The AI Pascal’s Wager
Should you accept contributions generated by LLMs in your Open Source project?
@ploum Compellingly argued. Feels like the best position to take is: "no, but we will review in n years".
@ploum Enforcement is an issue though. IME if you say "no AI", they don't stop coming, they just don't disclose it, and use other software to make the llm output look less like llm ( there are tons of sw to do that now ... ) and this become a cat and mouse game. Reviewing PR is painful enough, but trying to determine if it comes from a LLM makes it an order of magnitude more painful.
@tnt :Enforcement is yet another matter.
See https://destructured.net/negotiating-the-burden
My own strategy is to ask questions about the code to the contributor. I don’t explicitely hunt for LLMs: I just want to make sure that the contributor took conscious decisions and know why he did those.
Having a clear policy also mean that even if someone manage to slip slop in (lol), it’s their own fault. They willingly broke a clear rule.
@ploum Err, this feels a bit like you wrote the conclusion before the logical reasoning.
"LLM generated contributions will eventually slip in despite some guardrails such as "human review is mandatory". The longer a project stays in the "middle ground", the more ineluctable the slop infection becomes." This is a slippery slope argument. You don't know this, it's not obvious and needs to be demonstrated. Just as likely is that some AI contributions slip in, but otherwise go to other projects.
No. Close data centres in every town and country. Time for humans, all of us.
"Let’s put aside for a moment all the ethical/ecological/societal considerations about the use of generated code"
Let's not.
I don't know where this article is going but if it opens with the following lines, how can it be taken seriously?
"Let’s put aside for a moment all the ethical/ecological/societal considerations about the use of generated code"
It goes on to claim a pragmatic point of view and I wonder: would anyone accept an article that puts aside the pragmatic point of view beforehand?
It's a cheap trick, even if the article possibly tries to argue against accepting LLM PRs, it just ignores most of the central concerns with this and thus plays into the hands of the "rationalist" camp that will always choose the most efficient path, no matter the fallout.
@manu @ploum Maybe you try first reading the article before commenting on it? The article is not for you or for me. The article is for those who "put aside ethical/ecological/societal considerations". During the Debian GR there were many people who did exactly that. I wish that they could've read the article in the past because what the article does is to give those people who put ethical/ecological/societal considerations aside an argument that even if you do, even then, you should reject LLMs.
@ploum
And sorry if you're the author then this may be a bit confrontative - but that's exactly how reading the first lines of the article feels to me.
People spend so much energy on arguing the aspects so easily discarded here and that's disrespectful.
@manu : my point is that, from ethical/ecological/societal point of view, there’s no doubt you should avoid AI.
So the only reason to use it is the "pragmatic" one.
But as the article demonstrates, pragmatism should lead to reject AI (at least for now).
@ploum @manu The voting choices were difficult for many. It was difficult for me. During the discussion period, I changed my view on the matter. I believe that others changed their view during the discussion period as well. I think that yes, more arguments like this would've changed some people's minds.
Maybe the rust-style LLM policy could've passed if we had that? Very nice writeup, I 100% agree. The LLM fatigue is a real thing, and all proponents argue like it's nonexistent.
- replies
- 0
- announces
- 0
- likes
- 2
@ploum A contribution can be generated by a model and still require a maintainer who can explain every line. What review rule would you refuse to automate?
@ploum The unglamorous part is making the review trail legible months later. A maintainer should explain why each line stayed. What review record do you keep?
@ploum
Hi again and sorry again for jumping on it that way. It just really struck a nerve because so many arguments open that way these days and it is so tiresome to me and Fedi is a place where I see it a lot less than somewhere else, hence I was eager to comment. I wasn't very friendly but I hope it came across that I was trying to be kind of constructive in spite of being really annoyed. (i admit i didn't do very well) Hence my second comment.
I thought about it a bit more and I really see a difference between writing something along the lines of this: "ethical/ecological/societal points of view forbid the use of LLMs (insert references), so let's also look at the pragmatic perspective".
Versus "let's put them aside" because it makes it look like a valid approach.
It's not unique to the discussion of LLMs. This happens the same way in discussions about e.g. fossil/nuclear energy.
