Trying to discover the magic that people feel around LLMs
Trying to discover the magic that people feel around LLMs
I’m trying Aider with Qwen3.8-27B locally on an M1 MacBook Pro. It takes a bit of time, but it works through the task given. Going from natural language task to source code change translation is really neat. There is no getting around it.
If you look at what is happening, step away and think about what the tools we have available really mean for working software: You get into a sort of realization that some software developers (myself included) can be sort of thought of as boiled frogs. We have been using the tools for a while. I guess even the old tools from like IntelliCode you could see the seeds (2018-2019). I tried it. I found it distracting and useless. It was cool and I was not part of the target user group. Having inline auto complete still feels distracting and annoying, it feels like bad UX when you know what you intend to write. I do feel that the user experience of sending tasks as prompts is way better.
I asked ChatGPT for a timeline of the hype around AI that I’ve heard from some people talking about how most knowledge work will be replaced by AI and only managers will be left (2017-2019). ChatGPT told me that it started way earlier as can be seen in this MIT Research Brief in 2012. I knew about the AlexNet’s 2012 ImageNet result and thought it was neat. I really didn’t talk with the people in the know, the people that knew important people. You know the type? The person that name-drops some famous person now and then in order to make sure that you know you are out of the loop. People in the know told me that AI was the thing. Maybe this iteration of what is called AI will follow the same pattern that other AI algorithms have followed: The use becomes boring and invisible once people figure out what it is good for.
I’ve been trying to go further than that. Give into the vibes, try to see what happens when I give up and speed run agile process on agents for a non-trivial but known task. As an experiment it seems they might do well, but agents seem to follow the same patterns as software teams hounded by a manager that tries to get them to deliver on time. What you get, fast, is essentially brownfield-style code. The big assumptions went unchallenged. Even when using Fable, the code was kind of what you would expect someone to write while stressed. Perhaps we should see it as when you avoid looking at the code, you act the same as a manager without all of the messy details. Perhaps the probability space of such prompts implies that kind of solution because that is usually how it has been done. I can see how I will do it again for things where I do not expect to need regular maintenance.
What can I say? I’m not discovering any magic. People have told me that programming is kind of magical. The experience when done daily becomes ordinary. Tools that have been called AI in previous iterations of hype cycles became just tools, hidden away once people understood how to make them useful. The magic isn’t something to discover: it’s what a tool feels like before it becomes ordinary.
Tags
Do you want to send a comment or give me a hint about any issues with a blog post: Open up an issue on GitHub.
Do you want to fix an error or add a comment published on the blog? You can do a fork of this post and do a pull request on github.
Comments