(Bsky thread) "This turns the maintainer into an unwitting vibe coder"
27 points by isuffix
27 points by isuffix
(Lmk if there's a better way to mark/post a Bluesky thread like this or if I just shouldn't put it on Lobsters)
The main thing I do when reviewing PRs is assess the competence of the contributor, and try to predict to what degree they will choose to follow up should anything go wrong with their changes.
If the answer to these questions is "high" and "sufficiently", then I'll typically merge the PR even if I spot non-critical bugs or other issues.
Or, if I have the time and energy and the person seem to be looking for mentorship, then I'll engage that way.
With this framing I tend to detect a vibecoding situation pretty early and then I just block the user without giving them any feedback and move on. I don't worry about false positives either. There are plenty of nice, considerate people waiting in line for PR reviews, no sense in wasting even a moment trying to explain to an inconsiderate person why they are being inconsiderate, or to determine with certainty that they're vibe coding. Block and move on. I'd be interested in exploring something even more drastic, in fact.
I'm sure some communities can do fine embracing LLM coding as part of the workflow but certainly, forbidding it is a reasonable strategy for community building, particularly one governed by ZSF which includes in the mission statement:
to support and facilitate the growth of a diverse and international community of Zig programmers, and to provide education and guidance to students, teaching the next generation of programmers to be competent, ethical, and to hold each other to high standards.
Remember that if your programming community embraces LLM coding, there is now a division between rich and poor that you didn't used to have. Programming together as a hobby is not perfectly equitable - nothing is - but sans monthly AI subscription fees it was pretty damn good along this axis.
This seems exactly right. So much energy is being spent trying to decide if something is AI. But what we usually want to know is if someone is any good... If AI is bad then blocking bad stuff blocks the AI surely but also non AI stuff we don't want which is also valuable keep out. Etc.
This is a sort of gotcha argument. Yes this could happen, but as a maintainer you don’t have to just give comments if there’s too many rounds of this. Also people who use agents can also think about the comments.
Finally, the “familiarity drops” argument is a little much. When I first look at an open source codebase I won’t have much familiarity and an agent can help to read into it.
It's a good point, and a toxic situation. This dynamic really broke down my enthusiasm for helping people getting into programming after years of trying very hard to pay forward all the help I got when starting out. I used to be the kind of person who would skip lunch to hop on a call with someone and talk through a comment on their PR, my whole career up to now I assumed that time spent providing quality feedback to people was something that would pay off in the long run. In the LLM era I've more or less lost faith in this idea. There's little in the world more discouraging than going out of your way your explain something well, taking time to consider your audience, struggling with the tension of not talking over someone's head and not over explaining, just generally trying to show compassion for another human being and getting an obvious LLM generated patch back knowing that not only did you fail to help anyone learn anything, but that no human even attempted to read it.
I'll take the time to help someone who can demonstrate good faith, but I think I'm pretty much done with assuming good faith from coworkers and random contributors, empirically too much of the population that I interact with has apparently decided that my feedback is better shoveled into a chatbox than engaged with.