Transcription
Hello everyone. Today we have a very interesting episode. You have never seen anything like this from me before. And because today we will talk about 1C with Petr Tsap, who is the first 1C practitioner. And I know him as an ideological "vibe coder" in the 1C environment. That is, he is one of those who, first and foremost, brings programming with neural networks into the world of 1C. And today we will talk about this topic, talk about what 1C is, how it differs from regular programming, whether you should switch from your Python to this world, and so on. Petr, hello. >> Hello, Gleb. I have actually prepared a whole presentation on this matter. It so happened that I was indeed the first practitioner who started achieving some serious results, meaning not just playing around, but getting something serious. And I actually want to talk about many things today. In essence, I want to talk, globally, about two things. First, about how the pipeline and coding look in 1C, since it's not entirely standard, perhaps, but universal. Second, I want to talk about what to do next with "vibe-coded" products. Well, we don't "vibe-code" just for fun, right? Many people create their own pet projects, maybe larger projects, right. And I wanted to talk about this too, about what happens, how the 1C ecosystem is structured for those who haven't encountered it. So there. >> Let's start with this. This is probably the most crucial point for now, because when I tried to get into it, I realized there's just an abyss. And this abyss lives by its own laws. So, as a developer, I have Git, I have a development environment, everything is clear, I can release, deploy, and so on. But with 1C, some problems start at the stage of needing some keys, some versions of 1C itself. Then you download it and what do you do with it? Where? Give me, give me the familiar IDE, yes, give me Git. None of that is there. In short, a complete mind-blowing experience. Tell us about it. >> You know, that's a wonderful status, everything is complicated. Well, naturally, people have worked on this for a long time, trying to persuade 1C itself, which is known for its rather moderate friendliness towards outsiders. 1C's approach in this regard is very simple. Either you are a partner and you get everything, or you are not a partner and, well, before there was nothing at all, and now, well, something has appeared. Therefore, if you are just thinking about becoming a partner, delving deeper, or getting acquainted, then now there is something for you. There is an educational version, it can be downloaded from the website, well, from various websites. There is a Russian version, and there is an international version, which can be downloaded from the International division. Along with it, you will get the educational version of the platform, you will get the default configurator, which is built-in, and which is the main tool for the vast majority of 1C specialists in real life. And if necessary, you can download a special fork of Eclipse. There is such an idea, >> a terrible development environment, [laughter] yes, which 1C chose about 10 years ago as its prospective environment. And this is called EDT, Enterprise Development Kit. And, well, you can work in it, >> well, sometimes it can be very difficult, but it will work perfectly with the educational version. So, I'll say right away, if we're talking about Russian-speaking, Russian-speaking real practice, then according to my surveys, about 85% of installations are still where all current work happens in the good old configurator. Why do I say good old? Because the configurator was released in 2003 along with the first version 8.0. And since then, it has not fundamentally changed. >> Isn't constancy a sign of mastery, of quality, yes. >> Yes. The 1C company, in general, adheres to an ideology where they start to be interested in and notice only proven IT trends, right? So, if it has existed on the market for five years, then we will consider using it somehow. That is, when there is a request from clients, for partners, when partners agree, at conferences, in private, when a critical mass accumulates, then all this is added to the platform. Only then. And actually, since it's unusual, but since it has been working for over 30 years, and it works well, then, well, you can do it this way too. So, let's imagine that you have sat down and "vibe-coded" properly and "vibe-coded" something that you intend to sell to businesses in the former Soviet Union countries. This is a very important point, because this all concerns the former Soviet Union, in general, because Kazakhstan, and Uzbekistan, and many, many others. And even the Baltic countries, by the way, the penetration rate is slightly lower there, but it is also quite significant. And how is it all structured there. So, point number one, you see here, this is the key decision-maker, who for, I'll say right away, I will be talking about small and medium-sized businesses now. The most interesting thing is that for large businesses it works approximately the same way, only there are even more influence agents from the 1C side. >> So, it's really difficult there. There can also be different teams with conflicting interests. When you have some SMB, small and medium-sized business, yes, you need to clearly understand one thing. If, in general, it will be automated by some 1C product by 99%, then everything above self-employed, everything above that, there will be something based on 1C:Enterprise. And here I will talk specifically about 1C:Enterprise, because in fact, the 1C company is a very well-known, widely known in narrow circles strategic buyer that buys very, very, very many startups. Very, >> well, famous ones are Bitrix, yes, Megaplan, >> the most famous are Ama, Megaplan, Bitrix, and there are a million more of them, actually. So, technically, it's not the 1C, not the 1C:Enterprise platform, but there are many such businesses. So, the last thing they told me, there are over 350 subsidiary companies. They are, by the way, good strategic buyers. Their terms are known. Well, as you can imagine, when there have been 350 acquisitions, then all the terms are clear. They are good strategic buyers. They invest in reliable cash machines. So, if your SaaS has achieved some normal success, meaning you have users, they somehow intersect, normally intersect on the B2B side with products, and it fits into, let's call it, business automation, yes, in a broad sense, then 1C can be a great strategic buyer for you. I can't disclose the general details, of course. Of course, individual offers are made in all cases, but standardly 1C comes, buys a controlling stake from you, leaves you as director, pays you good money, and you go ahead to earn for yourself and the investor, the buyer. Everything, it's simple. Very, very many people work under such conditions, and they are all happy. Well, Nikovinin, and Ryzhikov are not fools, and they started working with 1C. Apparently, there's something normal there. >> Yes, yes. So, considering the exits that are currently happening on the Russian, specifically on the Russian market, well, let's say, 1C is a pretty good option. So, just keep in mind that this is a completely viable development path. And even then, you don't necessarily need your product to be based on the 1C:Enterprise platform. Not necessarily, absolutely not. So, for example, web technology products, this is perfectly demonstrated by CRM systems and other things, there are still other things, well, very specialized ones. The first key point. In small and medium-sized businesses, the key decision-maker for purchasing business software is the 1C specialist. Not the company director, not the company owner. It's the 1C specialist, sometimes the chief accountant. >> Everything is complicated. [laughter] >> This is a very, very important point. And you will actually be selling to him and bypassing him. I will show this further. I have prepared slides specifically for this. Next, problem number two. If we are doing B2B, something B2B, yes, and we are doing it with the intention of selling it to enterprises from the CIS, we must always clearly check if there isn't an analogue in the 1C price list. And this is not information that you will find immediately with Google or Yandex. So, you might find something, but I would first go to the 1C website, open the price list, look at your industry focus, for example, or client portrait focus. Well, in general, yes. At the same time, you can understand, in fact, how much the software is sold for. So, the prices are not taken from thin air, the prices are more or less working, yes. So, if a product costs, say, 300,000 rubles for an installation, and a subscription for it, they are ready to pay for a subscription, there are conditions, there are subscription conditions, what is included, well, then they are ready. So, you can try to sell more, for a higher price, yes, a higher subscription cost, higher. At least you will have a first reference point. But again, you will need to look at the analogue. If there is an analogue, even if it's worse, in the 1C price list, well, that's problem number one. Or rather, problem number two. Let's talk about it after the decision-makers. Now there's a key problem. I will talk about it separately. So, this is the probability of its presence in such a small and medium-sized business. If we are not talking about companies at the self-employed or micro-business level with up to five employees, it most likely won't have a hub based on 1C, but with a high probability, it will have a CRM system hub, yes, where everything falls and around which everything is built. Again, I say, 1C also buys this. So, for example, Ama, Bitrix, these are understandable examples. So, if you are digging into something in this direction, 1C is also your buyer. I mean a strategic buyer. Another thing is that in precisely this segment, micro-businesses with up to five users, you will not intersect with 1C with a 99% probability, since at most it will be 1C:Accounting. And 1C:Accounting is a whole separate world. So, if you are doing something for accounting, then you will intersect. If not, then that's it. It doesn't apply. But in companies that have grown a little bit more, even a little bit more, a hub product appears. And in 99% of cases, it's based on 1C. It's a certain product, a certain one. So, for a trading company, it could be "Trade Management," yes? For a transport company, it could be "Vehicle Management," or specialized ones like "Alfa-Auto" and many others. I assure you, if you open the 1C price list, there are very, very many niche products. Very many. >> Ah, so these are specialized industry solutions for this product, probably based on some platform, right? >> Yes, yes, based on the 1C platform. And a very important point is that we are talking in 2025, not 10 years ago, not 20 years ago, this product already exists, it is already installed, people have been working with it for a long time, and real work has been going on for a long time. If we are not talking about businesses that have just grown right now, meaning you had five people yesterday, tomorrow you have a large contract, your staff grows to 15, and when they grow to 15, all such businesses have the question: "What should we automate so that it works normally with accounting?" Because, I repeat, it opens up from accounting, which exists. So, a super important moment. Let me reiterate. Even where there are one or two directors, a couple of people, they will still have 1C accounting almost 100% of the time, with rare exceptions for all these services like Elba, even including a hired accountant. And the person who runs the business, half the time will be wondering how to run it, yes, and the second point, how to give the accountant the data, so that it's convenient, so as not to overpay, and so on. So, even at this level, this moment arises. If we are talking about a company with 15 employees, where you have grown, I don't know, trading, providing services, or something else, then the probability of an industry hub appearing is very high. And if we are talking about businesses with 50 or more employees, then there is a 99.9% probability that it exists and that it is a 1C product where everything revolves. There are really millions of them, all sorts of medical centers, transport, pick any industry, and you'll find something about it. If it's not in the 1C price list, then there's also a specialized marketplace, which is also, naturally, a subsidiary of 1C, called Infostart. And you'll find everything there too. So, here, a need arose, let's say, some time ago there was a very strong growth in sales of a specialized solution for state defense orders. Everyone came together, everyone bought it. Well, we're talking about many thousands of subscriptions. Well, because 1C also does subscriptions. >> Uh-huh. >> All updates and so on. And so, if such a product exists, then there are specialists who support it. This can be an invited specialist, an outsourcer, for example, a freelancer, it can be a partner company, a franchisee, yes, but mostly it will be a franchisee, but there are also many strong lone wolves on the market. If it's a larger company, well, say, 100 employees and, well, it depends, then there will definitely be a full-time specialist. If the company is even larger, then there will definitely be a full-time specialist. Entire departments start to appear. For example, I myself am currently working in a company, we have a little over 200 people. We have four specialists just for 1C, if we count. Well, naturally, there are also web programmers, naturally, there are administrators, that goes without saying, but there is a department. Accordingly, this operational product, the operational hub, is the heart of the company from the perspective of the CIS. You, in general, any solution, any solution that you bring and that somehow affects the business, and you probably, well, if you've even thought about it, then it will affect the business, right? Well, to sell it. Everything will be considered solely through the prism of how it will integrate with 1C. Can it integrate or not. If it integrates, plus, if it doesn't integrate, minus. So, it's very simple, in essence. >> Further, I repeat, on this hub product, I'm saying, there is a specialist responsible for it, who is like a tech priest, let's say, right? And this tech priest looks at this whole story. So, yes, you can sell, I repeat, you can sell to the director, or rather, attract the director's attention, excuse me, you can attract the attention of the sales department head. Well, or depending on the position, as it's called, yes, you can attract the attention of some production person, but if it's B2B, it will immediately, either instantly or after a meeting, be escalated to the 1C specialist in 99.9% of cases. There are exceptions, but probably. I've never heard of them. Well, I think exceptions are all sorts of project companies, some IT companies, where there is specialized software, very specialized, and they say: "I want exactly this." You somehow integrate it with your 1C. No, >> you know how large the 1C departments are at Yandex or Playrix or similar companies? When? A good department. [laughter] Departments. Well, of course, who calls them what, of course. Yes. If we're talking about retail chains, then we're talking about hundreds of people. So that the scales are clear. An exception might be if you bring a B2B solution for another IT company, and it affects their main product. Only then, perhaps, such a question won't arise. But this is a specific, well, a specific situation, it can happen, of course, it can happen easily, but it doesn't define the market, let's be honest. So, if you're going to delve into any, let's say, real industry, like the real world, then you immediately encounter this instantly. So, blocking is the right way to make the 1C specialist your partner. Well, the right way, in my opinion. Why did the 1C company acquire such gigantic influence and spread so well? They never started selling. When you sell any software from the 1C price list to a client, for the client there is a final price of 300,000, you get 150, well, I mean, as a partner, I apologize, you sell a subscription for 100 or how much, 150,000 per year, and you get 50%. This is their rule. Even when they created the cloud, colleagues, so that you understand, when they created the cloud, which, by all rights, they should have taken 100% for themselves, this is purely cloud accounting, there is cloud accounting, quite popular, I have many acquaintances who use it. No, everything was done beautifully. Each user is assigned a partner. Partners, so that they can fix sales, here you go, 50%. Well, this is a well-known thing. This is why 1C gained strength, why they became the de facto standard, and Bitrix followed the same path. Ama, well, I don't remember, but I think they also followed the same path in the end. In fact, it's a very cool scheme if you are thinking about partnership. In general, it works in many examples. Do you want to make a subscription for 5,000 per month? Make it for 10, give 50% to partners, consider the value proposition. I assure you, the partner will come up with something themselves, because they have access to the decision-makers' ears around the clock, well, daily, unlike you. So, you need to understand very clearly that 1C specialists, well, specialists servicing 1C, have at least constant access to ears. And at most, it's daily. If they are your ally, then you're done. Well, here I have written down the main fears, interests, and pains, specifically for specialists. Here it's very simple, in essence, the fear is, what if something breaks, if any integration is assumed, if anything at all is assumed to be connected, what if it breaks. Preserve the status quo. If this product has any benefit, what benefit, I don't know. You need to think about this in all thorough cases, yes, everything is fine. If not, well, then you get an enemy for no reason, who will grumble about why we installed this crap, we should have bought this from the price list. And I assure you that in a very large number of cases, they will grumble. They will grumble, and you will, well, and the installation will be lost. Well, churn is churn, all that, so there will be a disconnection. And B2B disconnection is an expensive story, actually. Well, and the pain, if integration is needed, then you must, and here's another important point, conditions of uncertainty. You can try to invent anything, do something about integration, but how it will actually be structured will be decided not by you, but by them. So, you not only need to invent something for integration as such, yes, come up with some solutions, study everything, but also, how to put it, leave it as open questions, provide documentation correctly, implement maximum flexibility so that they, within their sub-project of integration, determine how it is most profitable for them to do it locally. Don't make any rigid decisions. They will be useless here. They will rather harm you, you will just waste, well, excessive effort to do it. And the last thing I wrote down, if you touch the interests of accounting even a little bit, then you are on a minefield. It's just completely random. I have been a commercial programmer for 22 years, most of this time in 1C, 17 years in the ecosystem. And even my wife is a former chief accountant and financial director, and I can confidently say, I've seen hundreds of them. It's, in short, a minefield. I'm telling you seriously, it's a minefield. How it will work is just random. If you are trying to do something around a B2B thing, which, say, lawmakers have come up with something, and you quickly, we have "vibe-coding," we can do it quickly, right? You quickly find something, start trying to sell it. But if it's, again, legislative, and if accounting, and the enterprise is involved in this, then they will also look at you. And they are also, they are the harshest decision-makers, meaning, if it affects, if the chief accountant doesn't like you, then your software won't be there. Even one of them won't convince them. Well, that's how it really works. Well, if, well, if everything is fine, if you are closing some pain, and if you are closing it beautifully, the director might even, well, they will bring it to the director for signature, to pay here. I'm serious, this has happened hundreds of times. This has happened to me personally dozens of times. So, they will bring it to the director for signature, to pay here, everything. Everything else will be resolved, decided. How to integrate, what to where, who to launch. I'll say simply: "Everything is fine." Next, I've gathered simple survival rules here. This is what I've summarized. In general, what I'm saying out loud now. Just step by step. One, two, three. This is how it's structured in the ecosystem. >> Listen, I have the main question. You say: "Everything for 1C already exists, so it's like nothing can be done." What can I do? >> Well, for example, let's not even give examples of specific products, although that's also interesting, but how should a person think? So, we are listened to by, well, let's say, small programmers who often don't understand this world very well. They understand how to do something well, but don't understand what. And this big, complex, scary world of 1C with accountants and other general directors. What can they do? Well, look, there is, for example, the Infostart marketplace, which I've already mentioned, yes, if you go there, there are really thousands, if not tens of thousands, of paid products, each of which solves some niche problem. It can be ultra-niche, it can be aimed at an accountant, it can be aimed at a salesperson. For example. The simplest example off the top of my head. All ready-made integrations with marketplaces. I'm most involved with marketplaces now, working with them. They are all there. So, if you've developed some integration solution, list it there. That's optimal. That's optimal. Next, if you are doing something bigger, meaning you are trying to do something here, where there will be some kind of CRM system, well, I wouldn't really recommend competing with Ama and Bitrix in the Russian-speaking market, but you can always try. So, if you are trying to compete, you must be at least integrated, no worse than them. And present everything, otherwise they won't even consider you. So, there's no point. And if you are doing something completely new, well, for example, a simple example. Finding counterparties. So, some time ago, you know, this story, checking counterparties, with full data, with a bunch of services. When they all started about 10 years ago, a little earlier, some of them, 1C eventually bought half of them, and hid them. And the same, well, not hid,
In essence, they still work. You don't know that they are 1C. Here. But they belong to 1C. The same applies to all sorts of electronic document management systems, integrations of history, and everything else. You create a company, enter the market with a solution, connect gateways, sell to people, generate certificates, then 1C comes to you, buys you, makes you an offer, buys you, and you continue to work. >> Now, it's popular now, right, with artificial intelligence. For example, I made some kind of PDF parser with artificial intelligence. It needs to be integrated with 1C, am I understanding you correctly? >> Generally correct. 1C itself is working on such products. It has its own ITS. Well, in subscription, in the sense, it's called 1C Information Technology Support. That's what they called the subscription. Uh, it has such services, but you can always do even better. If it's good, then first, buyers will buy you, and second, then 1C can buy you. >> Well, excellent, excellent, yes, >> I repeat. So, all these features like smart PDF, everything you can think of with intelligence, wrap it in B2B, uh, and have it as a service, that's really a window >> some kind of voice assistant, right, on the phone, where a robot answers me about my nomenclature, for example. I've heard about something like that. Well, that's a classic, classic thing that is now entering people's lives. That is, if you do anything at all, right now we have a super mega revolution, and it's best, of course, in coding, like text management and everything else, we help ourselves, but, uh, from the perspective of end-users, all useful applications in the Russian space of the former Soviet Union will go through 1C, and you will have to integrate, you will have to provide data correctly, take data correctly. Well, and if you are going to integrate, you will definitely have to integrate, right, so why not go to the market right away so that everything is good for you? >> You can. By the way, you know what I thought? I would probably personally look for some popular products outside of 1C and look for those among them for which, for some reason, there is no normal connection with 1C, and just, well, look for such connections, I don't know, you say marketplaces are great, right? You can find a popular connection, but not only. I don't know, take some банально email, right. Uh. Some kind of connection. Chat GPT. Even Chat GPT is a product, right, uh, it's not banned in our country, in Russia. So you can use it, there are difficulties, but again, think of a connection, something GPT 1C. What could it be like? Yes, well, something, well, in short, something like that. Gleb, go to the Marketplace, to Fostart, there's already a lot of that stuff there. You'd better look at the world. I'm just sorry that I rolled my eyes here. I was just trying to imagine even a little bit of a serious product on the Russian-speaking market that doesn't have integration. And I don't know of anything like that. Well, if we look at something international, right, at something global, at the freshest trends, and then quickly localize it and build it in, then here will be fire. Here you can earn money, and earn it for a long time. That is, you'll sit there, do an install, somehow set up a subscription for updates. Cash machine, it's simple, a cash machine. You sit and rejoice, then 1C comes to you, if you occupy a sufficient market share and become more interesting to them, they come to you, make you an offer, you get additional funding and sit further, if you want, you earn, if you don't want to, you hire a director. I knew such cases too. Well. When, in the sense, the owner of the remaining non-controlling stake hires a hired director. So, you understand the landscape, what to do next? >> What to do next? Well, you know, I'll repeat myself, look at global trends, right, right now, that are emerging worldwide. That's what people are doing now. Then, of course, 99% of it is AI and everything related to it. That is, we take this, uh, start to understand how it is, we need to take exactly two steps, right? The first step, how will Russian-speaking users accept it, including technical issues, availability, unavailability, where will you proxy the API and everything else? Because, for example, Claude doesn't sell APIs to Russia. Chat GPT, between >> not just doesn't sell. Go find a VPN that works well with it. [laughter] Yes, yes. Well, so, yes, so now, among what works more or less freely, it's only Chinese, among what works more or less freely. Accordingly, you need to understand what neural networks work here, where we apply them, how you will proxy the API and how you will integrate with 1C so that everyone is happy. And considering that a very large number of IT specialists are still in denial, I understand that we, Gleb, are pioneers, fans, interested, talking, but according to my estimates, if you open Habr, everything is bad there. Well, if you open Habr, open the comments, and it's a nightmare. So, like in the Winnie-the-Pooh meme, right. You know, but in real life, when you talk to people, well, I'll say boldly now that 80% of people in the industry are in denial. >> Yes. So, yes, and of course, yes, I'm looking at it. Look, there are different opinions, they say hype, not hype. I've experienced a lot in my life, different hypes and non-hypes. I clearly understand the difference. >> And the difference is that if it changes your way of working so much that you don't even listen to these 80%, you know? >> You understand that these arguments are really pointless to listen to. It's like, uh, a well-worn, maybe, well, a difficult example for some, but it's like tank battles in 2022. It doesn't work, it's clear immediately. It's just that some people look a little further, understand a little better. There's no point in discussing it here. >> But the good news is that the remaining twenty need to be sold right now. Any solution that really provides benefit, that has added value, you want to take a part of this added value for yourself, really, they will already buy it. So I think that with further penetration, say, in the next 5 years, we will get full market acceptance. Well, that's my preliminary estimate for now. It might even be faster. Well. But for now, I'll conservatively budget for 5 years. In general, this is the fastest, in quotes, hype that has happened in my life, to be honest. Therefore, I would say that earlier, if we take previous, well, tectonic changes in IT, I would name two, actually. The first is the emergence of the internet, the second is the emergence of normal smartphones. Well, personally for me, the internet penetrated for a long time, about 10 years. from enthusiasts to more or less mass adoption. Smartphones penetrated faster, already about five years, and the first iPhone was in '08 or '09, and by '14 it was already mass-produced for almost everyone. >> Software, Gleb, even with a delay. >> Software with a significant delay. Yes, yes. And the current story? Well, I'll say it will be twice as short. Well, maybe 2.5 to 3 years, but it's just beginning. It seriously began in 2025. In early 2025, the first networks appeared that truly solve problems, not just for fun. So there is still time, yes. >> Tell me about the difference between pipelines and programming. >> Of course, everyone is waiting for that. >> In 1C. With 1C, it turns out. Especially since if you are going to do integration with 1C, you will most likely have to work with all >> everyone is eager. [laughter] So, look, what works. What works is Set 4, GMI Pro works, Opus 4.1 works super well. Generally, well, a gem, not Russian, just the most expensive. GPT5 works only for specific tasks. It generates 1C code, it looks plausible, but there's still too much hallucination, too many errors, too much, well, much better than before, so it's improved, and before they were worse, but you can't properly prompt GPT5 to write you 1,000 lines of code, a bunch of things you need, and have it work. Sonet can do that, Opus can do that. Jemi, well, with caveats, of course. In short, for me, personally for me, a good error rate is when I get no more than five, or rather, if I divide it like this, 1%, no more than one syntactic or hallucination error per 100 lines of code. So, this metric suits me more or less normally. When you generate a thousand lines, quickly put it in, check everything. Just yesterday I was testing this new third Kven, which is made with my architecture with sparsity. You've already heard about it, that it's 80, 80 and 235, but only 3 billion works simultaneously. In short, for it, of course, you need to be clever with semantics, but on the other hand, they didn't just make this sparsity, you understand, it's essentially made for multiplexing. That is, it's a model that should be on the office server, and it gives 3 billion to the accountant, 3 billion to the manager, 3 billion to the 1C specialist, 3 billion to someone else. All this simultaneously, more or less. The model itself is like that. So I really like this. You know, the Chinese are digging a lot for these models, specifically for Chinese businesses that have installed such servers and continue to install them. Well, local, that kind of local work, right. This Kven third, it really worked yesterday at the level, I tested it on real requirements, on a real work task, it generated about four hundred, I read it carefully. It really works at the level of Sonet 3.7. So, in the spring, I actively used Sonet 3.7. So, really, by feel, it works at the level of 3.7, specifically for 1C. So, I always check, when new neural networks come out. Chinese ones didn't impress me in this regard. People release something with deep learning, I start checking it, and it's not good, in short. But this last Kven, released a week ago, is really impressive. Yes, there are bugs, yes, there are glitches, but I quickly scanned it, it generated four hundred lines, I found only two errors that were immediately obvious. Well, of course, there might be something I missed logically when running it, but overall it's very decent. It's more than a working option to sit down and work with it. My colleagues have also managed to look at it for generating requirements. It's also very decent. It's not smart, of course, because it doesn't have reasoning. It's a generation machine, but overall, everything is well incorporated. So, it doesn't do very complex things, you know, it doesn't need conceptual architecture and all that. But, you know, working with one object, doing it locally, everything is beautiful, it does it well, normally. So, the quality is satisfactory. What pipeline have I developed? In fact, it developed, I don't know, about a month after I got the first good working code. And that was in winter. Still on Sonet 3.5. And I was very happy about it. So, the pipeline is, by and large, simple. Requirements in neural network form. I'll explain in a bit more detail what that is. Then, accordingly, we move to generation, and then we do testing. And in the context of 1C, scenario testing is our everything. The rest is optional, depending on who has it and who doesn't. And in the vast majority, other types of testing are not used. Well, if we look at real practice, then in the vast majority it's all scenario-based. So, you sit down, click through, you click through yourself or by machine for the particularly advanced, click through by machine, and that's it, and you look. Logs, accordingly, you return them, then you work with them. Optional, but ultra-useful specifically in the context of 1C. These are context describers for correctly stitching them together. Let me explain what that is. It's useful in other things too. Because I'm also slowly working on Verdem, I'm also involved in it. I worked as CTO in Gendef, which is also close to me. I've worked with it a lot and still do. What is it? Imagine, Gleb, you have a business process, and within this business process, you use, say, step one, step two, step three, which are processed by different modules. Understand? [laughter] Since 1C is large, you need to clearly understand that 1C is large. And what web programmers usually discuss, well, the size of projects, honestly, sometimes you chuckle, because, for example, my working configuration has 4 million lines. >> Uh-huh. Normal. >> And it's small, old, rewritten, specialized, right? So, somewhere, I don't know, 2 million, 2.5 are native to 1C. The rest is accumulated over 15 years. So, modern configurations, which are installed now, like the latest ERP, colleagues said, it's significantly more than 10 million. So, it's already around, it's closer to 50 million lines. Its text export weighs several gigabytes. Text XML. Well, XML is a rather redundant format, as you understand. But this means that it won't fit into the context and won't fit. It might never fit. I'm not sure they can achieve a thousand-fold increase in context. Therefore, uh, I encountered this from the very beginning, I understood. Therefore, all my work, the 1C work that I do, that people who listen to me, who learn from me, do, is done strictly through content chunking. Strictly. So, even the automation of this level is minimal. I first teach them to work with their hands. So, no cursor, no RAK, that's all, it doesn't work. >> Shotgun. [laughter] >> Shotgun. Uh, >> well, this is for the viewers. It's a software, the essence of which is that you just click which files are needed for the context, and it prepares them. But 1C has its own quirks, right. It's all arranged differently for you. >> In short, in fact, a 1C configuration is a kind of large, gigantic object, a configuration, right, which is loaded into the database and stored there, launched, it's launched as a whole, simplified, when you export it to files, it's exported, roughly speaking, into two types of files. There are declaratively described forms in XML format. Well, you open it, and there it is, where each position is, like, decorative markup. So, left, right, corner, corner, one after another, here's the order. And the attributes of all this history. Well, in short, it's mainly a UI story. And there are, in fact, the texts of the module code, which are grouped entirely according to 1C. And you can adapt Shotgun for it, think about it. So, Shotgun can be adapted for it. There's a specific order, you just need to take it into account, in the right way. But I think it can be adapted. Therefore, uh, so, and I work with these modules. So, what do I do? I do pure chunking. At the stage of thinking through requirements, etc., I determine, as an architect, in this case, my role is completely architectural, from business analyst to architect, I determine which objects, which database objects, which metadata objects will be used in the process, right, so where we get data from, where we put data, how we process it. So, all this context is prepared not by a cursor and the like. They can't handle it. This is 4 million lines. I repeat, people are now trying to make RAK, but it's still not working very well. It's really three orders of magnitude easier to assemble by hand when you get the hang of it. Preparation of context, such context for requirements takes 5-10 minutes, and you generate thousands and thousands. Well, I mean the result. And these context describers are precisely auxiliary, auxiliary things that take, for example, I have a business entity. For example, I have a real combat topic - it's an advertising management center for marketplaces. So, in marketplaces, there is internal promotion, right. And, naturally, the management center is natural for me, natural for all of us, it may seem unnatural to you, but it's natural for the market that this advertising management center for marketplaces is located within 1C, it's part of it, it runs on 1C Enterprise, because you have here, it's not some separate product, it's useless to come to us with such things, it must be a deeply integrated product. So, you have all the real digital data under the hood, shipments, and they are collected from all marketplaces, from all of them. And you calculate a million, a million indicators from them, put everything together, calculate it, and based on this, you start blindly calling API methods. Turn this on, change this here, modify this here. That's it, then you analyze the statistics. Well, conceptually, that's how it all works. So, the concept of 1C is a special metadata object called a processing. And technically, it consists of several modules. So, a module that is responsible for UI UX, and a module that, uh, there can be many such modules. And a module that is responsible for the main function, while also pulling in additional modules, part of its functionality is moved to separate ones, which, so, all this context, to work with it, I collect it by hand, determine what goes into the business, and what doesn't. And the trick is specifically with any neural networks so far, that they poorly understand, they don't understand the damn token predictors at all. Not at all. But when you have this bunch of context laid out, where to apply which method correctly? And, I mean methods, both standard 1C libraries and custom-written ones, well, called functions, well, calculate this, calculate that, they need examples. And for this, I use these context describers. So, I generate based on this processing, this processing, I put it all in its entirety with all its stuff into the fifth, because it does it well, it's cheap. Here I will praise the fifth. It's cool. Well, in the sense, GPT5. I put it into coding mode, and it calmly codes. Well, this is technically an export, it's like folders, right. So, you just do a project export, open the folders and say, "Okay, go, go, go, go." Well, so it has a protocol. That's it, you give it an example, it calmly goes through all of it and makes a full description, which, uh, differs, uh, by, first of all, all attributes are used, in short, all data is described, the business purpose is written, the usage is written, the cases are written, because smart neural networks always need the business purpose. I'm not a fan of any memory banks. I don't think code should document itself. I believe that all reference information should be stored in the context of the object where it resides. Uh, I believe that data models should also be self-describing, right. So, all this information that speaks about their business meaning should also be stored with them. It's just convenient in the end. Therefore, I take and create this describer, it collects the context describer for me, and I use it further in my work. It is used in requirements, it is used in generation, it is used in testing. Then, colleagues, it does not lose the context, this set of context and its connections. Then it is not lost. I use these separate stitchers, describers. For 1C metadata objects, they can be stored permanently if you don't overwrite them. And for a task, for a project, well, they can be easily regenerated by the fifth or Gemini. I actually like Gemini too, it's just too flattering. You are brilliant. This is absolutely the solution. >> I don't look at it at all, right? >> You, well, your eye has become jaded, right? So, when you turn up the temperature, it's like, well, lower, it starts to get stupid, unfortunately. So, creativity, so, temperature one, it has creativity, well, you understand that you're still subconsciously looking through it, subconsciously it works. So, look, you are essentially assembling pieces of 1C code >> and data. And you prepare such requirements, where the network itself builds business connections, right? >> Yes. When it analyzes, it writes business connections, and we record them. So, it's very similar to your architect stage. I'm just saying, if you slightly, slightly adapt it for 1C, for our pipeline, it will be a gem. There are nuances. I've already thought about it technically. I've looked at your repo many times, in short, while >> discussed, >> yes, discussed. [laughter] So, now let's talk about requirements. Requirements look like this. For any more or less minor thing, I do this. What does the machine write itself? What's the difference? There's an example. There's the same. So, Gleb, look what I have. It must be there. So, I even break it down into stages, because it would take two, or even three people to work here. So, a business analyst would work here, an architect would work here, and a technical writer or, I don't know, a senior QA, who, yes, although usually business analysts prepare it in our field. So, what must be present? Business purpose must be present. I don't know why, but there are a million theories, in short, that I know. But the fact remains. Whatever code you generate, the neural network needs to know why you are doing it. It needs a grand goal. So, the grand goal and how we achieve it. That's it, you give it, it works. You don't give it, it works half the time. It just works randomly. Then, the second point, use case. The third point, test case. So, what is a use case? Usage. So, I even wrap it in tags, because I actively use semantics. For example, here I have one, two, three. Uh, how to use it? Here. So, by example, they are easy to generate. I can actually write it in a crude way without markup, just in words, I always do it that way. How will we use it? How will we check it? Test cases. The fourth description, which I talked about before, but here it's the result for the requirements. So, what specifically will be taken from the database, where it will be put, and so on. So, not a general description of the business process. Well, so, I might have 40 columns in a table.
Yes, there, for example, but I need two for this business process. Well, I, accordingly, these two blocks, that is, table such and such, uh, attributes are fields such and such, that's all, nothing more is needed. Next, test data. That is, this whole topic, they don't really talk about it properly, in my opinion. I don't even understand why. Based on my experience, not only with 1C, it turned out that if you, in the initial prompt, that is, a neural network technical specification is your initial prompt, yes, which you load and then start interacting with for generation, then for debugging, everything else. While you are preparing it, I am preparing test data back and forth. I noticed, I checked this not in one, not in two, and not in three cases, that a prompt that does not contain examples of test data produces much worse code than a prompt that contains test data. That is, if you physically, I show people, uh, shove something like, well, I trade on the marketplace, yes, you give an example, I have a smartphone, the price is 25,000, and its profitability is such and such. This is the price, well, its price is such and such, its profitability is such and such, its cost price is such and such, advertising costs are such and such, other expenses are such and such. And you give an example back, like, here, for example, I need to recalculate the price from profitability or profitability from price, right? You immediately provide verification data. And this goes into the first prompt. And when you had a real tester, a real tester sitting, he would look at this element with such and such data, it should be like this. Right, Gleb? And the machine looks exactly the same. And I, once again, I repeat, it is right before your eyes, if you don't give it such data, it, well, at least okay. And everything that concerns work, it's work with code data, our SQL-like and direct SQL, generally all, all that concerns data, it changes right before your eyes. I even looked in the chains of thought, in short, oh, the user indicated, we also have a lot of data divided by time, that is, like on the first day we sold so much, on the second day we sold so much or received. We have a lot of such data with time fields, periodic data. I also, as it were, feed them. Well, naturally, I anonymize and feed them. I, you know, how I explain it to myself. In fact, it's always like this, but you are a programmer, you write code, do you even look at the database schema sometimes, at least you have a brain, right, but you look at the database, right, I look. And what, the network doesn't need it, right? [laughter] Is it worse than you, or what? That is, or better, maybe? Ah, yes-yes-yes. I also often mix in, well, at least the schema. sometimes I mix in individual objects from the database into prompts. And if I need specific places, >> I mix in the data, it really works. When it starts, as they say, with ORMs, well, we don't have them, of course, mostly, it's SQL and SQL-like things, it rebuilds queries right before your eyes, it's easier for it to understand. Even in the chains of thought, it's visible how it, as it were, specifically test data. Like, on the first day you have this much, on the eighth day this much, and on the fifteenth day it should be this much. Why isn't it working for me? Well, let me recalculate. In short, it rebuilds the algorithm for them. I've said this more than once in the chains of thought. Well, the sixth is to-do lists, like everyone else. That is, as it were, for agents, for the next window, for that. Formed. And we work. Well, and as it were, two optional, but desirable points. These are ER diagrams of interaction, how it moves back and forth. Semantic markup of the technical specification itself. Like with XML-like tags, as they willed, as they are willed to us by Anthropic, as they are willed to us by GPT5. This is, well, in the sense, OpenAI wills in the last one. I've been using this focus for a long time, I advised you. So. That is, these tags work perfectly. That is, combined use with tags. You fix a section. I, for example, how do I break down code examples? I explain, I break down with tags, I name, in short, the tag uniquely, but which contains the semantics of what I am giving, well, like, the simplest thing is that the main data query tag, then the cost price data query. So. You mark it up, inside you can have Markdown, inside you can have plain code. Well, in the sense, if you write Markdown, as it were, I always mark up test data. That is, there is a classic, there is a classic, for example, if I work with goods, then item one, or data, then the first day, or then I also mark it up with tags, I build it. Feed it. That's it, I have a neural network technical specification. Based on it, you can print out - ask it to give it any form you need with the client, for example, or with an internal client, or with an external client, if you are an outsourcer, you give it a form, it draws a working technical specification based on this neural network technical specification, which you sign and go. And this neural network technical specification is fed into the generation stage. So, in short, we put this into the generation stage, and we encounter the following problems. The configurator is an invention of 2003. Let's go. What is it? In a nutshell. It's like a giant text field where you paste your configuration. Right or wrong? >> The configurator is a very specialized idea that shows you the current and auxiliary configurations if necessary, and the current database you are working with, or rather, the database and all 1C code, the entire 1C configuration. There is code, it demonstrates all 1C metadata objects that have been invented there, directories, then processing, documents, registers, registers, then various http services and similar things, then scheduled tasks. So. And, accordingly, its key moment is, as it were, a text field where you have code, right, and you, accordingly, paste it back and forth. What's the problem? You transfer it there manually. That is, nothing connects to the configurator in any way. It's a monolithic application from 2003. What? What? What? What? I don't know how else to explain it. It's not integrated with anything, in short. And it won't be >> some auto-clickers. It's a desktop, right? It's a desktop. >> Of course, of course. What cloud services? Well, there's also a Mac build. It works. >> And is there one for Linux >> configurator? I don't remember. But, in principle, we can make an auto-clicker, probably with machine vision, which will run around these fields and forms. >> No, now, now they are going to approach the forms from a different angle. That is, there is a batch mode in the configurator. That is, you just load, unload files, and you can work with files in Visual Studio. Visual no. The enterprise language markup was added to GitHub, God grant me memory, in 2017. Visual Studio knows it perfectly, marks it up and displays it. And Claude perfectly understands the 1C format as is. You just upload it, in short, and that's it, it worked. It knows everything. In short, GPT5 doesn't know GPT5, but you have to explain. So it says: "What kind of file is BS?" You say: "Uh-huh, they didn't teach you at the index." Claude was taught. Everything that is in the index, I understand, it knows everything. You just put the file there and that's it. Well, it knows everything, where to look, where, as it were, formats, etc., etc. Ah, well, I'm generally not a fan of working with neural networks in IDEs, I just transfer code and everything else there. That is, I don't look at it much at the generation stage. It should work on its own. It throws out errors, you give it the errors, let it fix them itself. My opinion, my opinion is this. That is, >> well, I generally agree overall. to see what it has generated again. If it doesn't fall, it should pass scenario testing, it should first pass scenario testing, that it fully meets the customer's requirements, which is that. And then I will launch optimization with separate prompts, how to say it, when it's ready, if I want. I'll launch optimization, I'll ask where my weak spots are. Let's look, let's improve this. So. But first, it should all work, uh, as planned. Speed is prioritized over debugging technical debt and the like. Accordingly, yes, that is, I am currently, well, I am pleasantly pleased with agents, that is, IDE code, and that. I can't say that my entire pipeline is currently fulfilled by them due to the specifics of 1C, yes, all these things, but to the maximum I am now trying to use them. That is, describers, that is, analyzed, wrote. In the case of Claude code. My code doesn't work with 1C code. Ah, it only works with describers, yes, with analysis. Code, uh, Claude-code makes changes if necessary. I then load or transfer manually, because if there are many changes, the configurator can also load them for several minutes. It's easier, easier to do it yourself, easier to copy, paste, especially when, you know, you generate large modules, in short. Well, and then we check all this. Logs. It's very convenient to do semantic markup in logs. Very convenient. That is, first of all, on the neural network immediately, when it sees your code, it builds a tree, it already understands its execution. And when you give it the practical results of the algorithm's execution, with tags again, then it clearly sees, here is initialization. For example, here is initialization, initialization. And then, for example, the number 500 arrived, but 400 was supposed to arrive. And when you have a log marked up in this way, first of all, what's useful? You don't fall into this coding pitfall that still exists. This is a dead loop. This is when you do something, there are errors. You return it back, it corrects it, there are other errors. You return it back, and it returns to the first, to the essence of the code. And the errors that were there reappear. You start, in short, executing them. Although this cycle repeats again. This is a very common classic, and as I understand it, Gleb, it still exists. This classic has almost completely disappeared for me, with the use of describers and glueers, so that it always knows the context, where and when. And semantic markup of logs, that is, logs, that is, scenarios of execution of the entire branch. >> Well, you probably do that too. You do, here's your technical specification, here's what we did in the first run. Here are the logs, what happened. Here's the problem. Fix it. >> For this, your too. In the sense, shotgun is also good. Well >> I do it this way. >> I do it this way too. That is, yes, exactly. That is, here's the technical specification, which was the initial one, here's the code that has already been done, here are the current problems, go ahead and fix it. By the way, any not very large bugs during debugging, it just doesn't need to explain anything else. That is, here's the code, here's that. Here's the error, fix it. That's all. And in the case of agents and agent behavior, markup is also very convenient in our 1C. Well, it's also great for other code in this regard. When the agent sees it, it tries to change only that piece that goes from one branch to another. That is, it doesn't try to rewrite the entire function, it doesn't try. We have large functions, procedures, we have large listings. That is, it's normal for us when, for example, I have a working, there is, in short, a working function, not just a query 325 struct, to be clear. >> And there's still some post-processing of it. >> The query itself, that query, it's assembled, so that you understand, a lot of temporary tables are built, it's assembled from different places. It calculates Cartesian products internally, does a lot of other things, all this is packaged, given to the server from the client. Well, to work with it further. So. And this query alone of all this history is 325 lines. Logs cannot be put into the query itself, but in terms of further execution, logs are ultra-useful in this regard, because when it sees that something went wrong, it intervenes precisely in the block where it went wrong. It doesn't go into the code to rewrite the entire function, it doesn't go into rewriting part of the function, it goes first of all into that block. And often I had, I had situations with debugging, when the machine, as it were, looks at the algorithm and says: "Listen, we, in short, put the block in the wrong place. Here, in short, we need to cut it out here and move it here, and everything will work." Well, and it worked, naturally, right? So this is a very useful thing, in short. >> And here's another question, you've really told some hardcore things. Let's talk about absolute beginners. For them. 1C developers don't have Git. What do they have? They just have folders and files, right? >> 1C developers have Git. And in EDT there is Git, uh, 1C developers themselves have Git, that is, people export to Visual Studio to work with it. But it's not built-in by default and is not always used by everyone. If we talk about corporate development, then 1C has, of course, since the 2000s, its own version control system, which is worse than Git. >> They have something based on SVN, I think, right? >> Well, yes, SVN-like, in short. That is, there is a repository, that's it, and you pull data back and forth. It is intended for you to all work with a large configuration, but everyone, but each programmer with their own modules. That is, control of storing all this history, versioning, but not in such a way as to merge >> but without merges, >> yes, >> yes, without merges, without all this history. That is, if you want, then there is a standard export in the configurator or batch, there are scripts on GitHub, people have described all this long ago, you run it, export it, then it collects it back for you. That is, you can work with 1C code in Visual Studio. Only with code. You can't work with forms. Therefore, if you need to configure forms, you still go to the configurator and start dragging and dropping there. >> So, in fact, as I imagine it, you go to the 1C website as a developer, register, download the platform, and don't try to build with various things, you work in the configurator, and you write files in the configurator using neural networks, and as Peter will tell you in his course, right? And then, well, that's what you learn. So. And then you delve deeper and deeper. >> So, a little about me, my channel, which I run, and so that it's clear that I'm not just talking about sales. That is, it so happened that besides, well, besides programming and consulting itself, I also sold a lot on the side. That is, when you consult well, they buy from you, you understand, Gleb, right? I don't know, [laughter] I don't know, I've sold at least a hundred copies. My maximum bonus for those who were involved in sales, they will understand my level. Well, no, there are people who sell software for millions of dollars and who have, as it were, large checks. I can't do that. I haven't had the chance. At the small and medium business level, yes, please, I know, if anyone wants to try to interact with 1C, with the business part, I've tried to explain everything. From the technical side, Gleb has just explained, that is, we have the 1C website, there is registration, there is an educational version. And from the educational version, there are demo databases, God grant me memory, some kind of, where all this can be seen and deployed. By default, it's file-based. Well, in the sense, it's a file format and it's similar to SQLite, but different, of course. Everything is in the traditions of 1C, everything is its own. Yes, a very important point. 1C has its own. 1C has its own. We, as they say, don't use external tools much. The workers only write tools for themselves in 1C Enterprise, and work with it. External, completely external tools are actually rare in our infrastructure. Except, well, Git is an exception, in corporate development, that is, at the SVN level, few people use it. If we take large Russian-speaking companies, then Git will be there with a probability of 80 percent. >> I think this is actually a growth area for the ecosystem. If people familiar with other paradigms start coming in, will they bring something? You know, this discussion is from eight or nine years ago. I think it will change everything significantly for us now. Very much so. Well, and in connection with AI >> so, it will change everything so much that I, in short, wouldn't look back much, you know, even at the level of purely, how to say it, one of the key paradigms. I first started programming in school in 1995. And my teacher told me, so and so, look, one of the key rules is code reuse. I grew up with it. And now, that is, various libraries, reuse, that is, you are used to it, you try to write yourself, you study some libraries, framework approaches, you pull, and now you are in a situation where, why did I used to write 1000 lines, right? >> Sometimes it's easier to write now >> on the spot, and now it's easier to write on the spot. Often it's just easier to write on the spot. It's just there, it's exactly what's being executed with these signatures, and that's it, and don't integrate anymore and forget. And then write another version next to it, almost very similar, very similar, but with some differences. And don't forget, in short, about its reuse and don't call old calls, in short, don't draw any wrappers, just put it, right? So this is like >> the OOP guys are rolling over in their graves somewhere. >> It's bad form. Everyone is rolling over there. But when your function, for example, is redone, generated in a few minutes, then you don't need to delve into it, you don't need to redo it, it won't take you half a day, for example, to integrate correctly, and when you just do it, here, do it on the spot, it needs to be redone like this, hop, and it works. If the signature No, well, I try to reuse, of course. That is, if the signature doesn't need to be changed, then of course I integrate and redo the existing one, yes. But if, God forbid, the signature needs to be changed, then there will be a second version, because it will be literally in a couple of minutes. These practices, they, for example, I have a deep, deep conviction, the more I work and further, based on the glitches that 1C gives me, well, you understand, 1C is a language that is ranked below 100th place in the TIOBE index. That is, in TIOBE, when I last looked, it was something like 186 or 187 languages, it's somewhere below, somewhere there, closer to. And there was no place to get datasets, by and large. There isn't that much open code in this sphere, especially on GitHub. That is, there are some specific places like infostart where you can get it. But still, it's hard to get it anywhere. Of course, Anthropic are pirates in this regard, it's already proven by court. I'm sure they just unpacked some existing configurations and uploaded everything to their datasets. I'm personally sure. But, of course, it's not proven. Well, how do you prove it? In any case, the same Claude, the same Google, they write perfectly. And from the errors and hallucinations they sometimes make, I understand that in fact they don't write in 1C, and not even in C#, not in anything like that. They write in something of their own. They have some common, their own common programming semantics, and they substitute tokens for you in the global sense, they substitute tokens for you in accordance with what you said: "Well, I need the 1C language, or do you need the C# language?" >> If we, as it were, delve a little into this topic, but we won't go deep, I'll just throw in, I'll throw in that with the help of neural networks, you can write translators from one language to another very quickly. Yes, I'm trying. >> And you can, if you talk to, say, ChatGPT on this topic, it will offer you, it will say: "Let's write in YAML, and we'll also write a translator for your language, 1C, everything will work great." I think, as an option, we'll come to this, or something similar, >> yes? They somehow, these programming languages, as such, don't exist. They substitute some programming construct on the spot, all these approaches. >> Tell us a little, you created your course, who is it for? For beginners or for advanced? For whom? >> It's not for beginners now. That is, literally three weeks ago, I launched it, they've been asking me for a long time, in short, I have a channel, yes, I opened it not so long ago, in fact. How did it happen? I sat for six months, in short, I was just afraid. I understand that this is a terrible revolution, but it will sweep everything away, in short, everything. I looked in very narrow, very narrow circles, communicated, but I was afraid to talk about it to people, in general. Then, at one point, I stopped being afraid and decided to create my own channel. That's it, people started subscribing, colleagues started subscribing, yes, I post prompts on the channel, how it all works, examples, etc., etc. Just yesterday, Quentin was testing, and I started, literally, I don't know, the channel existed for a week, they started asking me: "Peter, are you going to organize courses, when will you teach? Let me sign up for you, etc., etc." I, in short, also walked around for two weeks, walked around, walked around. As it were, in fact, this is also scary. >> I also tell you: "Do courses." >> Well, yes, and you too, you too, as it were. Well, I, in short, assembled a test group. Now I'm working with the group further. What's the key moment that I'm getting now, that everything I've shown today is reproducible results for others. It's not that I'm saying something alone, that these are some things that I've... Well, for colleagues, of course, on the courses I tell even more, many times, an order of magnitude more, there are a lot of examples, etc., but these are all working tools. Specifically, they give results. People, well, people really like it. Only seniors are coming to me now, or, well, conditionally, well, in short, everyone with experience comes, that is, with great experience. My first group was formed from heads of departments, so that you understand. Ah, so no ordinary programmers came to me at all. Only team leads, heads of IT departments, who still write manually. Well, as it were, so as not to lose, as it were, so that people understand what to teach their own and so on. I can even say more, specifically in 1C now, in short, the maximum result is only for seniors, only for seniors or at least middle plus programmers. That is, it will be very difficult for a complete beginner. To get the maximum performance that is achieved, you need to know a lot. Moreover, well, you know this yourself. When you do AI coding, you become a small CTO, in short, everyone around you runs, agents, not agents, you, in short, play any role. And for this, colleagues, you need to have the moral readiness for this in principle. >> A broad outlook is important. >> A broad outlook, yes, a broad outlook is important. And I think this is the biggest challenge that currently exists. When you have worked in one role for 20 years, approximately, and your life starts demanding that you become a jack-of-all-trades. And now it demands that you become a jack-of-all-trades. It will help you everywhere, in everything. You have agents, but you still have to be a jack-of-all-trades. Such a shape, right, a very wide stick. So. And this is actually very difficult. I think that this restructuring is very difficult, but when you are a beginner, you don't have any of this. And it turns out, so I don't know how we will solve the problem of juniors, you know, about which they talk. That is, a senior, well, as it were, you have a T, you need to get a broad outlook, you need to delve into something, and now we have such a wide and broad spread. That is, for you to produce results, I need to know a lot in principle, in terms of broad outlook, know how to ask, at least, if you don't know, right? This is completely different. These are completely different requirements. These are completely different models, completely different behavioral patterns. I don't know how we will live with this, by and large. Now. Why it's only for seniors, because they, over their 15-20 years, have just had a wider stick, they've just seen more. And a twenty-year-old guy comes, for example, who has such a crossbar, I don't know what to do, I don't know what to tell him. It's clear that he can, instead of, as it were, go deep, but, I'm saying, it's not just a stick, there are also, you know, from it, as it were, hanging down sticks. It will look more like a power line support than the letter T. Well, we'll see. In short, if you are a senior, if you want to get the most out of 1C development, what kind of acceleration? 5-10 times, right? >> Well, realistically. Well, three times. >> Well, maybe more. [laughter] Well, three is already good, right? Instead of three developers, one, instead of three, thirteen. Well, and so on. >> Let's say, let's say, from 30%, if there are fewer approvals and everything. Well, in the sense, if you have a lot of calls, approvals, etc. in your work, then 30%, 40% acceleration. In general, if you have a dedicated role to write code and you have little interaction with these things, then x10 is possible. That is, from 30, from 30% you can safely say. This is, as it were, already the practice of those who studied with me, 30-40, again, within departments, if you're doing something yourself, building your own bicycle, then a pet project, or a startup that is already fully functional to sell, then the productivity is much higher. Well, or maybe you're a freelancer with small clients, they, >> yes, if you are a focused freelancer, >> a task that used to take a week, now takes one day. Well, something like that. >> Well, you understand, you won't earn much more money from this, >> because everyone will start doing it. [laughter] >> No, no, no. Now a very cool moment arises. You won't earn much more from this, because you need to sell. Selling is a certain equilibrium state of the market, which you disrupt, you can't just disrupt it, but your work-life balance shifts towards life very strongly. That is, if before you had, say, for a program, well, let's say, I was such a freelancer for many years, by the way, not only in the network, not only in corporations, but also a freelancer was good >> in my hometown. And I can tell you, you have a work week, you spend, say, 20 hours programming or consulting, and the remaining 20 hours you interact with clients in some way, that is, you have calls, you sort things out, well, all these interactions. And now, if these calls, sorting things out, don't decrease drastically, then these 20 hours that were programming can easily turn into five, and you get an additional 15 hours a week for life. Just for life, not for work. The money remains the same, and you do all the work on time, everyone is happy. It's not just that you're churning out code. Listen, for specific people, for specific needs, >> no, specific tasks, and they also interact. Deadlines are different for everyone. You save time. This, you can rest more, you can rest more, you will be better. You will pursue hobbies. I used to sit and stare at the screen for 15 hours, checking one thing, then another. Well, that's a normal hobby, an IT hobby for everyone. Well, for many. So, as long as it's here and now, it's primarily about work. >> Then what, are we finishing? Subscribe to Peter's channel. Very interesting, life-related topics. 1C is the blood and flesh, yes, of the industry, a truly hardcore one. Be interested in what's happening. >> Goodbye everyone. >> Goodbye everyone. M.