Back in february, I cancelled my Claude Code subscription. Then a few days later, I subscribed again, at a higher tier… For a few years now, my upstairs neighbor has been really noisy and instead of losing it, I did the next best thing, gather data. Over a few days I got Claude to generate an android application that I installed on an old phone, and started recording their noise while monitoring the background sound level. I might make a separate post about the app because it ended up pretty nice.
The app was developed in languages I do not know: Kotlin, Faust, and Java in a domain I was not familiar with, audio signal processing. Would my rusty engineering management skills be enough to get something out of the LLM? TL:DR; yes. Setting a frame of reference, automated quality tools, providing examples of how to do specific things, building a knowledge base for the project, making sure requirements are shared and there are no ambiguity left, that sort of things.
I could have learned all this, but the timescale would have been months instead of a few days. It's fair to say I wouldn't have done it without LLM, which brings me to the next few points:
- I'm not over any of the problems of LLMs about their training, their ressource needs, etc.
- I don't believe individual actions are enough to reduce the harm done by LLM, they're not pointless but let's make sure we're not delusional about the order of magnitudes,
- I don't think LLMs will be the solution to the problem they're causing, breakthroughs will come from smart or lucky folks (probably both)
- LLMs can make some parts of my job easier,
- Using LLMs to generate an average or below average solution is better than having no solution in some cases.
- LLMs are not deterministic, but in practice they are deterministic enough.
I mean humans are not deterministic either and we can get working code out of them, it's a process issue. On a project I'm working on, another developer is using Claude, the code is OK, better than some of the code I saw written professionally over the years. For me I was able to cut down implementation time of some bugfixes by a significant amount using LLMs. I also used it to configure, make various scripts necessary to populate and serve search.tresbien.tech on top of the android app.
There are good things that can come from the AI/Machine learning field, LLMs? probably not, specialized models might be a better answer. LLMs are centralizing power in the hands of a few while we argue amongst ourselves. How long until software development is treated as so called unskilled work? probably not long, and infighting is only accelerating the process. Let's celebrate (no) the fact that software development being treated as unskilled work probably means there is going to be a lot more diversity in tech!
I hate what's going on globally, while locally I can find positive impacts. Makes me with wish I paid more attention in philosophy classes to deal with the dissonance better. An arguably better written take can be found here: I used AI. It worked. I hated it.