Another conspicuous thing is the lack of vibe-coded PRs on mature open source projects. Maybe it's because these projects have erected policies limiting AI contributions, but given the high scores on SWEBench, you'd expect _something_ to come of it?

And yet in real world use you get stuff like https://github.com/scikit-learn/scikit-learn/pull/32101 (not to pick on that particular author since I pulled it completely at random. But it's notable that this also was not a fly-by-night PR by a complete newb. The author seems to have reasonable credentials and programming experience, and this is a fairly "self-contained" PR limited to just one function. Yet despite those advantages, the LLM output couldn't pass muster.)