Egregoros

@p @lumi Well LLM code is going to be statistically average at best, and at very least 90% of code is pure crap nobody wants to read (and so nobody reads).
But it's far from the only problem, like AT&T Unix code was (and quite few times still is) great, but nobody would touch it for anything legal back in the early 90's and even now it's still a bit unclear depending on the versions.

Plus outside of purely code, since then there's also the absolute spam of completely bogus security reports, same breed as people throwing a full set of static analyzers and fuzzers without any kind of thoughts in their brains, except now it's with stuff that doesn't even qualifies as tools.

@lanodan @lumi

> LLM code is going to be statistically average at best, and at very least 90% of code is pure crap nobody wants to read

Sure, yes. Reasonable.

> but nobody would touch it for anything legal back in the early 90's

Right, yeah. on FSE comes from a photo of bwk holding up his copy of Lion's Commentary when AT&T finally announced that they were going to stop suing people for using that book, and there was the whole SCO thing. AT&T got litigious as soon as they possibly could.

(I think copyright assignment mainly makes extraction/relicensing easier.)

> there's also the absolute spam of completely bogus security reports

This is real, yeah, but people that are putting clankers to your codebase are, by their nature, not paying attention to contribution guidelines; I don't think a rule against clanker code stops the spam, it just gives you a thing to point to when you ignore the spam, and you don't really need that. You can just put something about ignoring clanker-submissions and that's kind of a hint to people that are people to avoid looking like clankers.

@lanodan @p @lumi
>>And 9front's even more elegant:
>3. Submitted code is not computer-generated

Technically that bans also completely mundane stuff like auto-generated headers from things like device trees/spec sheet documentation or m4 macro expansions. Of course it will never be enforced that way.

>and at very least 90% of code is pure crap nobody wants to read (and so nobody reads).

Things got way better in the last year or so with new-ish models. If you take the time to set up a workflow for the new models, the results are still somewhat easy to distinguish from non-clanker contributions, but they aren't bad either and usually require some smaller fixups and removal of some overly defensive checks. If done by someone that knows stuff at least. And even with the "chat in a browser window" way, you can get decent-ish results if you keep things small.

>Plus outside of purely code, since then there's also the absolute spam of completely bogus security reports

On the other hand you also have security reports with really high severity that were found within a day or two of letting a model loose on a codebase. The recent nginx CVEs were all found by LLMs, most of the recent Linux LPEs as well. The way I see it is the same as static analyzers and fuzzers, a useful tool in the hands of someone knowing how to handle it.

@p @lanodan @sun @lumi I get where Moon is coming from. If it weren't for you and lanodan in this thread, I wouldn't have replied either, because it had one of the typical setups for the typical "all AI bad, you are an idiot if you think otherwise" threads that me and probably also Moon are tired of seeing on TLs every few days (picrel).

Both of you are the sole reason why I bothered writing my more or less unopinionated and emotion-less reply for the quest of having some discussion in the first place.
ai-llm-arguments.png

@phnt @lanodan @lumi @sun

> I get where Moon is coming from.

It is the nth time he made this specific incorrect assumption about this specific topic, and then if I say "No, I'm not going there." I get the "Well, I'm sick of talking about it before you even showed up so I'm going to ignore this input and will resume with exactly the same assumptions next time" and he's generally a reasonable guy to talk to which is why it is baffling to do it on this topic and moderately annoying that someone I have known for years will jump out from nowhere to do this accusation.

> If it weren't for you and lanodan in this thread, I wouldn't have replied either

Well, like, it was *just* me, you, and lanodan, right, and I think it is reasonable to try to guard against "All these jithub dipshits are putting their hallucinating clankers into my pull requests", so it's just three of us participating and he's warning one about the other two.

> Both of you are the sole reason why I bothered writing my more or less unopinionated and emotion-less reply for the quest of having some discussion in the first place.

Well, shucks, friend. Likewise, you are worth paying attention to and bouncing things off. (On the other hand, like, feel free to do opinionated.)

@p @phnt @lanodan @lumi

> It is the nth time he made this specific incorrect assumption about this specific topic

this wasn't an extension of a specific claim about machine learning from however long ago, btw I think I engaged with you one time about it. don't remember. Anyway it was a prediction that the conversation would go south in a specific way, and it did, a completely uncivil response to a casual observation.

> hey I've been using it and it got better in the last year
>> shove it up your ass, learn how machine learning even works

@sun @lanodan @lumi @phnt

> btw I think I engaged with you one time about it.

Casually while we were walking but the topic drifted and we ended up talking about your former coworker doing a bad job of vibe-coding instead. (Granted you were going through a thing or two last time you jumped in to say everyone was terrible but because you expected bad faith last time, you said "I did it and it worked and you people hate all of it" in the thread where we were poking at them and discussing uses and then ignored the question of "What model?" and this kind of derailed the discussion until ideclineu popped in and showed Kimi-K2 correctly answering a question that I had not gotten a good answer for from a model. In my case, it also kind of ignores my background since I have been playing with these things since the days when MegaHAL got best-in-show from Loebner.)

> it was a prediction that the conversation would go south in a specific way, and it did

There were three people in the thread and you cautioned one of them that he would not get a good-faith response. I think it's an understatement to say it's somewhat uncharitable to hop in just to make an accusation. phnt is also a friend and I think after this much time, he'd probably have noticed if lanodan or I didn't listen and speak in good faith. Bad faith is lying, it's the use of language as a weapon, it's the main tool of idiot slap-fights on the internet, you know, I don't do that shit (I have delegated it, in fact, to the aforementioned Markov bot), and I haven't seen lanodan do it either, so you know, "these guys are assholes about this topic" is at least not impugning anyone's character (though still incorrect, I think; "these LLMs suck at this task" is not "these LLMs suck" is not "all LLMs suck" is not "all possible LLMs suck" is not "no interesting machine-learning can be done and also something about the devil").

> >> shove it up your ass, learn how machine learning even works

That did appear in the thread, yes; "caustic" isn't quite the same as "bad faith" and maybe it wouldn't have appeared in the thread if "I predict you'll be shitty" hadn't appeared, though; there are some people that are persistently assholes and I don't think that applies to anyone in the thread. Like, one thing to note a bias and another to jump in out of nowhere just to say "Don't talk to these people, they'll be awful at you." I don't know how that was supposed to have any sort of positive effect.