Where we are - new

Why I Decided to Write This Now

Well... I think the time has finally come for me to write about artificial intelligence.

I want to talk about how I use it, how I see what is happening in the world, and where I think all of this may be going.

I will also explain why I decided to write these articles specifically now.

I had many thoughts from the beginning, and at some point I understood that they would not fit into one article.

So I decided to split the topic into several parts.

The first question is where we were.

The second is where we are now.

Then comes the question of where we are going.

And finally, where we may arrive in the much more distant future.

Until recently, I was not fully sure what was really happening.

I understood that many things would change. That part was clear.

But the direction, the shape of these changes and the lines between them were still not clear enough.

This year, however, something changed in my own work.

I can now see enough practical progress to start talking about it more seriously.

The Four Dots

To build any forecast, whether it is about economics, technology or something else, you need at least two points.

With one point, you only know where you are.

You still do not know where you are moving.

But when you have a second point, you can draw a line between them.

And once you have that line, you can try to imagine a third point and even a fourth one.

For more than three years, this was my problem.

I had the first point in my head.

I knew where we had started.

But I could not clearly see the second point.

Without it, I could not connect anything. I could not understand what was realistic, what was unrealistic, and which direction the whole process was taking.

In 2026, at least in my own practice, I finally felt that the second point had appeared.

There was enough real progress for me to say, “Yes, this is already an important checkpoint.”

Now we can try to draw the line.

This article is about the first and second dots.

Where we were, where we are now, and what changed between those two points.

The next article will be about the third dot.

Since I work in IT, it will be easier for me to imagine that third point through the software industry.

It will be my personal forecast of what may happen there.

Another article will be about the fourth dot.

That one will be more global.

It will not be such a precise medium-term forecast. It will be more about where humanity may eventually arrive.

Maybe it will sound a little utopian.

It may describe a period that none of us will see during our lifetime.

But I still think it is important to discuss it.

Of course, I may be wrong.

Something may happen earlier, something later, and some details may change completely.

But I think the general direction can still be visible.

And perhaps some parts of that future will already appear during our lives.

So this is the plan.

First, we talk about where we were, where we are now, and what happened between those two points.

Then we move further.

My First Contact with AI

Now let's talk about the first dot.

For me, that was 2022.

It was the year when I first met ChatGPT.

Maybe it appeared a little earlier, but it reached me somewhere around the middle of 2022, probably at around the same time as it reached many other people.

I still remember my first reaction.

My very first thought was something like:

“This must be some prepared answer.”

I remember that thought very clearly.

Along with it, I had several other thoughts.

One of them was that this could not really be happening.

It looked too close to science fiction.

But in reality, I think the people behind it did something that others had probably not managed to do.

They believed in the idea and brought it to the end.

The basic principle itself did not look impossible to me.

I could imagine that, based on a sequence of words, a system might calculate which word was most likely to come next.

The approach sounded understandable.

The real question was whether it could work well enough.

How Can a Model Predict Words?

This was the part my brain found difficult to accept.

Imagine a simple question:

“What is the capital of France?”

After the word “France,” the answer “Paris” can probably appear with a very high probability.

That is understandable.

This is knowledge that can be found and learned.

But imagine that the sentence is not finished.

Someone only writes:

“What is the capital...”

How can the system predict the next word?

It does not yet know which country the person means.

From my point of view, that was much harder.

But it turned out that there was enough existing language and knowledge to train a model to continue differently.

After the incomplete phrase “What is the capital...”, the model does not have to guess a country.

It can build a whole response:

“Please clarify which country you mean.”

That was the surprising part.

It was not only predicting one obvious word.

It was able to construct a complete and reasonable sentence.

Life showed that enough human knowledge already existed to teach a model to communicate.

And that changed the direction of everything.

Models Existed Long Before Chat

Of course, models themselves were not new.

Machine learning had existed for a long time.

I first came across it around 2013.

At that time, models were already being used by companies such as Google and Yandex, for example, to recognise images.

Even then, a system could look at a picture and try to understand what was in it.

It could recognise a car.

It could recognise a shirt.

It could even try to understand whether the shirt had a particular pattern, such as dots.

That is not an easy task.

Human vision feels natural to us because we use it every day.

But automating it is difficult.

Still, people had already learned how to do this more than ten years ago.

In a way, one of our senses had already been partly automated.

A technology could already say:

“This is a car.”

Once you receive that information, you can make another decision based on it.

You can build more software around it.

You can imagine a machine moving through physical space.

A camera looks around a kitchen and recognises where it is.

Then the system says:

“I am in the kitchen. I can move one metre forward, then turn.”

So this was not completely science fiction anymore.

I did not see image recognition itself as something cosmic.

What I could not believe was that a system could predict language so well.

Languages are different.

People use the same words in different situations.

Context changes everything.

Imagine saying only one word:

“Hello.”

What should the next word be?

Perhaps the most probable answer is also “Hello.”

Then the system waits for more instructions.

With every next word, it slowly understands more context and becomes better able to predict what should follow.

That was roughly how I understood it at the beginning.

Trying to Build a Junior Developer

So I started working with it.

My first practical goal was to turn it into a good junior developer.

I wanted to give it tasks and receive useful technical work.

But I did not succeed.

I tried several times.

Between those attempts, I waited around six months, expecting the technology to improve.

But at that time, I reached a simple conclusion.

For me, it was not yet much more than a very well-indexed internet.

It could understand my questions more precisely than a traditional search engine.

It could understand broader requests.

And it could answer in a form much closer to what I actually wanted.

That was already valuable.

These directions continued to develop, and they will keep developing.

But we will talk about that later.

At that moment, I introduced chat models into my life mainly as a replacement for Google.

After about six months, I almost completely stopped using traditional search.

And when I say completely, I mean almost completely.

The only reason I still opened Google was to find one exact page.

Even now, this is mostly how I use it.

It remains the default search engine in many browsers, so sometimes it is simply the fastest way to open a known website.

A New Form of Search

There is another interesting detail I remember about traditional search.

Google used to show the number of results for a query.

For example, you searched for “Titanic,” and Google told you that it had found one million results.

At some point, I read that this number was not always exact.

Perhaps it was calculated approximately.

Maybe the real count was one hundred thousand and another zero was added. Maybe it worked differently.

I do not know the exact technical details.

But the psychological effect was clear.

You saw a huge number and felt that an enormous amount of information had been found for you.

Now, even when I sometimes return to Google, I usually do not go beyond the first page.

Quite often, even the first-page results already feel irrelevant.

This tells me that Google probably does not worry much about what appears on the second or third page.

Maybe it has not worried about that for the last ten years.

At least, that is how it feels to me.

I think this type of search is slowly dying.

And I can say with almost complete confidence that, in its current form, it will eventually disappear.

The reason is simple.

Searching for information this way is inconvenient.

At one time, it was useful and fast.

But today, you open one website and find only one part of the information.

Then you open another site and find something else that was missing from the first source.

The second website looks different.

Its navigation is different.

Its text is structured differently.

Perhaps the information is hidden behind advertising, pop-ups, menus or a completely different page design.

Your brain has to adapt every time.

That requires energy.

And most people do not want to spend that energy.

When you ask a chat model, the process is different.

It already works with aggregated and indexed information.

It tries to understand you.

It gives you a more direct answer.

It can ask additional questions.

And it presents information in one consistent format.

You do not have to adapt to a new website every time.

The interface remains the same.

The way you ask questions remains the same.

The structure of the answer is also familiar.

For me, this format of search is simply much better.

It is more universal.

It is more standardised.

And the problem of every website presenting information differently almost disappears.

This was how I understood the technology at that time.

It was a new form of search.

A new form of indexing.

And that was how I used it.

I could not yet grow my own junior developer from it.

But it did help me with code.

I would say that I started saving around ten percent of my programming time.

That was useful, but it was not the biggest effect.

The bigger effect was somewhere else.

It was connected to personal growth and to learning in areas where I previously had less knowledge.

Where We Were

So this was the first dot.

We had a system that could communicate surprisingly well.

It could understand questions better than a traditional search engine.

It could collect information into one answer.

It could help with code and save some time.

But, at least for me, it was not yet a developer.

It was not something I could trust with an entire task.

It was not yet the kind of system that could take responsibility for technical work from beginning to end.

It was a powerful search and communication tool.

And that is where we were.