Matthias Mair
Hynek SchlawackAs someone who’s been maintaining FOSS projects of various levels of popularity for more than a decade, I need y’all to understand one thing: LLMs didn’t change the median PR quality. (1/6)
In fact, most PRs were flawed and needed significant work from my side before they could be merged. Often, more time than the contributor invested. So any argument that slop is going to ruin a project because it has an AGENTS.md is completely disconnected from reality. As far as we know, Jia Tan was not an LLM. We’ve been complaining about code quality for decades (remember discoveryd?). (2/6)
What SHOULD concern you is how processes and gatekeeping are set up and breaking down. A project is merging PRs/publishing new versions at a significant higher pace than before? THAT’S a red flag. LLM review can be used as a tool to improve existing review – or as a Potemkin-village-style faux process full of Markov-chain sycophantic glazing combined with hallucinated 0-days. (3/6)
Of course, circling back, thanks to LLMs, the absolute number of contributions skyrocketed which means the number of trash PRs went brrr. Because ultimately what the current crop of AI is, is LEVERAGE for existing skills. If you don’t know what you’re doing, you will cause lots of harm. As I wrote the other day, this is a huge problem. (4/6)
Firstly because it overwhelms hands-on maintainers like me. When I waste hours of my time on humans’s subpar PRs, I can at least lie to myself that I’ve made the world better by teaching impressionable souls useful skills. This is obviously not true if I’m talking to an emphatically soulless Claude. (5/6)
Secondly there is an insidious tendency to subvert review processes. We’re all just human, and sometimes hearing that clicking a button will solve our review queue is exactly what we need to hear. The pressure is right there. This is where the harm is.
I understand and appreciate that y’all believe that LLM-based technology is harmful and it certainly is in certain ways. And I understand the tendency to fight evil on all fronts. But this is friendly fire. (6/6)
(P.S. yes it’s a marvel that Claude Opus 4.5+ exists; no it’s not even close to production quality; if you think it is, you get a single-9 uptime like GitHub or Anthropic 🫳🎤⬇️)
Whine Spectator@hynek I'll give you five nines as long as I can put an eight in there somewhere
David Lord 
@hynek way to call us all out :-) sure we were mostly shouting into the void back then too with reviews on low quality stuff, but at least the void didn't reply back with an essay starting "you're absolutely right"
@davidism Heh a while ago I saw an allergic reaction of you to someone quoting ChatGPT to you in an issue and I was like “OK, that seems a strong reaction” and the other day someone did the same thing to me and I almost instantly banned that dude. It’s. So. Dehumanizing.
Chris Adams@hynek I was thinking about this earlier tonight when lookigng at a PR which had a generic set of claimed improvements but made none of them. Curious whether this was karma farming or the erstwhile next Jia Tan, I looked at the hundreds of PRs they’d opened. Many were abruptly closed but some were a sea of LLMs praising that user’s non-contributions or going back and forth on review points for hundreds of messages. I can only imagine how many person-hours of time they’ve burned in aggregate.
@acdha yeah i’m kinda grateful when people are brazen like that because it helps me to classify it immediately. the only problem is when something _looks_ useful but is buried in a mountain of slop