These are some of my thoughts on recent AI reasoning improvement (especially multimodal models, e.g., GPT, Fable, Grok, or Gemini) and what it could mean for young people getting first jobs in the post-2020 era - like me.
To make this very clear, these are just my thoughts on where things could go long term if models continue getting better and industry adapts. If you’ve found this article hoping to find an answer to how to be successful in the future career-wise, you won’t find the answer here.
This article is an outcome of a thought experiment, during which I tried to persuade myself of what skills and areas of competence could bring value in a business setting long-term, assuming programming, as we know it now, won’t exist.
Personal context
I am 21. I started coding around 9 years ago, when I was still in primary school. Attention Is All You Need was fresh and nobody in the industry even thought that software engineering would disappear.
I, or we, for the ones with a similar timeline, had literally been watching YouTube tutorials, reading blog posts on the internet, writing code and debugging. Building for fun, where fun was getting your hands dirty with UTF-8 symbols. When I got my first job in early ‘22, GPT could help you with generating a slop essay at high school. Basically, people were amazed that LLMs are capable of generating sensible sentences, but programming videos on YouTube still explained new JS frameworks (boy was it beautiful).
AI software engineering today
We are living in crazy times now. Since late 2025, models’ capability in generating code got really good, and it looks like AI isn’t going anywhere unless we hit some other geopolitical barriers or hardware issues.
Research has been done and weights have been deployed. Things are working in general. For me, it means that tech is there, and it can only get better and be adopted further.
I had a weird opportunity to observe a single workplace in 2024, 2025, and now 2026, interleaved with studying, as I am still pursuing my Bachelor’s degree in CS. As I wasn’t working during the semester, getting back each time felt a bit like having a new profession. So, when AI-backed software engineering accelerated the most, I could see how my work changed. And I want to give you a glimpse of it (keep in mind what I felt is highly subjective).
In 2024, we were writing commercial code almost only by hand. Typing UTF-8 characters. These were the times of GPT-4. Models were rather poor with keeping up to date with the bleeding-edge, blazingly fast web development ecosystem.
In late 2024, I noticed that learning maths with models started to work. When you hit a wall, you could explain what you have, what you know, and figure out the path to the answer with a model together. I am speaking of quite easy math, obviously, as I was a freshman (I’m thinking of linear algebra, analysis-like things, mostly mechanical stuff, or understanding definitions).
Then 2025 came. I believe that the introduction of Gemini 3 and further improvements of GPT were the turning point for me. In my studies, these models could easily teach you if you gave them the necessary context, like a photo of an almost-solved math problem that made you stuck. They were getting quite good at writing and explaining ideas behind complex algorithms or bachelor-level analysis II course material.
When I got back to work in 2025, prompting was more common. I was consulting bits of code I wrote with an agent or asking it for a review. While most people still wrote code by hand, models were getting noticeably better. They made sensible reviews and could write complex structured bits of logic using different well-known APIs. I remember asking a model how something works and how it can be implemented, hand-rolling my own implementation afterwards. Things like Claude Code were there, but for me it was weird. Like, I had my own editor and fingers.
Late 2025 and whatever happens now in 2026 was a huge acceleration in the tooling. Claude, Cursor, Antigravity. Building harnesses. Writing loops. People got into a frenzy. A ton of slop generated in the meantime. But research was getting there.
As of today, with the introduction of models like Fable or GPT 6, I am just explaining a feature to an agent and it gets done. The same with reviews, bug finding, and other stuff. More and more tooling gets shipped and agents can use their reasoning capabilities in a wider range of tasks. Keeping up to date with libraries is not a problem either, as agents just browse documentation when needed.
What’s interesting, though, I feel that I am using skills learned in the past heavily. Every decision I make is backed by some experience I had. Let it be how a project is organized or what tech stack to choose.
Of course, on top of this, the idea of agentic development is in quite a complicated state at the moment. New models release every 3 months, every industry has different security requirements, and every workplace adopts tech at a different speed. Some teams can accelerate immensely, some get buried in a pile of slop, and others just can’t use LLMs as they are working with highly sensitive data and it’s too risky.
How things could look in the future
I found myself in few situations recently when I HAD to write code by hand. It was more like guiding the agent. Telling it that it’s generated slop somewhere and that it should fix it in some way because it will work better. But ultimately things got done. I was there just orchestrating the process. Reading and thinking.
What I wanted to state here is that: In a few years from now it’s totally possible for us to just ask a model to do something and get it done. And it affects every job that has anything to do with software. Both jobs that create software and jobs that use software.
Again, there are a ton of variables in the economy and how companies operate. But this doesn’t change that hard research is already done. It’s just a matter of tooling around it or regulations.
That’s why I feel it’s a valid reason to rethink what we are spending time on (in terms of learning) and get comfortable with the idea that our work might change drastically.
I have a feeling that a lot of people my age pursuing Computer Science are lost. Some are, of course, deep into vibe coding, having fun, but I think that the worries graduates might have about the job market are justified looking at what is flying in the news or on social media.
You should learn
Now, the previous sections were a bit gloomy. But these were my observations. They made me depressed sometimes. And if it happened to me, it surely happened to others. So let’s get back on track now. Why do I think learning hard stuff is worth it?
I believe there will always be some level of complexity to the world we are living in and technology. In the end, AI is just a tool. It allows you to accelerate, but you are the one asking questions and deciding what to build.
Most people still won’t have the technical ability to understand fully what is happening. Exactly the same way that I am not aware of how a nuclear plant operates. This means that there is a need for people who understand and have expertise. Ones who build. Ones who are competent. Ones who can handle a larger part of development, making sure everything glues together. As always, with great power comes great responsibility.
I believe that, long term, building just won’t mean being a syntax engineer, but being an architect of a system that needs to be reliable and that needs to support the business. Or research. Or whatever.
If code quality produced by agents gets perfect, we might be able to build more robust and larger systems that will improve efficiency and reliability of many things, but also get complicated on other levels. Planning how the software we create behaves, how it integrates with other products, how it looks, and finally who is responsible for its maintenance (even if this just means scheduling what needs to be fixed and when). All of this brings human-level value that’s not possible to replace with an Agent. It looks like there’ll be a need for someone who has deep technical understanding of the system and knows exactly how it operates, but also is able to communicate this outside.
Try to open an agent and do work in a field you don’t know anything about, for example, physics (jk). If I wanted to build some new propulsion system, as of today, I don’t know how to start. Like what? Make a 3D model in Blender of a combustion chamber? But, .. like, what would even my prompt look like?
Maybe this works in the future, but I doubt it competes with a fleet of highly educated engineers who know their grit, know why the current propulsion system is flawed, have a factory necessary to make a physical copy, a test facility, and, on top of that, can make their work more reliable with AI.
You need to know what you’re doing really well at a certain level of abstraction. It’s just like most of us don’t know assembly these days or couldn’t write a program that runs on a computer from the 80s. We just might not know today’s programming in the future because AI will be the next advancement shifting complexity somewhere else.
Understanding tech abstraction you’re working with underneath is just essential and will always be required. AI is a tool that solves the pain of writing UTF-8 inside correct files and without typos. Tool that can browse documentation in 5 seconds in search of the answer, instead of spending a whole day reading. Tool that can tell you “This might be a better pick as a ton of people have done it in the past and it worked remarkably well. Do you want to use the same solution in this project?”. And finally, a tool that can brute force and test plenty of approaches leaning on top of already done research.
All of this accelerates development, but you are the one responsible. Unless models gain human consciousness and become a separate species.
So, just make some good pour-over coffee and hit a linear algebra or systems engineering book.