New blog post:

[82mhz@mastodon.bsd.cafe]

New blog post:

Thoughts about coding with AI
http://82mhz.net/posts/2026/01/thoughts-about-coding-with-ai/

#blogging #blog

[nina_kali_nina@tech.lgbt]

@82mhz beware that things are not so rosy when people decide to use genAI for addressing issues in complex projects, or trying to solve complicated issues. Reading AI-generated thousand-line code submitted for review often makes me feel like my brain is melting.

Imagine someone submitted an AI generated image with extra fingers and different eyes to you for review; you might want to try and understand why those things are "wrong" and "broken", but there's no why, they just are, that's how the thing works. And of course you're often told afterwards that details aren't important, the big picture is, and the big picture is human-shaped. After seeing enough misshapen hands and extra legs, your brain stops paying attention to them, it's just background noise. When a customer finally comes complaining about wrong number of fingers, you notice you've lost your painting skill; the number of fingers is right, because you're a human and know how to count, but the drawing doesn't really look right anymore. Kind of.

[mirabilos@toot.mirbsd.org]

@82mhz urgh, fucking slop, fucking burning down the planet for nothing, relying on fashtech built on stolen works. Do not do this.

@stefano I’m disappointed that you even boost this.

[irgndsondepp@gts.da-miez.de]

@82mhz @cinimodev
I'm still processing my thoughts on Dom's post. It hit me real hard, because it touched on points I observed and figured out too around Christmas, when we did that AI Hackathon at work.
But Dom's post, and yours too, go beyond that. I didn't get to the point where I felt guilty or maybe hollow, because I hadn't done any work. Maybe that's because I did it for the experiment, or maybe it's because I always knew that I'd out it down after two weeks. But I can imagine how that feels. I felt like an imposter for the first few years when I started working and it was rough. I only managed to break out, when "that one" coworker started asking me questions about the code and asking for my opinion. (I guess today I have become "the one" at work.) 😅

[nina_kali_nina@tech.lgbt]

@82mhz and that's just a slice of technical issues; there are far more 🥲

[stefano@mastodon.bsd.cafe]

@mirabilos @82mhz the post is against it - that's why I boosted it.

"But if I avoid doing everything that's hard and outsource all the thinking to AI, then I will pretty soon turn into a lazy and stupid blob who isn't capable of thinking for himself anymore, and that's a fate I want to avoid at all cost."

[mirabilos@toot.mirbsd.org]

@stefano @82mhz doesn’t matter because he used the planet-burning thing, in an actual attempt to solve a problem at that

[alecsargent@social.linux.pizza]

@82mhz I'm lazy and to me it generally feels like less effort using a search engine. The amount if time figuring out what is wrong with the answer the LLM gave you could be spent actually learning the thing or at least skimming documentation to get to the answer directly.

That is not mentioning the way you described in the post that makes you feel like you have cheated or like you have a butler that that goes to the library and does your own homework.

In my opinion, even if it was harder doing it by yourself, it will get easier and easier to do that particular thing. I've found that I'm searching less for solutions online or at least I know better what to look for. Nonetheless I think there is nothing inherently wrong with the tool and is useful under some contexts.

Good post man!

[tiotasram@kolektiva.social]

@alecsargent @82mhz the reputable studies I've seen back this up that for serious work (or even probably these kinds of projects if amortized) the "productivity gains" are negative, even though subjectively they seem real.

[alecsargent@social.linux.pizza]

@tiotasram @82mhz That is a relief, because I actually felt less productive and people were telling that I've must been using it wrong.

[tiotasram@kolektiva.social]

@alecsargent @82mhz here's the paper I was thinking of:

https://arxiv.org/pdf/2507.09089

There are a *ton* of low-quality studies that do things like "measure" productivity by asking developers how productive they were, or giving them toy problems to solve where there's no preexisting code and there's a known valid solution.

[srgower@ottawa.place]

@82mhz I love the feeling of doing it myself, for sure. At the same time I will delegate small coding tasks to LLMs without thinking about it, usually because I don't fully grasp the code needed to make things work. I'm thankful at least in most cases the bot will tell me "the code being used here is to do XYZ" etc. But I know it's not the same as doing it yourself, because I'm being told what it's doing rather than going through trial and error.

It's something I go back and forth on a lot.

[82mhz@mastodon.bsd.cafe]

@stefano I thought about responding, but I decided to just block the guy. He didn't seem interested in having a conversation anyway.