alt Hacker NewsAs a contributor who had to deal with 2 automated reviews in the past month, it's a bad experience. I really wanted to upstream my changes, but these bots made me question if i should at all.
Bots don't understand how their words can hurt people's feelings, and can even less understand that talking to a LLM is even more sad/infuriating for actual humans. I can handle CI/tests just fine, but LLMs playing on my human emotions will have me swearing in no time.
I contribute to free software projects because i want user control over the machine (that was the idea, long before LLMs). Slopware might technically be FLOSS, but it's certainly not in the spirit of FLOSS.
Replies
> Bots don't understand how their words can hurt people's feeling
That's a super weird way of putting it, which makes me hesitate to agree with you.
buuuut I agree that the LLM bot review experience is usually miserable, as the persona these things use is insufferable. It's a know-it-all only-speaks-in-imperatives thing, where the speaker lacks the actual merit that would grant them the right to that style.
LLMs can be very useful for code review, but the existing product implementations took a bit too many pages out of Kafka's works and are a bit too real in simulating a dysfunctional bureaucratic org.
I haven't been exposed AI reviews in OSS but I like them for work. Quality of PR review feedback is the most revolutionary bit of AI I think. More so than automated coding. I guess one issue with the AI reviews in OSS is that your dialog (I assume) is public. You know the interaction is with a bot but you also sense it may be observed by humans.
If I get a review at work from a bot that is annoying, misunderstands, etc I just ignore their reviews, or tell the bot to leave me alone. The bot is there to help me, and if it doesn't - that's not my problem. There's also always a human reviewer. For any automated feedback for PR's I think there should be "manual review" requests. If you disagree and the bot can decide that there is real human effort involved, then just let real humans also make the effort to review.
This is the key idea: to spend the human effort from maintainers, where there has been human effort from contributors. The AI should be able to determine the amount of human effort in a contribution. The bot should be an aid to make sure the human maintainer effort is spent only where there is human contributor effort.