Just seeing the list here: https://codeberg.org/small-hack/open-slopware
I don't know what to think about it, many incredible open-source projects went downhill, some worse than others, full AI permission usage and some of them even advertise AI providers on README.md. I'm even using many of them myself.
Even the good guys are falling, I'm not sure what to think about it. Am I overreacting maybe?
You might think, that's fine, not a big deal, some of them just allow AI usage, but not AI generated code, but for how long? If you allow use of AI for anything the tendency is that you'll be even more open about it in the future.
List of projects that personally draw my attention or I use eventually:
- Firefox: not unexpected, but still, I had hope on Mozilla bring more tech awareness on mainstream
- Spaceship prompt: I use this on my terminal for customization, why'd you need AI for such a simple project?
- VLC: just sad
- curl: sad x 10
- Vim: sad x 20
- zoxide: they literally promoting AI providers in the README, such a simple tool as well, why?
- CoMaps/Organic Maps/OsmAnd: the few ones providing a good alternative to Google Maps
- Element: that's literally the most used client for Matrix I guess?
- Python: I thought they were the good guys as well
- Lemmy: unexpected, code of conduct says it's allowed
- Linux: the final boss, unbelievable
Is there any hope at all? Or am I just overreacting?
I use Claude Code for work, and before that I used Cursor. I'd say I have about a year of total experience with the tooling, and I've used all the various AI ephemera: MCP integrations, RAG context management, skills, vector databases, custom orchestration laters, etc, etc, etc. I've built AI-powered tools like customer service chat bots as enterprise projects. I've even experimented with using AI in personal projects just to see if the experience is any better than working on large enterprise projects with years of accumulated technical warts.
It's all crap.
Actually it's worse than crap, because at least a turd has the decency to identify itself with a foul odor. AI code is insidious. An LLM is just a statistical model of grammar, so its output always has the same statistical features of correct code. It looks correct, and when it's wrong, it's wrong in ways that can be very hard to spot. Of course there are also times when it's just brazenly, comically, catastrophically wrong right from the jump, and it's always fun to point and laugh when that happens... but the other times, when it's almost right, when it's right enough to slip an immaculately subtle, server-killing memory leak past code review and past QA, it stops being funny.
Now your production app is dead in the water for no clear reason, and you've got a release candidate made of a dozen AI-generated PRs, any or all of which could be responsible. And at this point you realize that you don't actually know this code anymore. Yeah, you prompted it. You guided the agent through the process and you carefully reviewed its output... probably. Were you actually paying attention that day, or did you auto-approve Claude's changes while you were in the middle of something else? Either way, you didn't actually sit there hammering out the keystrokes and getting that reinforcement learning that only comes from physically engaging your brain with a problem, so now you're not even sure where to start debugging.
And by the end of it all, you've burned a zillion tokens — which probably isn't something that matters to you right now, but it will in 6 months or in a year when the big AI firms are forced to IPO or run out of money, and they all start scrambling to close their utterly massive revenue shortfalls by dumping subscriptions and switching to usage-based models. Suddenly the $200 a month you were paying for a Claude Max subscription is $200 a DAY, and now your passion project costs triple your mortgage.
Have you considered creating a skill telling it to make no mistakes?
People have unironically suggested this to me and I've tried it, because... fuck it, why not?
It doesn't work. :(
Oh I see you’re telling it to make code for entire features and just accepting what it made without going over it yourself, or … I’m gonna stop here. I know where the actual problem is.
Sometimes, yeah, when management is constantly breathing down your neck telling you that you need to be more "agentic". Just typing a prompt and waiting for the LLM to spit out a feature is the fantasy workflow that a lot of orgs are pushing for, and it's what companies like Anthropic keep pretending is possible.
But even the more careful and focused approaches still produce crap — or, after much hand-holding and angst, they give you something that actually does work while conforming to the styles and standards of your platform... but only after you've spent more time babysitting the AI than it would have taken you to just write the damn code yourself. And, again, you've burned a shitload of tokens in the process.
The only thing AI coding tools are reliably good at is writing "boilerplate", but non-AI tools have existed to handle that stuff for decades... and honestly, if your project has that much required boilerplate, it's probably not very well-engineered in the first place and the AI is just acting as a crunch holding up weak architecture.
The only way to go over something yourself correctly is by doing it yourself correctly.
You: fire all developers but yourself.
Got it. 👍
Only if you want to go over everything.
If you don't you either need to make someone else responsible and accountable for going over something, or be accountable for not being responsible.
Remember:
🥱