Как получить максимум от ChatGPT. Занятие №1. Александр Свет

Good evening everyone and welcome to the first session of our new course on getting the most out of ChatGPT and the coding agent Codex, which comes with ChatGPT. This will be a very interesting immersion into the real possibilities of AI. And today we will talk about ChatGPT itself. So today our entire session will be completely dedicated to capabilities that are, well, not exactly hidden, but most people really don't know about these ChatGPT capabilities. And I think you too will definitely learn a lot of new and interesting things today. Let's start by looking at how our sessions will proceed. What will be our course plan? And also, if this is your first time at Profail School, I'll tell you that simply in your personal account, you will, firstly, find recordings of all sessions. There will be homework assignments, and on this course, we will have both practical and theoretical work, so you need to complete them and get a grade of at least 75% to receive a certificate. Theoretical questions. I'll say right away that since you are the first group, theoretical questions will appear with a slight delay, approximately, well, usually the next day after the session, the theoretical question will appear, and the practical part should already be available now. Also, if you have any questions that you want to ask me, for example, when the session has already passed, you know, somewhere in the middle of the week, you can ask them in the questions section. And also in communication, we communicate. And I've actually provided a link here. I've created a new chat for our group on Telegram, because practice shows that it's just really very convenient to discuss everything in a chat, to resolve any operational issues. So, join us. The Telegram chat is very, very convenient. Now let's look at how our course will proceed and how it is structured overall. So, we'll start by saying that during the course, we will complete three projects, right? So, each session will be one project. For example, after our session today, you will have a project to create your own custom GPT for answering highly specialized topics with your data. And this is truly, actually, a very cool thing. We will create it today, all together. You will see how it's done, and most importantly, you will understand how powerful this thing actually is, because it appeared a long time ago, you know, these GPTs, and it kind of got lost in time, so to speak. So many people have forgotten that such a cool thing exists. It's genuinely very cool. So, each session will be one project. Then we will create an interactive website using Codex. This is the coding agent from ChatGPT. And we will even create a working Telegram bot, also using Codex. So, what will we talk about today, right? We will talk about all the capabilities of ChatGPT, you know, beyond, you know, the standard ones, we won't talk about some of the most, most standard things, but we will discuss what you really need to know for good professional work with ChatGPT. Firstly, we will talk about models and reasoning modes, because this is a very, very important topic. We will definitely discuss how memory works in ChatGPT, talk about how to use projects, the canvas mode, look at how the GPT agent works, and we will definitely talk about deep research, because there are now several different types. Now there is Deep Research of Capture One. This is very cool. Oh, children, good heavens. Something from the past just popped out. And, and accordingly, at the end of the session, we will create our personal assistant for specific tasks, right? So, we will create our own custom GPT. The second session will be completely dedicated to an introduction to development with Codex. So, we will talk about what Codex is, how it differs from ChatGPT, how it can be launched, because it exists in different forms. There are separate applications, there are extensions for other programs. And we will talk about why people who are not actually involved in development, you know, not programmers, people completely unrelated to programming, why everyone now needs to do this? So, every person now needs to be able to develop AI. And this is not an exaggeration. Believe me, if you are not doing this yet, when you, you know, figure this out, you will simply understand why I am so categorical. The third session is also dedicated to Codex, but rather not from the perspective of, you know, creating a website or, in general, introducing this whole story, but we will talk about very practical ways of using Codex. So, we will talk about how, for example, to work with your documents using Codex, to create, edit PDF files, Excel files. How to use such a very important thing as skills for this, skills, these are skills for typical tasks that significantly improve the quality of AI work. And we will also create our own Telegram bot, also completely without any programming on your part, right? So, it's important to understand that this course is designed for people who have never programmed in their lives and, in principle, do not plan to do it themselves, you know, manually. You will only manage your coding agent. And this is a very similar experience to, for example, if you have a programmer on your team, you need to communicate with them in the same way, discuss the project, and, accordingly, develop. So, he will develop everything for you. And, accordingly, our third project will be the creation of a Telegram bot with Codex. And in the last, fourth session, we will actually do a very important thing. You can already, you know, try to figure out what you would like to discuss in this session, because the fourth session is essentially dedicated to your specific tasks, solutions to your specific cases, how you wanted to solve them with AI. So, you can start sending me your ideas, tasks, what you would like to solve with artificial intelligence. And in the last, fourth session, I will try to give very specific practical advice on how to actually solve these tasks in practice using everything that we have learned in this course. Because, it seems to me, with AI, it's impossible to be a theorist, right? So, you just need to immerse yourself in practice and, in fact, not come up for air from it, because it will simply be an endless process. You will constantly be doing something, doing something in AI, every day something changes, you will try these new technologies. And perhaps it will draw you in just like it drew me in. And also in the fourth session, we will look at the use of some additional things that we will not talk about today. For example, we will talk about m t Atlas. This is an assistant built right into your browser. We will talk, for example, about how to record meetings with ChatGPT, about video generation. So, in general, if there are any topics that we don't discuss in these three sessions, we will definitely return to them in our last, fourth session. Well, and the most important thing is the analysis of your cases, so please, please send them. You can write them in the questions section in your personal account. This will probably be the most optimal way, because then I will definitely mark them for myself, that I will definitely analyze your case. And, well, again, you need to understand if it's a realistic case, because there are situations when, well, yes, AI can do a lot and can actually do it, but sometimes to build a complex AI system, the knowledge from our course will definitely not be enough, because then, for example, if you need to build a very complex AI agent, well, that requires a different level of knowledge, but with the knowledge you will gain, you will be able to do a lot. It's also important to say that you must have a ChatGPT Plus subscription. You can either buy it yourself, for example, or you can do it through Profail School. So, you can go to the course section, where it is. Here is our course section. And here you can select "Request Access". So, if you fill out the form, the school administrators will contact you and give you access. But you absolutely need a ChatGPT subscription, because without it, most of the things I will be talking about today will simply not be possible to do. And, in particular, you will not be able to complete the homework, and then you will not be able to work with Codex, because Codex also comes with a ChatGPT Plus subscription. So, this is, well, simply a mandatory condition for this course. You will not be able to get the most out of ChatGPT if you don't have it. And, and I also invite all of you, accordingly, to our chat. Let me open the link now. Alexander asks: "In which session will we do code review?" Oh, well, look, code review is already a slightly different story. It's more about development, right? But on this course, we are primarily learning to get the most out of ChatGPT and related tools. After all, code review is a slightly different level. It's something that we cover in the development course, and even more so in the second level of the development course. So, it's a slightly different story. We will use Codex, but we will use it not for such serious tasks where you really need to do code review. Although I will show you how it can be done in Codex, because Codex actually has a very simple command for this. And in general, GPT 53 Codex for code review is perhaps the coolest model. Because it really delves very, very deeply into the code and can find problems that other models miss. So, personally, for example, I love developing with Claude Opus, but I always double-check with Codex afterwards. I'll just show you how to do it, but we won't go into it deeply, because it's already more of a development-related story. Now that we've gone over the plan, I want to recommend one more thing. And it's directly related, actually, to our session today. It's my Telegram channel, and in it, I publish various articles, translations of articles, the best English-language articles. I'll send the link now too. And why is it related to our session today? Because this very GPT, right, our assistant, which we will create today, we will actually create it based on this Telegram channel. You will learn how it will work there. But it will be a really very cool thing, very useful. And I will genuinely continue to use it. Now let's start diving into the topic, the topic of ChatGPT. And, of course, first of all, we need to talk about models and reasoning modes, because this is actually the most controversial, and, I would say, ambiguous topic related to ChatGPT, because if you turn on your chat, right, here you have a model selection. And the thing is, for some reason, the developers of ChatGPT, with the release of the GPT5 model line, decided to do this: let's just show the user that this is, like, GPT5, well, now it's GPT 5.2, D. But in reality, in reality, a lot of different models will technically work inside. We just won't tell the user about it. And this was a slightly strange decision. And therefore, when, for example, the first reviews of GPT 5, right, models appeared, they were, on the one hand, some people were simply ecstatic, because they wrote that, you know, ChatGPT can now solve very, very complex problems, it's just amazing. On the other hand, people were just showing screenshots where GPT5, well, couldn't answer elementary logical questions. And no one understood how, why it works like that, why, on the one hand, it seems to be the most brilliant, and on the other hand, it really can't solve the simplest logical problems. And to this day, I think you see, periodically, such screenshots appear when, you know, some slightly tricky question is asked to the LM and, well, ChatGPT specifically, and you can see that it really doesn't give the correct answer. In many ways, this is precisely because these GPT models, right, are not just models, but also model routers. A router. This means that when you send, for example, a request, right, it's a system that decides which specific model to send this request to, right? So, for example, it can send it to a smarter model, it can send it to a dumber model, because there is, for example, GPT 52, mini 52 n. Well, we don't even see them in the list, but they are inside, within OpenAI itself. So, there are actually very, very many models in the JPT5 series. If you set the auto mode, you are essentially letting it decide for you. So, you are 100% just saying, okay, you decide yourself. For a simple question, let a simple model answer. For a complex question, let a stronger model answer. And in theory, it sounds great, right? So, in theory, it's great, but in practice, it turns out that sometimes complex questions go to a very dumb model, and questions that, well, could have been solved by a simple model, go to the smartest model. And because of this confusion, the quality of results in auto mode can be very so-so. And the same applies approximately to the instant mode. So, instant is a mode when you send one of the weakest models. And it's very fast, which is why it's called instant, like instantaneous, but you need to understand that it's incomparable, for example, with the thinking mode. And especially if you also set extended thinking. We'll talk about that too. But for now, let me just say that if I summarize everything I've said, just make it a rule. Always choose GPT 52 thinking. That's it. That's it. That should be your model to work with. You don't need someone to decide for you which model to choose, because in most cases, it ends up not very well, and people only get disappointed in AI. Therefore, thinking is your choice. There is also a section called Legacy Models. These are, like, models from the past. And here there are models, for example, GPT5, 5.1, 5.1 instant and thinking. And there's really no great sense in the 551 models, to be honest. I think they add them here more, you know, so as not to disturb people that a model they might have gotten used to could disappear so quickly, because some people actually get used to models. Now there was a big story about how GPT removed the 4O model, remember, there was such a thing as 4O. And many people really got used to that model, literally. They miss it a lot, its, like, personality. And there were even pickets to get GPT to bring back that model, but it has now completely removed it. The only model from the old ones that remains is the O3 model. And this is actually an interesting model, because this model differs from the GPT5 line, and for some tasks, it's amazingly cool. Especially O3 can determine, to a frightening extent, where a photograph was taken. So, for example, you can, for example, send a frame, and the model will analyze this frame for, say, 20 minutes. You can watch what it's doing. Very impressive. And after 20 minutes, it will actually tell you where this frame was taken. It's a frightening thing, really. Try it. O3 was really good at this. It's very impressive. And I see that Anna is writing: "Interesting, I don't have either of them, or O3 in my menu." Yes, you know why that might be? In the settings, I'll show you now. I'll periodically hide my screen for settings, because the only thing I don't like about the settings in chat is that there's a lot of personal information, like phone numbers, all sorts of things. Well, I don't think it's right to show it like that. Therefore, I will sometimes show you the settings, but only when I get to them. There's just a setting that allows you to see, yes, I'll show it to you now. You need to go to the general section, and here "Show additional models", meaning show, you know, these additional models. And, accordingly, then you will see these models. Actually, there used to be even more of them, and they have now slightly reduced this list. And if you have the subscription that costs 250 dollars, or rather, 200 dollars, then with this subscription, you should have access to the GPT 4.5 model. This is an absolutely amazing model for writing texts. This is the only GPT model that I really, really loved how it writes texts. But unfortunately, that's no longer the case. Bogdan writes: "I thought each next generation should be better than the previous one, but here the old version somehow performs better than the new one." This is a very interesting idea and thought, because if you ever try to train models, you will understand why this happens. The thing is, training a model is half science, but half art, in the literal sense. It's very difficult to explain, but these people, you know, who know how to train top models, they are not paid billions for nothing, literally. Their salaries are billions. This is because they are, yes, of course, they are great AI scientists and all that, but they have this, you know, like an artist, right? They have this feeling, the feeling of how to make this model work as it should. It's impossible to describe. It's very difficult to explain, but there really is such a moment. And if you just try to create a model yourself, it's actually not difficult to do anymore. Anyone can create a model now. And try changing its various parameters, and you will see how much art there is here, not science. And it's a very, very interesting thing. And therefore, it happens that some past models can be much better than previous ones in some aspects. This happens sometimes. So, yes, that's regarding models. Now about modes, right? So, let's just note here, right, that we always use thinking. I will tell you a little later. Well, not later, but when we actually create this GPT today, I will tell you in which cases it makes sense to disable thinking. But for a regular chat, a regular conversation, always enable it, because it will give you the best results. And O3. Well, you can try it for some tasks. Well, in particular, it copes well with the task of analyzing images. But for other tasks, I think 52 is no worse. And now about these reasoning options, right? So, this is, well, it translates into Russian as effort, how much effort they will apply, mental effort, right, thinking effort. And so, there is the standard reasoning time, and there is a longer one. This means that they will simply spend more tokens. Well, a token is, like, a unit of measurement of LM activity, let's call it that. And, accordingly, if they think longer, they will simply, essentially, be more expensive to operate. They will spend more energy and, well, computational, like, computational activity to process your request. Again, in most cases, you rarely come to ChatGPT with, well, very simple questions, so it's better for it to think properly. If you come with a super simple question, then of course, extended thinking is not needed, but in some situations, you cannot assess whether it's a simple question or not. It might seem simple to you, but it would be good to think about it properly. Therefore, personally, I always have standard 52 thinking and extended enabled. That's it. And I work with this mode. Yes, sometimes you have to wait a little longer. Well, I'll wait, I prefer better quality answers. Yes, that's regarding models. And it's also worth mentioning where it's better to work, right? In the web version, that is, which is available here, right, on the site, or in the application, right, the application looks like this, accordingly. This is also a slightly strange story, because in general, the web version is better. Better in the sense that it has all the newest things, it has all the most up-to-date new technologies, it supports most of the innovations that exist in ChatGPT. So, in general, the web version is more, you know, up-to-date. So, if you want something new to definitely work, it will definitely work in the web version, but the application is better for working with MCP, that is, for connecting to various other services. We will also do this today. And this is also important to understand, because, for some reason, the web version really works worse with this. And it's worse both technically, it glitches more, connects worse. And purely visually, it's not as convenient. Well, I'll show you how it's done later. Just for yourself, I advise you to decide that it makes sense to work in the web version most of the time, and the application is convenient for MCP. Plus, the application has several conveniences, for example, it can, on Mac, I don't know about Windows, but on Mac, it can read the content of your window. So, for example, if you have notes open, it can read the content of those notes. Or you can quickly press a button, and a search window will appear, to quickly type something into ChatGPT. So, these are small conveniences that the application provides, but in general, the web version has everything more, you know, up-to-date. Now let's talk about the first, the second most important thing after choosing a model, right? Because we, you know, we've chosen that we work with GPT5 thinking, we set extended thinking, so that it thinks longer, the smartest model, and gives the best answer. But besides this, it's very important to use internet search in your requests. So, when you make a request, in fact, this is often enabled automatically now, but just in case, you can set this web search mode, right, activate it, so that it definitely searches the internet. And this is a super important thing, because without internet search, you rely solely on the knowledge embedded in the model. And this is where people often have all these problems with hallucinations, when the model starts giving nonsense answers, when, for example, it starts making up obviously false facts. All of this is related to the fact that it simply doesn't know this. And it's still not possible, well, no one, at least, has managed to teach the model so that it clearly answers, "You know, I just don't know this." No LLM can answer this yet, because it goes somewhat against the very essence of LLM, because in reality, it doesn't know anything. If you think about it properly, LLM actually knows nothing. It doesn't have, you know, a brain that knows something, right? But it can, by essentially using probabilities, give the correct answer. This is a very complex process. When I talk about probabilities, it's a wild simplification. In reality, it's a huge simplification, but it

It's like it doesn't know anything in the sense of how we interpret knowledge. And therefore, none of them can tell you, "You know, I just don't know this." Well, you can push it in a certain way to get it to answer like that, but in general, it's very difficult for an LM to do such a thing, so it will rather invent something for you. It will just take some facts from another topic and apply them to your topic, right? And this is what very often happens if you don't use internet search. With internet search, everything is great, because firstly, it searches well. And plus, you know, many things just look great when searched on the internet. For example, if you ask: "Show me the current English Premier League table. I love English football." And look, we ask it like this, and it will show us now. Okay, thinking, of course, takes a little longer, it always works. Well, you just have to factor that in, and then it's fine. Okay, now it seems something didn't work for it. Let's see. You can see what it's doing there. It's checking everything. I think it didn't activate. It has such a cool tool to quickly show the table. For some reason, it didn't activate it. This happens sometimes. Especially since our classes are in the evening. And those who are on my course in the evening, European time, know that this is always a problem because Americans wake up. Let's see what it will answer in the end. But I don't think it will answer the way I would like it to. Let's see. Yes, this is not good. No, no, no, it's just a small glitch. Well, it's trying to make a table. Let's try again and see what it does. It should have a beautiful one for many such things. Wow, this is cool. See how beautifully everything is with logos and everything. And this is actually available for many things, not just the English Premier League, but for various other things too. Bogdan asks: "Including Web Search, do we get the same answer, but with references, or does the essence of the answer change?" The essence of the answer changes significantly. Because how does the search actually work? To show you how it works, look, for example, you write "compare prices, functions, differences, for example, Netflix, Disney Plus, and Amazon Prime," right? Different streaming services. And actually, what does ChatGPT do at this moment? It breaks down your request into structured queries for Bing. ChatGPT uses Bing, which is Microsoft's search engine. And it breaks it down into several specific queries. For example, it says, "Netflix prices and conditions, for example, in 2025," or "Disney prices and paid subscription conditions." Or "Netflix, for example, comparison of 4K plans," right? So, it makes several specific queries on the internet and then, based on what it gets, it prepares an answer for you with sources. So, it will be a fundamentally different answer. It's not that it just thought, "Hmm, maybe Netflix is more expensive, this one is cheaper. Let me back up my opinion with links." No, no, no, it first forms several series of queries, specifically from your request, it makes several queries, and then it gives you the result based on that. And therefore, the more specific your request, the better ChatGPT will do each individual search. Because if you just asked, "Compare prices of different streaming services," right, something like that, then yes, it would probably think that there's Netflix, and maybe it wouldn't remember that there's Disney Plus, for example, as a streaming service, and it wouldn't consider it. Therefore, if you want to get the best possible accurate answer from ChatGPT, always provide key phrases that you want it to consider. Names of key brands are essential, they will definitely improve the quality of the answer. So, definitely use this. I always do research on the internet. Especially anything related to any new events, any non-standard topics. Any question about AI, any question about AI should only be asked with internet search enabled. Any question about rare technologies. All of this must be done with search enabled. And if we talk about how to get GPT to search for quality sources, then, firstly, you can specify in the prompt, for example, "Use specific sources, for example, academic, official." So, you can directly tell it to search. You can say, "Search, for example, specific websites." So, you directly say, "Go and look on sites like Nature or MIT." So, more serious sites. It will go and search there. You can set a specific time filter, for example, "Search for events only." It's better not to write, for example, "in the last month," but rather write, "Search for events, for example, in January, let's say, of 2026." Because "in the last month" is also a weakness of LMs; they don't have a concept of time. And this is also a serious problem because when you start asking things like "in the last period of time" or "since when," these are complex concepts for LMs. It's better to write specifically "in January" or "February of 2026." And also, be sure to ask it to provide links to all sources. This also helps to have links. But actually, now they are almost always there automatically. And if you combine these four things, right, like "search on these sites," or "use these types of sources for this period of time, with links," then you will get a super high-quality output, and most likely, you won't get hallucinations, but really very accurate answers. Honestly, I stopped manually googling anything a long time ago. I do a gigantic amount of research with LLMs. I think I do about 30 research tasks a day for sure. So, I constantly say, "AI, do this research, AI, do this research, AI, do this, do this search." So, I constantly use LLMs for this. So, definitely use search. This will completely solve it. In principle, if you use a good, high-quality model and enable internet search for it, well, and actually, even without search, it will still find something. If you, well, if you say "search on the internet," you don't necessarily have to press this button, it will activate it itself. But the main thing is to say "search on the internet" and then clearly formulate your request. And that's it. In principle, a quality model with internet search will give you a much higher quality result than any other ordinary requests. You will practically completely protect yourself from hallucinations, unclear conclusions, and invented facts. Roman asks: "Each request in a new chat or in one?" For any, this doesn't just apply to ChatGPT, it applies to any LM. And not just in chat format, but when we work with coding agents, you always need to create a new chat for each new request. Because each new chat gives you a new context, right? Context is how much information goes into the message to the model. And, accordingly, if you put many, many different questions into one chat, they all accumulate, and with each new question, the model reads all your old questions, and this significantly degrades the output. Therefore, if the topic changes, always do it in a new chat. And it's also worth mentioning here that ChatGPT, because it's such a user-friendly thing, right? It's more designed for the mass user rather than a professional user. It hides some things from you. And one of the things it hides is that when your chat starts to overflow, for example, the GPT-4 model has a context window of 400,000 tokens. A context window is how much information it can hold, you know, in its "head." And if this 400,000 tokens overflows, it's actually a lot. 400,000 tokens. That's, well, almost War and Peace. Almost the book War and Peace, it's just a little bit more. Well, for a visual representation, it's about that much. And when it overflows, ChatGPT starts to discreetly, well, to clean up your chat history. It doesn't inform you about it. You might just notice that it starts to forget old facts that you discussed, and it starts to get dumber before your eyes on old topics. And therefore, it's always better to create new chats. In general, this is a general recommendation. Alexander asks: "And can we see anywhere how many tokens we have used?" Yes, this is another thing that ChatGPT hides from you. You won't be able to see it. That is, if we are talking about ChatGPT. With code, it's at least a little more transparent, not as good, but more transparent. ChatGPT hides everything from you, and it has its own internal mechanisms of operation, which are actually complex. I don't want to diminish the engineering capabilities of the people at OpenAI. There's actually a lot, you know, it seems to you that you have such a simple interface, like a chat, and you can chat like this. No, actually, very complex things are happening there, and many complex technologies are used. It orders your entire dialogue in a special way so that it works well. So, it's actually a complex system. But you still need to understand that this is a mass-market system, right? So, it's a tool designed for the mass user, and therefore, well, it won't work perfectly, let's say, as we sometimes would like. But if you create new chats for each topic, it will definitely work more reliably. Now, let's talk about chat memory, because this is one of the most underestimated and, in fact, a bit frightening things if you start to delve into it. How does GPT memory work? Firstly, let's start with the fact that ChatGPT's memory consists of two completely independent systems. The first system is Saved Memories. And this system is completely transparent and understandable for the user. And firstly, you see if ChatGPT saves something to its memory, and you can view these memories. So, look, for example, I love Tarantino's films. I have already watched Kill Bill and Pulp Fiction. And now we will watch. Let me send this. It might understand on its own that it needs to be saved to memory, but if not, we can ask it to save it to memory. Let's see. Let's see. In some cases, it can decide on its own, and in some cases, you can ask. Let's see what it will do with this knowledge now. See? Updated saved memory. The user likes Quentin Tarantino's films. He has already watched Kill Bill and Pulp Fiction. This is what it just saved to memory. Let's see. Let's see where it is. It's in the personalization section. And here is memory. And here, firstly, you can enable it to reference saved memories, saved browser history. This is if you use the GPT Atlas browser, we will talk about this in the last lesson. And chat history. We will talk about chat history now too. For now, just about these saved memories. And the most important thing is that you can see what you have saved. And before the start of this course, unfortunately, I completely erased the memories that were here. Because if you go into your memories, if you have been working with ChatGPT for a long time, and you go into these memories, I think you will be impressed by how much it knows about you. And not only about you, but maybe about your children or, well, such things. For example, at some point, I opened the memories, well, these chat memories, and there was literally a phrase describing one of my son's friends. And it's completely unclear where it came from. And it seems like it gathered even such a thing from some very unrelated things. So, it seems like we never directly discussed it with it, but apparently, from some requests, it made such a portrait for itself. A very impressive thing. Look at this. And it has, accordingly, now remembered that I watched this film. And you can, firstly, remove this memory, or you can clear all memories. And plus, they are also automatically organized, so some old ones are deleted. If you have "automatically manage" enabled, then it will also be automatically cleaned up, so it won't get super cluttered. And in practice, it works like this: for example, let's create a new chat. Again, we created a new chat, a completely new context. And therefore, I will write. I'm thinking, I want to watch a Tarantino film today. What would you recommend? And let's see now. By the way, I use Whisper for dictation, but you can also use ChatGPT's own dictation here. I just prefer Whisper, so I use it. And so, let's see now. Here, here's the key phrase. Look, it recommended different films, and then it writes: "Considering that you have already watched Kill Bill and Pulp Fiction, I would recommend today..." And then, accordingly, it advises. So, you see, its entire response was built on this knowledge of me. And the more such memories there are, the more you do this, the more it will know about you. I'm not kidding, there will be things that you just, well, you won't expect it to know. Really, don't expect it. Well, when I delved into these memories, I was just a little shocked because, well, it seems to you that it remembers what you tell it. No, it's more complicated than that. It draws certain conclusions from the things you tell it. So, memory at this level, well, it's like this, well, it's good that you remember everything, everything is very transparent, so it's clearly visible what was saved, what was remembered, so, for example, you can clearly control it. You can go and delete everything, all this memory. So, it's a completely transparent system. But there is an opaque system called Chat History. And, so, ChatGPT builds some kind of additional profile for you internally based on all your past chats. So, most likely, this is done for advertising. You know that free ChatGPT ads have already appeared. And this user, this profile, is hidden from you. You cannot view or delete it. This data is collected automatically, and you can only enable or disable it in the settings. So, here, if you go to settings, personalization, here you can find "reference chat history." You can turn it off if you don't want it. And actually, some researchers say that this is the most opaque thing about ChatGPT. So, most likely, ChatGPT has a very, very accurate profile of you, and they don't show it to you because it might be so scary that it's so accurate that they don't want to show it, but it exists. And therefore, for example, advertising in ChatGPT, which will appear soon, will be very expensive for advertisers. Actually, it's very expensive advertising. And most likely, it's so expensive because you can make such accurate targeting, so, I don't know, I think it can go down to some specific psychological traits and decision-making of people. So, I think you can do such targeting there that no other advertising system can do. And there is also something called temporary chats. It's called Temporary chat. So, when you enable a new chat here, you can enable this temporary chat. And it also, you know, it shows that this chat will not appear in your chat history, and it will not be used for training our models. And this is, in principle, all true. The chat is indeed not saved in history, it does not update or create your memories. Well, it's definitely not used for training models. But the thing is, despite this, firstly, OpenAI stores a copy of all these temporary chats for another 30 days for security. And by a court decision in June 2025, they now store all such chats indefinitely in a special secure storage. And here's what you need to understand: chat bots are not covered by the right to privacy of correspondence. This is very important to understand. So, correspondence with a chatbot is not like your letters to someone or such things. And indeed, to read your letters, it's a complex legal process, well, in most countries. And it's a serious, well, a serious matter. But understand that your communication with a chatbot is not covered by this. Moreover, you know, like there is, for example, that you cannot, let's say, read, for example, I don't know, correspondence, for example, with your doctor or psychologist, but if you use ChatGPT as your psychologist, then all these texts will easily fall into the hands of the authorities who request them at the first request. Just keep this in mind, even with Temporary Chat. So, think about this before discussing anything that you wouldn't want anyone to see. Yes. So, let me answer the questions. Oleg writes: "If possible, tell us about other LLMs in comparison later." I think we will do this when we have the second, third lesson, there will be some coding happening, and at that time, I think we will return to this topic. Evgeny asks: "Are projects just for sorting convenience or do they have a sacred meaning?" We will talk about projects today. Bogdan asks: "Record mode?" We will discuss this in the fourth lesson, because it's a bit of an additional function, and we don't have too much to discuss right now. We will discuss recode later. So, now, regarding projects and project memory, the thing is, well, first I'll just say that there is such a thing as project memory, and after that, we will talk about what projects are. So, project memory is a somewhat separate thing. So, mm, projects have, yes, let's do it this way, yes, let's do it a bit differently, it's better. Let's first talk about what projects are, and then I'll tell you about project memory. I think that will be more correct. So, what are projects? Projects are essentially smart spaces that combine all related chats, uploaded files, and special instructions in one place. So, for example, here you can find "projects," and you can create a new project. And let's name it, for example, "Course on ChatGPT," right? And I'll name it. So, and here you can choose how its memory will work. We'll look at this now. So, a project is, imagine, a grouping of chats by a specific topic. So, it's a convenient way to collect all your chats that relate to a specific topic, for example, "ChatGPT course," right? And upload unified files for them. So, for example, you can upload some PDFs, spreadsheets, documents to a project, and you create a new chat in this project, and it automatically has access to all these files. They have, projects have instructions at the project level. We'll talk about this now too. And there is this built-in project memory. What is it? So, project memory is that each project has two types of memory. Or, like, the default one, right? What does this mean? It means that, in principle, the project works the same way as any other chat, in the sense that its memory is shared with all other chats. So, for example, you might have project A, project B, and any other chats, but specifically memory, the memory will be shared, right? There can be different files, different contexts, but memory, what is saved to memory, right, as we saved about Tarantino to memory, right, and such things will be shared. So, in other words, if I create a project now, and I choose this default memory, it will know that I watched such and such Tarantino films, right? It will know this. And then, on the one hand, it's good, but on the other hand, everything gets mixed up, right? So, your entire memory, your memory about yourself as a person, and your memory about, maybe, your family or such things, will also be available in this project. This is not always necessary, if your project is purely work-related, for example. Therefore, there is a second mode, which is "project only." So, here you can enable "Project Only." It's important to say that this choice must be made at the beginning of project creation, because you cannot change it later. So, you need to decide at the beginning how this project's memory will work. Will it use all system memory or only what is related to this project? And then each project, in "Project Only," will have its own memory. So, all memories will be only related to this project, and no other memories will go there. So, this is complete isolation. And this is good in its own way, because, for example, imagine if you, well, imagine, for example, this: you use, let's say, GPT as a psychologist, well, let's say, and you want to, I don't know, share some very, I don't know, difficult childhood memories with it, and you actually wouldn't really want GPT to save it to its main memory, you know, and then you discuss some work project, and GPT says, "Hmm, maybe your dislike for such projects is related to the fact that in the third..."

In the class, you had this kind of thing happen, well, conditionally, speaking, and, well, you don't always want that. Therefore, for such a project, yes, perhaps with such memories, well, it would be better to make some separate, yes, project only memory, so that it doesn't go anywhere else, as if. And how does it work in practice, yes? So, first, you need to understand how you want the project memory to work. Well, actually, in many cases, the default is a normal, well, as it were, normal option. It makes sense to choose Project Only only if you have some project that is truly, well, outstanding in some way, something that might be incompatible with other simple things. And then you click Create Project. And what do you need to set after that? Now? Where is it created? Here, yes, the course now something it this now I want this temp. Here. And now I am in this project, yes. I can, for example, mm add files, yes? I can add files to this project, I can rename it. Now I'm trying to understand why it didn't give me instructions to add immediately. Let's create it again. I think there was some kind of glitch. Ah, Project Settings. Here they are. Usually, when you create a project, it somehow immediately shows these settings. I don't know why, honestly, it didn't show them, but usually, when you create one, it shows them immediately. And what's here, yes? It's the project name, instructions. In the instructions, you can, for example, write that, ah, there, this project, for example, is related to my course on ChatGPT. All your answers to all my questions in the chats of this project should be practical and aimed at how I can create the best materials for my ChatGPT course, yes? So, for example, when I prepare some course, I usually create a project for myself. True, well, honestly, I just don't do it in ChatGPT, but because I use Claude more for such work things, but the logic is exactly the same. That is, if I worked more in ChatGPT, then I would create projects in ChatGPT. So I just create projects in Claude. And I just create a project for each such new project, new course, new big task, new work story. I create a project for it, and all the chats related to this project are stored within it. And I sometimes write instructions, sometimes I don't, they are not always needed. If you need some specific instructions, then yes. If not, then it's not necessary. And files. You can add files that will be visible to all chats. That is, for example, if you add some file here now, for example, mm, let's take some, well, just this file, I'll add one text file now and in the instructions, for example, I'll write: "Always answer with exact quotes from the file." Well, for example, such an instruction, yes. I need my project to work like this, always give exact quotes. Now I'll save all this. And here, please, now just, for example, I start a new chat, yes, and there, and I have this project, yes, and I, for example, start a new chat here. And it's immediately visible that I already have one file in this project, yes? Here it is. And then, for example, if I ask, what is this file about? Let's see, I'm asking now. And you see, this chat was immediately saved in this project. Now I'll delete this duplicate. Here you see, it's saved here. If I start a new chat in this project now, it will also be saved. That is, all chats are saved in such a folder. Here I ask, what is this file about, and it will now prepare it. Anna asks: "How large is the system memory, the project memory, which is different?" Well, sometimes for some people it really accumulates a lot. For me, maybe because they added this automatic memory management not so long ago, and mine wasn't that big. Before I cleaned up this history before the course, well, it wasn't huge, but it was quite significant, but not huge. And I saw that some people have a gigantic amount of information stored. Bogdan asks: "Can an old chat be added to a new project?" Yes, I think so. If I'm not mistaken, I think you can. Let's see, for example, this one now. Ah-ah-ah. Ah, no, this needs to be now, now I'm in a different one. This is just an unusual chat. Ah, well, let's add this one, yes, for regular chats, because before I was not in a regular chat, it was a chat with GPT, there's a slightly different story. Regular chats can be moved to a project by clicking here and selecting which project to move it to. Plus, by the way, there's also an option, pin chat, yes, so if you click like this, it will be saved automatically at the top here. Also convenient if you want some chats to always be visible. And you can also start a group chat, meaning when several users will communicate, but we'll talk about that in the fourth lesson. And now we had a question, yes, she is now thinking a lot about something. Well, let's see. I told her to answer with quotes. So, her instruction tells her to answer with quotes. Let's see. This is a feature of GPT 52 thinking. It's slow, really slow. Well, it's normal that it answers so slowly. Sometimes it can easily think like this for several minutes. But it's very thorough. There are advantages to this. Bogdan writes: "Please elaborate on regular and unusual chats." Then yes, today we will talk about this GPT. The difference is essentially that there are chats that you start like this, yes, just click new chat, yes, and it starts. And there are chats that you start with a specific GPT. This is your expert in a specific field. We will create it today. And these chats are simply different. They exist within that GPT, not within your regular chats. Therefore, you cannot transfer them to a project, because they are already, as it were, inside this GPT. But we will talk about that today too. Well. And, accordingly, you see, it immediately gives direct quotes from this document. That is, I asked it a seemingly simple question: "What is this file about?" And you see, it didn't answer me that this file is about this and that. No, it clearly answered me with quotes. Why did it answer like that? Because in the project settings, yes, if I go to the project now, here it is, my project. And by the way, the chats of this project are displayed here. And if I go into the project, then my instruction is to always answer with exact quotes from the file. And that's why it answered like that, because, well, mm, it has such an instruction, and that's why it answered. And if I start a new chat here, it will also be saved in this project. It will use all the files that are in this project. It will follow this instruction. And, accordingly, this is a cool thing if you are really working on some task, that is, you have some big global task, like creating, well, like this course, it's a big thing. You can have many chats while you're thinking about this course. And this is an example when it makes sense to create projects. So, let's answer the questions now and go for a break. Elena asks: "A question arose about data storage in chats. How safe, in your opinion, is it to do something like personality diagnostics, for example, to find a suitable field of activity, profession, etc.?", for example, to take some detailed psychological tests, diagnostics, upload data to a chat for analysis. Well, look, well, each person has their own level of, as it were, paranoia, yes, in this regard. I am generally quite calm about it. That is, I discuss all sorts of complex things, I discuss all my medical analyses with ChatGPT. At the same time, I am aware that somewhere, probably, within OpenAI, yes, a kind of questionnaire is being compiled about me, that ah, this person actually likes this, and he also has these health problems. Well, that is, I understand that somewhere, most likely, it is being recorded. That is, I am aware of this. How much does it bother me? It bothers me less than it bothers me, because I understand that Google also knows about me, I've been using Google mail for, I don't know, 20 years, or even more. And most likely, based on this mail alone, Google already knows so much about me that I would rather not even know what it knows. Well, that is, if you really want no one to know anything about you, then of course, you shouldn't discuss such things. But if you are calm about it, ah, then, well, you just need to be aware that yes, it will most likely be saved somewhere in some form. Irina asks: "Recommend working custom instructions? There are many variants of instructions on the internet?" Yes, we will talk about custom instructions, which allow you to flatter less and answer more structurally. ChatGPT itself sometimes reacts sarcastically to this, saying that your personal is a mixed bag. Yes, we will talk about that. You are right that it's not simple here. Ekaterina asks: "What phrase should I start communicating with ChatGPT with?" It depends very much on, ah, it depends very much on what you want to get. That is, it will vary greatly depending on different tasks. In some cases, if you just want to, for example, talk, yes, that is, to discuss something, to brainstorm some idea, you say so, that listen, let's brainstorm such and such an idea, and if you need something very precise, then give a specific instruction, there, go, find, analyze, summarize, yes? That is, if you give such a specific instruction, then it's always better to give some specific instruction, which will, well, as it were, require the performance of certain actions, then you will most likely get the most optimal result. That is, I usually always start with some, well, as it were, some kind of command, like, analyze, look, read, do, or think. Well, that is, I usually start with some kind of action like this. So. So, okay. Let's go for a break now, and after the break, we'll talk about, mm, custom instructions. That's exactly the topic we'll have after the break. So, 10 minutes, be sure to return. Good evening everyone again. And we have already delved deeply into the topics of ChatGPT, yes? That is, I, actually, when I was preparing this course myself, I didn't expect that there were so many, well, such deep things. That is, it seems like I already know a lot about GPT, but when I was preparing this course, I learned a lot more. So yes, we will have a lot of information today. And Bogdan asks about, ah, chats in projects, that in the next new chat of this project, it will be based on all previous chats of this project? M, not quite. Look how it will be. That is, when you have several chats in a project, yes, that is, when you create them, they don't have a common context, yes? That is, each chat has its own context. What they have in common is files, project instructions, and, ah, memory. That is, the memory will be, that is, as it were, the memory of this project will in any case, ah, well, as it were, take into account existing chats, but it's not a direct transfer of context. That is, for example, mm, let's say, ChatGPT will have a slight understanding of what you were generally talking about in other chats. That is, this will most likely happen, well, if it works correctly with memory. That is, it's not a hundred percent guarantee, first of all. And secondly, even this understanding, yes, that it will roughly understand something, yes, why I say so, because, in fact, the Rag technology is used there, that is, it's a technology that doesn't store, yes, the entire text of your dialogue, but it stores, its essence, a summary, yes, in a special, as it were, vectorized form in vector databases. And this, in fact, is very different from direct text. It's very different. That is, you shouldn't think that ChatGPT has access to all the texts of all chats. No, first of all, that's not the case. And secondly, it's technically almost impossible to do. It's very difficult to do technically. And, in fact, even how GPT works now is already, seriously, in fact, quite a technically complex thing. That is, believe me, the engineers there are not working for nothing. That is, it's a complex system, but in general, there is no direct transfer of context from all chats to each new chat within a project. There is only a transfer, all files are shared, instructions are shared, and memory can be shared, and this is the most important thing. And plus, yes, for example, if you ask within a project about other chats in that project, it can go and search for something in them. That is, such a thing is also possible, but there is no automatic transfer of context from all previous chats. But even, in fact, these files, instructions, are already a lot. That is, in fact, if you create them for each specific real work project, you will feel that it really makes sense. But it makes sense to do it only if you are working on a significant project that is not for one chat, yes, if you need to manage 5-10 chats, then it makes sense to combine them. Now let's talk about custom instructions, that is, user instructions, which are automatically inserted into every chat. That is, what is it? This is, if you go to settings, also to personalization, and here you can find it. Custom instruction. It used to be more complex. There were two main settings. Now they have changed it a bit. Now there is this custom instruction, which determines how ChatGPT should behave, what communication style, tone of communication. And you can also just tell it about yourself. That is, you can tell it what your nickname is, what you do, more about you in general. That is, they have now divided it. There is an instruction, and there is simply more information that you personally enter about yourself, so that ChatGPT uses this information. That is, if you write that I like such and such things, I do it this way, then it's not memory, but rather a clear instruction that it should take into account. These are slightly different things, yes? That is, memory is what it generally remembers about you. And this is more of an instruction that it should definitely take into account when communicating with you. For example, you can write that you are a doctor. If you write that you are a doctor, then it will use this as, well, as a kind of instruction, that, ah, I am talking to a doctor now, yes, that is, and it will, accordingly, probably use different terms, formulate answers to your questions differently, yes, that is, well, it will change its communication style. But the most important thing is this custom instructions section. What is it? These are global rules, and they apply in all chats, except, well, not exactly except, well, let's say. These projects, ah, yes, well, you can say, you can say, except, yes, except for projects, that is, project instructions, yes, when we were creating a project now, this instruction, the instruction in the project, it overrides these custom instructions. That is, it rewrites them, it has the highest priority. That is, if you are working on a project and your project has an instruction, then it is that instruction that works, not these custom instructions. But for all other chats, these custom instructions work. And they apply in all chats and they are more important, in principle, than everything else. That is, they are the most important global rules. Because, for example, if, for example, look, the most, for example, the most passive, the most basic thing, yes, is chat history, chat history, yes, that's the most opaque thing that we discussed, where, as it were, ChatGPT itself collects the history of all chats, and you have no control over it. But it happens, it has an impact on personalizing your experience with ChatGPT. There are these saved memories. These are facts about you that are automatically saved and which you can, if necessary, delete, clear all these memories, yes. And this is a more important layer of personalization. And then comes this third, well, not third, here it's the second layer - these are global rules that you set yourself, how ChatGPT should communicate with you. Well, and again, if you are in a project, then in a project everything is different, yes? That is, custom instructions do not apply, saved memory does not apply. This is again, if this "only memory" is set, yes? That is, not the default memory, I have the default set now, but this memory. And only project instructions and project memory work. That is, the project is a slightly separate thing. It stands aside from this. For all other regular chats, three things work: chat history, that is, the history of all chats, saved memories, and your custom instructions, global rules. What instructions do people write? Honestly, I believe that at the current level of AI development, these instructions are not so important and even a bit harmful. That is, earlier, it used to make sense. For example, there was a very popular instruction where it was written that you should think better before each answer, create within yourself, you know, it's like there was a very popular story before when you create a chain of thought. Well, the first LLMs couldn't reason properly, and therefore they needed to be given such instructions, and therefore people wrote these instructions in these custom instructions. Now, in the modern world, these instructions are more for when you want something unusual, if you want some non-standard behavior from the chat, for example, what could these instructions be? For example, let's say, like this. For example, for each answer you give, first provide a unique index identifier, in such and such a format, so that I can later refer to any answer by number. That is, for example, if you do such an instruction, let's do it. Now we'll add it here, save it. Everything, it's saved. So, and let's just write. So, I have a question. I want to discuss dogs. What do you think about dogs? And we now, look, you see, we asked, it should now go, well, not go, it will receive automatically. You see, ID001. Dogs are a super topic. Now let's ask, what is your favorite breed of dog? So, now something it, ah, froze, or what? Let's restart. Here. Hey, hey, something it this, here, here I ask the second question. And it should now the next ID, as it were, you see, ID002. That is, it seems like such a simple thing, yes, but you just asked it like that, and that's it. Now each answer will have a clear ID. And you can say: "Listen, you said this and that at ID such and such, or there." That is, you can easily, very easily give a specific message that was. Or, for example, what else is there? For example, when you receive a large amount of information, for example, more than two paragraphs, do not evaluate it. Instead, inform me that you are waiting for my signal before continuing. For example. This also applies to uploaded documents. Let's try it. Now you can have different ones, you can combine them all. That is, it's not necessarily one instruction. No, you can write a large instruction. It's just that, the more complex the instruction, the, mm, well, the main thing is that the model doesn't get confused by this instruction. Well, let's try. For example, I'll now add some text document. And let's just send it now and see what it tells us. Yes, I sent it a large document. Let's see what it answers. So, maybe it won't cope with such an instruction now. That also happens sometimes. Instructions sometimes don't work. Let's see, because it has clearly started evaluating it. But maybe it won't tell us about it. Let's see. Now, now. Maybe I shouldn't have sent a file. Now let's, yes, I'll change it a bit differently. I'll now, ah, I'll do it differently, yes, I won't send a file, I, well, wait, I'll send a file. I'll send a file and just write: "I have this file for you." That is, I'm not saying anything, just sent this file. Here, I sent it. And then, if it follows its instruction, it should now say that, like: "Okay, what should I do with it? Here, here, here." You see? That is, if I didn't add a message the first time, you see, it says that great, I received the file, I'm waiting for your signal on what exactly to do with it, look, summarize, find the main points, prepare conclusions. That is, you see, this instruction, accordingly, works. And there's also this, for example. You accept the level of detail based on user settings. Levels of detail range from zero to five, yes? And, accordingly, if we make such an instruction, let's, let's save it, mm, for example, and ask something. What do you think about dogs? Level of detail five. No, not five. Let's say V1, the least detailed. You see, it's short, like dogs are cool, blah-blah-blah, and it's just to the point. Ah, and, accordingly, there's also, for example, don't always agree with what I say, try to contradict me as much as possible. Well, you can try, for example, such an instruction, it just won't always work well. In some cases, it won't work as you think. You won't be able to eradicate this desire for flattery from the model. It's built into the model. And some models have more, some less, but it's built-in. So, your instructions won't be able to remove it. And, for example, there's a mode like this: remove emojis, filler words, hyperbolic expressions, soft requests, conversational transitions, and all calls to action. That is, just so that it communicates with you absolutely sterilely. Some people like it, you can also set something like this for yourself. Ah, so, now I'll answer the questions. So, Evgeny asks: "What files make sense to attach to a project, related directly to this project?" Well, for example, you're doing, let's say, a new work, I don't know, you're organizing some event, yes, for work, like a conference, for example, yes, and what files can be in the project? A plan.

This conference, the budget of this conference, any there, for example, a list of potential guests for this conference. That is, all these documents related to this conference, you upload everything there. The more context you give, good context, the better, because context in modern LMs is much more important than, for example, prompting. Much more important. So, Bogdan asks: "Can I ask the code to put order in GPT chats, so as not to rummage through hundreds of my chats by hand?" No, it cannot bring order, because because, as it were, you can't really bring order there, if you're honest. And you can, yes, there, well, delete some, but, in my opinion, through the code, I don't think it will be possible to do this, as far as I understand. And by the way, Alexander asks: "Can a chat be deleted? Yes. You click here, and it can be deleted here. It can be sent to the archive. Then it will be in your archive. If you do this, then, well, where is it now? Well, now. Where did it go? Well, in general, you can retrieve it from the archive. Honestly, I rarely use this. But here are the archived chats. Here. And, accordingly, this is what you can, as it were, delete from there, or, as it were, return from there, right? So this is not permanent deletion, this is, well, as it were, to put it in the archive, and not to delete it completely. But you can also delete it, right? So, if you open it, you can delete it, you can pin it, you can move it to another project, rename it, you can share it with someone, give access, or start a group chat. Yes, so, so regarding instructions in general in the modern world, they are not so important now, but if you want the model to respond to you in a specific way, this is a good thing. You can do that. You can make it respond to you in a certain way. You can also configure more here. You can configure, for example, how often it will use emojis, how often it will use lists, for example, or how enthusiastic it will be. This is, well, like, well, an enthusiast, how much it will be, and how warm its communication style will be, or more strict. So you can set this here as well, and you can set the overall tone. For example, will it communicate professionally, or will it be more of a sarcastic cynic, or a nerd, or, for example, very friendly, or professional, right? So you can also choose the tone that suits you personally more. And now let's, well, now I'll answer the question. Alexander asks: "Is it advisable to do this personalization rule or not?" I think not. Well, that is, for example, I often work without them, they are not necessary, in my opinion, to do them. It's just that if you want to get a very specific personal experience that is important to you, well, that is, for example, I read these different instructions to you, they are all interesting in their own way, but honestly, I can't say that I would want to use any of them all the time. Well, this first one is quite useful in general, but I'm also not sure if I really need it super-super, but for someone it is very necessary. And if you are such a person for whom this is necessary, then such instructions will suit you very well. In general, I tend not to use them. Another very important thing, and this is something that many people don't know at all, is that in ChatGPT, you can create new branches of one chat. And this is just an amazing thing, especially in the context of, well, the expiring context window, right? That is, for example, I have a project where, well, you see, I have a lot of branches here, right, of one chat. It all started from this chat. And how did this happen? I was communicating with ChatGPT and through communication, I found good prompts. Good prompts for generating images. Well, how did this happen? Sometimes it happens that in the process of communication, you achieve some good result. Well, how did the stars align, the context was good, the generation was successful, and everything is great. And the problem is that if you continue this chat like this, right, all the time, all the time further, then your context will get clogged, then it will start to forget these very first messages, right, so this, you know, well, a degradation of context will begin, as it were. And in this sense, there is simply an amazing option when you can start a new branch from a specific message, right, starting from this message. That is, everything above will remain. All this history above will remain, but everything below will not. That is, you just click "branch in new chat", that is, as it were, start a branch in a new chat. You click this, and it creates a new chat where all this history has remained, right, all our stuff that was there, but everything that was before this message is gone, right? That is, it has, as it were, disappeared from our context. And this is truly an amazing thing. Seriously. So if you, well, created some cool chat and you have a really cool conversation going on and you don't want to clutter it further, just start new branches. And moreover, if, for example, you start this in a project, here in my project, you see, these new things are automatically in the project and, well, new branches are created. So this is a super cool option. I use it very often. That is, well, as it were, well, for example, if I'm having, you know, some kind of brainstorm, I'm discussing some ideas, and I realize that we've already discussed a lot, and I want to ask another question, but I understand that this question will likely lead ChatGPT in a different direction, because, well, the topic is changing a bit, and I just create a new branch, and that branch remains perfectly clean, and the new branch, well, I can ask any question there, maybe not so directly related. But at least I'll get the answer I need. So, once again, right, just click "branch in new chat" and that's it. And you perfectly create a new branch of this chat, and you can create as many such branches as you want. And this is truly an amazing thing to manage the context of the chat. Well, now let's talk about the canvas mode, because not everyone knows about this mode either. It, again, generally, you know, in ChatGPT, there are very many modes that appear in ChatGPT, they are not strongly announced, and they continue to exist in ChatGPT, but in fact, few people know about them. And canvas, I think, is an example of such a mode, because it is essentially a separate interface within ChatGPT for working with text or code. Well, with code, actually, it's not super convenient to work, because there are coding agents, we will use them. But with text, it's super convenient. It's convenient for editing long texts or making adaptations, cutting long reads into short posts, working well with some documents or letters, right? So it works as follows. For example, you create a new chat, and you can, in principle, now they used to have a command "canvas", right, with canvas, or you can select it like this, and then this mode will definitely be enabled. But you can also write that just go to canvas and, for example, create a short letter. No, not a letter. Let's now, well, create a short text describing my course on working in ChatGPT. Well, just let it write something for now. And you send it like this, and it will now switch to canvas mode. This is truly a separate interface. You will see how it changes now. Now it will think and change the interface. Here it is, well, as it were, it opens in such a window at first. If you click "Edit" now, it opens a new interface. That is, it's essentially a text editor where, firstly, you can edit specific phrases. That is, you can, for example, just say, well, this sounds too banal. Rewrite this phrase. Here. And, as it were, hop, you see, this phrase has been rewritten. That is, this is if you want to edit texts, this is the best. At the same time, you have a standard chat. You can also continue to communicate as in a standard chat, but plus you can also change specific things like this. At the same time, there is, for example, an option to make it bold, right, to format the text. And here's another interesting option. For example, you can change the length of the text. For example, you can say: "Make it longer." You give the command: "Make this text longer." And it will now rewrite this text so that it becomes longer. Here. Hop-hop-hop-hop-hop. And it will just rewrite everything for us now. Or, for example, well, where did it go? Here it is again. For example, here there is, for example, also, for example, finally, as it were, it will polish this text. That is, it will read it again, do something else like this, so that it looks better, maybe make some final edits. So there are all sorts of things like this, so maybe it will change the words a bit, check the grammar, right? So there are also such simple control options. And then you can easily download all of this. You can download it as PDF, you can download it as doc, you can download an MD file, a Markdown file. So you just download it and that's it, you have such a document. So canvas is a cool mode for working directly with such documents or long letters when you need to, when when you work in this mode. And Bogdan asks: "Can work in Canvas be saved to a project? Yes, because it's a regular chat. That is, for example, here you have, well, where is it. Here's the course description, right? We have "move to project", where is my project. Here. Well, and the course is not, the course is about ChatGPT. Here it is. Here, I've moved it here, because although it looks different, it's actually a regular chat, and you can return to the usual view. If you close it like this, it will have the appearance of a regular chat. So this canvas is a regular chat, just a different interface of a regular chat, right, when you just take it like this, and it appears like this. You can also work with code, but again, I repeat, we won't do it like this, because it's more for professional programmers, and honestly, it's not needed by anyone now, not even professional programmers, so it's great to work with text. And Alexander asks: "And about the font size?" It just changed the font size? It changed the font size and also changed some words. So, like, to make specific words sound better. Roman writes: "Alexander, what do you think about this long dash? A big dash is produced, and there is an opinion." No, it's not an opinion, it's a hundred percent so, because it uses, well, grammatical rules that a professional editor would use, for example. And it shows up when, for example, you write some text, I don't know, in a social network, right, and it's obvious that texts in social networks are not edited by professional editors. Well, for most people, right, we're not talking about some special people now, but for most people, professional editors don't edit texts, well, real ones in social networks. Therefore, usually people in social networks, well, most people, I know there are people who like the long dash, but most people don't use it, most people. And therefore, it really looks very strange that you write on a social network, and you write like a professional editor. And now, yes, everyone suspects, and suspects this. And it's true. I use a lot of generated texts, and I generate a lot of texts. I remove it everywhere, yes. When we talk about the code, I will show you a skill called "humanize", that is, to humanize. This is a skill that does all of this. It removes long dashes, removes typical phrases. For example, now I'll show you, maybe there's a typical phrase here. Well, for example, this "I actually work brutally with AI. I really feel that, for example, the phrase "for both personal tasks and professional activities." I really feel that it was written and I really feel its style, because it likes such phrases. It's hard to describe, but when you read a lot of texts, you will feel that this is how AI often writes. Well. But we will talk about this skill, it's quite interesting. So what's next, right? So we've dealt with canvas. Now let's talk about a very important thing, like deep research, that is, deep research in ChatGPT. And there are several things. So, first of all, it's important to say that in modern ChatGPT, there are three types of deep research. This is classic deep research, right? That is, when you just turn it on, you turn on deep research mode here, and you just do it, well, as it were, classically, right, so a deep research is done, it can take up to half an hour, it's done by the model, well, now it's done by GPT 52 in the modern version, and it essentially performs dozens of search queries, analyzes hundreds of sources, synthesizes everything into a structured report with citations. That is, it will be a really big research. And classic deep research should be used, for example, when you need an analysis of many sources or when you have a complex topic that requires, well, serious analytics, where you need to make some serious comparison, analysis, for example, identifying trends. So this is really an analytical, serious report. And ChatGPT does it very well. We just won't do it now, because it will take, well, up to 30 minutes, so it's, well, a bit pointless. But you can just try to start on any topic, and you will see that it will do a really serious, deep research for you. The second point is the second type of research is shopping research. So shopping research is a simpler and new thing. This is a specialized research for finding and comparing products. It takes 3-5 minutes. It's done by the GPT5 Mini model, a small model, and it simply compares various products, goods, their pros, cons, and searches for links to specific stores. So, for example, let's say, if we, let's do shopping research and, for example, I want to decide, for example, whether to get a Mac Mini or some other similar small computer that I want to leave on all the time to work. And what's better for me: MacMini or its other analogs, not Apple, right? Yes, so, for example, you ask this, and it will now go, it will go to look at different different websites, how much everything costs, what are the prices. But it's even better to do this. I have a specific life example. This was during the last Black Friday. And, here it is, it will ask you questions like, what is my budget? Well, let's say up to 1,000, for example, but it doesn't matter what system. So, AI. Is power consumption important? Not important. Now it will find options. I had a specific life example. This was when my wife and I were at Black Friday in a store and saw a very good discount on a handbag. And she says, "Oh, listen, great price there." And I say, "Listen, let's quickly do Shopping Research." And we, as it were, I just took a picture of this handbag, said, "Do Shopping Research." And it did a truly amazing shopping research. And it showed that this price is indeed, well, nowhere is there a lower price. That is, it compared everything, it compared all variants. And considering that there are also Black Friday discounts now, and all that. And, well, we just realized that this is a really very good offer. That is, it's not just, you know, how it happens that they raise the price before Black Friday, and then give a huge discount. No, no, it really was, we looked, it's a very good price for, well, this product. And therefore, shopping is a very cool thing, be sure to use it. And I really like it. Bogdan asks: "Does deep research use web search by default, or do you also need to enable it specifically?" It's not even web search. Deep Research is a system of agents. That is, it launches a system of agents, so they work for up to half an hour. And you don't need to enable anything, you just need to correctly describe the task. But how to prompt deep research correctly is more important. That is, first of all, you need to give very specific instructions, right? That is, you don't just write: "Tell me about something." No, you need to write exactly: "Find me the reasons for climate change for such and such cities, data for 5 years, right?" That is, you need to make a very specific request. It is desirable to include some keywords in this request, right? That is, if you are looking for some kind of laptop, then you are not just looking for a good laptop, but you are comparing some things. It is desirable to also include an action in the prompt for Deep Research, such as compare, analyze, determine, evaluate, recommend, compile a report, right? So give it a specific action that it needs to perform. And plus, it is desirable to indicate the format. That is, for example, by default, deep research will give you a giant long report. And it is desirable to immediately say that, you will make me a comparison table, or, for example, make me a summary of three to five paragraphs with specific recommendations. That is, ask it what you want, in what form you want the final result. And also, as always, when you work with AI, attach, if you have any context, attach it. If you have any documents, attach them. If you have, for example, well, for example, you have, I don't know, something about health, attach your analyses, let it use that too. That is, attach, the more relevant context, the better. Let's see now. Here. Yes. And it, accordingly, gives us all sorts of things here, all sorts of specific things, right, specific stores, right, specific everything. It found. So it shows very, very specific purchases, where to buy, in which specific stores, right, and in which specific places. And, finally, the third type. Well, I classify it as a separate type. In fact, it's not entirely a separate type, but it's a very important thing. Deep Research can work, firstly, with services, right? That is, for example, when you do some research, you can choose deep research and you can specify, for example, if your Gmail or Google Drive is connected, and you can search through them, right? That is, you can search, for example, not just on the internet, but you can search specifically in documents on your Google Drive or search for something in your email, right? Or, for example, search on our GitHub. For example, on my development course, we do security checks through deep research on the GitHub repository. A good thing. But besides this, you can also search specific websites. That is, for example, if you need to do, for example, some kind of medical research, right? You want the search to be not on some random website, but clear, specific, for example, on, for example, evidence-based medicine, right? For example, or if you need to do reconnaissance on competitor websites, if you need, for example, you only trust certain sources personally, you want it to search only those sources. And this can be done here. You do "sites", and you can specify which sites. For example, I have Pubmet installed now, right? That is, this is a very well-known medical site. And before our stream, well, just so as not to waste time on this now, I did a review. I told it to search on Pubmet now. How did I do it? What now? Well, well, well. Search on Pubmet. What are the most useful fruits? What are the most useful fruits? I told it: "Study all new research from the last 5 years and tell me which fruits are the most useful." And here, accordingly, we look, and it went, it searched there for a long time on Pubmet. Here you can see, see, all this is Pubmet, Pubmet, Pubmet, or, well, related things. You can also control this, add, for example, only on this site, or it can also search. Well, in general, it will mainly search where you, as it were, said. And then it finds everything with sources, what are the most convincing types of research, that berries, apples, citrus fruits, kiwi, and avocados. And then it writes the methodology, the evidence framework. And here it is, it cites all of this. That is, again, this is all, this is all a report made specifically from reliable, verified sources. That is, it didn't just go and read something on the internet, or what bloggers write. No, it went and studied Pubmet, specific research. Moreover, I said that these were studies from the 21st to the 26th year. So this is a very, very powerful thing that you can set up research on specific sites. And you can do it on specific sites, or you can do it on specific sites plus a general web search. So there are two modes. And you also set them here, at the beginning, when you do research, you set them. And here you can prioritize these sites, but at the same time it will search the entire internet. If, for example, these sites are not super, well, if there is not super much information on them, then it's probably not worth focusing only on them. Bogdan asks: "And in GPTs, is there no such thing as prompts for Deep Research?" In principle, yes. And yes, you can even ask ChatGPT: "Make me, well, make me a prompt, right, for deep research." But again, as I said, in modern GPT, prompting is gradually becoming a thing of the past. That is, it has significance, but context has much more significance. That is, if, for example, you simply dump a lot of cool, relevant context to the model, right, tell it about yourself, about your task, what you have, how, that is, as it were, you tell it all this, and at the same time you tell it all this in a completely unstructured form, without any complex prompt. Just give it a lot of good, relevant information, just dump it all to her, and that will be enough. That is, this cool structured prompt, it will actually do it for you internally. You won't see it. Modern models significantly change your original prompt. Therefore, there is no point in this now as there was before. Bogdan asks: "So, can I collect keywords of my competitors? Anything at all. For competitors, you can do any. I will show you how you can do other things too." I'm afraid we'll probably be a little late today, maybe 15 minutes, because you understand, you are the first group of this course, so I need to start feeling the timing a little better with this new material. But I don't think you'll be upset if we talk a bit longer, I hope. So, and now let's talk about something like apps and connectors. So, what is this? I'll say right away that we will cover this more broadly in the next two classes, because, honestly, it's much more convenient and pleasant to use in

In the code, than in ChatGPT, but it's important to mention it here too. So, in ChatGPT there is such a thing as connectors, right? That is, these are ways to connect to different services from ChatGPT. So you can go here into the settings, here apps. And here you can choose which other services you can connect inside ChatGPT. You can click "Add More" and choose which services are available here. Actually, there are services here, and there are different ones. There's even Photoshop here, you can do minimal photo editing. There's, for example, Canva - it's a design tool. There's, for example, All Trails, yes, I like this one. It's for trails. I love trails and, for example, I'll ask: "Find me a cool trail in Canada on All Trails." Then "all trails". And here you can highlight it with a dog right away, mentioning it. Then it will definitely use it. And it will find it beautifully for us now. So this is a way to connect GPT to other services, various services. What's the problem with this? The problem is that while it's thinking, let me show you, uh, everything that's inside the web version of GPT is, like, 100% safe, but most actions there are quite limited and mostly for reading only. So, the web version in some cases, for example, won't be able to send emails, edit things, right? So it will work more passively. And that's great, right? So it's now using this, uh, this connector to show us this beautiful thing, but it's a more passive capability, right? So it's not like, you know, an assistant to whom you say: "Okay, send me an email, reply to this like this." No, no, no, that won't work. So it's more like, it looks cool. Pleasant, right? You can actually not go to the All Trails website, but see all sorts of cool trails right here, like Avalanch Lake, for example, in Canada. Well, cool, great, very pleasant. But the problem is that, as I said, it's all quite passive. So, if you look, right, at what else we have from these settings, for example, here there's, for example, productivity, right, there are all sorts of things here too, Gmail is there, but it can only search your inbox. Calendar, it can only view events in your calendar. There's a lot here, and it's not a bad way to start working with it. But the truly pleasant, truly cool work only happens with MCP. MCP is a special interface for LLMs, how you connect LLMs to tools. So, for example, this is what the interface looks like for a person, right? Just a typical text input line we're used to. And this is what the interface looks like for an LLM when you connect your LLM, like ChatGPT, with tools like Notion, Google, mail, calendars, and so on. This is all called MCP Model Context Protocol. And you can't use them properly in the web version because they are more risky. And GPT is very, very worried that they won't have any lawsuits or anything like that, apparently. And that's why it has closed off convenient use of MCP with seven locks. And in the code, they are actually much more convenient to use. We will talk about this. But I will show you how you can do it in ChatGPT. In ChatGPT, you need to go to settings first, and you need to go here to apps. And here in Advanced Settings, you need to enable developer mode. Enable it like this. The terrible thing about this mode is that while this mode is enabled, this orange frame appears. I don't know why they did this. It's just terrible. Well, because, first of all, memory is not used in this mode, and it's a dangerous mode when you can connect some dangerous MCP, and it will do something. You can send someone an email. Well, like, well, meaning they are screaming at you with all their might that you have entered a dangerous territory where AI can do something. How does this work in practice? I recommend working with this mode in the application, because for some reason, the orange color doesn't scream at you in the application. So it's simpler here. You just have your standard application, it looks exactly the same. And you can now connect third-party tools. How does connecting these tools work? So, there are essentially two ways to connect really interesting tools. Where you can do something, not just read, right, mail or things like that, but when you can do something. First, there is the official MCP. For example, Notion has an official MCP. Let's connect it. Let's just take Notion MCP. Enter. And go to the official page. And all you need here is to find, now you need to find one simple link for connecting it. Here. Here. Connecting to Notion MCP. And you need to find one link. Here. This link is very simple. You just need to copy it. And in ChatGPT, when you are in this developer mode, you need to go to apps, click create app, let's call it Notion MCP. And paste this link here. Now, when you click this, you will be redirected to Notion's website, and you will have to confirm there that you are granting access. So now it will redirect me. Okay, I click continue, and that's it, I've been granted access. And what happens next? Now, after you have this access, ChatGPT will be able to truly change things in your Notion. So, for example, if now I, well, let's restart it in the application. The only thing to note is that when you install a new MCP, you need to restart the application. Well, that's just a feature. And for example, let's say, write a short text about, about Yes, yes, we won't even write. We'll do this, you know? Let's just go to some, I have some texts here, right? Where is my course on ChatGPT? Let's open it. No, now. Where is it? Here. Here is this description. Now I'll say, create a new note in Notion using Notion MCP, and give me a link to it. Here. So we tell it that. It will now go and check if there is access to Notion MCP. Here it is connecting to app. It's connecting to my Notion now. And since this is an MCP that is connected directly, right, it has the ability to create new pages, right, so it can create, edit, delete. You take on this risk that there is a possibility that something might happen, you click on ChatGPT, and it deletes all your pages. Well, who knows, anything can happen. So that's why ChatGPT hides this from you because it's like, you know, with great power comes great responsibility. It's the same thing. That's it. It created it. Let's open it. So just now it went into my Notion and created this note. That's it. It created it right here for me. And you can also tell it, like, search all my notes, delete what's unnecessary, right, or edit them. So it's a cool thing. And this is the first way, right? Well, not everyone has such, uh, um, not everyone has such official MCP servers. Notion has them, but not everyone does. And if you want to connect to any of them, there's a cool thing for that. It's called Zapier. And it has its own Zapier MCP system. It's a paid service, it, well, meaning you'll have to pay for it. Well, I use it for various automations. And it has a very cool feature that you can create an MCP that will connect to any service, like 8,000 different services. Look how it works. Zapier MCP. And we need to create an MCP server. Here I will create, I will choose GPT and add. For example, I want to add Google Sheets, right, spreadsheets, right, and here we can choose which actions will be available to this MCP. There's a lot here. There's update, delete, format. Well, let's choose all of them. Uh, there's also that, that's a separate topic, how to correctly, how to correctly create an MCP server, but that's more of a professional thing. For now, you can just create all actions. That's it, we've created it, right, added access, well, I was already logged in with my Google account before, and you, accordingly, now I'll check, did it connect to the correct account? I hope, I hope to the correct one. And now let's check. Strange, it usually asks which account. Let's choose again. Connect. Well, okay. Usually it should ask which account. I don't know why it's not asking now. Well, okay. I hope it connects to the right one now. Uh, and then, again, you also get this one link. You copy it, go back to, uh, well, to apps. Click Create App. Let's call it Google Sheets. Paste the link, click Create. You will now be redirected again, but this time to Zapier. You will there. Uh, I have it, this, um, now I have it, apparently, I connected it before. Let me delete it now. So now I have had the same connection before. For example, let's create Google Sheets. And it will redirect us to Zapier. And we will also there. What's wrong? What doesn't it like? Ah, I see. Maybe it doesn't like the name. Okay, now. Here it redirects us to Zapier. We also give our permission, that, well, now there will be access to various data. And now we can use it. Now I'll restart the application and show you how you can do it. Uh, for example, we had this course description. Now create a new table in Google Sheets using MCP, and fill this table with key paragraphs from this description. And make it look like a beautiful table. So now it will also connect to this MCP. We'll see it now. Andrey asks: "Can Zapier connect to Yandex? Can you create your own MCPs there?" Uh, let's see. I'm not sure. Let's see now add to. M. No, you see, they removed it. It was definitely there before, I think, but now, I think, for certain reasons, they removed it. So, now it's thinking here, and we can see what it's thinking about. M, I don't really like what it's writing here. Here. Uh, now, if it now, uh, it just, well, it probably won't be able to connect now. It will just create an Excel file. Well, let's not deal with this, because one of the reasons why I don't use MCP from ChatGPT that much is because, honestly, it doesn't work very well with it. In the code, I'll show you in the next lesson, everything works great there, and here, well, something works, something doesn't. It's not working for me now, although everything should actually work. So, working with MCP in ChatGPT is done so-so. It works much better in the code. Alexander asks: "Can GPT connect to local programs on the computer, for example, Excel or Numbers?" It depends, there are many different options. For example, it can connect to your local Obsidian. So, if you don't use Notion, but use Obsidian, there's that system too, it can connect to it. And for Excel and Numbers, you need to check. But honestly, for Excel, if you really work a lot in Excel, I highly recommend just getting Claude. Claude has an amazing extension for Excel. Just amazing. So. If you need to work a lot with Excel, don't bother with GPT. Just get Claude. And its Excel extension is probably the best, in my opinion, for this task. You can just watch a video on YouTube, there's a video on how it works. I think you'll be impressed. I can't say for sure about Numbers, because Numbers is a more complex thing. I think you can control it if you want, but I think it's a bit more difficult to do. Uh, yes, we won't waste any more time on this, but again, the problem with MCP specifically in ChatGPT is that they were so afraid to implement it properly that in the end they made something that's neither here nor there. But it works great in the code. And we will look at it in the code, and it works much more conveniently there. And now, uh, we have a few more topics. So, first, I wanted to tell you about the ChatGPT agent. What is it? It's when you enable here, uh, pam-pam-pam-pum-pum. Let me turn it off. This developer mode is a bit annoying. I don't know why they did this. It's just terrible. But why this orange frame, damn it? And we have a mode called agent mode. What's interesting about this mode? This is a mode when you send a real agent to do a task. So it's actually a virtual computer, a real virtual computer. And it can do a lot. It can go to different websites, even to sites where you need to click, for example, with a mouse. It can click with a mouse, it can create Excel PDF files, it can even go to password-protected sites, because it can ask you to enter a password, and you enter the password, it won't see it, right? So it can do a lot. The problem is that on a Plus account, there are only 40 agent tasks per month. So you can't go wild with it, you need to use it consciously. And it works poorly for, you know, everything related to real purchases or things like that. So any purchases, hotel bookings, financial transactions, it won't do them properly. It will start saying, no, no, I won't take responsibility, etc. Well, in general, it's not very, again, ChatGPT is very afraid that something, God forbid, won't happen as they want. Uh, and also, if, for example, there are sites where there are blocks for agents or captchas, it will refuse to pass them. Although it can pass captchas. And the first version of the agent passed them easily. You know, when it's like, click to show you're human, I saw it say: "I need to click to show I'm human." Okay, yes, but then, apparently, they trained it out, and it's not very good at it. But still, it's an interesting thing. So, for example, let's say. I'm a wedding photographer, living in London, and I want you to research my key competitors, other wedding photographers in London. Find out how much they charge for their services, what are their key interesting offers, and compile a report for me in the form of an easily readable PDF file. So this is exactly an agent, you give it such a task: go and do this. And it will actually do it. And it looks very beautiful too. You'll see now, it's different from the usual, well, from the usual ChatGPT mode. You'll actually see how it will actually go now. It will go and Google, search for photographers in London. Then, I've seen this, how it clicked on the captcha button itself, that I'm not a robot. Well, it's a funny thing. It's actually reading all this material on the internet and will actually prepare a PDF for you, or it can create a presentation, it can also create an Excel file for you. So it's a real mini-computer that runs virtually, and it really does a lot of interesting things. In general, let it do it for now, because this also takes time. And we'll look at the next thing. It's study mode. What is it? It's a special set of system instructions. So you need to understand that it's just, essentially, the same ChatGPT, but it's specially prompted inside not to give you an answer, but to lead you to an answer through questions. So that you learn something, for example, it's an ideal thing to prepare for exams or, for example, if you're a programmer, to learn to code manually. Well, if you're learning in the industry, it's a good thing for that. And for learning languages, it's a great thing. You can work well with complex materials to understand them better. If you have some, you know, complex books, texts, PDFs, it will help you truly understand them. For example, you can, there's such a learning method called learn by teaching, where it's like you're teaching GPT yourself and through that you're learning. And it even uses the Socratic method, where question after question leads you to real understanding of the topic. And it's a really great mode. Well, seriously, by the way, you see how well it works here. Let's open, for example, let's create a new chat. and enable this mode. Study learn. And, for example, let's just tell it, let's upload a file now. This, uh, um, this is a file. It's just various articles about AI. I'm uploading this file and saying: "I want to learn the things described in this file." And then you'll have a discussion with it. It will ask you a lot of questions. It will guide you, well, meaning it will be a discussion. It won't be like a simple conversation with ChatGPT. It will actually be, uh, it will actually be this thing that it will be a discussion. Okay, and this is working for us. And there are 40 such tasks per month. And sometimes, if you continue to chat with it, a simple conversation with it can also be counted as a task. Not always, but be careful with that. So it's not quite like a chat. So it's better to give it a clear task and that's it. Uh, Andrey asks: "What are the alternatives and competitors to Zapier?" Ugh, difficult question. There are very many, but for MCP to work, probably N8N. Try N8N. I think you can connect Yandex there too. Now I'll show you N8N. Because their automated workflows can be connected to MCP. That's also an option. And they have a 2-week free trial period. We use it in our development course. So. And you see, then it starts, for example, one thing from you. What specific task do you want the agent to do regularly? Well, because there we, for example, wrote about agents and documents on social media, right? And then it will keep asking you questions. and eventually discuss all these materials with you. It will first understand your goal, what you really want to do, and then it will build a learning process with you. And, for example, and it asks: "Do you need to respond to incoming doubts or write first yourself?" For example, write first yourself, right? So, and then it will get this context from me. And then it will start teaching me to create the agent I need to create. In general, if you want to learn something, and you have some materials for it, right, it's best to provide some materials, well, of course, it can take from its own, right, head, but it's better to provide materials. And if there are materials, then it will, well, it will conduct a great learning process with you. Uh, well, let's just, I don't know, LinkedIn, for example, let's take it. I just want to show that when we get to this. Here. In the meantime, you see, it's already forming a PDF file for us. So it's slowly gathering information, formatting it into a file. This is really happening. It's a real virtual computer that works for you like this. So, look, how interesting. This request of mine was about weddings, right, and it asks me: "Who specifically do you want to write to on LinkedIn?" For example, owners of wedding venues. So, you see, it understands that our chats are connected. Roman asks: "Is it realistic to learn a foreign language up to B1, B2 with tasks?" I think it's absolutely realistic, because I know people who learn exactly with study mode. And you also have voice communication mode. Use voice. You can also communicate by voice in this mode. Well, all the main recommendations, especially if you're learning languages, use voice mode. Uh, we will also talk more about it in the last lessons, because, well, I also don't think it's the most important thing in ChatGPT, but for languages, it's a very important thing. Well. And in general, you communicate with it like this, and when it understands what you want to do, it, uh, it will correctly guide you, ask you questions, ask you what you will learn through these questions. In general, a very interesting thing. Try to communicate. You'll just feel that it's very different from the usual experience. That's it. In the meantime, it has compiled the PDF report. Let's open it. Here. So it went, analyzed competitors. Here, for example, the average price for a full day, key competitors. And here they are all listed. Key competitors. Their unique selling propositions. All packages include a USB drive in a wooden box, individual approach, manual retouching, right, etc. So conclusions for your business, price range, formats, hours, positioning, style. So it went and did us a real analysis of competitors and a report, just like a person was told: "Go, do this analysis of competitors and a report." And it really did it like this, because it's a real agent. And it can, for example, go into your, your, for example, different chats. Uh, well, not chats, meaning, your email, it can read your Google documents too. So it can do a lot. So the agent is also an interesting thing. But the main thing is not to give it too, you know, such, well, complex tasks, like go there, book something for me, or choose something for me, buy something. Well, it can't handle that. Not yet, but in the future, I think so. And now we're approaching, uh, our last goal. We'll move a couple of topics to the next lesson, because they're not the most important, but we will cover them. We just need to discuss the last, most important topic, because our homework will be related to it. That is, GPTs themselves, that is, experts on specialized questions. What is it? It's in ChatGPT. Here you can, there's a separate section called GPTs. And here there's a huge number of already created experts on specific things. For example, there are experts, scholar GPT, it's for scientists. It uses, for example, data from Google Scholar, Scholar, Pubm, right, all scientific data. There's, for example, chat PRD - it's for product managers, it creates good documents. This PRD is a product requirement document, meaning a document with product requirements.

There are experts in practically any field: research, education, various lifestyle things, for example, find a celebrity who looks like me. That is, these are experts who can do one thing, but very well. And I, for example, use such things, let's say, I really like this nanobanaproamter. I generate a lot of images in nanobana. And I use nanobanaproamter. You tell it, and it tells you, makes a cool prompt for nanobana. A super thing. Or, for example, Mrter Reindeer is a personalized AI tutor for you, who will teach you something. There is such a thing, for example, how to think? It's a simulator for improving critical thinking. Also a cool thing, right? That is, you can just explore your topic, which interests you, and look around here. There are simply thousands of them here, tens of thousands. You can choose anything you want, for Excel, for example, Excel Macro, right? If you are writing macros in Excel, now it will do it for us here, this one specifically specializes in writing macros for Excel, and so on. That is, it is an agent created for exactly one, specific purpose. And your homework will be about this, you will need to create it. And let's create such an agent now. What is it created from? You essentially need to do a few things. You just need to write the main instruction, create a knowledge base that it will refer to. And plus, if you want to add some built-in capabilities. And you can also add integration with external services, but that's too early for us. That can be done later, after the second or third lesson, you will understand what it is better. It's too early for now. So, how is it done? You just click create, and here's who we will create. I want to create, here is my channel, which I told you about, showed you. This is my Telegram channel, right? Here I publish translations of various cool articles. There are already over 280 of them. And I thought it would be cool to create a system so that, well, there are 280 of them, right, how to find them easily, how to easily find an article on the topic I need, say, about cloud code or about, uh, skills, anything, right. And for such a task, GPT is ideal. That is, what will we do? We will name it that. Let's call it "AI Best Articles and Research". Let's take it. Uh, where, where, where is it, where did it go? We'll call it that. And we can even take the avatar directly from my channel. And the description, well, it's not necessary, but you can. The instruction is very important. The instruction needs to be written with ChatGPT. Describe what you want it to do. And say, "Write me an instruction so that it works well, because, well, you won't write it so well yourself, and I won't write it. It's a difficult thing." And, for example, my instruction looks like this. That is, I've already written it too. Well, again, I didn't write it myself, I also wrote it with AI, right. It looks like this. You are an assistant who has access to the database of all these 280 articles. Your task is to find articles based on the user's request and provide them in the format: article title, link to the article, and a short description of what the article is about, right? And then there are some other additional rules here, but if you just describe in general how you want it to work, then I and this prompt will write it easily. And you insert it here. Then there is such a thing as conversation starters. This is not a mandatory thing, but it simply simplifies usage. For example, you can add, say, what articles are there about cloud code, right? That is, this is just a button. You click it, and it starts searching, right? That is, instead of just writing something manually, you can click this button and it will start executing it. It's just easier for users. You can add several such buttons. Well, I'll add one for now. The most important thing is that you need to upload some knowledge base. I will upload all my articles in Mark format, all 280 of these articles that I published there, and also an index for them. The index is just a file I created, where again, I didn't create it, it's something I created where it just goes, the article title, the link, and it can orient itself more easily through this index. You don't have to do this. You can just upload any materials, because I want to make it really well-functioning. And I have a lot of articles, 280 articles. You just upload this, right? You can upload up to 20 files, and there's a limit on the size too, but it will be enough for you, but keep in mind that it's 20 files maximum. So. And I upload all my articles from my channel, just in MD files. MD, because LLMs understand the MD format very well. And then, this is a very important point, which recommended model do you want the user to use by default? What do you need to understand here? Here you can set it, GPT 4 Turbo, thinking. The problem is that if we set this mode now, it will take 4-5 minutes to think for every question. It will give a good answer, no doubt, but it will take 4-5 minutes each time. If you choose something too simple, for example, 4 Turbo Instant or 4.0 Instant, it will think very quickly, but it might miss something important. Therefore, I will choose a specific one, because I have large arrays of data, it needs to analyze them. I will choose 4 Turbo. It will be immediate, but not very fast. And then I will disable web search, because I don't need it to search the internet. I need it to search my files. I will disable canvas, because I don't need it for this task. I will disable image generation. I don't need that either. Although, if you want yours to generate images, you can enable it. What you need to enable if you are working with files like these, you need to enable code interpreter and data analyst. This is how you give access to all these files. In fact, it will analyze your files, this knowledge base, through code. Therefore, the last one must be enabled. And then Actions is a more advanced feature. This is essentially, you can connect external services for it to access, but this is a more complex story. If you understand what an API call or something like that is, you can add it, but you don't need it yet. This is after the second or third lesson, I think you will rethink this. Everything, click create. And here you need to choose whether it can be available only to you, but for your homework, you need to do this. Anyone with a link, anyone with a link. That is, so that you can send me the links. Save. And now it will be created, and let's test it. That is, we gave it a lot of knowledge base and an instruction on what kind of answer we expect, that the answer should be a link to a specific article. And let's ask, what articles are there about cloud code. It will now go, look through all my 280 articles and from them, now it might take some time, while it analyzes all these articles. This can happen. Let's see. And be sure to check, right, that this checkbox is enabled. But usually it takes a little time. Let's wait a bit. Because there is a process, it has many files, it needs to process them. Let's, while it's doing this, I'll show you what your homework will be. So, homework is practice. You will need to create a GPT like this, which will be an AI expert in a topic you choose. Choose any topic that interests you. You need to set up the instructions correctly. Talk to ChatGPT about it. That is, it will tell you how to do it. So don't worry, just enable internet search. Say, for example, search the internet on how to correctly create an instruction for a GPT that I want, for example, on this topic. It will go, search, and help you with everything. That is, it's important for me that you start working with AI independently like this, so that you don't follow my instructions, but that you find these instructions from AI. And give it access to various files that it will use as a knowledge base. Publish this GPT. And this is just what the school system requires. Upload any file, just a screenshot, for example, anything. Just upload, maybe a screenshot of your conversation with this agent. This is just what the school requires. But the most important thing in this field, right, in your comment. Be sure to write the link to your GPT. That is, the link is like this. You can copy the link and it will use it. Let's try again. But maybe it will take some time. It still doesn't want to, it's still not picking it up. But let me just show you how it works on something where it has definitely picked it up, because I've already, well, when I was testing, it was exactly the same. Everything is exactly the same. Uh, the same exact instruction, all the same files, also GPT 4 Turbo code. It's just that more time has passed, it has managed to process all these, all these files. And let's also ask it: "Find articles about cloud code." Let's ask it. Now it will go and search, because it has already indexed a large number of these files, and it will find them for us. 4 Turbo doesn't work super fast, but since it's without thinking, it works much faster than 4 Turbo thinking. So, now it will find and show us. So you will have the task of creating a similar GPT, which will help you do this. Now it will think here. And this is about chats, right? So these chats that you have with GPT are not ordinary chats, right? You can't put them into a project, because these are chats that happen within your expert. And it's just, well, a slightly separate story. Right. Right. And you see, it, it went, looked at my 280 articles and is now taking data from them. Moreover, each specific link can be opened like this without any problems and you can read a whole article from mine. You can try it yourself. Now I will, well, let's just do this. Copy link, I've sent it to you in the chat, please. You can also find it on any topic that you need, that you want. You can take it and just find it. It found a lot of things. And now let's check if this one has started working already, if it has picked up the data. Let's check. It's already analyzing. That's good. That means it has probably picked up the data. Let's see. Alexander asks: "Can I add more files after saving?" Yes, no problem. You can add, remove, there will be an update button at the top. What is this? Something, I think, because it's evening, it thinks longer. It picked it up faster for me. But you can do it without any problems. You can remove anything here, add anything, then just click update, and everything will work. Something, because, I think, because it's evening, it's somehow longer. Let's, let's ask something else. For example, find articles about Skills. Let's see. Ah, now it's quickly starting to pick up. No, not now. No, yes, it found it correctly. Only it didn't give the links properly for some reason. These things sometimes work worse in the evening. Let's start a new chat and ask it again. Now, find articles about Open Cloud. Let's see what it finds now. Well, I think it hasn't indexed everything completely yet. Now it, now it has started thinking normally again. So, if this happens to you, don't worry, just wait a little. Evgeny asks: "If it's a public GPT, will the changes apply to everyone?" Well, yes, because everyone else connects to yours, right? They don't have their own, they connect to yours, if you change something, then everyone who connects to it will also see that change. Everything, you see, it all worked. It needed about five minutes, probably, to process all this, and it did everything. It found all these articles. So you can also use it to find many cool articles and read them. Yes, this was a very intense first lesson. Sorry that we went a little over. The next ones won't be like this, I'll just skip through some topics a bit. But apparently, I thought we would go through some topics faster. But this time, how long generation actually takes, it still takes a little longer than I thought. In any case, I hope you enjoyed our lesson today. And now, you will have a theoretical test and plus, and plus you need to create your own GPT. If you have any questions, no problem, please write. Either in the personal account, or better, join our Telegram group chat, so that if you have any questions, I will help you, I will answer everything. But for now, at our current level, we specifically did an overview of tools to understand what it can do. In the next lesson, we will start diving deeper into code, we will use MCP with code more, we will, well, dive deeper into more serious capabilities, because everything we learned today is great, but code is even cooler. So, see you in a week. And now, the last thing. Will you provide us with these article files for homework? No, you will create your own. You need to create your own. Create it on a topic that interests you. This is just a topic that I created, that interests me. I don't need you to just repeat after me. The point is not that. The point is that you suffer a little. That is, you go to ChatGPT, ask: "Listen, ChatGPT, I need to create such an expert. I want to create one on this topic. What should I give such an expert?" And ChatGPT will tell you. So, I want you to start communicating with ChatGPT yourself, to think about what files to upload, and what kind of expert to create. It can be absolutely anything. Just think about it. It doesn't matter at all what it is, but I want you to solve this task for yourself. Alexander writes: "I haven't read many articles yet, and it would be useful to have one like this." So, please, I sent you the link, use it. It's completely public, it's fully open, so use mine without any problems. And you need to create your own for homework. So, see you in a week, and I will be waiting for your GPT questions, and next week we will dive even deeper into the capabilities of HG GPT and code. Goodbye.