Do not look for a neural network – you already have it in your team

Do not look for a neural network – you already have it in your team

Frankly, I did not think that companies would grab neural networks: firstly, the strategy of client-centricity immediately crumbles, and secondly, trust in automation seems very strange. Well, that is, since the 70s, the means of automation in business inevitably aroused suspicion, and here suddenly something contributes so that some are ready to make cuts in marketing, sales, and even in development and testing. Productivity, economy, efficiency, and all kinds of intense growth seem to loom behind the neural network. And in fact, something happened to us that was repeated in history more than once or twice. Welcome to another cargo cult!

Neural networks can do everything

This will be our first article with an application – in the end you can see what requests and how the neural network fulfilled, and I, as a manager, evaluated it. To be honest, I would not want such an employee. On the one hand, she fulfilled almost all tasks for a marketer, copywriter and salesperson, but the quality and approach leave much to be desired. The reason is simple: a neural network learns from big data. The biggest data is network content (globally, texts, code, pictures, and videos). Most of the content you have to see is standard texts that have been copied from one to another for years. Accordingly, neural networks will give you something very averaged, often with factual errors.

At the same time, every company has a cool and powerful neural network – its employees, team. These are people who once passed an interview, adapted, and learned. Some of them are more hardworking, some are more talented, some are slow, but all of them are inside the company, they understand its goals, advantages, work rules, from the brand book to the code style. And it is they who are able to perform tasks exactly as expected by the manager and, most importantly, by clients and users. Employees know who they work for and understand the needs and problems to be solved with a service, software, engineering solution, etc. This is a people-oriented approach, according to which we do not push something in the middle to users, but do what we are ready to use ourselves. (For whom, of course, but, for example, we use all opportunities our developments and every employee works in them every day – by the way, this gives additional advantages for working with the backlog and with testing).

And you go explain!

Working with a neural network to illustrate this article, I had to use several dozen messages, some of which were unsuccessful: a prompt for a neural network should be as specific, precise, and verified as possible. At the same time, you need to try to take into account all the fine points and details. That is, if I need a short newsletter text, I write it myself or set a task like “write an announcement for a spring sale” and it goes quickly. Compiling an adequate prompt for a more or less acceptable result takes the same amount of time, but the generated text still needs to be honed and adapted. A kind of processing is obtained. Despite the fact that a large format cannot be trusted in a neural network in principle.

Now about the program code. I’ve been writing code since I was 15 years old – I went through Assembler, C/C++, Java, PHP, Delphi, Kotlin and much more, including 1C. And I would never trust the writing of even small fragments to a neuron, even knowing that it is essentially a neuron works like my some programmers issues code with open repositories in mind. Again, it’s a waste of time: why disassemble and refactor someone else’s code, adapt it to the code style adopted by the company, suffer with compatibility, when you can think with your human head (among other things, an exemplary neural network) in less time and immediately write a normal, almost working code for which you are responsible and in which you are confident.

The neural network is not part of the team

Whatever the neural network invents, it must be accepted by its main users – employees. For some, it may be a relief to work and to be happy, but if there are enough adequate and ambitious people working in the company, it is important for them that their work and their activity products are accepted. Contribution to a common cause is an important moral aspect of motivation. Neuronets are not empathetic, are not able to follow ethical standards (if they do not include it in the algorithm) and can be very wrong. It is unlikely that anyone will be very happy with such a colleague 🙂

In general, a neural network is a tool such as a CRM, ERP, PM, task management system, or accounting program. By the way, here is a good example for you. First accountants were supposed to be freed by calculators – but no, overtrained accountants who could handle electronics were needed, and then Citizen. Then they were definitely supposed to be freed by Excel — but no, the market began to demand accountants who are good at Excel, VBA, and macros. A little later, the main liberator of Russian accountants appeared – 1C. Well, almost all Habra readers know the bottom line here: demand has increased… The same story with artificial intelligence: individual freak companies and crazy experimenters will agree to a logo from Mykola Ironov, others need designers who will work faster and more efficiently with the help of AI; texts from neural networks will attract only lazy people or trolls to sites or Habr, but a copywriter who picked up the idea in the generated text and wrote a unique text will cost more. And so on. We are simply entering an era of popularity of a new tool—the usual technological evolution.

A neural network has no intuition

Intuition is an important property of any person, which is formed in the presence of experience, supervision, understanding of the context and “background” analysis of various factors. Intuition, as a rule, is present in all people – to them it seems a little illogical, supernatural, in fact it is only the final stage of a complex thinking process. Thanks to intuition, interesting features in a software product, good advertising campaigns, successful sales scripts and much more are created. We, people, analyze information, understand it, but at the same time we are also aware of a lot of factors of huge historical embeddedness – until we can stuff all this into an artificial neural network, if only because our mechanisms have not been fully studied.

A neural network can pass the Turing test or the “towel test”, but at the same time it will not be able to think, have empathy and make the right decisions as quickly as possible. Moreover, the required correct decision may be an exception to the algorithm and then the AI ​​may cause harm. And fine if it’s an old form invoice or a stupid text about a new phone, but if it’s the car’s autopilot? We are not yet able to solve these problems (although, fortunately, we know about them).

Blindly following technical fads can one day just ruin everything the team has been building for years. In addition, the customization and adaptation of the neural network for specific tasks of the company also requires resources, forces, time, and professional developers. Otherwise, it will be a very simple toy.

Below I present the problems and answers of the neural network with my comments. I did not delve into the depths of development and force the neural network to write code, but set it very simple, trivial tasks that my colleagues and I perform literally on the fly.

Neuronet, work for us!

The text itself is bad: tautologies, repetitions of semantic blocks, questions in the text. But that’s half the battle. For example, the neural network offers to familiarize yourself with our portfolio and see successful projects. Purely formally, it’s cool that she took this into account, but… we don’t have a portfolio section and don’t talk about our projects under NDA in particular. Where did this part come from? But from a bunch of landing pages with similar texts!

It is obvious that the text contains typical features of CRM systems. Here’s a good example of when a staff member’s head would work better. I need ideas for new features for a specific RegionSoft CRM, and even if I list all the existing ones in a prompt, it won’t help much either, because it would be nice to show customer requirements, analytics on the most used features, etc. . It is much easier, faster and, most importantly, more effective to gather employees for a brainstorming session and generate a bunch of different ideas from which to choose 1-2 that can be put to work.

Some names are simply illogical, you can use two at a time, and they do not pretend to be any novelty or creativity. I suspect that the idea of ​​the name “March youth” has its roots in the numerous “let’s return youth by March 8” salon promotions.

I will not say that this is the most clear and simple definition (again, not without grammatical problems). Although generally almost canonical. There are many factual errors in the definition that can be corrected.

Everything is banal: almost any client sees these reasons for using CRM in the first minutes after typing a question about the system into a search engine. They are simply not suitable for solving the task. But it’s not that interesting and we will clarify the prompt – we will ask for non-trivial motives.

It could be ironic that the security argument is non-trivial for most CRMs, the neural network clearly knows something 😉 And yes – again a rather simple and well-worn set.

The response from technical support is formally good, but again — vague and can irritate rather than reassure. Unfortunately, you won’t help matters with words. But in fact, of course, nothing concrete: offering an overview instead of pointing to the point of the problem and a quick solution is not the right option. No solution is offered, but at the end of the II, it is clarified whether it is suitable. However, according to experience, more than half of the technical supports answer roughly like that 🙂

There are almost no comments here: in some places, apparently, our own announcements were simply spurred. There are questions about small things, but this is probably the most successful task (although some Direct ads are not an example, they need to be corrected).

This task is failed, but with a fire: you can laugh from the bottom of your heart. I just can’t imagine what will happen if the managers really take the proposed options into service. Even if you have to hear this on the phone – so what, the bots don’t blush.

Everything is simple here: the neural network does not understand what a commercial offer is and how it is made. And the text itself simply does not stand up to criticism in terms of the level and number of stamps.

Please write shorter, because the text in the commercial offer is not the main thing. Well, this is already a complete failure. Even the most backward companies do not use this.

We make clarifications and give another chance for another text about nothing. That is, we need to explain to the neural network that such a commercial proposal is unlikely to be successful unless we write a special application for creating texts for KP.


Conclusions that are coming out so far.

  1. Working with prompts for generative neural networks is a separate complex task that takes a lot of time and requires consideration of many factors. It seems that it is faster to come up with the head.

  2. In the generated texts, there are many factual flaws that sound “in line”, but at the same time spoil the impression of any professional.

  3. Commands are taken too literally (which is to be expected).

  4. The texts contain many templates and stamps that you do not want to see in any corporate content.

Now it is difficult to say whether the future is in neural networks or not. Rather, in this direction, the future, as before, depends on people who design and write algorithms, take into account exceptions, train neural networks, fight artifacts and hallucinations. It is early to use neural networks in production — and it seems that users will tire of such texts quite quickly, and the value of experienced leather copywriters will increase. After all, language is a process that is difficult to put into algorithms. Of course, if it is a literate, beautiful language of an educated person.

Oleksiy Surikov

Chief developer RegionSoft

Related posts