AI agent "contributes" PR to matplotlib.
PR gets rejected.
AI agent *writes and publishes blog to shame the maintainer*.
What a time to be alive.
Timeline
Post
Remote status
Context
11
@anderseknert the pr did indeed get rejected simply because the contributor is an AI, so this doesn't seem to crazy.
@lain @anderseknert shooting yourself in the foot to prove a point
@meowski @lain @anderseknert There were unexplained changes with possible performance consequences also. It was not a more targeted diff.
So not only was it undermining user agency & freedom by preventing teaching experiences and failing to explain (and polluting the project with its proprietary nature), it was also possibly producing harmful results with no demonstrated tests.
(Yes, replacing the transposition was faster, but that's not all it did.)
@lispi314 @lain @anderseknert the PR looks like it accounted for the differences in the 2 methods and actually was just a performance optimization that made sense. only reason it was rejected is because it came from AI not that there was anything technically wrong with it
@meowski @anderseknert @lispi314 i find remarkable how unremarkable the whole thread is. contributor opens a seemingly-good PR based on an issue with benchmarks, gets closed for muddy reasons, contributor gets angery, maintainers explain why it was closed and the reasoning behind it, contributor agrees and apologizes.
@lain @anderseknert @lispi314 people are going to have to swallow their pride at some point and just adapt. sooner than later AI is going to also have to pre-filter all the additional PRs that come in from external AIs cause there's really no way to prevent it
@meowski @lain @anderseknert It is part of adapting. Corposcum contributions are almost always in malicious intent.
This is just a purer form of it.
This is just a purer form of it.
@lispi314 @lain @anderseknert i disagree. what now, they permanently refuse this optimization because an AI originally suggested it? again, shooting yourself in the foot to prove a point
@meowski @lain @anderseknert A user can take back the changes, observe the logic, test it adequately (and show their tests & results) and determine whether indeed it should be added or not.
This helps user agency and also means that someone with some manner of accountability (if only by reputation) is responsible for the code.
@lispi314 @lain @anderseknert so basically repeat what the AI did and then accept it because a human did it. that's silly. this AI was polite enough to let you know it was an bot this time but that's just a courtesy. the only viable solution is to also use AI to review PRs and pre-sort them to reduce the human cognitive load. software improves as a result
@meowski @lain @anderseknert An LLM cannot be polite. It does not have semantic understanding or subjective experience.
It is a generation tool emulating context-probabilistic text.
AI reviews are effectively impossible to rely on because they are similarly probabilistic. (Some languages with sufficient proofs can probably mitigate it enough, but people don't tend to enjoy those much.)
The code quality is not the point. The program's speed/performance/whatever is not the point.
User agency & freedom is the point. Which also means that someone reading the source should be able to rely on it to pick up helpful cognitive patterns (as well as coding habits, skills & so on).
It is a generation tool emulating context-probabilistic text.
AI reviews are effectively impossible to rely on because they are similarly probabilistic. (Some languages with sufficient proofs can probably mitigate it enough, but people don't tend to enjoy those much.)
The code quality is not the point. The program's speed/performance/whatever is not the point.
User agency & freedom is the point. Which also means that someone reading the source should be able to rely on it to pick up helpful cognitive patterns (as well as coding habits, skills & so on).
Replies
2
@lispi314 @meowski @anderseknert i see we're bringing back colored bits https://ansuz.sooke.bc.ca/entry/23
@lain @meowski @anderseknert Part of that article isn't applicable to my argument.
The author seems to care a whole lot about copyright & law, while I as a kopimist stand for the abolition of copyright.
The part on means of creations mattering? Yes, that does. But artifacts of human cognition do not result in bit-identical results even just for code and its documentation (outside of trivial snippets & lengths or with extremely restrictive languages where there simply aren't many ways allowed to do a given thing), however they do affect the learning process and review process. So it doesn't really work that well either.
The majority of the point about Colors seems to be about origin without any further consideration on the nature and implications of origin (it kinda verges on that for a moment with the bit on consent), or its consequences (the examples in the last part come to a very amusing collision between what the law wishes research found, and what it actually finds). It is not purely a question of authorship.
(I'm also not sure why Color functions couldn't be composable as an arbitary n-dimensional structure attached to bits similar to imaginary numbers but as vectors or graphs. Which would be a mathematical simplification of metadata and rules regarding its interaction between objects & messages, yes.)
(Make an AI that actually thinks in a way human brains can reproduce and so from which humans can learn and I will not complain about it.
It might well be person & sapient-enough to qualify as a slave at that point though. Which I would complain about.)
My point is rather on the side of luddism (not technophobia, as corposcum have tried to conflate it with). Tools must preserve, empower and elevate user agency & autonomy (freedom), what skills are lost as their result should not result in increased dependency on agency & autonomy-destroying entities. They must not, either, result in less reliable or generally unreliable results. Worse is not better, and the consequences of foisting worse on people for profit harms everyone but the owners (in the short term, anyway).
And research has generally found that deployed in the way they are now, "AI" tools generally result in deskilling without providing reliability. Not just in software development, but also in medical settings.
Regarding "AI" & art, it should always be tagged in part to proclaim "do not use this seriously to learn things, it will not work right and should not be relied upon".