When it's the maintainer who's AI-pilled

Thomas Karpiniec normally steers clear of projects that lean hard on generative AI. Too much slop out there, he says, and he'd rather bet on the mature, hand-crafted option when one exists. But dirge, a small CLI coding harness, caught his eye through a blog post about squeezing better performance out of smaller LLMs. So he installed it anyway.

He hit two bugs almost immediately. One was a simple unfinished feature. The other was messier — his model kept losing track of state after a compaction step, the kind of fuzzy problem that's hard to describe precisely, let alone fix. Karpiniec filed both as GitHub issues, half-apologizing for how vague his reports were, figuring he'd eventually circle back with a patch or better logs.

He didn't need to. The easy bug was fixed in an hour and a half. The gnarly one, the one he wasn't even sure how to reproduce cleanly, got fixed in three hours, complete with a regression test. The maintainer, it turns out, is at least as AI-pilled as any contributor showing up with an agent-written PR — maybe more so.

That flips a dynamic most open source discourse hasn't caught up to yet. The usual worry is about AI-generated pull requests landing on old-school projects whose maintainers have to manually untangle machine-written code they didn't ask for. Karpiniec's experience was the mirror image: a maintainer already fluent in AI-assisted development, treating a well-written bug report as essentially a prompt. No PR needed, no context-switching to evaluate someone else's contribution. Just describe the problem, and the fix appears.

He's careful to note this only works because the maintainer clearly supervises and curates whatever the model produces — same standard he holds his own AI-assisted code to. But if this becomes more common, and he thinks it will, the etiquette of contributing to open source quietly changes. Writing a sharp issue might soon be worth more than writing the patch yourself.