AI Journal 1: The Slow Surrender to AI
2026-05-27
I have wanted to write about my AI journey for years. But each time I sat down and gave it some thought, I was unsure about how to express my thoughts about it. Every draft I made also become obsolete in a matter of months, sometimes faster, because AI just progressed that fast.
I'm an experienced software developer that has been coding professionally since the late 2000s, and for fun since the early 90's. I was an early adopter of ChatGPT, and have been using it for code related stuff since it came out late 2022. I also vaguely remember and time before that when you could play around with early GPTs in the OpenAI sandbox and make it write simple code for you. Back then, it was a novelty. We never thought it would become good at it in just a couple of years time.
What I want to do with this series is document for myself (this is a journal, after all) my progress from toying around with early GPT to basically using Claude Code (and similar) for all my coding needs today. I'm not at all comfortable with the current state of things, and looking back and reflecting on this journey is my attempt at coming to terms with what the future might bring. On one hand, I really enjoy using agentic tools. On the other hand, I'm worried what that means for the state of software development in the future.
The early days of wonderment
When ChatGPT came out, it was a really neat and fun toy. Like many, I saw the potential quickly. I was a teacher at the time, teaching software development at a local college. I discovered ChatGPT before my students and was quick to integrate it into my teaching. It was fun seeing the students attempting to prompt their way through solutions for simple programming exercises. I deliberately gave the students some problems that ChatGPT could solve easily, and some that I had verified would make it run around in circles. I was confident ChatGPT would never be able to replace a real programmer despite being useful for certain simple tasks.
Using AI as a shortcut
The year after, I quit my teaching job and went back into software development. Some today might call it "software engineering". In the future, I'm sure it will be called "agentic engineering" or something like that. To be honest, I'm not sure how to accurately label my work anymore. Job titles come and go.
Anyway, I was a bit rusty, but eager to get back into the game. Certain things had changed in the last decade and I wasn't fully up to speed with all the new developments. Long story short, I used ChatGPT to brief me on simple stuff like how to write a proper docker-compse.yaml file, how to write WSDL files (for SOAP services) that could validate with the proper tools -- simple stuff like that. I felt like I was cheating. I never read a book or asked anyone for help. I just used a convenient tool that gave me all the answers. I also quickly ran into problems trying to mangle a certain clusterfuck of xsd-files into something that could generate a proper Java client for a service I had to work with. That stumped me, and ChatGPT as well. It just went back and forth with code that never quite did what I wanted. One of my new colleagues took over and saved me. He made it work by copying something from another project he had recently worked on. That irritated me. I wish I could have found the answer on my own or that ChatGPT hadn't been so useless. I quickly got over it though, and in less than eight months, I improved to the extent where I started to take full responsability for many new services and deployments. I could feel how everything got easier as I started to get comfortable yet again in the role as a software developer. In less than 18 months, I became one of the most well-rounded developers on the team. Other team members would often come to me to discuss tasks and problems they encountered in their work, and more often than not, I was able to help one way or another.
I still used ChatGPT quite a lot, but mostly to discuss small pieces of code, architectural decisions, and other simple matters. Like, when I had a certain snippet of code, but I wanted to rewrite it in a more functional style using Java streams? ChatGPT was good for that. It was a partner I could bounce ideas off, a rubber duck, a human-like linter. It was many things, but it never took over any actual work. It just wasn't capable enough. Yet.
Using AI to speed up coding
I've had my eyes on a better "auto complete" tool since the early days of GPT, so when the beta for "GitHub Copilot" was announced, I was quick to sign up. Remember when Copilot was just a plugin for your IDE from GitHub? All it did was autocomplete code in your IDE where your cursor was, and it was fairly good at guessing what code you wanted. I quickly learned to optimize my "context", writing good code comments before letting copilot fill in "the blanks".
For the first few weeks, I didn't tell anyone. Again, I felt like I was cheating. Here I was, writing code faster than anyone else. Especially tests. I hated those, and copilot was especially good at churning out unit tests when some were already written. Code coverage was at an all time high when I was in charge. I think it only took a month before I fessed up and let my team mates in on my little secret. They all thought it was a pretty neat tool, and soon after, I did a small presentation for the whole company (around 20 people including the CEO) with my experiments and findings using ChatGPT and Copilot to assist me with coding. Everyone liked what they saw, and 20 minutes later, copilot licenses were bought for the whole company. Up until then, I had been paying out of my own pocket for access.
ChatGPT becomes better
For a while, everyone uses copilot, and more people start to rely on ChatGPT for more complex matters when they got stuck on something. Even the CEO uses ChatGPT for his writing (sales materials, etc). It becomes normal to ask "have you tried to solve it with ChatGPT?"
I no longer feel like I'm cheating or holding back a secret. I'm one of the believers, and despite everyone having access to the same tools, it doesn't change the fact that some people are simply better at using those tools than others. I thought AI would be an equalizer, but an inexperienced developer will still make bad descisions despite having access to the exact same ChatGPT and Copilot I had access to. Team mate would still ask me for help occasionally when ChatGPT didn't deliver. I was still performing well despite no longer haven the "AI edge" over everyone else. As it turns out, using AI is not cheating if you are using your skills and experiences to utilize it in the right way. AI doesn't turn a newbie into a pro. My performance is still because I'm actually good at what I do. AI is still just a tool.
The honeymoon phase is over
I don't remember when I first heard the term "vibe coding", but I knew instantly it wasn't for me. I used AI to become a better developer, not to offload my work to a computer. Also, I knew from previous experiences that an AI could never perform and write code as well as a human like me. I actually believed that well into 2026. Looking back, I was in denial. I truely believed that vibe coding was a dead end -- and I still kinda think that today -- but that belief had the side effect of making me blind to another recent development: agentic coding. You see, whenever people talked about using AI tools based on LLMs to write code for them ("cursor" being and early example), I instantly associated that with non-programmers producing slob by vibe coding. And I knew that "this was not the way".
I've been a reader of Hacker News for many years, since 2011 at least. Early 2026, I was complaining to a friend that Hacker News had been taken over by vibe coders. Everyone was into it, and all the sane people had left the site. You almost couldn't open a single news story without someone talking about how they did this or that using "claude code", and everyone else would applaud. I was baffled. Couldn't they all see that all they were doing was producing unmaintainble slob? Was I getting old?
I made a conscience to downloade claude one that to see what is was all about. If I wanted to critize it, I had to get some experience using it first. My first attempt was stopped because I couldn't find the guide to install the CLI tool. The desktop app was only for Windows and Mac, and I was using Linux. Only later did I realize my mistake (claude code is both CLI and desktop -- two different products, it seems). In the meantime, I got OpenAI Codex to experiment with. It was an eye-opener. Despite being an inferior experience to claude code (at the time), it worked much better than I had expected. It was quick to open, easy to use, and so smooth. It looked at my files, came with suggestions, did a few edits on it's own after I gave it permission. That day, I realized that it wasn't just all hype. This was actually pretty useful. It took me another few weeks until I realized that I could never go back. Agents had arrived. And for once, I wasn't the one holding back a secret.