Hello, everyone! We've got a pilot run today. Oh, I'll try to tell you one interesting, in my opinion, and a fabulous story. It's, uh, what I want to share with you is what we all want to share with you, but what's not fitting in the sub-catester format where we're talking about the news four times. If it turns out you're interested in this format, write about it in the comments, we'd love to continue to take these episodes off.
And tonight we'll talk about the beasts. So, look, you and I know computers work with numbers, but they can't work with anything else. If we need a computer to work with something, we need to get this to the numeral. In this way, practically every issue of our sub-cash, we're discussing various chat-bots, LLM models and everything, and we see that computers are doing great with text and meaning, and now they can even do it.
I'm not sure. How is that going? Let's try to figure this out. We need to translate the text into numbers. Aaah, there are many different approaches. We'll take the most intuitive simple and understandable option. Let's say we have some text. Says, "The table is orange. And we need to somehow replace each word by number. Then the computer can work it out somehow. How can we do that? We'll take a picture of the orthographic dictionary. You've probably all seen them in your life. And every word we say is numbered. Let's take the first word, number one, second, number two, and so on. We'll get some numbers, and instead of, uh, the "on the table, the orange" text, we'll get some sort of chisel. With this computer set, something can be done. But look what you're doing. Oh, we want to, or, like, scientist, we want maths to work with numbers like numbers. What does that mean? There's one plus two equals three or three minus two equals one. They're working on simple arithmetic operations. Can we do that with the set of numbers we can make on the basis of the orthographic vocabulary? For example, we'll assume that the first three numbers are apricots, orange and appendicitis. That's how we put them in the orthographic order: number one, number two, number three. And if we take these numbers, one plus two equals three, math is right. And if we take the meanings we get, the text itself, what do we get? If we add apricots to the orange, we'll get an appendicitis.
It's like there's nothing you can't work with. Or we'll take the appendix, we'll take the apricos out of it, I don't think it's gonna be an orange. This approach, as it may be, can generally work, but does not give us what we would like to do. But it turns out that in 2013 scientists developed an algorithm that called word2vec. And this algorithm can keep meaning, and it allows them to work further, like the numbers. We'll leave behind the brackets that this isn't about the simple numbers you all know at school, but it's about the vectors.
But it doesn't matter, just because the same arithmetics work with the vectors: we can fold them, we can count and get different, and, uh, equations. So when the scientists developed this word2vec algorithm, and everyone's shaken by the example you may have even heard. The example was: we take the king, we deduct the man from him, add to what remains, the woman and, strangely, the meaning that is very close to the queen. Let's try to see what's inside, huh? The king is some man, a man, a king and a man with power. If we deduct a man from him, we'll get something in the crown and have power. Now we add to what's left, a woman, and we'll get a woman in a crown with power. And they get that this works at the level of understanding. We can put this in our head, right? And we're doing a vector-tracking field of math work. This example was so amazing, he was so surprised that I wanted to tell you a story from my life. About two, two and a half years ago, when chat, boots, ChatGPT, there, three and a half, four just got popular, all about them just started talking, and I got a call from one person who's doing business. Innovations that are involved in the so-called digitalization of business. And he's calling me and saying, "Ilnar, and you can tell me on your fingers how big language models work, what is it that's under the hood?" Well, I say, "Let's try it." I'm starting to tell him. I'm telling you how words are translated into numbers, into a vector, how can they work with these vectors? Well, obviously, I'm telling you about this example: King minus man plus woman is the queen. I'm still telling you what's going on next with the prediction of each word. He listened to me very carefully, and when I finished, he said, "Ilnar, you're the third date of the syentist I've been calling for today, and the third time I hear this example of a king, a man, plus a woman, Queen." What do you want to say? That this example just turned out to be so amazing.
He showed that with the texts, with the meaning, words could work exactly the same way as math can work. So the same math laws work, turn around, and with words, with meaning, most important. But it's not the only example. He's the most famous, popular, but he's not the only one. And today I want to share with you one very interesting resource. He was developed by Elena Voito. It's a great researcher, a date-synthesist, even more a reserver, yes, who works very hard in this field. And now she's created her own site, on this website, she's got some great lectures on NLP - Natural Processing, yeah, that's, natural language processing. That's what, uh, is based on everything that's about the ChatGPT, chatbots, and so on. A, and accordingly, in one section, she placed a very interesting quiz game, which gives you various examples. And, uh, examples of this kind, we'll take some object, we'll take one out of it, we'll add another to it, and what happens? You need to know what's really going on. And here I want to say there are a lot of interesting and very invisible cases I'd like to tell you. But the format of our short video will only contain a few. So I'll show you the most interesting, in my opinion. You try, uh, play this game yourself, play this quiz and write the comments, and what's the most amazing thing for you. Okay, let's see. So, we'll take this example that's on the screen. We have a hamburger. From the hamburger, we're gonna graduate America and add Italy. I'm offering, yes, I mean, we have some food, a hamburger, a m special for America. We're deducting from the wrong-- what we have America, just like we did the king minus the man, and adding another country to Italy. We have four options: there's a hot dog, there's a lasagna, a baghett and there's a cabbage, respectively. You can put it on the pause, think what's gonna happen. I will choose one of the options, and we will see if this is correct in terms of the Word2Vec model used on this resource. Well, we're gonna pick lasagna. I explained my logic to you. Uh, I wonder what you chose and how you explained your decision. We'll see the next option. We have Potter, that Harry Potter. Maybe you've seen movies and read books. We're reading out of Potter Radcliffe, adding David Tennant, and we're gonna have four options right in the Word2Vec model: Hermione, Sherlock, Dr. Who or Batman. Oh, think about it. You can put it on the pause. I'll try to explain the logic that's laid here. Potter is some character from a work or a movie. Radcliffe is the name of the actor who played Harry Potter in Warn-- Warner Brothers movies. Accordingly, we're reading out this actor's character that played this role and adding David Tenant. David Tennant is a British actor. Oh, for me at least the most famous part on "Doctor Who." So, accordingly, I assume that the right answer here should be, uh, Doctor Who, Doctor Who. Okay, well, according to the website, that's the right answer. In fact, let's see the next one. And you can also draw attention to the fact that in every field, and, uh, the meanings are tied together. I mean, on the one hand, we had a king, queen, on the other hand, we had food and countries that they might be specific. We're looking at the characters from the films and books and actors that they're connected to. See, the same arithmetic works in completely different fields. It may look a little like magic, but it needs to be understood, and what is under the hood of these models. These are statistical models that learn from texts that are available in Wikipedia, on the Internet, in books and so on. And the point that can be statistically removed from there is this arithmetic, that math is available. Yeah? Let's see the next example. We'll take the sun, we'll take it hot and we'll add it cold. There are four options for answering: the moon, rain, land or winter. Choose your option. The right answer here is the moon. Try it.
No, see, I thought the moon was, but actually--
I thought the moon was, but I was wrong. The right version in the Word2vec model that was built here will be the answer for the rain. See? I hope you chose correctly and not wrongly like me. We'll take the next option. Oh, he's amazing, too, just because the meaning of it is preserved, so, as much as possible. We'll take a green color. From green, we'll take the blue and the rest of it, add red. The correct version of the Word2vec model is that one of the four options is pink, yellow, brown or orange. I'm guessing that green gamma is like a combination of blue and yellow. If we remove blue and yellow from this combination, it's only yellow. If we add red to yellow, it's got to be orange. Let's see if my thoughts are consistent. Yeah. Oh, again, it doesn't mean that the model can speak in colours, but it means that the same result is that the statistically obtained from the textbook on which Word2vec was trained, I've been able to explain this logical reasoning. Doesn't mean it's under the hood, but the coincidence already talks about a lot and quite interesting. We'll take another example that I think is quite curious. There's Elena's name. We'll take Russia out of this name and add Scotland. And I need to understand, and what's gonna happen. Some options are presented here. There's Bonnie, there's Stana, there's Lucy, there's Phoebe. Some of these options, in Word2vec, are considered to be correct. Think about what? Yeah. Well, I think those who wanted to put it on the pause. I'll pick Bonnie as a pretty special option for Scotland. And that's amazing, too, right? We take a name that is specific to one country, we read the country out of there, add another and get a name that is popular or specific to another country. Here's Bonnie for Scotland, a pretty popular name.
We'll go to the final version I want to share with you. There's something strange - gato. We're gonna get a cat out of there and add a dog. What do we have to do? There are four different options. Oh, not all of them I can read, I think, right. I won't read. Read on the screen what's written. Is that what that is? Oh, this example, he seems as strange as possible and unfamiliar as long as you don't understand that there's a word in another language. The words are spoken in Portuguese. I mean cat and dog in English, and all the other words are Portuguese. So gato is a cat or a cat in Portuguese. Accordingly, we take the word in Portuguese, deduct the word in English cat, add the word in English dog. What do you think should happen? You can try to translate all these four options from Portuguese to Russian or English and see what should eventually be done. Well, I've moved that cachoro into Russian is translated as a puppy, you know, or a dog. That is the option that is chosen, and it is here that it is correct. And this option is the most striking thing. If you and I see it all work in one language before, it works in different languages. We have a word in Portuguese, we read English, add English and have the word in Portuguese, but that corresponds to the word we were supposed to have in the English version. English. So we took the cat to the Portuguese, took the cat, added the dog and got the dog in Portuguese. I'm, of course, impressed to the depths of my soul. Accordingly, these are the examples I have been able to tell you today. Oh, please write as many points as you can earn in this game. Next, write the examples you've been struck and why they've struck you. Why do you think it's amazing? Have you ever waited to see this? In addition, write if you're interested in this format. It's not a sub-cadast, it's a little fascinating story. I hope you like her. If you like it, you're gonna put the liki on the phone and you're gonna sign it in the commentaries. If it turns out that the audience that is interested is a lot, we can continue, uh, to take down small numbers in this format on different topics. We have a lot of things we want to share with you. If it turns out you're interested, we'd love to continue this case. Stay on the canal. News and subcasts are about to be in the middle of a few minutes. I hope you liked it. Bye-bye.