Transcription
Okay, well, colleagues, good day once again to everyone, everyone, everyone. And right away, let me clarify, we will introduce ourselves a little later and on the slides. My name is Maxim, I will be presenting the blog roughly in the middle. And we have Katya, Ekaterina Gorbacheva, she will be communicating at the beginning and at the end. We will try to monitor questions in the chat as we go. If we don't manage, there is time at the end to answer questions. Katya, I'll probably hand over the word to you. >> Yes, thank you. Can we have the next slide. So, I want to introduce you to our company. We are the largest integrator of IT solutions in Russia and are actively developing this direction abroad. For over 28 years, the company "Perviy Bit" has been helping clients from various industries with their IT solutions. We don't just work with 1C, we work with a wide range of fields. Our company has over 100 offices worldwide. We have over 50 of our own IT developments. We operate in nine countries and are not going to stop there. Next, please. So, what do we do? We automate accounting, of course. This is necessary for the vast majority of our clients. Business process automation is the next level of digitalization for clients, using artificial intelligence, robots, various directions that can simplify and improve our clients' businesses. Because our goal, our mission, is to make our clients' businesses stronger through IT solutions and create opportunities for their successful development. Next, please. So, what will we discuss today? Today we will talk about what artificial intelligence models currently exist, their features, and how to use them. We will show several practical examples. We will show our development, an assistant with a library of convenient, working prompts to simplify routine business processes. We will tell you where to learn artificial intelligence and answer our listeners' questions. The duration of our event will be about an hour. After the event ends, the recording will be converted, and within the next few days, all participants, both registered attendees and those who were not at the event, will receive the recording. Further, if you have any questions, please write them in the chat. We will track them and either answer them in the chat ourselves or leave separate time for questions. And please, don't hesitate to ask questions, we will be happy to answer them. Next, our main speaker is Maxim Kolmakov. He is one of our leading implementers, including in artificial intelligence, who is currently leading several projects for the implementation of this mechanism in 1C for our major clients. My name is Ekaterina Gorbacheva. I am the head of the training center. So, I will tell you exactly where and how you can learn to work with it in the modern world. Next. And next, we hand over the word to Maxim, and he will now tell us what artificial intelligence is, how it's used, and how to live with it in our >> Thank you. >> active paradigm. >> Yes, that's all. I'm taking the floor. Colleagues, for the next approximately 40 minutes, I will be communicating with you. Let's move on to this right away. Artificial intelligence is now something that, on the one hand, is trending. I saw comments about the sound, by the way. I will try to speak louder. If it's really bad, please write; if there are many such comments, we'll figure something out. But if not, please try. Well, accordingly, a local solution. But I will speak louder, honestly. As for artificial intelligence, I said at the first webinar we held in April, I think, that soon it will be like a computer. The moment when we, you know, realize that those who don't use artificial intelligence will be left behind. If not in their job role, then at least in terms of efficiency compared to colleagues. Simply because, well, it's everywhere now, it has developed too much, and it's too easy to apply. I really hope that as a result of the webinar, you will see how to apply it for free, right here, right now, for work tasks. You download it and start using it immediately. Therefore, my goal is, on the one hand, to inspire, on the other hand, to show, and on the third hand, to conduct a discussion. Oh, by the way, here's an excellent question. If possible, Alexander, I will also postpone it until the end. My favorite topic is how to apply it in the 1C environment. Over the past 5 months, I've had four projects, let's say, two quite large, two small ones, specifically for implementing full-fledged business contours, closing one specific task. Well, like a task, contours. It's always a complex of tasks. Nevertheless, let's move on to the theoretical part. There will be little theory, but I must give it because there are definitely more people than before. And some colleagues are not familiar with my terminology or the terminology of artificial intelligence. Look, types of artificial intelligence, please pay attention to these three words. I divide it into the following: free, trained, and integrated. I think it's very convenient to use them in this breakdown from a business perspective. So, free artificial intelligence is the artificial intelligence that is directly accessible to us in any open tool. This is when a neural network, well, you know, AI is a neural network, right? Behind any artificial intelligence, it's just like that. Colleagues are writing that they can't hear well. I apologize. Let me try something. So, 1 2 3 1 2 3. Ah, thank you. So, if there are many, then, thank you. Other colleagues are writing that everything is fine. Colleagues, those who have problems, please try to adjust your volume or reconnect. I'm honestly trying. I've reconnected, it should be a little better. So, I won't get distracted anymore. Free artificial intelligence, which operates on knowledge about your request solely within the scope of your request. You write something to it in the instructions, and it only knows that about your request. You apply it in business, for example, you want to apply it to medicine, recognize patient analysis, or give a diagnosis, and it knows about the diagnosis and medicine only what is written in your prompt instructions and what it initially had in its knowledge base. I ask not in the knowledge base at all, but in its, let's say, ability to understand people. Because when neural networks are trained, there is a huge amount of information. And one way or another, there are some parts about medicine and about 1C business, but there are very few of them. So, free artificial intelligence is ChatGPT, it's our Deep Assistant, which I will talk about in its classic form, when the system only knows about you what you wrote in the instructions and it goes further. I will scroll myself here. Thank you, colleagues. So. Free artificial intelligence, what the system knows about you directly from general data, you could say. Here are examples of what you can do with it. The most classic example. You go to DeepS, you write something. You go to ChatGPT, Yandex GPT, Sber's Gigachat, you write something, the system answers something. It doesn't know about you, it doesn't know about your business, except for what you wrote in the instructions, large or small, but nevertheless, in any case, this is paradoxically the simplest, the fastest, and it covers 80% of tasks. Well, with minor reservations, which we will discuss later, with minor clarifications and so on, but today free artificial intelligence works very well simply because it has large capacities behind it now. I will not read all these examples of applications, simply because they were at previous webinars, and here as examples for possible discussion, possible inspiration. Practice will be separate later. There I will show the screen where we will do something. The second level is trained artificial intelligence. As soon as we give the concept of either a knowledge base or model retraining. That is, as soon as we, for example, write not just a question about accounting, but about accounting, and also load information into the model, into the knowledge base, about what accounting is, what accounting regulations are currently in effect in Russia. If you write a request in Russia, for example, if there are documents or full retraining. If there are questions about this, I will explain in more detail. We can look at what retraining a neural network is for a long time. We say it like this: a knowledge base of terms, about your request, about your company, about the industry, when we provide it, artificial intelligence obviously starts to work better. This is the classic business application. In its simplest form now. This is the classic business application, business use, when you have provided information about yourself, your company, industry, about the task in one way or another, and it then provides all answers to questions, well, depending on what you are using. It's a sales chatbot, it's automatic recognition, automatic generation, and so on. It performs its task based on this data. It's more effective. The third level, when we talk about the next step specifically for business, is integration. Integrated. And I highlight it here. Why? When we don't manually write requests to a chatbot, or go to a web form on a website and write there. When we have integrated artificial intelligence into our application, the integration options can be different. We added a button "Send to Artificial Intelligence," "Send for Analysis." Here you have a report, for example, in accounting, and above there is a button "Analyze with AI," and this is integrated artificial intelligence. In the background, there is a trained AI, meaning the request is sent to a trained AI that is integrated into 1C. Another example. In the background, automatically every day at 8 PM, for example, an analysis of the registration log is compiled, how the database worked, how the system worked, were there any slowdowns, or how users performed their tasks, if we are talking about document management, how receipts and sales were, if we are talking about trade management. Yes, all data from 1C can be transferred to 1C. Oh, excuse me, well, it can be to 1C, and then to artificial intelligence. It can be directly to artificial intelligence. Again, if there is time, if there is a desire to discuss this, I will explain the three options for integrating artificial intelligence with 1C from the perspective of analyzing data within the database. Nevertheless, integrated artificial intelligence is when it is built into your process, built into your product. And we are talking about 1C, so these are buttons, background work, automatically triggered procedures. That's it, we've almost finished the main theoretical part. Just a couple of important words for context, because I will refer to these terms. Everything I said earlier can be implemented in two ways. Two halves of a whole: cloud models and local models. I will call them that. The second is a closed loop. Cloud models are those that are available to us via, let's say, an unsecured or not fully secured communication channel. It can be HTTPS from a technical perspective, but the key idea is that you don't know where this model is located, where its servers are, where the data center is where everything is processed. Classic examples are ChatGPT, DeepS, and in some cases Yandex GPT can also be included, despite the fact that we officially know which country its servers are located in, you still provide information in an open format. In what sense, I mean? So, cloud models are what is available to you now if you just go online, type in artificial intelligence, chat. And one of the options is offered to you. They usually work very fast, usually very modern, usually not in Russia. Well, if we talk about ChatGPT, the most classic model, with which everything began and which, I can't say it's completely leading, but it holds very good positions in many arenas, in many model comparisons. In terms of communicating with ChatGPT, you need to understand, you send it a request, this request is saved on its data, you are not guaranteed that if you give it personal data, information related to trade secrets, that you can guarantee that it won't end up anywhere. In what context, won't it end up in the dataset for training this model? Officially, many models now promise not to use this, but nevertheless, for some, this is the case. So, I'm highlighting a very clear pitfall of using cloud models, which is that they don't guarantee non-use of your data further. Not all of them, at least. And in terms of advantages, why are they convenient to use then? Why am I not talking about it? Because it's very cheap and very fast to build a prototype on them. You build a prototype of your idea, business solutions on a cloud model in literally a day, two, a few hours, sometimes depending on the complexity. Well, the only thing is, you send some anonymized data there, and it works. To understand, prototype, and generally decide whether to launch a process or not, it's ideal. But in terms of disadvantages, you've probably already managed to read them. It's difficult to customize it for yourself, to retrain it. Fine-tuning is retraining, if we use our terms. Difficulties in retraining for yourself, and security is a real pain. An alternative, if you have a business solution, because for personal tasks, of course, it's great. For business solutions, it's not always suitable, meaning there's always the issue of security. It's just, well, a cornerstone. I will, unfortunately, repeat and refer to it several times, because, yes, it's a key alternative when we talk about commercial secrets, personal data, or in general, work in a closed loop is required, it's so-called local models. Roughly speaking, we take and install the model ourselves. The same DeepS, the same, say, Claude, or Google's model, LaMDA. There are many of them now, very early. Over a million models, by the way, exist now, so I'm naming the most popular ones. But to find one for yourself, for your task, is quite easy. Well, how to say, you can, you need to spend time, but it's possible. In any case, it's best to start with large, powerful models, those that weigh a lot, have many weights inside, but then you can take a simpler, but more specialized one for your task. So, when you take this model, install it yourself, meaning yourself, either physically on a server, or rent a server from someone, from 1C, from Perviy Bit, from some other provider, a company located in Russia, that's also possible, right? Well, just like there's cloud 1C, and you or server rental for 1C, you can also rent a separate server for AI if you need a closed loop but don't want to deploy it yourself. Or, well, in the case of large projects or large companies, servers are usually deployed internally. The main advantages are full control. That is, it doesn't go out to the internet. All requests are executed here and now. Full logging, as detailed as possible. That is, what came in, how the model thought, what results it produced. Aha. At this point, we probably need to give it more information or something. Retrain it, and these models are much easier to retrain because they are yours, they are with you, you have full access to them. What retraining options are there, also, if there's time and desire to discuss. This topic is just a bit broader than our webinar, so we can discuss it at the end. The key thing is that a local model is usually much more expensive than a cloud model. Orders of magnitude more expensive to start with. The minimum server cost for a local model, if you install it yourself, is about 1.5 million, it's unlikely to be less because the graphics cards are very expensive, and these CPUs are very demanding. So, where can we go next? Based on what I've explained, this is probably almost the last theoretical slide, and then we'll move on to presenting the practice. There are existing models, right? Here I have, let's say, not necessarily the best, but probably those that are well-known and seem acceptable to me for use. These are GPT, DPК, GRК, Yandex, GPT, and so on. Tinkoff, by the way, has a decent AI for some tasks too. Next, if these general cloud models don't suit you, you go to the next level, for example, 1C services. Within 1C programs, about five services are currently being developed. I've highlighted three that you can try to start using: document recognition, speech recognition. Well, it's clear, right, what it does, and 1C Assistant. With the Assistant, there's a peculiarity that it's currently only available to partners and in beta testing, but at some point, I believe, when this webinar is recorded, a little later, it will be fully available. Document recognition is an analog of standard recognition that external resources have for scanning, for scanning a document, recognizing it, creating a card. Speech recognition, voice input of data into the system. They also promise, um, task management, for example, in the system, in document management, for example, setting a task by voice, completing a task by voice. But so far, when I've tested it, it doesn't work very well. They are still developing. So, the third level, for example, the webinar, right, from Perviy Bit. I, accordingly, cannot not mention it, and I personally like it because I use them. These are three services, which I will mention one way or another, they are quite simple to start with. Also, Bit Assistant, our chatbot, which we will work with today primarily. Bit Newton, a transcription tool. If you work with audio, you have meeting recordings, conversation recordings. If you want to analyze phone calls, you have a sales department, for example, and you want to check the quality of their work or compliance with scripts, or identify negativity from calls. Well, Bit Newton. Bit Platform is my favorite place, probably. I showed it at the last webinar, we did something there. If there's time, I'll show it today too. We have something else. There you can create a trained artificial intelligence, that is, the simplest way to connect a prompt, meaning instructions, and a knowledge base. That's it. Literally in one place, you write how the system should behave, provide a set of documents, regulations, instructions that the system should rely on. You have a trained artificial intelligence ready, working with your data. My colleagues from departments, within my office, ask me to create something or help them do it. We do it specifically on the platform. First, what we will work with today is Bit Assistant. Again, I've talked about it before, now I'll try to be super brief. What's the key idea? We at Perviy Bit have a laboratory division, and it's quite young, more than a year, but less than two, as far as I remember, it's functioning, and colleagues are promoting and popularizing artificial intelligence. So, Bit Assistant is a free chatbot on Telegram that can solve, in general, most work tasks. And specifically, 1C tasks. Today we will see what can be done, how it can be done, why. Why am I talking about it? Because some of the services I mentioned about cloud models are blocked by VPN now, right? That is, they are not available from our country, and you need some corporate VPN if you work with them officially, and you require it. Some local models that you can install yourself are expensive to start with, and therefore difficult to use. Therefore, well, some compromise suits me personally for most, say, 70-80% of my work tasks, and that's Bit Assistant, in the context that it's, on the one hand, quick and easy, and on the other hand, it's free for the client. A few words about it, and then we'll go directly to looking at it. This is what it looks like. I will show it right now in Telegram. It's a bot. Further, there is a QR code here, you can scan it and start using it. At the end, as far as I remember, we will also have it. You use a set of tools that are available there. There is a standard universal assistant for code search, writing, or analysis. There is a translator. I also use image generation. So, and let's talk about our artificial intelligence then. I think I want to bring up the demonstration now. And first, we will try to work directly with So, now there will be a mirror effect. So, here is our Telegram bot. The image should be visible. Well, colleagues will correct me if it's not visible. Pay attention, I have two Telegram bots at the top. Black and white, white is Bit Assistant. This is what it looks like. This is our internal Bit one, exactly the same, but black. This is for clients. They are currently no different. Simply, one is for internal use at Bit, the second is for clients. So, the Telegram bot leads here. What's here and why I will personally talk about it further and use it. Simply because you connect it and start using artificial intelligence. There are limits, but they are quite large. As far as I remember, I know, well, I don't even have many requests when colleagues, clients exceed these limits, except for images. For images, you can generate three pieces, but again, here it's more for work things. So, we generate all sorts of cute pictures elsewhere. For work, three is not enough, but it might be enough to try. So, what will we do with it? We have a choice of modes here. I will have the universal assistant selected. It's currently on. The standard model is 4.1 mini. There are three models you can choose if you want. I'll leave this one for now. And my first case will probably be the following. Those who know, those who work, there is a document management system, and work with internal documentation tasks is done there, right? That is, if here are tasks, collaborative work, all tasks are open, employees perform certain tasks. I want to show an example based on this system, well, in fact, it's replicable to any other. There is a lot of data, in particular, tasks that many employees have performed over a long period of time. In any 1C, you have reports. For example, I am interested in a report on performance discipline. This is what this report looks like. If you open it directly in Excel, it's a report that indicates which employees performed which tasks, whether they were overdue or not. Their tasks. It's quite large, employee, number completed on time, not on time, and so on. I'm now urging you to look at this. If you don't work with document management, then from the perspective of an example of working with a report, if you do, then as an example of use. So, look, it's actually difficult to analyze such a report simply because, well, I understand, I have total completed and on time. Ideally, they should match and there shouldn't be too much discrepancy between the numbers. Here it's 122, 128, but it's difficult to identify any general trends, any tendencies, or even recommendations. Here's what I've displayed in columns, what can help me analyze performance discipline in certain departments. But I want to ask artificial intelligence to analyze this report that I showed you, which I have saved here in Excel. I will draw your attention to the fact that I generated the report in 1C, exported it to Excel. The simplest automation looks exactly like this. You generate something in 1C, export it, and then analyze it. Why not? This is also a semi-manual process, but again, you'll get 80% of the result this way. Okay. I've written a more or less large prompt. Act. Here is the report on performance discipline. I will not write the prompt manually, I will copy it purely to save time. If you wish, you can look at the prompt itself on the recording later. But you are an expert in data analysis and experienced leaders. Conduct a deep report, analysis of the report on performance discipline. Identify systemic patterns, bottlenecks, practical recommendations. Those who notice a peculiarity, I will say a few words later, it will be interesting for you. What is a bit non-standard here is the composition of the prompt, if you've seen my previous webinars. And this is what I want, trends, a brief summary, conclusions, recommendations, and so on. This will be insightful, constructive. The key thing is that the report will be recognized by the following. Let's try. That is, I can do any report exported from 1C like this. Now I'm interested in the production department. I just send it. Note that I am currently working with the client. Maybe I'll switch to the Bit one if the load on the client one is too high. But usually, I haven't had such problems. What did the system give me? Yes, excellent. I'm waiting for the report file from you. I send this file directly, the performance discipline certificate, and I'm waiting to see what it will tell me, what it will analyze. I repeat, here I have a maximally general, on the one hand, and maximally thorough prompt, on the other hand, asking it to find everything it can, identify any patterns. For you, as experts, when you work with your reports, it's obvious that you already know what's more important to you, what you want to highlight. If it's a report on receipts, then something that is non-standard relative to periods or a comparison for a month, quarter, year, and so on. If it's a financial report, then, probably, trends related to money, including when we will have a cash flow gap, whether there will be one by chance, and so on. So, I won't read everything, because the report is quite large. What does the system see?
What is our situation regarding discipline? The state, the tension. Only 371 out of 560 tasks were completed on time. And quite a lot are either not completed or overdue. Systemic complexities. Okay, that's clear. Is there anything that goes into specifics? Let's see. In a number of departments, yes. the level exceeds 30%. Well, that's probably a reason to pay attention. Perhaps one of the departments is overloaded, or we've set up the routes incorrectly. Well, in document management, this is also, unfortunately, a common situation where many tasks fall onto a role, and consequently, you simply cannot control who does how much, simply because there are three people in the role, and 300 tasks a day. Well, when it's visible in a report, it's clear that you need to unload or supplement. Okay, I couldn't analyze by weekly months because the report I have is indeed generated for a specific period, I think it's for a month. And specifics by performer. Click-click-click. Here are examples of our departments, yes, what the system considers a bottleneck, that is, where I have a significant delay. Well, in this format, you generate and export any report. You can have several reports. Yes, I personally like to generate several reports for the first month, second, third, and write a prompt. I'll give you several reports. Make them in the form of, analyze the trend, analyze the dynamics. That also worked great. I'll return to my prompt file for a second. Pay attention to this. So, previously I copied the prompt in Russian, and those who watched carefully might have noticed that some words were in English. Why is that? The system, in general, the vast majority of neural networks are initially trained in English, yes, and for them, well, the datasets, the initial, data for training were prepared in English. Consequently, when you write in English, it's less now, simply because systems have become quite smart, well, powerful, smart are synonyms in this context. But before, sometimes by writing a prompt in English, you would get a slightly more effective result. Keep this in mind, especially in business, it can be important that when you've written an excellent prompt, you can try to translate it into English and give it to the system. This can also provide an unexpected benefit in accuracy. Okay, the next task too. Let's go. My next example in the task, well, in the document management base. I am currently working with it. Further, we will move on to heterogeneous systems. There is the concept of approval. In approval, certain employees look at the files sent to them, well, within the document, attached, and analyze and approve or not approve. And I asked artificial intelligence to take on the role of approval now. Acting as an experienced lawyer, concentrating on a contractual project, because I asked it to check the contract. I have such a simple contract. Here, if you look, I want it preliminarily, well, okay. Such a maximally simple result from technology, yes, one of the first bits with such and such a company. Everything is anonymized there, don't worry. With such and such a company [music] to conclude an agreement for the provision of services. I want to analyze this request, this contract from the perspective of artificial intelligence. I send it a prompt. At the end, I will also have a file, it will be sent in the next message. I'll take this contract, send it now. Yes, everything, it's waiting for me. Please provide the file with the contract for analysis. Send. What's the peculiarity here? This is either an approval assistant, yes, because not all projects, processes can be fully handed over to artificial intelligence. Right now, I still believe that artificial intelligence can help well now, but it cannot completely replace a person in most cases. Perhaps in the future, considering how quickly it is developing, it is very likely that it will be able to take on this role in some functional areas. But for now, it's a help. Therefore, I would add such a prompt and such work to the approval task. Here, for example, I would have to approve a certain warranty letter and below the text from artificial intelligence. I can approve, not approve, but I can look at what artificial intelligence has written to me. By the way, here I chose this example because I wanted to confirm a certain point. Pay attention, I have written as follows: you must make a final verdict: approved, not approved, or conditionally approved. Conditionally means approved with reservations. And the system really dislikes being strict, yes? That is, at least for now, those that are trained in standard access, they really like to be kind. And when you ask it approved, not approved, it will more likely say: "Well, rather approved, but there are important reservations." Well, this is rather conditionally approved, but there are critical reservations. Although in fact, from the perspective of this contract, yes, if you look at what it highlighted as critical, there is a lack of detail in approvals regarding volumes and deadlines. There are no clearly defined criteria for accepting the result at the stage. There is no section regulating responsibility at all. I actually deleted it. The procedure for returning advances is also not specified. That is, in fact, from the perspective of an experienced lawyer, it is not approved. But our artificial intelligence writes conditionally approved. And this is important. And this needs to be strengthened, that you read carefully. And if you find critical reservations, write not approved, yes, because it is important from the perspective of understanding the work. In general, by default, artificial intelligence likes to be kind, and it is ready for anything, including hallucinating, to help you. Okay, everything, let's move on. I won't go back to the presentations, in general, I want to spend more time on practice. What else did I want to show, tell? Next is a slightly smaller process of prompts, perhaps in terms of volume, but more in terms of premises. What I personally like to use in the assistant in terms of why I decided to talk about it at the webinar. I often come up with, well, I have work and personal tasks that I use with artificial intelligence, but at some point, I feel like I'm not using it enough, or I feel like I'd like to try something new. And when I entered the prompt library in such a mood, pay attention, here are ready-made requests to the assistant, that is, I clicked on the ready-made requests menu. I really liked it and was inspired. That is, this is a set of prompts that colleagues created, well, initially filled by Bit employees, we had a competition for the best prompts, and now it is supplemented by both us and clients. There is a "suggest a request" at the top. In fact, these are ready-made prompts that allow, well, they are divided into categories. We will now look at the most interesting ones, in my opinion, and how to work with them. Or let's first see what's there, and then how to work with them. Generate a commercial proposal. And here is a prompt that needs to be copied and something of your own inserted in curly braces. A script for a call, a negotiation scenario, generate objections. I wanted to work with this. I personally really liked it. Look, you are an experienced sales manager. Prepare for a meeting with a client, the goal of which is to insert your meeting. Generate what objections the client might raise. For each objection, provide your variant of the answer. Give variants of objections to these answers. Counter-arguments. I click copy. I go here, paste it. Then, again, not to write by hand, I already have this prompt written. I just inserted the topic in the description. You will like it, I hope. Prepare for a meeting with a client, the goal of which is to implement into existing processes based on 1x. Well, there were such meetings, yes. It's a shame I saw this prompt from colleagues later. It hadn't occurred to me, let's say, to prepare for a meeting with artificial intelligence before. Although, it would seem, by the way, for this webinar, I tried to generate questions from active listeners, and asked it to give pros and cons. We had some debate, it was quite unusual. I found it useful. Let me read a couple of examples to understand how well it understands both the specifics and the formulation in general. Objection one. It is difficult for us to understand what real benefit artificial intelligence will bring to our specific processes. Proposed answer. And capable of automating routine operations, increasing the accuracy of forecasts, and improving decision-making, which saves time and reduces errors. We can start with a pilot project on a key process to demonstrate clearly. Client objections. A pilot project involves additional costs and risks. It's not a fact that we will get, well, let's put it this way. In Russian, it would be "it's not a fact that we will get value or a business-profitable thing." Counter-argument. Provide a clear pilot plan with measurable KPIs and budget control. Customer experience shows that implementation pays off due to the following: cost reduction, increased efficiency. By the way, general points. I actually fully agree here. Now, in all artificial intelligence projects that we start discussing, entering, no, I try to lead everything through a pilot, through a maximally simple prototype, simply because it's in the air, and everyone understands what artificial intelligence is, but everyone understands it in their own way. And the results are often, well, not disappointing, but surprising and force us to slightly change our approach. So, employees are not ready to work with new technologies. The second objection will be resistance. We respond, we organize training, staff support, showing the advantages of new tools for their tasks, which will reduce stress and increase motivation. Moreover, AI will make their work easier. Counter-objection. Training will require time. We don't have resources for a long process. Response to the answer. Training sessions can be adapted to the schedule and conducted in stages. Technical support is also considered, which minimizes the burden on employees, allows for quick, and so on. Yes. So, I personally liked it because this is a topic I've encountered, and in general, such questions really exist. And plus or minus, I actually give such answers. Well, it's clear that somewhere a little more, well, not to say experienced, but with more examples, but no one prevents me from taking this as an example. And again, this is just one of the prompts from the library. And personally, when I read the prompts, read the topics that are in the library, I thought that it's very good for everyone to look at it at least once, simply because, what if it really resonates with you. Okay, um, make a summary. Well, I probably won't show it. Make a summary of a book or a file, by the way. But for a file, I use it for meetings when they send them. For example, I could have sent the same contract now and asked for a summary of this contract, highlight key features, and so on. Some colleagues really use it, well, I have acquaintances who use "make a summary of a book." Also useful, it can be quite useful for those who have their own website or track sales on marketplaces or analyze reviews. That is, the role is marketing analyst, analyze customer reviews of the product, insert the specific product and link. Our company is like this: determine the positive, negative, identify repetitions. Problems, yes, and analysis, suggest improvements. Well, this is what marketers do. If there is such a thing, you can simply provide a link to the website. If the website is closed, you simply download the information from the website, you can even download it in screenshots and give it for analysis. Again, this can be, I ask you now to look at the options I propose more broadly, yes, because I am talking about what I found interesting, useful for, well, at least someone from the current several hundred participants plus recordings will be more, but look at it as broadly as possible, because no one prevents you from applying it later for yourself, for your products, if you are analyzing not reviews, but, for example, a simple product description. Why not? The next piece, I want to have time to show more, what might be interesting to someone is to create a series of posts. That is, if there is promotion on online resources on certain platforms, well, let's write it here with you, a brand promotion specialist for the company First on social networks, who is responsible for the product, well, let's say BitAssistant, we don't promote it, actually, because it's free, but why not? create a series of posts from Let's emphasize five pieces. Topic: implementation, usage. Here's great. Using the Bit Assistant for application in work tasks. Again, regarding my first webinar. Colleagues, if you write a prompt and the prompt works, then you have written a good prompt. There are many examples of how to write good prompts. We will also talk about this in our course, for example, but for me, perhaps the key, the most important thing from everything I say there, is iteration. You wrote a prompt, tried it, it didn't work, or it didn't work as well as you wanted. The next iteration improves the prompt. You simply write: "Oh, now do it this way, or no, you misunderstood, do this and that." It works. Customer engagement and business information. Use a conversational style, for example, let's not, I still like business style. Be creative, unique, that is, generate something, again, this is probably the most classic and most popular. Commercial proposal, requirements list, technical specification, series of posts, reviews, and so on. Not in the sense of reviews that are published, but in the sense of helping you prepare a review if you are writing one for something in a business style. You wrote literally three sentences, and the system turned it into a full one. That is, what can be generated will be well generated with artificial intelligence. Well, here are examples, yes, with various emojis, structured, with hashtags. Well, a fairly classic post structure. We have more. I don't know if it will be useful, but I think based on the registrations I've seen from some colleagues, it might be useful. Those who are more involved with IT or as a client, or as a specialist in various functional roles. Development of project documentation. We recently, by the way, even held a competition within the company. I conducted a small activity where we asked our internal employees to generate client cases based on project documentation. What is meant by this? Here, for example, the development of project documentation itself, and a specific one, yes, a report, modeling, possibly a project plan, a roadmap, and so on. The second option, you have a set of documentation, protocols, reports, possibly technical specifications. You send them and ask to formulate a case or a brief summary of the project, or a brief outcome of the project. This will also work well. By the way, we had three winners in the one I was talking about. For those who are really IT IT colleagues, I will probably show you this code refactoring - it's interesting, but try it, depending on what language you are writing in. On 1C it will be good, on something like Python it will be excellent, because for 1C you still need to use slightly different services. But I personally like this. You are a qualified developer, perform code review and writing, yes, and recommendations for correction. No, wait, this is completely for techies. I wanted to show something else. I use this. This is an indicator. Find out what this code does. Again, if there are developer colleagues, again, I saw that there are 1C developers related to this. You know that in standard systems, in standard configurations, there is a lot of code that is written, well, using, let's say, internal BSP procedures or is quite complex to understand. Well, I think I even wrote this code separately. No, I didn't write it, because here, in principle, I will delete it here. In principle, I won't write anything. Code, describe what the code does, which I will send in the next message. It is written in the language. Here I won't even ask it to show. Sometimes this is useful for me. I wouldn't say very often, but for example, I have such a complex, scary piece of code, again, taken from document management. It's related to tasks. What exactly it does, I won't tell you yet. Let's see if artificial intelligence understands it. I throw it a large piece of code, and it analyzes it. Again, not for everyone, but again, explain what the code does, explain what this term means, explain how to understand what the director said, if there are any complex terminologies used, and you are, for example, a beginner in this industry, it also works excellently. So, what I found is a procedure for accepting a task for execution with condition checks, blocking, and errors. Yes, that's true. So, indeed, I took the code from a function called "accept task for execution," and a lot happens there. Well, the rest doesn't interest us, it's already very technical, but nevertheless, to find out something simple, from something complex, why not. You can even ask, by the way, to formulate it in your own way, well, in a conversational style, for example, yes. It sometimes helps me if I ask to write in the simplest language, in Russian, of course, if I'm talking about an English term. So, I wanted to show this piece. I personally liked it. Pay attention, yes, I am using search now. In fact, I could have easily gone through these sections that are here. HR, management, operational work. There's a lot here, but there's really a lot, and we wouldn't fit. There is a block called "Accounting Questions." Well, let's say. It's called "Accounting Questions." To save time, I'll show it here. The prompt sounds like this: "You are a 1C analyst, helping users understand accounting specifics." It seems personally useful for accounting, but it can also be used in any operational or management context. Users form requests in the spirit of. And a little lower, I write a request based on the request: "How do I create a nomenclature 'nails' in a quantity of 50 pieces from the nomenclature 'pack of nails'?" Well, those who work in accounting know, or in trade know, that this is related, let's say, to simple production. In fact, this is a request on how to do this in accounting. Everything below is directly taken from the prompt in the library. It's exactly like this there: "Determine which accounting area, what to do, what are the inconveniences, and so on." My specific question might be more general: "How to process VAT at 0% for import, yes, or how to reflect the transition to UN in accounting enterprise 3.0." I won't pay attention, I haven't specified in the system what accounting system I have, whether it's GP complex or something else. Consequently, the system writes that, ah, this is warehouse accounting, partly management. I recommend using something from this. Assembly-disassembly, receipts, invoices, warehouse, and so on. In fact, if I had specified accounting, then this document, "assembly and disassembly," would have been there, but the system suggested it as one of the options. Here is more detail on how to do what, and so on. So again, the system immediately looks for what problems might arise. Well, because my prompt is maximally thorough, again. If you don't need this, shorten it, try it, see. At the previous webinar, for example, I showed a prompt that sounded like: "Create a checklist for a new employee to check the acceptance of a fixed asset into the system, I think without VAT or something like that. Or not a fixed asset, but just an imported item, I don't remember exactly. Well, the key idea is that you write what you want the system to do, and you directly receive an answer and then try to understand if it suits you or not, improve the prompt. Improve the prompt once, then use it simply because you need to adapt it to yourself. I also like to use it for creating surveys and tests. I also showed this at the last one. I probably won't spend time on it now. Here I am scrolling slowly, so that if anyone is interested, they can read it, surveys and tests. Create a survey. Let's just say it anyway. Create a survey for clients on the topic of whether they need to implement AI in 1C. No, this is a maximally simple prompt. Pay attention, yes, I have large prompts that are written well, and the system writes well. There are small prompts, they are written, well, maximally simply. And the system also writes well. This is precisely the feature that neural networks are becoming smarter and smarter, and it is not so important how well, well, the further the opinion is important, how well you write, percentages at high levels of business requirements will especially play a role. That is, 79% and 80% accuracy, this is where the prompt must be polished almost to perfection. But for most of our work tasks, even this is enough. It is clear that if we want to write well, then we either go for training in our course, not ours, it doesn't matter, or we write many, many prompts and understand it ourselves. But nevertheless, you can write, it will work in any case. And here, how familiar you are with the possibility, it's a bit like the survey that was at the beginning, yes, what business processes, in what controls would it be useful, and so on. Next, from the simplest, perhaps the shortest prompt will be for the following. And pay attention, it says "top." Well, this means that it is done most often. Generate a QR code from a link. That is, simply, you have a link, there are many such, and here it's just a test link. And you have many external resources on the internet, but why not use artificial intelligence? The goal in the future, I think, is that eventually most of the different tools that you use will be combined into one tool. Well, for those who have time, you can try QR. If, if you don't have time, then you can later. In general, it works, I checked it before. If now, for example, it didn't work, or when you try it, it didn't work, then you need to play around, rephrase, check the correctness of the link at the moment. I'll show you, but I probably won't show the small one anymore, but here I talked about working and semi-working options. To-do list, why not? A prompt that will improve operational work, in fact, yes, you can use artificial intelligence in life, including Bit Assistant. I really use it for personal tasks in some moments. Why not? Take and dictate a to-do list and then write a prompt. Then I need to do the following. For example, I will send you in the next message or write directly here. Help me create a to-do list, prioritize tasks by importance. Don't forget to schedule time for breaks and rest, and make sure it's done. What I personally like about interacting with artificial intelligence via Telegram is that voice messages are also processed. Not everything has to be written by hand. For example, my example when I was preparing for the webinar this morning, as if it was acting as a listener, I gave answers by voice, yes? That is, on the phone, I turned on the voice input mode, here, and gave it an answer. You can do the same, dictate the prompt, the task itself, or the requirement. So, there is little time left so that I can also hand over the word to Katya. Let's look at one more thing I mentioned about changing modes, I'll say a little more detail. Besides the universal assistant, probably the second most used is image generation. In the same ready-made prompt library, the same ready-made blocks, I mean the library that I'm showing, I have "create a picture, an image." I don't see it. Well, okay. It's somewhere in operational work. I won't scroll through it now. The prompt looks like this. I've also written it already, I just inserted the description of what needs to be created, the style, and then the style is listed, composition, and so on. You don't always need to use and generate artificial intelligence in your work. I mean, images through artificial intelligence. Yes, few people work with this, but often you need to create either a logo, or a slide for a presentation, or some kind of block in a style for some social activity. I'm also working on a large project. And there was a mass project relatively recently where departments provided their logo, their vision of the logo. The vast majority did it through artificial intelligence. Is it a work task? No, it's done as part of work tasks, yes. So why not use it here? So, I just asked to create a conditionally cartoonish cat. And it's difficult, but simple, well, it means difficult to understand, simple to apply. I thought it would be a good slogan. It generates well. There is a limit on the number, so I think three of them now, per day, but for some logo, why not? Or check your concept, to then go to paid services and do something there. What I won't have time to show, but I'll mention, sometimes it suits me and I like development.
KPI, that is, when you need to develop or improve something from scratch within the framework, well, my work is largely project-based, and setting a task according to SMART, there is indeed a section below. Set a task according to SMART, and separately, if anything, if you use this, make a MindMap, that is, a kind of map. Here, text can be transformed, a MindMap based on some kind of directory, organization, pictures, and so on. Yes. So, colleagues, there's a lot. I understand that I went through it quickly on purpose. I understand that we had a rather dense flow of all this. So, I will probably ask to display the presentation on the screen again, so that we have time to finish a little. My message, probably, was the following. That's all, thank you. I'll leave it here. It can be applied to many things. The problem now, in my opinion, and from the experience of past webinars, and from the experience of working with clients on real tasks, is that this is a tool, but to apply the tool, you need to use it. To apply it well, you need at least 10 of your own tasks or one task 10 times, if you want to specialize in improving it. And to apply it, because if, say, a builder comes to you, you're renovating a house, a builder comes and says: "I have the best hammer, I have the strongest nails and so on," you're unlikely to hire him. But if he says: "I know how to build your type of houses since the tenth year, I will make sure that your house increases in value by half a million rubles and I will do it twice as fast as competitors." Well, you'll probably be more interested than in the first case." And here is a classic example with artificial intelligence of the first case - we are talking about neural networks, cool Open or Chat GPT or DPCIN in one form or another. Direct work with artificial intelligence is more important, even trying out super simple tasks. And our simple, working tasks, in the vast majority of cases, can also be done through bit assistants. I urge you to try it. And if there is a desire to understand deeper, stronger, then now we have already rushed through it, I didn't show it to save time. And there are my contacts anyway. Now I will hand over the word to Katya. Katya, sorry for taking a little more time than planned. If you want to understand deeper, we either go to our course, or try to search on the internet. If you want to do a project, a full-fledged business project, try either to assemble a prototype on any of the listed tools, on a tool you like, or go, again, well, we have experience where we do projects for someone ourselves, we do them with someone, yes, because it's important and they want to understand. And so on. Here. But now, the answer to the question that I hinted at at the beginning about training. Katya, please, I hand over the word. >> Thank you, Maxim. It's okay that you took time. Well, yes, it was important. There are many questions. I will also try to be quick so that we can dedicate time to answering our colleagues' questions. Next slide, please. Ah, thank you. So, the training center of the first bit has been operating for over 20 years, and for over 20 years we have been successfully training clients in 1C. And when last year the artificial intelligence trend began to rise so actively, the question arose: why not train clients in the use of artificial intelligence in relation to 1C? In our center, Maxim and I have developed two completely practical courses, on which you will all be engaged with a methodologist. A course for users, who will learn to correctly use artificial intelligence without using code. In essence, they will learn to write prompts, they will create their own Telegram bot, if they wish. That is, they will come with a specific task and leave with at least an approximate understanding of the solution, and sometimes with the solution to that task, depending on how complex it will be. And the second course is a course for 1C programmers, who already know 1C from the inside and want to learn how to integrate it into their processes, into their system. In essence, the second course is longer, and it will also end with solving a task, building a model. So, next. In essence, why study? I have just spoken it. All the advantages of studying with us are on the slide. You can study online without leaving home. We provide all the necessary access for training. That is, there is no need to buy anything to study. Further, you will make a decision based on your tasks and needs. The next course starts on Tuesday, next week. This is the course "Practical Artificial Intelligence for 1C Users". And in 2 weeks, on Thursday, the course on the practical application of artificial intelligence by 1C programmers starts. You still have the opportunity to sign up for these streams and start learning to apply artificial intelligence in practice right now. There is a special price for the next two streams. You can scan the QR code with a link to the courses. Yes, there is a special price for the next two streams, which is highlighted in pink on the website. You can sign up for the group at this price. Our groups are small. Usually, there are a maximum of 10 people in a group. That is, the teacher's attention is sufficient for each listener. Next, please. For those who are ready to try bitassistant in action right now, go back, please. You can use the QR code to go to our assistant and start using it for free right now, look at the prompt library that Maxim talked about, and try to solve your routine tasks right now with our assistant. For free use, there are quite good limits for testing, definitely enough. If you need to expand them, you can also contact our employees in bitassistant, who will help solve this issue. Next. And I believe we can move on to answering questions. There were quite a lot of them in the chat. >> Yes. So, I'll take over then. Katya, if anything, join in too. Thank you, colleagues, super active. Today I am sincerely grateful, because usually we don't even get ten questions, and now I am personally very interested, I will try to answer all of them. So, the key question I saw not for many. So, will there be a recording? This is not the key question, I just saw it now. There will be a recording, the manager will send it. The key question is about the connection between artificial intelligence and 1C. I tried, on the one hand, to show the chat. It's clear that what I showed is applicable to business tasks. And 1C belongs to business. This is the third webinar in the series, if you noticed, after the first two. And in the first two, there was more demonstration within 1C, we did something, we exported it immediately, and so on. Now the goal was to show what can be done with what we exported, what we took from 1C, and how we then apply it directly in artificial intelligence, because, as a rule, understanding, it's not so difficult to get data from 1C for a simple task. I sincerely urge you to look at two things. In 1C, maximally deep. The first is reports. A large number of specialized reports that can be exported for a period with specific filters, tabs, and so on. and give them for analysis. And the second, I remind you, colleagues, in your list of documents, there is a button "Show all", "Show list". This will display the list. What you see on the screen will be exported to Excel and can also be given. This is when working with a large number of, for example, sales or cash flows, if we are talking about finance. Well, for example, export all receipts and expenses without accounting and analyze further by counterparty. Again, questions, specific ones, for me personally, some were even, unfortunately, excessively deep for this course, from a financial point of view, for example, or some accounting ones. I see it this way. And I would ask you to first formulate for yourself: what do you often or routinely do now? And then, usually, the answer will naturally appear: "Oh, this can be transferred to artificial intelligence, because, well, here are recent practical examples. Colleagues of one client received a lot of primary documents in accounting, what accounts they have. Either for some reason, well, the business was structured that way, it was written by hand, or it was an Excel file, not always formalized, far from always, or a Word document, or a message in Telegram, well, very fast work, which recognizes from any incoming file, well, of any type, whether it's Excel, a file, text, a picture, nomenclature, quantity, articles, if any, characteristics, if any. Well, because it looks in the database and so on, yes, a banal example with this. Entering primary documents is a service of 1C document management, for example. So, the video question is about entering. Plus, you can add any other, but primary documents are such a thing, you need to see how much the current service suits you. Perhaps, for example, based on scanned documents, it will be more effective to use the current existing service, but add it as an intermediary that will improve quality, yes, for example, or look at some controversial points itself, so as not to give everything away, although no one will stop you from giving everything away. Financial flows, in my opinion, personally fit in there perfectly. Simply because, again, the question of what level of financial analytics you are setting, you look at whether it's budget analysis, budget comparison, drawing up cash flow statements, depending on what you do, if it's bitfinance especially. By the way, in the new Bitfinance, colleagues from another of our divisions have integrated artificial intelligence as an entity, that is, you can create a specific model that you want to communicate with, and it will communicate. So, financial analysis is also great, but here personal data is very important, or rather, commercial secrecy becomes more important. I saw an interesting question about dialogue. About dialogue. Can the system conduct a dialogue with, well, and the system conduct a dialogue with the user and build the context of the conversation based on the context of the conversation, build it further, based on the prompt? Yes, it can, it depends on the model's power. Cloud-based models can do most of it, local ones too, but you need a more powerful server. Plus, it's very important to describe this correctly in the prompt, that your goal is to get answers to these questions. If you haven't received answers to these questions within the first few messages, ask additional questions. Don't stop asking additional questions until you find out. Well, and now the second question, when I asked if the client would communicate with AI. Yes. That is, there are cases when implementing it directly, so to speak, head-on, we just put a chatbot placeholder on the site, remove operators, it worked poorly. Therefore, it needs to be implemented very carefully. And that's why I say that a chatbot can help excellently. And by the way, a very frequent request is to create a salesperson chatbot on 24/7. I personally like that algorithm, and I try to offer it to all colleagues, to clients, when the bot has communicated, identified the data we need to create an order in 1C, it creates a draft order in 1C and simultaneously notifies the manager so that the manager can then handle this order. Because the answer to the question, can it create a document in 1C and fully complete the operation, yes, we can. Is it safe to do so now? No, it's not safe, because verification and validation are needed. And, well, in the context of document management, for example, recognizing emails is a trivial task. I also talked about it, I saw an example that they were present at the last one, I'll say it again, for a client, we did, well, we helped, and participated in the project for the full automation of the incoming email process. It's still being improved and so on, but the point is, there are many incoming emails, more than hundreds a day. They need to be processed by artificial intelligence to read the text, read who sent it, who it was sent to, read the attached files and collect 15-20 attributes, create a card in document management and send it further along the route. So, the first step of the route is a person. This is normalization and the process is called that. And yes, it works, and it could be fully automated, but for now, I'm not ready to do that. I think a person is always needed for control. Let's move on to the rest. I saw several questions about bit assistant. What models are in bit assistant? The models are open. That's why I showed that there is GPT 4.1, for example, if you open the model list. There are three of them, there is also 4.0. For example, therefore, we do not send confidential data and personal data there either, right? These are strictly local models. Bit assistant can be installed locally on your server. That is, we have cases where colleagues asked for it, and it's called, well, a company assistant, so to speak, it's deployed on your server, either physical or rented, and then it works without the internet. All. This is the closed bit system now, which is here. It's free because they are open models. An important point. Why is this not always important? Because it's very easy to add a de-identifier either through another artificial model or programmatically, there are already existing things that will replace, for example, the names of all organizations with Daisy, Peony, and so on, with flowers. All full names and passport data with, I don't know, cartoon characters. That is, you work with contracts, for example, but you first anonymize them, and then work with them in a cloud network. Why not? This is also a real case, it's done this way. So, can the 1C constructor be combined with E or GPT5, for example? Chat GPT and other models have APIs. Many of them do not have all the functions that you see in user mode. It can be combined globally, but I don't know of a ready-made solution. More precisely, well, we are still developing it, so that it's possible to connect to the assistant base within 1C, which is more or less local or controlled, and not on Open servers. Well, colleagues from the laboratory promise to help with this, well, to create the infrastructure so that we can do this by the end of September. And then perhaps it will be interesting to talk about integration. So, I answered about dialogue. Will there be a presentation on accounting? I tried to show examples with postings plus talk about exporting reports. In my opinion, regarding accounting, personally, from what I've heard most often, it's questions about reports, about data that has already been formed in some way. For those who haven't used it, let me remind you, in 1C there is an interesting thing called a universal report. It allows you to build any report on any system data. Well, the only thing is that it's specifically for a particular, as a rule, it's a single data source, that is, it's only for sales documents or only for receipts or for the tabular part of sales, for goods, and so on. And there is a developed 1C report, there is an opportunity to develop your own external ordinary report in the 1C language and then export it to artificial intelligence. Because from what I showed and said at previous webinars about accounting, it's checklists for junior employees, it's instructions, today I also showed additional analysis of reports and work with, well, hints in complex accounting. In my opinion, it's far from being covered by this, but again, you need to find the routine or find what takes time, because it's different for different clients. A great question, I personally really liked it. Can a model form a language, or rather, a query in the 1C language? For those who don't know, 1C has its own query language. In Russian, it sounds like select sales of goods and services.link, .name and so on. Yes, it can. Short answer. We are testing it. Can it work fully, universally, not yet. Very high sensitivity to syntax, very high sensitivity to changes in the structure of the configuration in updates, but yes, it can make queries, but you need to check, correct, supplement it a little. So, how is the text recognition from images going? Excellent. I also recently spoke at a gathering of our internal bit employees, there was a gathering of thousands of employees, and we talked a lot about artificial intelligence. Three out of four projects that were presented at this gathering were about text recognition. Text recognition works excellently. In my example, it's recognizing email text. There is an example of recognizing nomenclature from very diverse sources. There was an example of universal recognition from UPO, invoice, goods T-13, well, Torg 12, I apologize, or T5, if we are talking about ZUP. Recognizing and creating primary documents is also possible. Well, there are solutions, there are semi-ready ones, there are ready ones, but ready ones are more specialized, you can assemble them yourself. Well, here it's convenient. What happens when a document doesn't fit into the LLM context? So, we haven't talked about this, but I understand that colleagues are already clearly aware, yes, there is context, but the window where we write text is not infinite, it's usually limited. A year ago, yes, even 8 months ago, I would have told you to write carefully, the context is, well, four to five A4 pages, well, Word pages, so you need to fit in there. Now I will say this, the context has increased significantly. The context is hundreds of A4 pages. If we are talking about cloud models, it's clear that for local ones, the context is super important. And if the context doesn't fit, it's the same as you're talking, you receive a letter from someone who didn't finish writing, right? You see the first part, you can guess what's next, but the cut-off part, well, you haven't received it, and therefore you don't work with it. How to estimate the size of a PDF file in tokens? Good question, again, colleagues are clearly aware, tokens, yes, this is what artificial models count when interacting with them. That is, you write, and they spend so-called tokens on writing a response to you, on analyzing your response, on writing a response. Roughly speaking, in some cases, a token equals one character, but not always. So, the size of a PDF file. Here the question is more interesting, because, in my opinion, it's a question of recognition, right? That is, you first recognize, well, as a rule, all text from it, and accordingly, you initially convert it to text, and therefore there is more counting of tokens for the text. But if we are talking about audio models, yes, or video models that work exclusively with images, they have their own mechanism. Here, well, I doubt you'll send a PDF file there. In business applications, I've more often seen that PDF files mean you extract text from them. Well, and the size of text tokens, we can estimate plus or minus. Roughly speaking, it's one token per character. You can estimate with a small margin, because it's not always like that. Can LLM understand text with formatting, i.e., headings, bolding, etc.? I see your comment. I'll answer it right away. As a rule, yes, modern cloud models can, local models can, depending on which ones you choose, but there's a small trick, a nuance, when you write text for the model to process, it's better to write it in a structured way, i.e., with bulleted lists, with indents, with paragraph separation, well, paragraph separation is not so important, it's the indents of paragraphs that are important. This is usually understood by almost all models now. So, well, bit assistant can be deployed on a computer, but it will be local and it's more expensive. That is, you can deploy it locally on your computer, but it requires the work of colleagues from the laboratory. It's more expensive. It's like installing a local model, because installing bit assistant means installing a local model, that's the main problem. And all this is what I said earlier, yes, a graphics card requires a sufficiently powerful server. So, a specific task. So, if we have time, we'll discuss a specific task. So, the discussion is already going on among themselves. So, let's probably wrap up. Regarding, probably, I can answer two things about a specific task. And I've hung a hook about working with data and 1C. There are three approaches to working with data. When you have a database of artificial intelligence, oh, a 1C database and artificial intelligence, and you want to get something in the end, for analysis, compilation, recommendations, and so on. My first option, shown to you, is that you manually form data in 1C, export it. As surprisingly as it may seem, in the vast majority of cases, this is enough to understand whether this task works at all, whether the idea is viable. The second situation. The second example, when you say, um, how to put it, you call a model from 1C at a specific moment. This is already integrated artificial intelligence. You use APIs of open models or closed ones, if we are talking about local models, you can use the same APIs, well, essentially the same, roughly speaking, the same dialogue. But you write your context there, that is, the message, what role it plays, what you exported from 1C, for example, it's a large XML file or it's a set of Excel files. As a rule, they exchange specifically JSON files of a specialized structure, if you are not involved in programming, it's not so important. The key idea is that code is written within 1C that accesses artificial intelligence. Through its published API. It works well, the complexity is obvious, that for a specific task you write specific code, right? That is, this is what we do in the course with programmers. We are analyzing three approaches. Simple semi-manual, when you need to make a good report, well, enough in 1C to make a good report to send it, to analyze it well. The second option, code within 1C, which allows specific data, this is the key, specific sales, goods, well, prices of goods, balances for a period, to send to artificial intelligence and write some prompt and communicate from 1C, call artificial intelligence with your data. And the third way, on the contrary, you say to artificial intelligence, take, for example, a local model, here's a link or some other access options to the entire 1C database, yes, this is another approach, when you take and not from 1C call artificial intelligence, but from AI call 1C. This is a more complex task in terms of implementation. Plus, it's quite complex in terms of authentication and access control. Well, that is, simply put, as soon as we give access to the entire database, well, obviously, not all users should see everything, right? And 1C rights no longer work in the classic sense. Well, if we talk about the extreme, that I just published the entire database in a certain form that artificial intelligence can read. This can be done, but it works for very specialized tasks. It should never be released for open chat access and so on. And you need to balance between these three approaches. This is manual or semi-manual data formation in 1C in the form of reports or something similar. This is calling from 1C and writing code for the logic of working with it. And the third is direct calling from AI. But here there are many pitfalls, especially with data access restrictions. So, it seems. So, let's see if I can answer a specific task, time is running out. We need to analyze documents in 1C Trade Management, shipments and advances for each client. If there is simultaneously a debit balance for shipments and a credit balance for advances, it is necessary to reclose the documents between themselves, display a consolidated balance, either debit or credit. Is this realistic? Yes, it is realistic. And by any of the three methods I listed. The simplest way is to create three reports that sequentially show each of these parts. That is, the first report shows, oh, I apologize, debit. The second shows, for example, credit. The third, shipments documents, you can say, the fourth advances. That is, to highlight the data in a form, because you can use standard Trade Management reports to output this, collect this data. Then download it and send it to artificial intelligence for analysis. The question is how accurate it will be, how suitable this simple option will be for you. This is already development, this is something to look at, try, check. But in general, yes, the task is realistic. This is a prototype. Reports were made in 1C and then collected, directly. Oh, colleagues have even made requests, yes, I see. In the 1C platform, there is a built-in text function that can be sent for analysis. And, well, why not? I'll highlight, there is also this 1C speech recognition service. It's specifically from 1C colleagues. Well, it's published, and they are developing it, an interesting thing. It's enabled in the settings, linked to the 1C contract. If you want, you can try it. So, well, it seems I've answered everything. So, colleagues, let's summarize. In general, yes, I had a webinar and tried to continue the first two topics that I discussed with someone who is present now, with someone I haven't discussed, but I tried to talk about some things so as not to lose them. This is about an example of a simple, but maximally broad application of bit assistant. In my opinion, it's a good starting point, and if necessary, then develop it further, or not it, but move on to other products and tasks in a more serious way. But to start using artificial intelligence, to at least apply it in your life, this is the simplest way, in my opinion. Why not? So, if there is time, colleagues, opportunity, yes, I saw the question about the course cost, I honestly don't remember it, I don't know, you need to go through the QR code, it indicates current discounts and so on. So, if possible, please take the survey, those who stayed, who sat through it, those who are our most persistent guys, I hope it was interesting. So, if something else flashes in the questions, I'll answer. Yes, Evgeny writes that, well, this relates to the request about the structure of directories and registers. I'll also add, there's an interesting option to program in 1C through artificial intelligence. This also requires specifying the structure, exporting the configuration as a metadata structure, and then using it. Again, it can be used in different, for example, in the cloud tool. So, colleagues, by the way, some people wrote prompts, a different question. I suggest you try this trick. You can write prompts yourself, and it will work well. I am more than sure you will succeed. Or you can write to artificial intelligence, because bit assistant or any other. Write me a prompt for the task. 1 2 3 Consider this and that. This works quite well. So, by the way, I didn't really want to dwell on this topic, but nevertheless, there is such a profession, it has been published quite a lot. And in general, it's still alive. Prompt engineer. Prompt engineer in Russian, a person who knows how to compose professional or more or less professional prompts. Well, in my opinion, 90% of such work is now covered. Well, let's say, even my course listeners will be able to do this, simply because if you know how to formulate your thoughts, if you can clearly set a task and not give up, this is the most important thing, on the first prompt [music] tell it: "Improve the prompt to this extent or do this next." You will succeed well in any case. Well, and in particular, in fact, in particular, to write a prompt through artificial intelligence, and then edit it manually, is also a working strategy. I sometimes do it one way, sometimes the other. >> Colleagues, I answered about the course cost in the chat. Now, with the promotion, the course for users is 41,900, for programmers it's 55,900. >> Yes, thank you very much, Kaz. >> So, I'm looking at the distribution of ratings. Those who answered, thank you very much. It's pleasant to see. So, it seems it was interesting, but at least I tried to explain what can be taken and applied here and now, and then transform it for your tasks, with a small, in my opinion, abstract approach, that if I can do this, then I can do this for any 1C system. Yes, thank you all very much. Many participated in the questions. Not all, but in fact, it has never been possible to answer all questions. Oh, there is one person whose question I did not answer. Colleagues, I will try to analyze and see whose question I missed. That's all, thank you all very much. If there are new webinars, we will write. Well, then, with some we will meet at events, with some at courses. Well, in any case, I wish you to apply artificial intelligence in such a way that it helps you. It is a tool in any case. The most important thing is your expertise and your capabilities. All the best, >> colleagues. Thank you for participating, thank you for your activity.