Aleksei Ivanov

Systems thinkers are the future

2026-04-25T22:57:17.847883+00:00

Have you ever wondered how many people taking on software engineering jobs actually enjoy them? How many of them are passionate about the tasks they are doing?

It is not a question about being good or bad developer. In fact, it might actually have reverse correlation.

The point I’m trying to convey is that not everyone is cut out to do everything and that’s more than okay–that’s actually healthy. Because it makes society diverse.

Now, continuing my previous post on junior developers I want to drop another idea: juniors will not become obsolete, the point of entry will shift.

What do I mean by this? Well, previously when you wanted to pursue career in software development you would start with an ever-green question of what programming language to learn, what framework to pick up and so on. Minor technical details.

With the advancement of LLMs this is shifting to a higher level of abstraction: what do you want to build, not how.

You know how big IT consulting firms have positions of systems and business analytics? They aren’t strictly required to know programming. It is another job.

So here is my take: with the advent of AI, junior positions will shift from deciding what language or framework to learn to learning analytics basics.

You will not build the systems yourself, but you will build and communicate and debate vision of the systems. Why they should work the way they should, regardless of the technical details.

And if you are curious, you can advance later and deepen your knowledge about particular technology.

In a sense, it is an inverse approach: instead of starting with a programming language (bottom-to-top), you start with a vision of the system (top-to-bottom).

Vision requires intuition and that’s something that LLMs cannot provide.

The future of juniors in the “age of AI”

2026-04-25T17:22:55.890772+00:00

It is said that junior developers are no more: companies don’t need them because it’s “easy” to replace them with language models (I would not go so far as to call it AI).

At the same time, new people coming to the industry are questioning themselves whether they should pursue it or not.

I think both angles are incorrect.

You see, companies never really needed just there programmers. They required problem solvers. And for LLMs to solve problems elegantly and eloquently they need guidance. Simply because it is an algorithm that lacks insight and intuition. And that won’t change.

So, it is better to pose a question differently: do you need problem solvers?. Because not every business do. Sometimes you just need a pair of hands and that’s where language models are capable.

At the same time, from the perspective of a student–who is wondering what to pursue–the question is not whether or not you will be replaced. That’s not a correct way to think about it.

The question is: do you enjoy solving problems? Because if you do, it doesn’t matter whether it is programming, architecture, data modelling, analytics or anything in between. You can leverage LLMs to give you a hand and at the same time teach yourself.

You have to approach things more fundamentally. It is not a question of what pays more in the moment. It is a question of what you truly enjoy doing for the rest of your life. Given, that is not a singular thing, but a whole range of different things. All of them allow you to compound experience.

The choice you really have to make is to pick the thing you will not get bored from. And if you don’t pick right on the first try–just try once more, there is no limit on the number of attempts.

Co-authored by whom?

2026-04-24T18:30:26.135210+00:00

Claude Code is a tool that you paid for. Usually, you don’t get advertisement in a tool you have paid for. So when Claude appends “Co-authored by Claude” to every commit message, it is only natural it would raise an eyebrow.

It’s not a simple matter, either: disclosing AI generated content might be a good idea, but it has so many edge cases you can cut yourself if you are not careful.

First of all, it might generate this for a commit it didn’t even contribute to (part from the commit message). That’s a stretch to call it co-authoring.

Then–at least for me–this disclaimer suggests that Claude did more than half of the work, while in reality it was merely instructed to execute a set of tasks. The idea was mine, the control is mine, I might very well be responsible for the commit and take credit for it.

You don’t credit Microsoft PowerPoint for the keynote you made or say that built in calculator app “co-authored” you financial analysis. LLMs are tools, not sentient beings. In fact, they have a private owner–the company that rents it to you.

And this is where the third thing comes into the light: intellectual property rights. Oh boy, is it a grey area with this one.

Models providers had to source huge amounts of data to train their models and as far as I know each mode weight doesn’t have “co-authored by this YouTuber” or “co-authored by this blogger” on them. Furthermore, what happens when you make a commit–which is signed off by you–that mentions such co-authoring?

Would that legally transfer some of the rights to it to the company that rented you the model? We don’t know yet, and when there is ambiguity it is better to play safe.

I started writing this post with no strict position on this matter. But in the process of writing it I think I see that there are really no redeeming qualities for doing this the way it is being done currently.

Interfaces as documentation

2026-04-23T20:10:37.311524+00:00

https://www.youtube.com/watch?v=v4F1gFy-hqg

This short talk by Matt Pocock has some great insights about how to work with AI and code on a professional level. Obviously, vibe-coding is not an option when you are doing mission critical tasks. And any project becomes mission critical once it reaches commercial success (that is, if you care about keeping paying customers).

Interfaces are a great way to separate the responsibility between design and implementation: let AI do the "dirty work" of writing the implementation, whilst we focus on designing the interfaces.

To be fair, I have not really tried this myself—I prefer to audit the implementation as well, because it is just too risky. However, I can see how this can be a better approach than spec-driven development, which is very verbose (albeit, sometimes there are simply no other options).

Special treatment

2026-04-22T04:43:56.663131+00:00

I have noticed that I get unusual feeling sometimes while meeting a person or visiting a place. It doesn’t really have to do with the person or the place, but with my comprehension of it.

On one occasion it can feel official or even hostile, while other times it is chill or friendly–with the same person or the same place. Maybe even things.

I can’t put my finger on exactly why this happens, maybe it is my inner feeling. But I suspect it is something else, something that I carried through all my life.

Where it came from – I don’t know, and I am not even sure whether I should get rid of this. It does bother me somewhat and started to increasingly more so after I started to notice these moments.

Maybe there is a name for this phenomenon. Or maybe I’m just tired. But I really wonder where does this come from.

They simply don’t have the time to test it

2026-04-21T02:19:01.846533+00:00

Leading AI labs have been churning out models and features left and right for the past couple of years, that’s because they consider it an arms race.

Anthropic in particular has been speeding up recently with Claude Code and ecosystem around it.

Code, Cowork, Excel, Chrome extension… and then for Claude Code itself they started creating more sophisticated features: Remote control, Dispatch, and now–Design

Did you keep the count?

It is understandable, they race to grow to a trillion dollar valuation. But such speed comes at a cost.

They want users to use their ecosystem, but it starts to crack like drying clay. Because the foundation is not properly cured–they simply don’t have the time.

Or maybe–and that’s a bit controversial–their coding agents can’t keep up with churning out so many features. Maybe their developers can’t keep up with the coding agents?

Whatever it is, stable and solid product is part of the brand. And after they sped up it started to erode.

Why would I use their Routines (yes, another recent feature) when they aren’t guaranteed to work? I gave it a spin, created a simple daily recurring task. It failed at least once with no reason apart from infinite spinning:

Meanwhile on my own server, task scheduler written in Python (by Claude) runs consistently for several weeks now.

Yes, I know, different scale and complexity. I also know that my setup isn’t valued at multiple billions.

I have been hard on OpenAI and was “praising” Anthropic. I really enjoy their model, even with their weekly quotas. But credit is where credit due: OpenAI doesn’t have multiple outages in a month. And they have a sizeable audience as well.

Your life is your life’s savings

2026-04-20T05:58:17.040714+00:00

An interesting thought I got while listening to one of Alex's Hormozi videos. He might not be always right or you might not like him, but the premise is true and simple: you either put effort into things things that matter to you or you put effort into things that matter to others. There is no third option.

It’s so obvious, yet most people would confront this with “it’s not that simple”. And you would be right–it is not. And that’s the point, sort of.

The idea of our limited time on this planet is actually an ancient meditation monks or buddhists have been practicing for a while now (for decades, to be more precise). And the didn’t watch Alex’s video.

By living through our lives we get experience and we can influence what kind of experience that is. This affects all of our future, even if slightly.

Therefore, it makes a great deal of sense to put focus and effort into things you truly enjoy. Because whatever you choose you would have to overcome it.

And if you think that in your current position it is hard or impossible to do, look at it from another perspective: would a person a hundred years ago feel lucky if they traded places with you? What one (out of numerous good) things that surround you, you can be grateful for?

Everything you choose to do is an experience that will aid you in your life, one way or another. It is a piggy bag of life savings you cannot get taken away from.

Future hackathons

2026-04-18T21:21:03.249868+00:00

I have stumbled upon a university project someone is doing for their diploma. The idea of the product is alright, but the site was clearly create by AI.

Is it good or bad? This got me thinking: prior, excellent teams have shown their mastery during hackathons. A bunch of people gathered together and dug through the tunnel, delivering often simplistic—but working and original product in 24 or 48 hours.

Do we ban AI use during hackathons now? Or do we encourage them?

How do we rate the work of a team? Does technical prowess not matter simply because the product needed web UI but team expertize is not quite there yet?

These are difficult questions.

On one hand AI empowers: it allows teams to leverage their creativity and expertize in other areas more freely. On the other, it levels out the surface somewhat. Literally.

If you don't put effort—and you usually don't have the time for that during a hackathon—then you end up with a generic design at best.

Is it important? Depends on the field, I suppose. After all, even games during game jams (analogous to hackathons, but for game development) usually lack high quality assets.

But before AI even those low fidelity assets had a chance to get created with love and care—even if hastily and not thoroughly.

I think we need to think long and hard why do we do hackathons. Because if it is for the product or the end result, then we can allow AI aid there and get performance improvements. If it is more like a sport, however, then AI is no more interesting than watching a computer play chess.

Student or apprentice?

2026-04-18T18:17:20.314759+00:00

Students nowadays have it hard: they to study more and they don't have time to dive deep into anything they study. Moreover, they can't even always evaluate prior to committing. Simply because the field is so vast that by the time they realize it might not be for them — it is too late.

Higher education is important, but most of the time not as a career but in the same way we read literature or study history in our leisure time. To broaden horizons of our thinking.

And of course, with generative tooling it is so tempting to take a shortcut, undermining your own learning in the process. It is a complex issue.

But I had a thought: what if current tools could act somewhat like a mentor? Of course, they make mistakes and then there is an issue of not knowing what you don't know.

Still, it is much better for a curious mind to have a 24/7 counselor which you can ask anything, and which could guide you through your first steps.

Simply to discover what you would need to know. Acting as a bootstrap. And after that, you can figure it out on your own — looking for people in groups related to your new topic of interest.

I wonder if a more widespread era of apprenticeship is upon us...

Speed and complexity; now and then

2026-04-16T21:27:02.402843+00:00

Many talk about how AI helps speed up development. 10x this, 100x that. But what if the real value of the new technology is not the speed itself, but complexity that is available to us?

Think about it: when you were doing everything by yourself, you had to study & learn the tooling. Then uncover edge cases (often by googling, usually ending up on stack overflow). Then lay out the plan in your head and implement it, often iteratively.

This would take so much time if you were to do everything "properly" that it would be often unreasonable to do it on your own.

Now with AI you can surface above the abstraction—keep your vision and delegate often tricky implementation details to the LLMs. Compiler flags? Project setup? Heck, I am so glad I don't have to pick out of 10 new TypeScript transpilers—each of them "faster" than the other.

I can focus on more complex and broad things. I can also do more proper engineering now: cover things with automated tests, run benchmarks. Actually validate the work.

Sometimes AI code generation tools can be overwhelming by the amount of code they produce, but I feel like with regard to managing complexity they can actually help and improve the workflow.