All right, gentlemen, hello, everybody! It's good to see you all, to hear from our special edition on the Two's Moon Channel. Let us try to discuss the topic today: can we use chat-bot and agents at your particular work? Obviously, I can't answer each particular, but I'll try to show you the company scale I saw on the IT company case, how it all happened. And you'll try to find on this scale your company you work in, and first of all, you'll see where and how you work. Second, you can see, respectively, and where is likely to be the case with the AI in your companies. I'm telling you, like, big tech IT companies, for two reasons. The first reason is because I know many people who work in these companies, and I saw all these processes from inside. And the second is the IT company, after all, on top of this AI-Hip. Everybody tried, tried to get it all in. So, as a result, I think it's good to see where all the other companies are going and where they're likely to come.
Accordingly, let us begin from the zero level of AI implementation when the company either formally prohibits or consequently detracts from the process. And you're either using or not using it. The company doesn't cover this side. These are companies that either inert or, accordingly, try to ignore this wave of Hippa involving agents with LLM, and that's all that's relevant. What's that got to do with it? This leads to the fact that, in the company, agents or chat-bots are at least used, but they use it in the shadow mode, when the underdress is quiet, and so many staff are trying. Do something. The company itself does not regulate it. So, that's why he's the one who's done it, and he's the one who's using it. It's like, I don't know, the camera is taking pictures of a laptop or a monitor, throwing it in a plain-the-card ChatGPT or Claude, and saying, "Here, help me solve this problem." Yeah, and I'll be more on the IT company and all that stuff that's gonna be connected, all the examples will be a plus-minus with IT, the design, data science, and that's all that's related. Consequently, such a shadow use within companies often leads to data leaks, often leads to undesirable effects that the company could avoid if it did not ignore the process. Because in any case, virtually everyone I know from IT Wednesdays is used by chat-bots. At least in chat-bot format, it's a parole, to go to ChatGPT, to drop the code of error or error that happened, and to find out, and then what to do with it, how to fix it. If you just ask for a code of error or reprint the lane, then as you said, you can just take a picture of the screen, try to drop it. If there's any way to get some code out and put it in the model's context window, they do it, too, but they do, like, titchy, different ways that might be. So, on the one hand, it's like a company, if it's forbidden to use all of this, it can think, yes, that protects itself from negative effects. In fact, on the contrary, negative effects are only more pronounced because staff members have no understanding within the company that it is possible that they cannot. Everybody's trying to find the limits of this freed man.
Cybersecurity inside, information security, it may not follow all of this, and there may be different unpleasant situations.
This story, it's really fundamental. So you've raised a very cool subject. You're not just raising the subject, you're just asking how important a person is to know what's going on if he's using artificial intelligence. I've been out here a week ago, and in this issue a week ago, I was talking about how I created a system there for CRM, like using artificial intelligence in sales and marketing. And there were many comments, and how did you give the artificial intelligence, so the personal data of the clients? What I wrote people's comments inside. And when you're in Google Drive, you're putting something out or in the mail, or you're going to a regular CRM system, any way you're selling Force, I don't know, in a Russian-speaking market of amoCRM or in a mission management system. You're loading up, what do you think it's different? I mean, it feels like, on the one hand, people have that artificial intelligence is some kind of special system that can't be sent for anything that's reworking and has something. Some unique feature, huh? And there's no way that it's just a separate system that has, for which you have to make certain rules in the company, to sign them. You can add that your data are being processed with artificial intelligence in terms of law. But I mean, it's just a little something from the law perspective that matters relatively, right? Because to date what is called artificial intelligence is being used, well, I don't know what systems don't add any modules there or some micro-declars. And I think that, well, for me, this is, uh, this little addition that is in the interesting discussion you, Ilnar, are giving and giving today. And to draw attention to this side, well, all of us.
And in parallel, of course, to listen to this part that concerns a very important part of what you can do and what you can't do, as you realize.
The second level, which then becomes more civilized when the company has rules and understanding and what can be done with chat-bots, LLM. Both internal and external. What can't be done? There is often a clear separation, which, for example, there is no way personal data can be leaked from the company into the open LLMK and should be processed only internally or anonymously and anonymously before they can be processed before they can be anonymously and anonymously. These data will, uh, be downloaded. Sometimes, tools are available that help clean up personal data from the documents. In addition, there is a clear distinction that, for example, there is a conditionality for the generation of ideas, uh, chat-bots can be used for, there, I don't know, text corrections can be used. Again, there's a distinction that you can't. So, here we have this legal purity, and, uh, the implication is that the use of AI is going to be legal, uh, inside the company, yes, when there are clear rules, and staff members are somehow. I'm starting to understand, yeah, that's what we can use for, that's not what we can use. This may have been included in the official instruction, and the staff member knows that her violation, there, has some consequences, and so on. Yeah, there's already a very clear environment in which to work. Uh, it's still, uh, very weak AI penetration, but it's a huge step compared to the previous level. Yeah, the last one we had was a little rascal, uh, quiet, so no one could see, and no one was in a position to be regulated. Absolutely different things can happen. There's already regulation here. Staff understand that you can't. And for the company, it's much more transparent, understandable, managed. But I'm not talking about infiltration until we're even here.
That's how you were allowed to use Guglom and said you could googly that you can't googling. The next level is when some infrastructure arises within the company when MCP servers are located, i.e., internal consignors. Then your chat-boats from this book, the question-and-response system from the Internet is becoming a question-and-response system within your company. Let's get an example from, uh, athys, again, because I'm getting closer. If you have Confluence, uh, internal documents, so, say, the connctor, then, uh, chat-bot, in answering your question, can look at internal documents, the rules that are there to decide something. problems, description of internal processes, description of the internal libraries that were written for your company, and on that basis, you should be answered. I mean, you don't have to-- in the past, you'd have to put it in the whole context, put it in a conditional, there, ChatGPT or Claude, and then, accordingly, get a response that's close to you. Well, for example, you have, uh, well, let's take an example from the analysts then. You have some, uh, meters inside that are considered to be specific to your company. There may be some abbreviations associated with these meters that are introduced inside your company but not used outside. If that's true, then the conditional ChatGPT on these metrics is probably not aware of anything, because he was studying, there, on the open Internet. And your internal docks are closed outside, and they're impossible to access. Then trying to clarify something about these metrics within your company with some data will result in the system going to start or hallucinate, i.e., to figure out how to do it, well, at best. She won't answer that question. If you have internal conjectors, yes, if you have the possibility of ChatGPT through, now, a little technology, MCP server to contact your Confluence, your Jira, to the task training, your code base, For example, some parts of access, documentation, and, uh, I don't know, databases, instructions, libraries that you have inside. Your system is much more integrated with, uh, chat-bots. We're talking chat-bot format. And, uh, the efficiency of staff in this sense is starting to grow dramatic.
Uh-huh.
Steadyly grow. Uh, the staff member doesn't need to collect the whole context and download it. And here we have, like, all these conjectors within the company. Again, we say that we have the previous phase, too. I mean, internal regulation, which can't be done. Accordingly, the system is also set up within these conjectors. It means that documents that cannot be accessed, well, that should not be accessible, and they are likely to be separated, and through this MCP server, uh, chat-bot will not get information. This is why this is another step, compared to the previous one, when we have all this already in place. It's very convenient. Well, let's try some simple example, shall we? Uh, many big techs need to be tested. That means that once, six months, once a quarter, you need to describe what you did, uh, with all the examples, and so on. And for these situations in companies where the level of penetration on this one has already moved to level three, you can ask for chat-bot to gather all the information on the tasks you have in your possession. A certain period, right? And if you have a Jira conctor, he can look at all your tasks. If these tasks were well described, the results you did were described, he could summarise all this information into structurally understandable blocks. That you were working on this project this quarter, working on this project, on this project, and that was a challenge. The project from this stage of stadium has reached this stage. And further on this basis, it is much easier for you to write this performance review. Uh, the task might look a little artificial, but actually, who worked in big tech knows that it takes a long time, uh, it seems bureaucratic, but it's really very important. I'm-- because it depends on how much you, uh, get, there, I don't know, a bonus, we're gonna get on the next grade, growth inside the company, and so on.
So we've taken three levels so far, yes, when we're all in the wrong field when we have rules, it's the next level. And the next level of integration. Next level I've seen when we're starting to have these comunities inside our companies, when we're not only going on top of this movement, but it's coming down and the staff are starting to be active, help companies introduce AI What does that mean? Besides, uh, prompot ingining is clearly a yes, so when you write your request into chat boot, you can add more squills. I'm sure many, um, already know that this is what they've heard, maybe even used or built their own. But for those who don't know, it's a directive to solve some problem. Like, uh, you have some kind of task you often decide. Well, let's say, count the metrics for a period, taking into account some extra water. If you just, in the plain ChatGPT, put that promp in, and he won't know what database is, uh, some kind of data you need, what kind of metrics are considered, and he'll just have it. free field with access to all this information. And if the company is large, the information is probably very much, no context window is enough to make this task first. But if you write a manual that is understandable, detailed for, there, ChatGPT, that the data-- e-- what if you've been given this task, then the data should be taken here. The database is connected as follows. The metrics that need to be counted, they're here. And the man who gave you the assignment needs to be clarified, there, such questions. These are parameters that are further used in the next steps. And this clear guidance with steps for, for chat-bot, or for an agent, it's called squill, well, or skills. As a result, staff members are actively engaged in these skills at this level. It's usually everyone doing it on some part of their own. The date-sientists do theirs, the sales do their own, the testers write their own. And there is a large number of staff involved in this process. first, he' he's-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a Some squills are written in good and optimistic, some squills duplicate the previous ones, some are not very well written. But most importantly, the process begins. I'm monitoring that in a month, a half or two such a little chaotic activity, uh, there are already teams of staff that start to systematize. And there's this internal, uh, open source of the company, yes, when we have a lot of staff working on one squillium repository, make sure we don't have any intersections inside, I want squills written correctly and understandably to work on probation and with Codex, and Claude, and with other systems that are needed. Consequently, in a few months, this has resulted in many model tasks within companies being addressed through, uh, agents. When a conditional product that may not be very well understood in the technical content, it may, without analyst, accomplish any task when the date-indictist needs to collect some data, but there is a need to do so. Skill prepared, and it can achieve this task much faster. So in this case, you're already getting a question and a response from the expert system, so you're going to have specific tasks that are specific to your company. It is quickly and in the way that some staff members have put up. Most often, staff members who are often responsible for this task, yes, they know how to deal with it and what to do with it.
Here. And then the next step is going to be made. A-a-a-a-a-a-a-a-bags and chat-bots are not used as, uh, staff tools, right? So, at the previous stage, which we have been describing for quite some time now, it's still tools. So the product is needed to ask this system, it will find the squill necessary and, on the basis of this, it will return the product or, there, the next person. And the next step is, we're already in phase five, yes, if I've thought correctly, we're already in a situation where agents are almost full of staff when they're big. The project, uh, is being conducted with the help of the agent systems, when large rations are being created, where agents themselves set targets, agents write tests themselves, agents write the code, which, in the case of the agents, They pass these tests, then they all shrimp and so on. This is the system I only saw in, uh-oh, start-ups, active start-ups. But it's clear that they can afford it, they don't have the same responsibility for mistakes, and they can experiment. It's the first thing. The second dimension, uh, is often that they still have a little money flow to keep a large staff, and it's much more advantageous to try it quickly and in part cheaper with a large number of staff. chat-bot help. In short, such a graduation from, uh, systems where-- from companies where chat-bots are banned, to, uh, companies where chat-bots are, respectively, almost full-time employees and performers. I wonder if you've been working in these companies. Have you seen such examples? Perhaps you still see some answers that could be added here. Here. You can give some examples of companies you've either worked in or something. I wonder what you're thinking.