My many conflicting thoughts on AI: where I stand as of late 2026

This post is quite different from what I usually write on this blog. I’m sure many people are tired of constantly hearing and reading about AI. Heck, so am I. Yet AI has such a persistent presence in my daily life, be it at work, in online discussions, or in the communities around my hobbies, that there’s just no escaping it. I have so many thoughts about it and have felt its impact so strongly that I feel the need to get it all out by writing everything down, ironically contributing to the exact thing I’m tired of.

If you’d rather skip this one, that’s totally understandable. Screaming into the void is part of the point of this write-up. I’ve got some interesting retro posts in the works, which I'll get back to after this. For those who are interested, I hope I can at least offer an angle or two you hadn’t considered.

💡
When I say "AI" in this post, I mostly mean generative AI and particularly LLMs, not other forms of machine learning.

Where it started and the years of disappointment

During the final years of my bachelor's degree in computer science and application development, we had classes on AI where we would train small machine learning models to recognise objects in photos. I remember one of the lecturers, one of my favourites and someone I really respected, telling us something along the lines of: "Guys, this whole AI thing might not seem like much right now, but it's going to be huge in the coming years. Major progress is being made, and it's about to blow up."

At the time, I didn't think things would evolve and blow up that quickly or that it would be that huge. Yet here we are, about four years later, and I'm astonished at how far the technology has come.

It was around that same period that I became aware of OpenAI and its GPT models. I remember trying the demos on its website at the time, where you'd enter a prompt and it would generate text: one prompt, one result, no conversation. Very basic compared to what we have now, and I didn't think much of it. It was neat, but full of errors, hallucinations, and weird words or sentences.

In June 2021, GitHub Copilot was announced. I was very intrigued, so I signed up for the "technical preview", got access not much later, and started using it for day-to-day coding. My impressions at the time mostly came down to this: it was very powerful as a context-aware autocomplete tool but made a ton of mistakes when writing even slightly large portions of new code. This continued for a very long time, years, in fact. New models and tools would get released, the hype cycle would praise them to the high heavens, I would give them another try, and I would ultimately find myself disappointed despite the improvements. It took a very long time for them to become reliable enough for me to use on a daily basis. For the longest time, I found they got in the way more than they actually helped.

Many models came out, and I tried several: OpenAI's ChatGPT, Anthropic's Claude, Google's Bard / Gemini, Mistral's Le Chat / Vibe, Proton's Lumo, as well as several Chinese models such as DeepSeek, GLM, Qwen, and Kimi through Kagi Assistant. With every one of them, I went through the same cycle of initial short-lived excitement before feeling underwhelmed and no longer using it.

Throughout this entire period, I was equal parts intrigued and annoyed: annoyed by the ubiquitous presence and endless hype I didn't understand, yet intrigued by the improvements over time.

There were only a few applications where I found the technology to be genuinely useful back then. Most of them came down to the language part of large language models. For translations, proofreading, and turning draft notes into proper documents, it was pretty great.

For almost anything else, be it coding or research involving factual information, it was simply making too many mistakes to be useful or time-saving.

Annoyances and ethics

From 2023 onward, I started to notice AI increasingly creeping into daily life. Family members started to use ChatGPT and would treat its output like gospel and absolute fact. AI-generated flyers, posters, and logos started to pop up everywhere I went. There was also a noticeable uptick in low-quality, low-effort AI-generated content online that persists to this day. This has remained a major annoyance of mine ever since. I'll come back to this later in the post.

As popularity and public awareness grew, more and more ethical issues came to light. I won't go into detail about everything being discussed, mostly because my opinions aren't particularly unique, but I do have a few things to say.

One thing you seem to hear less about these days, but that was very present in earlier years, is the discourse around AI labs scraping and essentially stealing all the text, websites, images, videos, code, and other material they could get their hands on, all to train their AI models. I, too, find this very problematic, and it's one of the biggest ethical issues I have with this tech to this day. I'm very disappointed that, after all this time, there still do not seem to be proper regulations or repercussions, despite good but (in my opinion) insufficient attempts.

Another issue that gets brought up constantly is the environmental impact of data centres and their electricity and water usage. I'll be honest, I'm getting a bit tired of this one. Not because I don't think it's valid or worrying, but because I don't see it as a problem that was caused by AI. AI only accelerated it and made it more visible. The problem would not magically vanish if AI were to disappear overnight. Your photos in the cloud need to be hosted somewhere. YouTube and Netflix would continue to transcode, store and serve incomprehensible amounts of data. I could go on. I'm not trying to give AI a pass here; I simply think it's part of a larger problem that you cannot blame solely on AI, and I find it a weak argument against AI because of that. It's something that we need to solve for all digital infrastructure, not just AI.

Despite these and many other issues I have with AI, given the field that I study and work in, I find myself in a position where it would simply not be wise to boycott it all, even though I’ve gone through periods when I really wanted to. Staying up to date and learning to work with new technologies is very important to me, no matter how I feel about the tech itself.

The tipping point

At the end of 2025, things started shifting for me. Yet again, the cycle began with new frontier models releasing and people hyping them up. The difference this time, however, was that I saw the opinions of people in the industry who were as sceptical as me start to change.

I was fairly stubborn for a few more months, not really wanting to spend the time adopting a new workflow again just to end up underwhelmed once more.

But as the praise continued, I eventually tried GitHub Copilot CLI and started using a fully agentic coding workflow for the first time. This time around, I was really impressed. Unlike in previous experiments, the AI models seemed to hallucinate very little, understood context much better, and actually produced solutions very much in line with what I had in mind. Now, even though I still had to steer and adjust many things, it was really saving me time.

The excitement turned into fear that lasted for a number of months. My worries about the future, particularly around my job, career and hobbies, went into overdrive. More on this later.

It took a while before I was able to take a step back, calm down and put things into perspective. While still quite scared and uncertain, I eventually reached the conclusion that, like it or not, this technology is here to stay. It will have a major impact on the software engineering field, with many uncertainties that will likely remain for years. While easier said than done, I decided to try and stick to my principle of always learning new techniques instead of fighting them.

Just half a year later, my workflow looks very different.

How my workflow has shifted

Throughout 2026, I've been actively incorporating agentic AI coding tools into my daily workflow. After a short period of using GitHub Copilot CLI, I eventually moved to Claude Code CLI. Both are very similar in terms of user experience and quality, as far as I could tell. Claude Code just happened to be the one chosen as the go-to tool within my team at work.

With AI now producing code up to a standard I find acceptable or even great, I regularly hand off tasks to it that would be considered chores. Think writing tests, large code refactors that are simple in concept but take a long time to do by hand, writing boilerplate code, and so on. It can do these things in a fraction of the time it would take me, with results that are almost always spot on.

When it comes to looking for information and learning new things, like many others, I've started to use chatbots as a Google replacement or companion. I find it incredibly helpful to ask a specific question and get a tailored answer. One important consideration, however, is the accuracy of the information. I added a few general instructions to my Claude account that tell it to verify factual information with a source whenever possible or, at the very least, be upfront when information comes from training data. This works great and makes it very easy for me to quickly verify that the information is correct.

AI has also let me approach things in better ways than I otherwise would have. When writing code, there are almost always several ways of approaching a problem. Before having AI tools, I regularly had to weigh the trade-off between two solutions that would both work equally well: one would be much more time-consuming but cleaner and more future-proof, while the other would be much faster to implement but leave more technical debt for later. Now, thanks to AI, there is almost never an excuse to choose the latter option, as the time needed to write and refactor code has been cut so significantly.

For writing entirely new features with new code, I start by carefully planning things out. I take a lot of time to think about how something should be built and why. This way, I'm still the one making the architectural choices. Having a clear plan allows me to break things down into small, reviewable steps that I can hand off to an AI agent. This is an essential part of my current workflow. I do not hand off large, vague tasks to AI and let it do its thing.

By using small, reviewable steps, I can catch fundamental mistakes early and correct them, avoiding having to rewrite large parts of a mountain of generated code. Working this way also has the advantage of helping me maintain a much better mental model of the codebase.

I make it a point to read (and understand) all code generated by AI at least once. How thoroughly and critically I review a piece of code depends on how critical that piece is. If it's part of, say, an authentication flow, I will look at it very carefully, as opposed to a simple helper function to handle some data.

What it always comes down to for me is this: when I contribute code to a project, be it at work or for an open-source project, I am always the one responsible for that code. That being the case, I refuse to let go of the principle that I need to have read, understood, and tested any code that ends up pushed to production and that people rely on. When something goes wrong, I need to be able to explain what went wrong and why, as well as justify any choices I made that may have led to a problem.

Personal projects

Most of what I expressed in the section above applies to professional work. When it comes to personal projects, it's a different story.

If a personal project of mine mostly serves as a way to learn something new, like a programming language, framework, or tool, I avoid letting an AI agent do the work for me as much as possible. However, I do still regularly use Claude's chat interface as a search engine alternative to get quick answers. When I get stuck on a project, I often use Claude to get a nudge in the right direction. I specifically instruct it not to give me the answer outright, but instead to let me know what next step I should try or provide me with a resource I can read to gain the knowledge required to continue. A good teacher does not hand you exact answers; they give you the knowledge and tools to find the answers yourself. And that's the exact mindset I keep when using AI to learn new things.

A different kind of personal project is one where I just need something that works and don't care about the process of building it myself. The perfect example would be one from a few months ago. I wanted to migrate my gaming backlog from HowLongToBeat to Backloggd. HowLongToBeat has a CSV export feature, but Backloggd does not have an import feature. Backloggd has no official API, only undocumented endpoints. I used Claude to figure out how the endpoints worked and how the data from the exported CSV should be transformed to fit them, then turn it all into a Python script. If I'd had to do this by hand, it would probably have taken me a few weeks of spending entire evenings on it. With Claude, I was able to do it in two evenings. For throwaway scripts like this, where I gain no valuable knowledge from doing the work myself, using Claude is an absolute no-brainer to me at this point.

There's also a middle ground. A few weeks back, I needed a shell script to back up a number of directories from my main server to a smaller Synology NAS using rsync. It wasn't a particularly difficult task, but it would have cost me some time nonetheless. I had Claude generate an initial script, and I reviewed and refined it iteratively. In this case, I made sure to understand the final script but skipped learning the exact syntax and different rsync options. I needed something that worked quickly but still wanted to verify and understand the solution, in part because there was a bit more at stake here.

There are many more examples of small side projects like this that, before AI agents got as good as they are now, I would never have tackled. Not because I wouldn't have been able to do them, but simply because the sheer time investment required would not have been worth it.

Insights and worries

Vibe coding vs AI-assisted coding

There is a lot of nuance in the difference between vibe coding and AI-assisted coding. Vibe coding is often used as an umbrella term for any code written with the help of AI, but it's not quite that simple.

Vibe coding, to me, is when you let AI generate all the code and make (nearly) every technical decision. You do not look at the code whatsoever; you only look at the end result. That's it.

There's also skimming, where you barely look at the code, but I would consider that vibe coding as well.

AI-assisted coding, on the other hand, is the kind I use and consider a much healthier way of using AI. I would define it this way: AI is a collaborator, and the person using it has strong technical knowledge and coding skills. Everything is clearly thought out beforehand, AI-generated code gets reviewed as if it were a pull request from a colleague, and you verify the solution on a deeper level, beyond checking whether the end result "just works".

The reason this difference is so important has to do with the long-term viability of a solution and the code behind it. To me, anything that's vibe-coded should be treated as a functional prototype and little more. As soon as a vibe-coded project becomes something that actual users rely on, someone needs to own the project and its code. An owner needs to understand the design, be able to fix issues, and know how to make it scale and expand it in a future-proof manner.

For anything beyond small personal projects, that difference cannot be overstated. A vibe-coded solution works right now; an AI-assisted solution built with expertise keeps working as it grows, because it can scale and be expanded and maintained.

Loss of perceived value as a software engineer

In my experience, it has always been rather difficult to convey to non-technical or less technical people what value a software engineer truly has. This is because the work of a software engineer is inherently judged by the end result, with only fellow technical colleagues being able to truly understand the long and often difficult process leading to that end result.

I'm afraid this will only get worse over time, as AI is incredibly good at producing an end result but still poor at producing a well-thought-out, maintainable one.

It's very demoralising to see people who proclaim software engineering is on its deathbed, as if coding is the only thing we do.

Sure, writing code has always been a large part of the job, and it is the part that has now become the most trivialised by AI. Having spent so much time learning to code, it is very painful to see this once hard-earned and unique skill lose so much of its value. But there is so much more to software engineering than just writing code. Lots of decisions have to be made and design work has to be done before even a single line of code is written. There still needs to be a very deep understanding of the practical business problem you're trying to solve, as well as broad technical knowledge, to be able to direct an AI agent to build a solution that will work in the long run.

AI agents are extremely good at creating what you ask them to. But so many people seem to completely overlook that knowing what to ask them to build, how, and why is not a trivial part of building good software.

Loss of enjoyment and satisfaction

Writing code by hand used to be a fun part of the software engineering job for me. I'm not going to pretend that writing my 80th CRUD endpoint and its five associated tests brought me a lot of enjoyment. However, writing a clean, efficient function or optimising an existing one to be much faster never got old. Thinking through a problem iteratively and getting your code to work by trial and error before finally seeing it work never failed to give me a rush of excitement.

Now, with AI being able to achieve in seconds what, in some cases, used to take me several hours, that excitement is all but gone. It's difficult to be proud of or excited about a piece of code when it isn't truly my code anymore.

New sources of enjoyment

Fortunately, I'm finding new sources of value and enjoyment as time goes on.

The software design and architecture aspects of the job are more important than ever, and I find myself spending much more time on them now that AI is able to do most of the coding work.

I really enjoy being able to test several approaches by having AI generate proofs of concept for each, so I can pick what truly works best.

For a moment, I questioned whether I ever really enjoyed coding in the first place. I wondered whether the feeling of lost enjoyment was actually real or just a vague thought. So I tested this by deliberately doing things by hand again for a while. For things like boilerplate, routine coding work, tests, and the like, I can't deny that I've lost almost all patience for doing them by hand. I gain very little enjoyment or value from writing them by hand, so there's little reason not to let AI do it.

Previously, I would rarely code in my spare time simply because I didn't have the energy for it after doing it for a full day at work. But now, I find myself enjoying working on coding projects outside of work far more than I used to. I've found that I very much do still enjoy the puzzle-solving aspect of coding. Thus, side projects can provide the aspects of coding that I now miss at work. It doesn't compensate for it completely, but it's better than nothing.

In other words, I've concluded that the enjoyment has mostly just shifted.

My interests within the tech field have always been very broad. I tinker a lot with security, servers, and infrastructure through my homelab. My retro gaming hobby involves a lot of hardware and electronics tinkering. AI as a technology is fascinating to explore in itself. These niches have fortunately not been impacted as strongly, and I find myself still getting just as much enjoyment out of them as I always have.

Uncertain future of software engineering career

The uncertainty around software engineering as a career has been discussed broadly, and it's a topic that's been one of my biggest sources of worry.

Nobody, including me, knows where things will end up. Anyone who says they do is trying to sell you something or is full of it.

I went through a short, panicky phase where I was beginning to believe the doomer mentality that the field was truly going to disappear. Right now, I'm leaning towards believing that the field will not disappear, but it will continue to shift significantly.

As previously mentioned, there are still a lot of skills a software engineer has that, no matter how good AI becomes, will always be valuable. Being able to deeply understand, debug, and fix code is something I don't think will ever disappear entirely. If your business grinds to a halt the moment AI agents are unavailable or get stuck, that’s a problem. Code review, diligent QA, and security audits will also continue to require these skills.

What I do think will be true, if it isn't already, is that the same work will require much smaller teams. Or if you look at it more positively: current teams will be able to do much more. Reviewing will likely become the biggest bottleneck rather than writing the actual code.

I can see a scenario where both generalists and specialists still have their place. Generalists can see the bigger picture and do systems and architectural thinking without knowing every technology at the deepest level, making them great orchestrators for AI doing the bulk of the grunt work. Specialists, then, will be the only ones who can truly verify that AI is doing things right and fix things when AI breaks something and can't fix it itself.

The junior problem and beginner’s paradox

As of right now, entry-level jobs are under serious pressure. Yet without hiring juniors today, we will have no new seniors tomorrow. Everyone seems aware of this problem, but no one seems to have a solution.

Newcomers will still need to gain the expertise to properly direct and review AI output. In my opinion, there's no way to truly understand how all these systems work without having reasoned through everything yourself at least once. Or, at the very least, without having gained a very good understanding, which I think only comes from reading, writing, and learning with minimal or no help from AI. You cannot learn how to review code if you've never written it. AI can still help a learner the way it helps me during side projects: by answering specific questions, providing explanations, or offering a pointer in the right direction. But at the end of the day, you still need to understand how everything works yourself.

This creates a weird paradox where you still need to be able to write code by hand to understand it properly, but it seems fairly pointless to learn to write code by hand if you're going to have AI generate it at work.

I've seen similar counterarguments several times, with AI being compared to a calculator or a compiler and people saying that you don't need to be able to do the complex calculations a calculator does or know how compilers work to use them. I strongly disagree. Calculators and compilers are deterministic. You can rely on them to produce the same correct result every time for a given input. With AI, that is simply not the case. And even if it were at some point, wouldn't it still make sense to learn how to code, in the same way a mathematician (or anyone, really) starts by learning basic arithmetic?

Now, assuming that newcomers will continue to need to go through an extensive learning phase, that raises the question: who's going to pay for it? A lot of the learning used to happen on the job, but now companies are reluctant to hire juniors when AI can already do much of the work juniors once did.

Schools will remain essential in this. I truly believe that schools have to keep recognising the importance of learning to write code by hand before moving on to AI tools, no matter how good those tools get. I can only imagine how difficult it must be for schools and teachers to navigate these AI times.

In addition, employers might at some point need to start seeing juniors as a long-term investment to create the seniors they need later. As unusual as that may seem, it's the only real solution I can come up with for the moment.

Double standards and trust issues

While I like to believe I personally have a healthy and well-thought-out way of using and dealing with AI, I find myself holding others to a different standard when they use AI, even when they do the exact same things I do.

Nearly every time I see someone vibe coding or using AI, I feel annoyed.

I'm still not quite sure why this is, but I'll try to explain what I think is the source:

It's difficult to judge whether someone went through a process of thoughtful design, iterative review, and proper testing by looking only at the end result. This has made me distrustful of projects using AI, and I have a tendency to assume the worst. Before AI, working code was proof that someone had gone through that process, but that's not the case anymore.

It feels hypocritical. My own projects made with the help of AI I trust, because I've verified the code to the best of my ability. With someone else's code that trust is gone. I assume they don't look at the code, don't know what they're doing, and haven't produced something that meets the standard I hold myself to.

That makes me distrust both random projects on the internet and code contributed by my own colleagues far more than I used to.

Similar to how I almost never feel proud of code I generated with AI, I find it difficult to appreciate and value someone else's work when a lot of AI was involved. Even when I know that, as is the case with my own work, a lot of human thought and review likely went into it.

What I look for in codebases these days is a clear understanding of the work shown in issues, forum conversations, or PR descriptions, small commits instead of huge ones, and the ability to answer questions about the code and the reasoning behind it. That already tells me a lot more than just looking at the code itself. It's often easy to see when a PR or code snippet is AI-generated, so small signs of human involvement help me trust a random project much more.

To circle back to the many AI-generated logos, flyers, etc.: For these use cases, I find myself being mostly anti-AI, despite using AI in my own field. Related to this, it greatly disappoints me how quickly people are satisfied with subpar or mediocre results. Be it visuals, code, you name it. This ties into the time-saving aspect I mentioned earlier. Not everyone has the time to learn how to code or use Photoshop, but it's disappointing to me that people take these shortcuts and either don't recognise that the end result is subpar or recognise it and don't care enough to improve it or look for ways to get better results. I feel like this is part of the reason for my general grudge against others' AI use.

Impact on the retro gaming / computing scenes and other hobbyist communities

In case you hadn't noticed what site you're on: retro games, consoles and computers are a major hobby of mine. I've noticed the impact of AI in the different retro communities as well.

Over the past few months, there has been a significant uptick in the number of vibe-coded games and projects for retro consoles. Vibe-coded decompilations and PC ports of games are also popping up everywhere.

Whenever a project like this turns out to be fully vibe-coded, I've noticed the reactions split into two camps: people praising and loving it, and people tearing it to shreds for being vibe-coded. There seems to be very little in between. The main complaint I've seen is the lack of quality control and the number of bugs, as well as a general dislike for AI out of principle.

I agree that quality is the biggest issue, especially with PC ports of decompiled games. Teams used to spend ages polishing these to infinity, with a lot of technical knowledge and skill. Now they're vibe-coded in days. They work, but are far from perfect quality-wise.

This circles back to what I said before about people seeming to be easily satisfied even when the results are subpar. I don't inherently have a problem with these projects using AI. What bothers me is the lack of passion and care. I personally have little to no interest in playing homebrew games or ports if they're completely vibe-coded and not backed by any sort of knowledgeable team.

What I think could be incredible is if those knowledgeable teams used AI to their advantage. I think long, tedious processes such as decompilation could be sped up significantly and still maintain a high standard of quality if skilled developers used AI tools. Unfortunately, this rarely seems to happen, as many of the skilled engineers in hobbyist communities seem to be in the camp of wanting nothing to do with AI at all.

Like I've mentioned before, when you do these things as a small personal project, that's perfectly fine in my eyes. In fact, just as I do with some of my own side projects, I think it's great that people can now see their ideas materialise without needing to invest huge amounts of time learning how to code or how specific console hardware works.

But as soon as you publish it as an open-source project and tout it as this amazing thing you created, it goes beyond a simple personal project in my opinion. At that point, I'd at the very least expect to see proper credit for other work that your project is based on, which is often lacking with these AI projects. Like with any piece of software, if you intend for others to use the project, you're often expected to have at least a decent understanding of what you built so that you're able to maintain it properly.

One problem that I've seen mentioned a lot is that it's seriously demotivating for people who want to make an expertly ported, bug-free version to see someone get ninety percent of the way there in a day or a few days when doing it properly takes months or years. I understand this sentiment and have felt it too. I've been trying to learn C and how to develop something for the CD-i. It has been a slow, difficult process so far. Seeing a flood of vibe-coded CD-i projects pop up has a very demotivating effect on me as well.

Of course, this is not the intention of the people creating these projects, but it is an unfortunate unintended side effect nonetheless.

One of the first posts on this blog was about Nobelia, a homebrew game for the CD-i that was developed during the COVID pandemic by one person with some help from the small but dedicated CD-i community. It went on to become the first truly original, substantial CD-i homebrew game in years. Before that, there had been merely a few other attempts at making a CD-i homebrew game. Only one, Frog Feast, was finished; it was a very simple game based on an old Atari/Intellivision title.

I watched a lot of the process. Months upon months of trial and error went into figuring out ancient C tools and libraries. I was amazed at the end result and appreciated the effort so much. It ended up as one of the best games ever made for the CD-i. I would personally still appreciate and look up to such an effort today, but I'm not sure many others will.

As with coding in general, making these projects that once required deep programming and systems knowledge has essentially become a commodity. Anyone can vibe code a CD-i game now. Almost anyone can get a decompilation project going now. Sure, not everyone can code it themselves. But this circles back to the perceived loss of value. A part of the community no longer seems to consider being able to build these retro projects a unique skill worth appreciating. Thus, most of the motivation and satisfaction for people still doing things by hand has to come from the process of learning itself. The end result has lost a lot of perceived value because it can be achieved so easily.

Saying "I achieved this by doing it myself, by understanding everything" is now mostly valuable to you alone. I don't think many people will look up to that the way they used to, especially as time goes on.

Personally, however, I will always appreciate the people figuring all this out. More than that, I really miss this feeling of being in awe of something that someone was able to create on their own. That feeling simply doesn't occur when something is vibe-coded.

Hopes for the future

For this final section, I'd like to express some hopes for what the future will bring. These aren't predictions per se. They may not even be rational or likely outcomes. But if AI is going to stick around, this is how I hope things will turn out.

Given the increased productivity afforded by AI, I hope we'll be solving more problems and providing more value with the same number of people or more, rather than solving the same number of problems and providing the same amount of value with fewer people. Mass job losses are always tragic and should be avoided at all costs. More output per person doesn't have to mean smaller teams or firing your entire engineering department and replacing it with one person and a laptop with Claude. Cheaper coding could raise the demand for software, so the same number of people, or more, could keep working and simply produce a lot more.

As time goes on and AI models become more efficient while hardware becomes more powerful and accessible, I hope we'll see a shift away from the current handful of big tech AI giants towards more self-hosted models.

I hope the overblown hype comes down to a more reasonable level. The hype is not entirely unwarranted, as AI is a new and exciting development that has many great use cases. But I'm so tired of seeing it praised as the solution to everything. Not everything needs AI.

Somewhat related to this, I hope the breakneck pace of new model releases and improvements will slow down. Not because I'm against progress, but because I think the world could use some much-needed breathing room to catch up with the rapid development of this technology. It takes a lot of time to figure out how to deal with tech that is as disruptive as AI.

Perhaps more than anything, I hope that people will not let their creativity, agency, and unique skills get lost or go to waste. AI will hopefully continue to be used mainly as an amplifier for people’s own work and productivity. It shouldn't be a means to replace them, nor should they lean on it so much that their own value and skill get lost. I'd also be very disappointed if we stopped appreciating the amazing work of human artists, creatives, and other people continuing to do incredible work.

I've seen several developers I looked up to go all-in on vibe coding, which is very disappointing. I hope others will be able to maintain a good balance and will not completely give in to the AI hype like some have. We need these voices in the industry to continue being vocal about the value of proper design, review, and human involvement.

For my own work, I hope that there will continue to be challenging and interesting projects in the tech industry that I find enjoyable. I don't want to become a manager of bots that do all of the fun parts.

What it mostly comes down to is that I hope AI will continue to be a tool for those who choose to use it for tasks where it makes sense, not a tool to replace all human skill and creativity.

Wrap up

It's safe to say that my stance on AI as of late 2026 consists of a lot of mixed feelings and uncertainty. There are enough thoughts and topics left to at least double the article's length, but I think this will do for now. Even if nobody ends up reading all or any of it, it will serve as an interesting time capsule for me to look back at in the coming years.