code, and this

applies not only to the domestic cultural code, this applies to any types of data that we want to show to the models, and for this, of course, we need to make our own development. we continue the creative industry podcast, our guest today is denis dimitrov, and the man who created the kandinsky neural network is still with you. elena kiper and roman karmanov. and what successes are recognized by zakandinsky already? that's what he does better than, let's say, analogues that are wrong according to his word? well, it looks like neural networks that solve the same problem, generating images from text, for example, where we are cooler, although kandinsky videos are generated now, yeah, now, well, that is, it’s like, we’ve recently presented the model , which generates the video, was called kandinsky video, although, probably, necessary.

from the internet, that is, on open image description pairs, does it seem to capture data at the time of the request? no, naturally, there is a team that deals with data, and that is, the data is already contained in it, it is not searching on the internet at this moment, no, no, no, in no case, it is creating it, this is kind of unique, where are they stored ? this volumetric quantity is a huge amount of data, kandinsky the third studied on one and a half billion pairs picture text, and you can imagine how much data there is storage is needed, well, there is a supercomputer, it’s all learned on the supercomputer and stored nearby in special storages,

they are called c3, but these are these big, big data storages that are really for, let’s say, the dataset model has the opportunity to see, they need to be put there, and not just put, you need to remove all low-quality pairs, that is, filter and remove all low-quality ones. if you go to the internet, then in general half of the data there will probably not be very good good quality, it’s full of all sorts of rubbish, that’s it, the picture doesn’t correspond to the description that is written down there, or it may be that the picture contains a watermark, yeah, such data is probably also bad to use, because the neural network will reproduce this watermark, it will , now they say that artists are already fighting, that is, they are spoiling their image so that it won’t be too, that’s also true. we have to fight, that is, there are a huge number of filters that need to be written before giving data to the model, because everything is everything of course, you can’t give them away in a row, it’s like

, you know, showing a child everything in a row. it is not clear what he will learn in the end, the data development team is, of course, responsible for this, and this is super important, both the activity and the team in itself, because not only half, but more than half of all success depends on it, and after that, how the data is stored, here it is in this c3 storage, you can actually train the model, yeah, training the model is just an iterative viewing of this data. and here's the change this huge number of parameters that i talked about, these billions of parameters that need to be adjusted so that when a new description appears at the input, the picture would correspond to it, we have a video that was created by kandinsky, but before you watch it , the question is, we are conservative people in general beings in our own right, and well, well, even if some person is watching you right now, he

thinks, well, it’s interesting, denis, you have to try, then the internal struggle begins, to let you into your life this artificial intelligence, not to allow it, these are the areas that are already actively using artificial intelligence, who have already grabbed onto it first of all, designers, we have a special tool called fusionbrain.ai and there it is such a developing photo editor, it’s actually very popular lately make photo editors based on or rethink current photo editors like... photoshop, integrating ai there, why is this useful, because it is clear that using the text to create pictures and edit pictures are much, much easier than actually drawing yourself, what people always did before, it took a huge amount of time, now it’s enough just to write in text, a picture will be generated, you don’t like some area, you covered it up, finished drawing what you want , really simplifies the work of designers, and very much so, that is, the ability to not only generate an image, but also

edit it with themes. if we automated their work, in fact it’s not like that at all, because artificial some designers are afraid of this, that we, as an intellect, are a tool that allows us to simply optimize work, this is true for any profession; design is no exception at all, there is a routine part of this work. it needs to be automated as quickly as possible, in my opinion, as high-quality as possible, this will only save people’s time there and will increase the quality of the content, and animation, for example, how many routines are there, animation is generally the creation of videos - in general, how many films are made there, well they can film there for years now this can happen faster and faster, that is, since well, until the creation of full-fledged films, technology has not matured to be honest, but the progress is such that...

kandinsky did it, as if how he does it, he draws the first frame according to the text, and then you choose, that is, how animation differs from video, in that animation is a camera flying around a static object, or some kind of camera movement, and video is a full-fledged movement of everything, here after all, animation, because there is a first frame, and then we select the camera movement, well, that is , the monotonous camera movement, and we

select the first one to draw. he had exactly what industrial products he had, that is, he doesn’t need a lot of memory, it’s all created in the cloud somewhere, yes, it’s all created on a supercomputer, oh supercomputer, this is an interesting topic, the training of the model itself and the application of the model, well, it’s called inference , this requires computing power, the model is huge, after all, on an iphone, roughly speaking, it will not start on your smartphone, although progress is also being made towards this, so that it starts smartphone, but at the moment kandinsky

is running on a supercomputer, safari, this is our supercomputer, on which we teach both language models and these generative ones, which turn text into pictures and videos, and that ’s where we actually use them when you.. this request, it flies to the supercomputer, it is processed there, it is there that kanzinsky lives, yeah, it is there that he makes a picture from the text, the picture is returned to you, where you use a telegram bot or bot is a certain number of iterations. , that is, i go into the bot, i talk about what kind of picture i want to create, the picture comes, then the next stage i say what i want to see for the movement, and it gives me an animation, i stand in line, waiting for it all to happen. that’s right, yes, but we actually have ways of scaling, after all, a supercomputer contains not one computer, but many, of course, if a million people come, there will be a queue, this is

inevitable, in this sense, how to do inference on a device, then there is a smartphone that is much better because you don't stand in queues, you do it there, but we try to keep a lot of resources on the supercomputer, so that the queue is still satisfied, so that it doesn’t happen. do you feel that your competitors are spying on you, we feel it, that is, you feel this attention on yourself, yes, and since they are spying on you, it means you are doing something that apparently causes them, i don’t know, fear to me it seems that with everyone who is creative and creates something, competitors are peeking, to be honest, and this is not artificial intelligence, this is not an exception, of course, but... just how far you are ahead of your competitors, this is a competitive struggle, relatively speaking, you read the news there in the morning in the evening in any case, and you probably also note this to yourself, wondering if something happened , what

will affect who beat whom, so to speak , at the turn, this is, of course, yes, like, say, the area of ​​​​creating large models, although it requires a huge number of computing resources, yes... and specialists who in fact, they will teach all this, but in principle, the large bitechists have they certainly have such teams, they are developing these models, for example, yandex has a masterpiece, and the guys are making a fairly similar technology, the only thing is that they don’t have video generation, but today we only watched animation, it’s not a video, after all a full-fledged video is the movement of everything, i have a question that concerns. copyrights, intellectual property, it’s clear that leo tolstoy, roman said at the beginning, everything is leo tolstoy, he didn’t inherit it from anyone, he’s open history, yes, when you upload data to a supercomputer, you license it, or it

’s kind of in the public domain, or it’s not necessary, well, it’s not like it’s in the public domain, that is, it’s on the internet, most likely in the public domain, yeah, when we uploaded them, after all, how would we them... artificial intelligence or the person who ultimately generated all this, all this, oh, this is this, this is a very good question, in fact, who owns the copyright, as it were including at the legislative level , all this is being developed and, let’s say, this is in

at some point, it will probably be fixed, although this is a very difficult question, you know, as they used to say, i have an angel on my shoulder, an angel was sitting and he told me, when the creators of someone have something, what is the difficulty here? business? determining who the author is, there is data, of course, behind each photograph there is some kind of author who took this photograph, that is, there is an author of the photograph, it’s another matter when there are one and a half billion of them, then it ’s as if these authors are all somehow mixed up, and most likely the author, there is no author of one and a half billion photographs, that is, this a huge number of people, there are millions of people, on the other hand there is a company that purchased a supercomputer with its own money, this is a very super expensive undertaking indeed. all the calculations took place on this supercomputer, that is, it can claim ownership, there are developers who, the people who sat, sort of converted their knowledge into this uniros, that is,

they can say that in fact they have the right in the picture, on the other side there is a person, one person who is sitting and prompt writes, in the end, and he can say, in fact, i am the author, that is, ... there are four sides, maybe even more, which in general, well, some participation in the creation of the final picture, they did made their contribution, but so far the story is that... if a person, most likely, if a person, well, logically, each company does it in its own way in the end, if a person pays for using the neural network and received the picture, then he is the author after all, that’s most likely the case, but if there motives are usually guessed from motives, if something is created, a license is issued anyway, i just want to say that such disputes arise only when the picture begins to bring a lot of money to the author, and then it arises. a large number and other people who claim authorship, but

there is actually still a point that the neural network itself can be, as if companies, developers, and so on can claim its authorship, the model itself has a license, for example kandinsky and this also depends on, that is, the masterpiece is a model that is not posted and is not in the public domain, kandinsky is the opposite story, since ours is one of our first missions, i have already talked about this after all.

any, any image, any painting,

he has some kind of white canvas, well, not necessarily white, just white paper, some kind of empty, napkin, napkin, yes, there is a tablet screen, he has some kind of idea, surely, he starts sketching out some details, with a pen, a stylus, whatever, adds details, maybe removes something, well, a person also has patterns in his head, that is, he has memorized, so he has a pattern, this is after training, here it is, step by step, it’s simple. how the kandinsky model creates a picture, in principle, an analogy can be drawn, he begins to draw a picture with the so -called white noise, what is it? well , this is when there is no signal on the tv, there is such interference, just some kind of matrix, naturally there is your request, this is the same idea, as if in a person it is expressed by these signals of neurons, in a neural network it expressed simply by numbers, which are the text into which the text is encoded, and from this matrix. step by step in the process of removing

noise in the desired direction, an image is created, that is, creating images from this white, this white noise, from this matrix you can simply visualize, that is, there are 100 steps or 1.00 steps, for which from any of this noise the final picture is created and visualized, so to speak, the final picture appears, for example, you want to make a mob in space, yeah, you always have white noise in beginning, and step by step a mob is obtained. in space at the end of this process, if the noise is a little different, it is clear that the noise can be whatever you want, swap two pixels there, there will be a different one, but also a mob in space, this matrix, in a sense, it is for the imagination of the model answers, this can be visualized, but this process, it’s called denoising or denoising, so that it happens correctly, so we need to show this to the neural network one and a half billion pairs of picture text, that is, so that she taught.

right away from your head to generate your own images, well, let’s say on some kind of canvas, i don’t know, well, in the digital sense, of course, there is actually research. i wouldn’t say that this is a super big practice there and that it is actively used there, but there are studies that make it possible to decode brain signals, that is, you know, they put a reader of electrical signals in the head, for example, it is possible there to without without words you, roughly speaking, understand what you want,

or you close your eyes, imagine how something, i don’t know, for example, some kind of palm tree there and it’s possible purely from brain signals at the moment you imagine, but using these time series , the brain signal makes the same picture as a whole , as if you connect to an electrode like this, you can actually train a model that understands you without words and draws a picture, if you need to draw a picture, but the limitation of these models is that under each. person, for now you still need to teach your own model, that is, for each person, when he imagines a palm tree, the signals are slightly different, that is , it’s not yet possible to train one model for all people, and it’s expensive to teach a model for each person, and you need a supercomputer, and you need a supercomputer, until there is some kind of universal thing that everyone

understands, in general, i hope, i hope that not soon, and moreover, it will still remain, that is, there is no need to be afraid of it, even though it sounds futuristic somehow. but there is no need to be afraid of this either, because in general this will be used by people who, for some reason, probably cannot write, and if you you know how to write, then you won’t attach electrodes to your head, you will, it will be easier for you to write, were there moments when you realized that the neural network that you created was doing something that you did not predict, second, is it necessary to think about limiting it somehow? generally speaking, these are all generative models. which are being created for now, and which will be created, they tend to hallucinate at some point, this is the official term, it means that if doesn’t know something, but you still asked her to create it, she can’t create it, let’s put it this way, that is, she creates something, well , not what you want and something wrong, but

she creates it anyway , and this is probably best seen with and can be traced, as it were, with language models, you know, when gpt came out, there were especially. these problems are visible and they asked what is heavier, a kilogram of fluff or a kilogram of lead, well, it is clear that the neural network, it saw in the training data, as if the word fluff with the word light, it saw much more often than fluff with the word heavy, well, on the contrary, which is logical, and of course, she answers that a kilogram of lead means heavier, a kilogram of fluff is lighter, yeah, and explains why, well, this is fluff for perfection, there are three three paragraphs there, why this it’s actually dangerous, because... a schoolboy, for example, who looks at the answer of this model, he will believe it, then he will go on to broadcast this thought further to his peers, and so on, and what to do, we need to fight hallucinations , in principle, this is the area that is now the most actively developing is how

to teach this large language model, the generative model, to hallucinate less, well, we started with the fact that we, in general, so to speak. now, if our digital images were created, created by artificial intelligence, it would be interesting to look at them, of course, here are our models, lena was thinking about what we could look like, we just have something about this, oh, yezhkin, yoshkinkot, no , nothing, it’s some kind of oh, you know, how kandinsky sees us like that, yeah, wow, wow, why did he make us look younger? look, he’s giving compliments, yes, he’s hallucinating, and i’m you, can you see in the form of prom in my head, you look exactly like that too, and i’ll actually tell you how photorealistic portraits are made, we have a team that deals with detecting fakes, you know, now...