Hello, everybody! Ah, we have a small issue today on agents and the current stage of development of this technology. You, given that you are the viewers of our ToTheMoon channel, obviously you already know about chat-bots. Almost all of you used them, but many of you heard that it's fashionable, it's popular with AI agents. I'm sure you heard there's some MCP protocol that created Aic. But, uh, not all of them, I think they tried to start these agents. You don't know what they look like. I really want to tell you a little bit today. But so you can answer the question immediately: how much would that be of interest to you? Let's start with the demonstration.
Uh, I've prepared a little, uh, an example. I have a file on my computer. The file is called Images, and it actually contains six paintings. Three of them have cats, other objects. So what do I want? Well, uh, the most careful people have seen what they've got, uh, some window where the prompt goes. But let me tell you what I want the program to do. I want the system, uh, I know what's in every picture. And if there's a cat in the picture, I want her to take these pictures to a special "Cats." If there's no cat in the pictures, I want her to take these photos to the other file. And so that we can get better, and why she did, let's ask her to make a presentation and, as part of this presentation, tell us, well, that's why one, one, one, one, I'm not sure that this picture was in my daddy. What is this, uh, system where I put the prompot in? It's a regular Claude. Claude from Anthropic.
The only thing you used, for example, Claude, is that you probably used a browser version, just like a GPT chat. But there is a dishtop version, which is a program that can be downloaded, installed on its computer and, accordingly, used as usual Brauser Claude. The only, in the framework of the dextop Claude, is much easier to connect different tools. So let me get this straight. While the process is going, we're a little bit with you, and, uh, we'll talk. Yeah, there's some kind of mistake here, but I think we're gonna have it all going down. Yeah, it's all set. What's different from the usual chat? The customary chat we've been using for years, therefore, has the opportunity to answer your messages.
Now there's a possibility of searching the Internet, now it's possible to start a code at the same GPT chat inside your ecosystem, but with your computer, with the files on your computer system, it's working. They can't yet. And now that we say that we want our, uh, assistant, our AI agent to have an opportunity to interact with some other systems, we must give him these tools accordingly. So if we're talking about AI agent, and, uh, we're taking a regular LLM and telling him that if you need to, like, transfer some file from the folder to the folder, then use this one. instrument. If you need to know what's on this picture, you need this tool and so on. And, accordingly, that's how much we give him, he's, uh, gonna get his powers so far. Here, uh, we need to talk a little about MCP protocol. It was possible to do this before, but it was only, well, just a few of the undeclared programmers who like to deal with it.
And they would have done some system that could work with a specific LLM, for example, only Anthropic or only with OpenAI on API, or there's another system. So, Anthropic, and that company-- created a chat-bot Claude, where I'm doing all this now. They issued MCP protocol in November of the twenty-fourth year. MCP protocol standardized, a, the LLM communication between the chat-bot we're communicating with and the tools you can provide. Thus, the community has received a protocol on which to come up with some tools. And with all the LLM that support this protocol, they will work equally well. Ah, many companies are now supporting this protocol. Accordingly, there has been a large enthusiastic community that is creating these tools, the so-called MCP servers. But, uh, we're not really looking at this deep-tech thing on the channel, so I'm gonna put this on the table. Let's see what's going on with you and me. You see, there's some movement here, yeah, there's a picture missing, there's a picture of it, Daddy's coming, some files are here. Let's just go through what we really are. I told you all that. So, you've seen the prompot on the screen.
The process itself is proceeding further, and, as a consequence, the Claude text is generated. And when he realizes that he needs to use some tool and that instrument he has, he's actually handing the ball to that instrument. Let's take a step by step. He's planning on the plan at the beginning, and how he's gonna do it. Actually, the user is asking. He wrote everything in the steps that I asked him how he saw it and he came up with a plan of action that he would do. Well, first, I'll see what files are available, find the Images file, and so on. And in fact, he's starting to act according to that plan. Yeah, he's actually, uh, starting to see what directories are. Next, they watch what's inside the directory. Got a list, right? There's a directory Images that's inside. Next, he's watching that there's a number of files inside this directory. They were called one, two, three, four, five, six, and so on. He's doing a little more analysis. Next, he's, uh, well, there's a movement on the list. As you can see, he's done quite a lot of things. But most importantly, even here, the text was written and what the pictures are, respectively. The most important thing is, we'll see the results. Item one. Our cats. Ours, uh, Tephtelka's her name is this cat. Our cat was in the same daddy, respectively. See, here's all three pictures of the cat that were in, uh, Cats's daddy. And there's a folder of others, respectively. All the other photos were taken to her, respectively. The paragraph first is therefore correct. The system was able to examine the files that are available to her directory, understand what is inside them and, accordingly, do some action. Let's see if the presentation made it. But, as you can see, with some heats, because he created a sheet of contents, but the contents didn't fill. But it's just a prom editing, you know, you can fix it. And then he says there's a motorcycle and a man here, respectively. Oh, yeah, yeah, there's a man I haven't seen here, but the system has, accordingly, found. There's a cat in the second picture, there's a cat, there's a chair, but it's not a chair, it's actually a smooth board. But the heat is a very light recognition system. I used YOLO here, they're allowed. Well, that's the slippers, not Teddy's bear. But the most important thing is that the Tephtelka was found. And then there was enough, too, a cat found. There's a cup, a dinner table, as you can see, that's how the cat didn't get it. And there's a bit of a potato plant here. As you can see, we've taken a regular LLM, well, not the most ordinary, Claude Sonnet four are still large, but we, by providing her with a few tools, have managed to make her act as an agent. While I was telling you something, something was going on in my system.
There's actually a lot of tools you can add. Uh, there are instruments on June 3, literally, and, uh, the day before yesterday, it was possible to add connies without any programming skills at all. I mean, like Drive Search, Gmail Search, Calendar Search. There are other services that can be connected, for which no need is, uh, to be able to handle the programming. And there are, respectively, self-written tools that I have been added for this demonstration. There's a File System. This is MCP-server, respectively, an instrument that can be connected to MCP, which has created Anthropic. It's because of him, uh, the system might know, what files are inside, right? Here's a list directory there or directory tree can cause, can move files, can read the contents of the file, can record files, edit, and so on. And, accordingly, a tool kit, yes, this is a program from Anthropic. And there are two MCP-Servers I created, I wrote it myself. LaTeX Beamer is just, um, a system that generates the presentation of this one, and maybe not the most modern type, but those who studied in this mechma, the physical or similar faculties are sure, and, uh, well, Not that they were surprised, but with skill, they remembered these presentations because, after all, many such technical presentations were written in Tex, the articles are written in Tex. I'm actually math-- I'm just saying, "Is that gonna work? Turns out it's gonna work out and it's not hard. And there's an Object Detector, respectively. That's, uh, the service that keeps a picture-a-a-box under his hood.
That's the card recognition, uh, YOLO, which can cause and get what's inside, right? There's a cat, a cat, a chair, a Teddy bear, a motorcycle, a man, all the work of that I.D. So it's not Claude doing it on his own, and Claude says, "I have an object, I want you to do this to him." This action is in the instrument. The tool is returning the result. Claude, it's working on it. It seems interesting and promising. Aah, we'll add a little bit of a degt's cover in this story.
Ah, that's very much increasing the number of tokens you're using., how can I imagine the work, uh, MCP, server and, consequently-- not MCP server, let's just say, your LLM work with a set of tools that are available. If we're talking about relationships, uh, man and chat-bot, this is a dialogue, where, respectively, a man writes, chat-bot as an assistant helps him answer questions and waits for a man to write. And you have some kind of dialogue, respectively. He's in control. You wrote one message, you got answered by one message, then you answered the next. And that's how the ball rolls. If we say we have some tools, we have, uh, a dialogue that turns into a group chat where we have some more participants than me as a user and LLM.K. And that leads to the following: that we have a system saying, "I want to know which directories we have." And she's sending this message to a tool that says which directors are available. In fact, it's like the answer to my question is over here. Next, the question went to the instrument. The tool's back. So he physically wrote what directories, what objects are inside, uh, what is the name of the father's name? And then Claude re-examines the next message starts to generate all this. Everything that was above is transmitted as the context of this communication and the message is further under way. And imagine that, uh, I didn't write one message, but actually here, one, two, three, four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, four, fifteen, Sixteen, seventeen, eighteen, nineteen, twenty, twenty-one, twenty-two, twenty-three, twenty-four, twenty-four, twenty-five, twenty-six, twenty-seven, twenty-eight. Twenty-eight, uh, chat messages went to make it all work. This results in a very large, direct explosive increase in the number of tokens that are spent on tasks. So, uh, the concept itself is very promising, but on the other hand, it will be viable and very active only if it is cheaper, the cost of the LLM system. Just because there's a lot more to spend on the Tokens to do all this. If you have Claude Desktop or any other, uh, subscription LLM systems you can use, uh, not the number of tokens, but even the number of messages, you can build these. systems. If you work on API, uh, you mean that the cost of these operations might be very high. So please do it carefully. Well, I think that as such a first demonstration, it's gonna be enough. Oh, if you want me to, uh, expose this story from a technical perspective, and how I did it, and how can I make my MCP servers, how can they be located so that they can be arranged They've been working the system-- put the likies on, write the comments. Oh, that's what it is, yeah. Liki, put it on the canal anyway. Here. I will take it off if there is sufficient interest in this subject. So, before we meet in the next editions. Every Sunday, as Sasha likes to talk, four people with news from the Silicon Valley and all over the world. Bye-bye.