This is a version of something I said in the #godot contributor chat, my views on #ai or #aislop
I don't want to wave away the various ethical concerns, particularly around using "public" AIs, given that a lot of them are literally powered by gas. But I want to push back against the idea that an AI authored PR would be good if the code was good. It is not.
A thread 1/x
PRs should be donations of someone's time and expertise. Not someone's donation of tokens bought from Anthropic/Google/OpenAI/whomever. *IF* we wanted to AI generate PRs at all (and we do not) then the useful way would be for someone to give use money to use an AI.
What is step 2 after the PR is submitted? I review and then the person "donating tokens" copy/pastes that into their Claude Code session? If I wanted to use Claude Code I'd be 1000x better off just using the tool.
2/x
"I have a PR here, if you have any suggestions just telnet into my Emacs session to make the changes you want."
It's just a lose/lose proposition, there will never be an upside to the project. There's only two possibilities for a substantilly AI generated PR:
1) It is perfect and we merge it as-is.
2) It is garbage and should be closed.
1) rarely happens, especially not today. People who heavily use AI for their PRs do so because they don't (want to) know the engine deeply enough.
3/x
"But humans also make bad PRs" and this is obviously true. But if I spend hours/days helping a human get a PR over the finish line I've taught that person how to do future PRs better. Maybe they never come back, but at a minimum I've taught another human something useful.
If I'm talking to an AI agent, there's a 100% chance that next session every nanosecond I put in reviewing the PR is wasted. The next PR even by the same "author", will have the same issue.
It's not sustainable.
4/4
HP van Braam
thomas_shone
Uilebheist
Tom Forsyth