Since late 2025 I had the feeling that I am losing my shit. While I did publish “Kann KI verschlafen werden” (Can AI be slept away) in September, the people of the church of AI got louder by the day. And at some point, that leaves marks. I don’t doubt my skills and abilities, I know how I work, I know what I find useful and what not, but still.
Things I never doubted:
- AI tooling is a tool of the few, deskilling a major part of the workforce
- AI has tremendous costs, obviously environmental, but on a same scale, socially when training models and later, leaving ecosystems in shamble
- AI as such—local model give or take—is costly as fuck and raises the barrier of entry and sustainability.
There is more and luckily, other people wrote about it. Like “Feeding the Machine”, from James Muldoon, Mark Graham and Callum Grant, or Resisting AI, an anti-fascist approach to artificial intelligence by Dan McQuillan and essentially anything tante wrote the last months.
However, that doesn’t help me with uncertainty and self-doubt. Ignoring and hating is cheap, but at some point, I need to know and understand what makes people rave about AI coding tools that much, first hand. Late to the party is an understatement here. In January 26 i tried a couple of things: I used Cursor with scrobbles4j and created three changes: 1, 2 and 3. These changes are indistinguishable of code I would have written myself, the tool even aligned well with my testate (commit 2). Fun fact: That was the change I needed to control the least amount, speaking for the fact that those tools are text rehashing machines.
Did Cursor make me faster? Hell no. Here’s the transcript of the session. Note how I did give very specific instructions, pointed several times to CONTRIBUTING.md, how the tool struggles with my Maven and Git setup (they are working well, believe it or not and thank you very much…).
My normal mo here would be: Let’s go for a run or ride, think about stuff and when back home and showered, just write it down. That is literally how my head works. Both in writing code, texts and books.
Now I knew how a tool would work on a small, yet a bit unusual code base (I doubt that the combination of Quarkus, Cute, JDBI and very SQL focused programs are that common). I honestly felt: Relieved. If I abstract away the inherit problems, I can think about it as “yet another tool”, that is easy to use and you don’t need special knowledge for it (put in another way: Don’t believe the evangelists of “Prompt engineering”, if you are able to describe a problem to a human, than you are just fine).
Second test was using IntelliJ Junie with a gigantic codebase that is Neo4j (Yes, I did use the company account for that, obviously) and help me getting my feet wet with Scala. It worked to some extend, yet I did not know the specific parts I was working with as well as my own pet-project, so my success of describing the problem was not as good. What helped me much more was, drum roll, the quite excellent Tour of Scala, old fashioned reading and later on, my new colleague who patiently described and explained my mistakes.
So that’s the me part: I won’t change the way I work, really. The “tool xyz one shotted my fab and never heard of app idea in 3 hours” is still bullshit. Of course, something runnable will come out of it, but only if you live this as Fast fashion. If you guide the tool, you invest the time you would probably use to hone your skills into the machine (and probably it will also be used as training), but you get decent output.
That brings me to the next part: AI assisted contributions such as this. I liked the interaction with Oliver a lot, it was positive before, and positive in this issue again. He disclosed that he used Claude, and apparently, the changes are not invasive at all. I did rewrite about a third (see this commit) so that I a) know what is in there and b) are happy to own it.
I posted this. Some people starred (liked?) it, some had questions. This is an answer to the latter. I don’t live in isolation, and the world keeps on changing. I don’t like all changes (in fact, most of them are terrible), but at some point, one has only so much energy to fight all of them. And honestly, that’s not the point: I will not integrate AI as my daily driver. It is definitely not as good as the church of AI promise it. But I will not condemn people that made decent use of it and even disclose it. Berating them is in my opinion on the same level as berating people on their travel and other lifestyle choices. I can only lead by example here. Obviously, this is not speaking of garbage PRs that consists of 20% emojis and 80% junk changing 50% of a project because someone had fun “one shooting” a feature.
No comments yet
Post a Comment