Why I Am Writing This Now
I think the time has finally come for me to write about artificial intelligence.
I want to talk about AI itself, how I use it, how my understanding of it has changed, and how I see what is now happening in the world.
I also want to explain why I decided to write about it specifically now.
At first, I thought I could fit all my thoughts into one article. But there was simply too much to say, so I decided to divide the story into several parts.
I want to describe where we were, where we are now, where we are going and, finally, where we may eventually arrive.
Until recently, I was not confident enough to make predictions.
I understood that many things would change. That part was already clear. But the exact lines and shapes of those changes were still difficult to see.
In 2026, something changed in my own practical experience.
For the first time, I felt that enough progress had happened for me to stop looking only at separate events and begin talking about a direction.
To Draw a Line, You Need Two Points
To build any forecast, whether it is technological, economic or something else, you need at least two points.
The first point represents where you were.
The second point represents where you have arrived.
Once you have both points, you can connect them and see a line. That line gives you a chance to estimate where the third and fourth points may appear.
With only one point, it is extremely difficult to understand where you are moving.
That was my problem for more than three years.
I had one point in my head, but I kept asking:
Where is the other one?
I needed something I could connect to the first point. Only then could I begin to understand what was realistic, what was unrealistic and which direction AI was actually taking.
By 2026, my own experience finally told me that the second point had appeared.
There had been enough meaningful progress to call it an intermediate checkpoint. And once the checkpoint became visible, it became possible to draw a line and begin making a forecast.
This article therefore begins with the first point: where we were.
Then it moves to the second point: where we are now.
From there, I will try to describe a possible third point, mainly in the IT industry, because that is the field I know best.
Later, I want to move further and discuss the fourth point: not only where the IT industry may go, but where humanity as a whole may eventually arrive.
That final point will be much less precise. It may describe a period that we will not see during our own lives. Some events may happen earlier, some later, and some may take a different form.
But even when the exact timing is uncertain, the direction can still be worth discussing.
The First Dot
For me, the first point appeared in 2022.
That was when I first encountered ChatGPT.
My very first reaction was something like:
This must be a prepared answer.
Alongside that thought, I also had another one:
This cannot really work.
I understood the general idea behind predicting the next word in a sentence. The approach itself seemed relatively clear.
You take a sequence of words and try to predict which word is most likely to come next.
The question was whether this could work well enough to create a real conversation.
For example, in the phrase “What is the capital of France?”, the answer “Paris” has a very high probability. That connection is understandable.
But imagine that the sentence contains only:
“What is the capital…”
How can a machine know the next word?
It could be the capital of France, Germany, Poland or any other country.
As it turned out, the model did not need to guess a country immediately. It could continue with:
“Please clarify which country you mean.”
That was the part that surprised me.
It turned out that the amount of existing human knowledge and written language was large enough to train a model not only to predict isolated words, but to build meaningful sentences and maintain a conversation.
Machine-learning models themselves were not new to me.
I had encountered them around 2013. They were already being used by companies such as Google and Yandex for image recognition and similar tasks.
Even then, a system could analyse an image and identify a car, a shirt, a pattern or another object.
That was already impressive.
One of the main human ways of receiving information about the world—vision—was already being partially automated.
A camera could look at a room and identify that it was looking at a kitchen. Based on that information, another system could make decisions about movement or actions.
So I did not think AI itself was something completely new or magical.
What I found difficult to believe was that language could be processed well enough to create a natural conversation.
Human language is complicated. Languages are different. The same word can have different meanings depending on context.
If I say “Hello”, the next likely word may also be “hello”. But what comes after that?
The system has to wait for another instruction, understand more context and then continue predicting word after word.
What the team behind ChatGPT achieved was something that many others probably did not fully believe was practical.
They believed in the idea strongly enough to take it to the end.
My First Experience with ChatGPT
Once I started using ChatGPT, I tried to turn it into a good junior developer.
It did not work.
I made several attempts, usually with intervals of around six months. Each time, I tested whether it could take a task, understand it and produce a reliable result.
At that stage, my conclusion was that ChatGPT was not much more than a very well-indexed version of the Internet.
It could understand questions much more broadly and precisely than a traditional search engine. It could also present an answer in the exact form that I requested.
That was already very useful, but it was not yet a junior developer.
It helped me with code, but only to a limited degree.
I would estimate that it saved me around ten per cent of my programming time during that first period.
The larger effect was somewhere else.
It helped with learning, personal development and quickly entering areas where I previously lacked knowledge.
That was the first point.
A New Form of Search
Within roughly six months, ChatGPT had almost completely replaced Google for me.
I stopped opening Google for normal informational searches.
Today, I mainly use it when I need to find one specific page or website, partly because it remains the default search engine in most browsers.
Years ago, Google displayed the approximate number of results for a query.
You could search for something like “Titanic” and see that Google had found hundreds of thousands or millions of pages.
I once read that these numbers were not necessarily precise. Perhaps they were estimates, or perhaps the system counted a smaller number and extrapolated it.
The exact technical mechanism is not important here.
What mattered was the psychological effect.
You saw a huge number and felt that Google had access to an almost unlimited amount of information.
But after moving to ChatGPT, I noticed something else.
When I occasionally returned to Google, I rarely went beyond the first page. In many cases, even the first-page results felt irrelevant.
It seemed that Google had not seriously cared about the second or third page for many years.
Traditional search in its existing form is dying.
I believe that with almost complete certainty.
It is simply inconvenient.
In the past, opening several websites was normal and relatively efficient.
Today, you open one website and find part of the information. Then you open another website and find a different part. Each website presents information differently.
You have to find the relevant paragraph, understand the page structure, close pop-ups, ignore advertising and constantly adapt to another format.
All of this consumes mental energy.
Most people do not want to spend that energy.
When you ask a language model, it works with aggregated and indexed information. It tries to understand what you actually mean, provides a direct answer and can ask clarifying questions.
The interface is universal.
You do not need to adapt to a new website every time.
I therefore began to see ChatGPT as a new form of search and a new form of indexing information.
That was how I treated it during the first stage.
I could not build a reliable junior developer from it. It saved some programming time, but its biggest value was as a tool for searching, explaining and learning.
That is where we were.
The Second Dot
Now we arrive at the second part of the story.
This is where the forecast begins.
Having considered the first point and looked at where the technology has arrived, we can start asking what a possible third point may look like.
My forecast will concentrate mainly on the IT industry because it is the field closest to me and the one in which I can examine the changes in more detail.
Before making a forecast, however, we need to understand how software development works today.
How Software Development Works Today
Software development can be divided into several major layers.
The first layer is responsible for ideas and requirements.
People analyse a problem, discuss it, document it and prepare tasks.
The second layer is the development itself.
Developers take those requirements and turn them into working software.
There is also a testing and acceptance layer, of course, but for this discussion, the first two are the most important.
Ideas and requirements are created in the first layer and transferred to the second. Developers then interpret those requirements, make technical decisions and write code.
Around five years ago, while ChatGPT was still emerging, there were already predictions from large companies that software development would soon work through a single command.
You would say:
“Add this feature.”
And the system would simply create it.
We have now reached that predicted future, and this has not happened.
What has happened is different.
AI-assisted editors can already write significant amounts of code. In many cases, that code works.
There is clearly something real and valuable here.
Developers will successfully adopt these tools. The main factor slowing adoption today is company security policy.
But those policies will gradually change.
When companies see how much productivity can increase, they will have to find controlled ways of allowing these tools.
AI will therefore enter almost every software company.
The larger question is how developers and companies will use it.
During the first stage, developers may use AI mainly to improve their own efficiency.
A person who previously needed six hours for a task may complete it in four. Later, perhaps in two.
This may even lead to more people working for two or three companies at the same time, trying to balance several positions.
A developer may finish the real work in a few hours while spending the rest of the day attending meetings.
Companies will not like that.
They will naturally try to convert the gained productivity into more output for the company.
But that creates a motivation problem.
When someone finds a tool that makes their work easier, they do not automatically go to their manager and ask for a second task.
People may also be afraid that by helping a company introduce AI, they are helping build the system that may later remove their own position.
They will not always participate enthusiastically.
This may explain some failed attempts by companies to replace or reorganise teams around AI.
The remaining employees may not have been sufficiently motivated to support the change.
Eventually, companies will find a compromise.
Processes will be created. Responsibilities will become clearer. The dust will settle, and companies will understand which specialists they need and which skills those specialists should have.
Some people may remain in well-paid positions inside large organisations.
Others may move into independent work and serve several companies.
Why One Phrase Will Not Build a Serious Product
To understand what those future roles may look like, we first need to discuss why serious software will not be created from one sentence.
Imagine that you have a machine capable of producing any car.
You tell it:
“Make me a car.”
It creates a BMW.
Assume that one day we can arrange physical materials and molecules in almost any required form. Technically, the machine may be capable of producing a BMW, an Audi or even a Rolls-Royce.
At first, that sounds perfect.
The weaker, cheaper brands may disappear, while the strongest products remain.
But what happens after you have driven the same BMW for two years?
Even a person who loves BMW may eventually say:
“I am tired of this. I want something different.”
That is human nature.
Most people do not want to follow exactly the same road forever.
You may ask the system for an Audi. Later, you may become tired of that too.
Eventually, you will ask for something unique.
Perhaps you wake up on a rainy morning and say:
“Make me a sad BMW.”
The system may lower the headlights and make the car look slightly unhappy.
Technically, it will satisfy the request.
But it will probably produce an average interpretation of what a “sad BMW” should look like.
You may respond:
“No, not that sad. Raise the headlights slightly.”
Then something else becomes wrong.
You correct one part and damage another.
It becomes like editing a plasticine toy: you try to improve one detail and gradually deform the entire object.
This is why a single phrase will not produce a serious product.
It may create a standard website, a simple shop or a digital business card.
Such products will exist, and people will use them.
But serious products require uniqueness, careful decisions and deep knowledge.
AI tends to move towards an average result because it is trained on existing examples.
A human being is still needed to push that result beyond the average.
Good art, good design and good products often appear through a concentration of experience, taste, accidental inspiration and personal vision.
A person may sit somewhere, hear a bird singing and suddenly imagine a melody.
That is how something genuinely new can appear.
AI-generated products may reach a high general quality, but uniqueness will remain valuable.
People will continue to pay for it.
AI Is the Engine, Not the Whole Car
The same principle applies to software.
A good application requires much more than code.
It requires requirements, architecture, integration, configuration, testing, deployment, monitoring and ongoing support.
I saw this clearly while working with a project related to robotics.
I will not name the company or describe its size.
At first glance, the robot’s main foundation appeared to be artificial intelligence.
And technically, that was true.
But when I saw how the complete company and product worked in reality, I realised that the AI component represented perhaps ten or fifteen per cent of the total system.
AI was the engine.
But a car contains much more than an engine.
The other eighty-five or ninety per cent consisted of management, configuration, integration, tools, testing, debugging and operational processes.
All of those parts required people and knowledge.
My own work on that project was not directly related to the robot’s AI model.
It was related to management and configuration.
Those functions required ordinary software tools, but the tools were still complex, diverse and constantly changing.
They had to be maintained.
And this was only one part of a much larger system.
Large company projects will continue to grow into a zoo of connected programs, technologies, internal systems and satellite products.
AI may create a small website with one command.
It may create a standard online shop.
But complex configurable products become difficult from the moment you start describing the task.
Every new button adds more possible user actions.
A user can click one button before another, use a function with or without a particular option and combine different paths.
The number of possibilities constantly grows.
You cannot simply say:
“Add this feature.”
You immediately receive questions:
How should it behave?
What happens in this situation?
Why should it work that way?
Would another approach be more convenient?
Answering those questions takes time.
That is why, in my view, AI will assist people extremely well—perhaps by fifty or seventy per cent, and eventually even more.
But people will remain deeply involved.
Requirements Will Change First
The first major changes will happen in the requirements layer.
Product managers, business analysts and designers will use AI to formulate and improve requirements.
A person may show the system an interface and explain:
“I want to add this possibility here.”
They will then have a conversation with the model.
The model may explain that another solution would be better, identify missing cases or help organise the idea.
Through prompts, discussion and additional context, people will move the product away from an average solution and towards something more suitable and unique.
AI may quickly provide the first seventy per cent of the result.
The remaining percentage will be more difficult.
This is similar to the Pareto principle: the first large part comes relatively easily, while the final improvement requires much more effort.
The main benefit will be time.
A task that previously required a week of analysis and documentation may be prepared in several hours, half a day or perhaps one or two days including approval.
The requirement will then move to the development layer.
What Happens to Developers
I cannot predict the exact process, but I believe there will be developers who translate product requirements into more technical instructions for AI.
They will explain:
Create this in a separate module. Make it reusable. It must support this integration. In the future, it may need to support another integration. It must handle this amount of traffic. Add unit tests.
The developer will have a technical conversation with the model.
The model will generate code and probably create a pull request.
I am not sure that even medium-sized products will trust code generated directly from a high-level business description.
There will probably remain an intermediate technical layer.
It may include junior developers, although I am not convinced that juniors will be the best fit.
More likely, these roles will require at least some practical experience.
Another layer will then review the generated code.
This may become the most important engineering role.
One group will guide AI and generate code.
Another group will review, validate and approve that code.
The closest comparison is an aircraft autopilot.
Autopilot did not eliminate pilots.
It reduced their workload and helped reduce the number of people required in the cockpit.
A pilot is still present because someone must understand the whole situation and take responsibility.
Software engineering may develop in a similar way.
Experienced engineers will validate whether generated code actually meets the project’s needs.
And they will do this based on experience.
Why Experience Will Become More Valuable
I believe that experience will become one of the main criteria for selecting senior engineers.
Today, a person with five years of experience and a person with fifteen years of experience may answer the same twenty Angular, React, JavaScript or Java interview questions.
Their answers may be slightly different, but both may be considered acceptable.
Real life, however, does not consist of twenty interview questions.
A person who has worked for twenty years has objectively seen more rainy days and more sunny days.
They have worked with more projects, more people, more failures and more unusual situations.
Sometimes they can feel that one solution is dangerous and another direction would be better, even before they can fully explain why.
That is experience.
It is also what we often call intuition.
In my view, the human brain is more powerful than any computer currently available, but we cannot consciously use all of its power at every moment.
That would probably require too much energy.
The brain therefore optimises.
We do not consciously analyse every step when walking to work. We simply follow a familiar route.
Many simple decisions happen automatically.
Intuition works outside the conscious part of our thinking.
For a very short moment, the brain may compare a huge number of previous experiences and send a signal:
“This direction feels wrong.”
An experienced engineer may not immediately provide a perfect explanation, but that signal can save a project from a serious mistake.
This is why I believe highly experienced engineers will remain valuable.
Companies will find new ways to test and select them.
That is the good news.
The bad news is that fewer of them may be required.
AI will remove a significant part of the work they previously performed directly.
They will move towards review, decision-making and responsibility.
I cannot say how much demand will decrease.
Perhaps economic growth will create enough new software that the market will remain balanced.
But I suspect that fewer engineers will be needed inside traditional company structures.
For individual people, luck may therefore become a larger factor.
Even a highly experienced engineer may sometimes struggle to find the right position at the right moment.
But that is not the end of the story.
Small Teams and Boutique Companies
Another group of engineers may move into independent work and create a competitive alternative to large outsourcing companies.
This could lead to a strong new freelance and boutique-company market.
Businesses will continue to need software.
Whether the economy is growing, slowing or moving through a crisis, companies still try to optimise their work.
They will also want to introduce AI.
Moving away from manual processes requires new software.
Today, a small or medium-sized business may approach a large outsourcing company and receive a very high estimate.
In the future, the same business may approach a local boutique team consisting of two, three or perhaps ten experienced engineers.
A group of engineers with ten or fifteen years of experience may be able to deliver a project within two or three weeks.
A project that currently costs £30,000, £50,000 or £100,000 may eventually cost £10,000 or £20,000.
It may even cost less.
The exact price is not the main point.
The point is that when productivity increases several times, the project price can fall several times as well.
Demand may then grow.
Many businesses that cannot afford custom software today will finally be able to buy it.
A local car dealership, restaurant, warehouse or service company may need a specialised internal system.
Instead of approaching a huge corporation, it could work with a small professional team and receive a real product at an affordable price.
There may be enough such customers for small teams to compete seriously with large companies.
I therefore do not exclude the possibility that some large outsourcing businesses will begin to split apart.
Today, one or two large projects can support a significant part of a company.
A “large project” does not necessarily mean work for a globally famous corporation. It may simply be a project priced at £50,000 or £100,000, with a significant company margin included.
If market prices fall, that margin also falls.
At some point, maintaining a large permanent workforce may become unrealistic.
Large companies can hire engineers who use AI to produce code more cheaply, but the margin problem will remain.
They still need management, sales, offices, internal processes and experienced people.
The market may therefore begin to resemble billiard balls breaking apart.
Some large companies may split into many smaller teams.
Projects may become smaller, faster and more clearly defined.
A strong online shop or another advanced product might one day be created by two or three engineers in two weeks for £2,000–£5,000.
This may sound extreme today, but it is one of the directions we may reach.
And for many people, this could be good news.
What Happened to Stack Overflow
To understand how powerful this change can be, consider Stack Overflow.
A few weeks after I started using ChatGPT seriously, I noticed that I no longer needed Stack Overflow for most of the questions I had previously taken there.
At least for me, its value fell dramatically.
I suspect that many other developers had the same experience.
This is my personal observation rather than a formal traffic analysis, but the change felt immediate.
Stack Overflow had existed for many years and created an extremely valuable knowledge base.
Its strict moderation forced people to formulate questions clearly.
Answers were also moderated and ranked according to usefulness.
One question could receive five, ten or more answers, and users collectively moved the strongest answers to the top.
Think about how much money, work and human time went into creating that structure.
The result was almost an ideal dataset:
clear technical questions, multiple answers and quantitative signals showing which answers were considered best.
For machine learning, it is difficult to imagine a better source of structured programming knowledge.
I cannot say exactly how much of that data was used to train particular models, but resources such as Stack Overflow clearly helped create the world in which language models can now answer technical questions.
In a strange way, Stack Overflow may have completed one of its historical roles.
It helped organise programming knowledge so well that AI could learn to present that knowledge in a more convenient format.
This demonstrates the scale of the force we are discussing.
AI can seriously weaken companies and platforms that once appeared extremely stable.
And it can do the same to many others.
Failed AI Replacements Prove Very Little
I have also seen stories about companies dismissing employees, trying to replace them with AI and later hiring people back.
Such stories are usually presented in a comforting way:
“Look, AI could not replace people after all.”
But I think we should examine these situations without emotion.
Perhaps the replacement attempt really failed.
Companies make mistakes.
They may replace the wrong people, choose the wrong process, use the wrong technology or fail to motivate the employees who remain.
A failed first attempt does not mean the company will abandon the idea forever.
It will analyse the mistake and try again later.
The next technology will be better.
The process will be better.
Management will understand more.
And the second attempt may remove even more positions.
I therefore would not interpret one failed replacement as proof that AI cannot reduce teams.
Reductions will happen.
At the same time, this does not mean that companies can simply remove every person.
There are serious limitations that will slow complete replacement.
The Knowledge AI Does Not Have
Every AI system needs a strong knowledge base.
The Internet is large, but it does not contain everything.
A huge amount of information remains inside people’s heads.
It also remains inside private books, company procedures, internal documents, personal experience and undocumented habits.
I suspect that the amount of inaccessible human knowledge may be many times larger than the information publicly available online.
Consider an experienced carpenter.
They have eyes, hands and physical sensations.
They can look at a piece of wood, examine its fibres, feel its resistance and understand that something is wrong.
An experienced mechanic can hear an engine and recognise a problem.
They may combine the sound with vibration, smell, appearance and memories of similar failures.
Much of this knowledge has never been documented.
We have blogs, documentation and technical manuals, but an enormous amount of practical knowledge lives only in people.
AI does not have direct access to it.
It can therefore provide a probability:
“The problem is probably this.”
But without the missing information, it may not know enough to make the final decision.
There is also the question of responsibility.
An AI system can easily say:
“Sorry, I was wrong.”
It makes decisions through probabilities.
When one option has a probability of fifty-one per cent and another forty-nine per cent, the difference is extremely fragile.
In a real engineering, medical or mechanical situation, that small difference may not be enough.
Another limitation is the lack of human senses.
AI can process camera images, audio and sensor data, and those capabilities will improve.
But it still does not experience the world in the same combined way as a person.
A mechanic sees the engine, hears it, touches a component and remembers similar cases.
A carpenter sees the wood in detail and adjusts pressure based on what they feel.
A human decision combines formal knowledge, undocumented experience and physical perception.
This may also explain why some attempts to replace experienced engineers fail.
A technical specification does not contain all the relevant knowledge.
Someone who has worked with a material for thirty years may know that a particular alloy eventually develops cracks when used in a certain component.
Perhaps they learned it during an old university experiment or from a failure twenty years ago.
That information may not be available to the AI.
Reading an encyclopaedia does not teach you every detail of correctly driving a screw, selecting the right screw or recognising when the material is about to break.
That knowledge comes from practice.
Until human experience and sensory information are digitised much more completely, replacing people will remain difficult.
This process may take fifteen or twenty years, perhaps longer.
The same applies to software companies.
Complex products contain internal connections and specific knowledge that cannot easily be found in public documentation.
A company can train or customise a specialised model, but that requires additional resources.
The Resource Problem
The world may also face a resource limitation.
AI requires energy, processors, memory, data centres and specialised components.
Traditional search was relatively cheap.
When thousands of people searched Google for “Titanic”, much of the work had already been done.
The same indexed results could be reused.
The system did not need to perform a completely new, detailed analysis for every person.
AI conversations are different.
Each person asks a question in a slightly different way.
The context is different.
The requested format is different.
One person asks for a brief explanation, another requests code, and someone else wants an image.
It is much more difficult to cache and reuse the exact result.
The processors must work again.
Image and video generation require even more computation.
If AI traffic grows to an enormous scale, the pressure on energy and hardware resources will also grow.
This will slow some forms of adoption and force companies to choose carefully where advanced AI is truly valuable.
When the Dust Settles
Many engineers are currently afraid of changes in the labour market.
They see professions being reduced, teams becoming smaller and familiar career paths becoming less stable.
In my picture of the world, we are now inside a tornado.
There is dust everywhere.
Something will be destroyed.
Some companies will disappear.
Some positions will no longer exist in their current form.
But eventually the dust will settle.
New roads will become visible.
The outlines of a new world will appear.
In fact, they are already starting to appear through the dust.
People will find new places for themselves.
Human work will still be needed.
We should therefore try to see these changes not only as threats, but also as opportunities.
The work that feels normal today may disappear, but it may simply return in another form.
People may need to learn different skills, balance several projects, work faster or move between roles more often.
The IT industry will certainly use artificial intelligence.
AI will enter every level: design, requirements, analysis, documentation, development and testing.
This will reduce the number of people needed for some tasks.
Some developers will guide AI systems and generate code through technical instructions.
Other, more experienced engineers will review that code, validate decisions and take responsibility for the result.
I do not think this will be a temporary stage lasting only a year or two.
It may become a relatively stable structure for five, ten, fifteen or even twenty years.
Completely removing engineers will remain difficult.
Control will still be required.
Projects will continue to differ.
Many standard products will be generated quickly, but people will still want something special.
As long as people want uniqueness, experienced engineers, designers and creators will have a role.
I cannot say exactly how many companies or positions will remain.
There will probably be gaps, shocks and sudden changes.
At one moment, specialists in a particular area may have serious difficulty finding work. Another group may experience the same thing later.
Everyone will go through some form of adaptation.
But the appearance of the second point has convinced me that this is no longer a theoretical discussion.
The changes have already begun.
We can now see enough of the line to start thinking about where the next point may be.
And despite all the uncertainty, I believe the right response is not to stop.
It is to prepare, experiment and learn how to use the opportunities appearing inside this new world.