You can hardly get online these days without hearing some AI booster talk about how AI coding is going to replace human programmers. AI code is absolutely up to production quality! Also, you’re all…
I’m a pretty big proponent of FOSS AI, but none of the models I’ve ever used are good enough to work without a human treating it like a tool to automate small tasks. In my workflow there is no difference between LLMs and fucking grep for me.
People who think AI codes well are shit at their job
(I don’t mean to take aim at you with this despite how irked it’ll sound)
I really fucking hate how many computer types go “ugh I can’t” at regex. the full spectrum of it, sure, gets hairy. but so many people could be well served by decently learning grouping/backrefs/greedy match/char-classes (which is a lot of what most people seem to reach for[0])
that said, pomsky is an interesting thing that might in fact help a lot of people go from “I want $x” as a human expression of intent, to “I have $y” as a regex expression
[0] - yeah okay sometimes you also actually need a parser. that’s a whole other conversation. I’m talking about “quickly hacking shit up in a text editor buffer in 30s” type cases here
Hey. I can do regex. It’s specifically grep I have beef with. I never know off the top of my head how to invoke it. Is it -e? -r? -i? man grep? More like, man, get grep the hell outta here!
The promptfarmers can push the hallucination rates incrementally lower by spending 10x compute on training (and training on 10x the data and spending 10x on runtime cost) but they’re already consuming a plurality of all VC funding so they can’t 10x many more times without going bust entirely. And they aren’t going to get them down to 0%, hallucinations are intrinsic to how LLMs operate, no patch with run-time inference or multiple tries or RAG will eliminate that.
And as for newer models… o3 actually had a higher hallucination rate because trying to squeeze rational logic out of the models with fine-tuning just breaks them in a different direction.
I will acknowledge in domains with analytically verifiable answers you can check the LLMs that way, but in that case, its no longer primarily an LLM, you’ve got an entire expert system or proof assistant or whatever that can operate independently of the LLM and the LLM is just providing creative input.
We should maximise hallucinations, actually. That is, we should hack the environmental controls of the data centers to be conducive for fungi growth, and flood them with magic mushrooms spores. We can probably get the rats on board by selling it as a different version of nuking the data centers.
I’m guessing if it would actually work for that, somebody would have done it by now.
But it probably just does it’s usual thing of bullshitting something that looks like code, only now you’re wasting the time of maintainers as well who have to confirm that it is bobbins.
It’s already doing that, some FOSS projects regularly get weird PRs that on first glance look good, but if you look closer are either total nonsense or riddled with bugs. Especially awful are security-related PRs; although those are never made in good faith, that’s usually grifting (throwing AI at the wall trying to cash in as many bounties as possible). The project lead of curl recently announced that anyone who posts a PR that’s obviously AI, or is made with AI, will get banned.
Like, it’s really good as a learning tool as long as you don’t blindly believe everything it says given you can ask stuff in natural language and it will resolve possible knowledge dependencies for you that you’d otherwise get stuck on in official docs, and since you can ask contextual questions receiving contextual answers (no logical abstraction). But code generation… please don’t.
Nice conversation you had right there in your head. I assume you also took a closer look at it to get a neutral opinion and didn’t just ride one of the two waves “blind AI hype” or “blind AI hate”?
I’ve taken a closer look at Codestral (which is locally hostable), threw stuff at it and got a sense for what it can and can’t do. The general gist is that its (Python) syntax is basically always correct, however it sometimes messes up the actual code logic or gets the user request wrong. That makes it a good tool for code questions aimed at specific features, how certain syntax in a language works or to look up potential alternative solutions for smaller code snippets. However it should absolutely not be used to create huge chunks of your code logic, that will always backfire.
And since some people will read this and think I’m some AI worshipper, fuck no. They’re amoral as fuck, the only models not screwed up through their creation process are those very few truly FOSS ones. But if you hate on something you have to actually know shit about it and understand its appeal and non-hyped usecases (they do have them, even LLMs). Otherwise you’ll end up in a social corner filled with bitterness and, depending on the topic, perhaps even increasingly extreme opinions (not saying we shouldn’t smash OpenAI and other corposcum into tiny pieces, we absolutely should).
There are technologies that are utter bullshit like NFTs. However (unfortunately?) that isn’t the case for AI. We just live in an economy that’s good in abusing everything and everyone.
Nice conversation you had right there in your head
that you recognize none of this is telling. that someone else got it, more so.
I assume
you could just ask, you know. since you seem so comfortable fondling prompts, not sure why you wouldn’t ask a person. is it because they might tell you to fuck off?
I’ve taken a closer look…
fuck off with the unrequested advertising. never mind that no-one asked you for how you felt for some fucking piece of shit. oh, you feel happy that the logo is a certain tint of <colour>? bully for you, now fuck off and do something worthwhile
That makes it a good tool
a tool you say? wow, sure glad you’re going to replace your *spins the wheel* Punctured Car Tyre with *spins the wheel again* Needlenose Pliers!
think I’m some AI worshipper, fuck no. They’re amoral as fuck
so, you think there’s moral problems, but only sometimes? it’s supes okay to do your version of leveraged exploitation? cool, thanks for letting us know
those very few truly FOSS ones
oh yeah, right, the “truly FOSS ones”! tell me again how those are trained - who’s funding that compute? are the licenses contextually included in the model definition?
wait, hold on! why are you squealing away like a deflating balloon?! those are actual questions! you’re the one who brought up morals!
Otherwise you’ll end up in a social corner filled with bitterness
I’ve met people like you at parties. they’re often popular, but they’re never fun. and I always regret it.
There are technologies that are utter bullshit like NFTs. However (unfortunately?) that isn’t the case for AI
Bro, sneerclub and techtakes are for sneering at bad technology and those that worship it, not for engaging in apologia for it (or worse yet, tone policing the sneering). If you don’t like it, you can ask the mods for an exit pass out (if they haven’t generously given you one already).
if you can’t make a good that’s a you problem. if people end up poking holes in your shit and you suddenly can’t keep your incoherent nonsense together, still a you problem. but:
nonsensically off-the-rails
take your abuser bullshit and fuck right off, thanks
Otherwise you’ll end up in a social corner filled with bitterness
This is a standard Internet phenomenon (I generalize) called a Sneer Club, i.e. people who enjoy getting together and picking on designated targets. Sneer Clubs (I expect) attract people with high Dark Triad characteristics, which is (I suspect) where Asshole Internet Atheists come from - if you get a club together for the purpose of sneering at religious people, it doesn’t matter that God doesn’t actually exist, the club attracts psychologically f’d-up people. Bullies, in a word, people who are powerfully reinforced by getting in what feels like good hits on Designated Targets, in the company of others doing the same and congratulating each other on it.
Hey, Devin! Really impressive that the product best known for literally lying about all of its functionality in its release video still somehow exists and you can pay it money. Isn’t the free market great.
No the fuck it’s not
I’m a pretty big proponent of FOSS AI, but none of the models I’ve ever used are good enough to work without a human treating it like a tool to automate small tasks. In my workflow there is no difference between LLMs and fucking
grepfor me.People who think AI codes well are shit at their job
Well grep doesn’t hallucinate things that are not actually in the logs I’m grepping so I think I’ll stick to grep.
(Or ripgrep rather)
With grep it’s me who hallucinates that I can right good regex :,)
(I don’t mean to take aim at you with this despite how irked it’ll sound)
I really fucking hate how many computer types go “ugh I can’t” at regex. the full spectrum of it, sure, gets hairy. but so many people could be well served by decently learning grouping/backrefs/greedy match/char-classes (which is a lot of what most people seem to reach for[0])
that said, pomsky is an interesting thing that might in fact help a lot of people go from “I want $x” as a human expression of intent, to “I have $y” as a regex expression
[0] - yeah okay sometimes you also actually need a parser. that’s a whole other conversation. I’m talking about “quickly hacking shit up in a text editor buffer in 30s” type cases here
Hey. I can do regex. It’s specifically grep I have beef with. I never know off the top of my head how to invoke it. Is it
-e?-r?-i?man grep? More like,man, get grep the hell outta here!now listen, you might think gnu tools are offensively inconsistent, and to that I can only say
find(1)find(1)? You betterfind(1)some other place to be, buster. In this house, we use the file explorer search barHallucinations become almost a non issue when working with newer models, custom inference, multishot prompting and RAG
But the models themselves fundamentally can’t write good, new code, even if they’re perfectly factual
The promptfarmers can push the hallucination rates incrementally lower by spending 10x compute on training (and training on 10x the data and spending 10x on runtime cost) but they’re already consuming a plurality of all VC funding so they can’t 10x many more times without going bust entirely. And they aren’t going to get them down to 0%, hallucinations are intrinsic to how LLMs operate, no patch with run-time inference or multiple tries or RAG will eliminate that.
And as for newer models… o3 actually had a higher hallucination rate because trying to squeeze rational logic out of the models with fine-tuning just breaks them in a different direction.
I will acknowledge in domains with analytically verifiable answers you can check the LLMs that way, but in that case, its no longer primarily an LLM, you’ve got an entire expert system or proof assistant or whatever that can operate independently of the LLM and the LLM is just providing creative input.
We should maximise hallucinations, actually. That is, we should hack the environmental controls of the data centers to be conducive for fungi growth, and flood them with magic mushrooms spores. We can probably get the rats on board by selling it as a different version of nuking the data centers.
What if [tokes joint] hallucinations are actually, like, proof the models are almost at human level man!
stopping this bit here because I don’t want to continue writing a JRE episode
@swlabr @scruiser Java Runtime Environment?
no the worse one
If LLM hallucinations ever become a non-issue I doubt I’ll be needing to read a deeply nested buzzword laden lemmy post to first hear about it.
@vivendi @V0ldek * hallucinations are a fundamental trait of LLM tech, they’re not going anywhere
There are plenty of open issues on open source repos it could open PRs for though?
I’m guessing if it would actually work for that, somebody would have done it by now.
But it probably just does it’s usual thing of bullshitting something that looks like code, only now you’re wasting the time of maintainers as well who have to confirm that it is bobbins.
It’s already doing that, some FOSS projects regularly get weird PRs that on first glance look good, but if you look closer are either total nonsense or riddled with bugs. Especially awful are security-related PRs; although those are never made in good faith, that’s usually grifting (throwing AI at the wall trying to cash in as many bounties as possible). The project lead of curl recently announced that anyone who posts a PR that’s obviously AI, or is made with AI, will get banned.
Like, it’s really good as a learning tool as long as you don’t blindly believe everything it says given you can ask stuff in natural language and it will resolve possible knowledge dependencies for you that you’d otherwise get stuck on in official docs, and since you can ask contextual questions receiving contextual answers (no logical abstraction). But code generation… please don’t.
Fuck you were doing so well in the first half, ahhh,
the poster: “it’s really good as a learning tool”
the poster: “but don’t blindly believe it”
the learner: “how should I know when to believe it?”
the poster: “check everything”
the learner: “so you’re saying I should just read the actual documentation and/or source?”
the poster: “how are you going to ask that anything? how can you fondle something that isn’t a prompt?!”
the learner: “thanks for your time, I think I’m going to find another class”
Nice conversation you had right there in your head. I assume you also took a closer look at it to get a neutral opinion and didn’t just ride one of the two waves “blind AI hype” or “blind AI hate”?
I’ve taken a closer look at Codestral (which is locally hostable), threw stuff at it and got a sense for what it can and can’t do. The general gist is that its (Python) syntax is basically always correct, however it sometimes messes up the actual code logic or gets the user request wrong. That makes it a good tool for code questions aimed at specific features, how certain syntax in a language works or to look up potential alternative solutions for smaller code snippets. However it should absolutely not be used to create huge chunks of your code logic, that will always backfire.
And since some people will read this and think I’m some AI worshipper, fuck no. They’re amoral as fuck, the only models not screwed up through their creation process are those very few truly FOSS ones. But if you hate on something you have to actually know shit about it and understand its appeal and non-hyped usecases (they do have them, even LLMs). Otherwise you’ll end up in a social corner filled with bitterness and, depending on the topic, perhaps even increasingly extreme opinions (not saying we shouldn’t smash OpenAI and other corposcum into tiny pieces, we absolutely should).
There are technologies that are utter bullshit like NFTs. However (unfortunately?) that isn’t the case for AI. We just live in an economy that’s good in abusing everything and everyone.
that you recognize none of this is telling. that someone else got it, more so.
you could just ask, you know. since you seem so comfortable fondling prompts, not sure why you wouldn’t ask a person. is it because they might tell you to fuck off?
fuck off with the unrequested advertising. never mind that no-one asked you for how you felt for some fucking piece of shit. oh, you feel happy that the logo is a certain tint of <colour>? bully for you, now fuck off and do something worthwhile
a tool you say? wow, sure glad you’re going to replace your *spins the wheel* Punctured Car Tyre with *spins the wheel again* Needlenose Pliers!
so, you think there’s moral problems, but only sometimes? it’s supes okay to do your version of leveraged exploitation? cool, thanks for letting us know
oh yeah, right, the “truly FOSS ones”! tell me again how those are trained - who’s funding that compute? are the licenses contextually included in the model definition?
wait, hold on! why are you squealing away like a deflating balloon?! those are actual questions! you’re the one who brought up morals!
I’ve met people like you at parties. they’re often popular, but they’re never fun. and I always regret it.
citation. fucking. needed.
Holy shit, get some help. Given how nonsensically off-the-rails you just went you clearly need it.
Bro, sneerclub and techtakes are for sneering at bad technology and those that worship it, not for engaging in apologia for it (or worse yet, tone policing the sneering). If you don’t like it, you can ask the mods for an exit pass out (if they haven’t generously given you one already).
if you can’t make a good that’s a you problem. if people end up poking holes in your shit and you suddenly can’t keep your incoherent nonsense together, still a you problem. but:
take your abuser bullshit and fuck right off, thanks
This is a standard Internet phenomenon (I generalize) called a Sneer Club, i.e. people who enjoy getting together and picking on designated targets. Sneer Clubs (I expect) attract people with high Dark Triad characteristics, which is (I suspect) where Asshole Internet Atheists come from - if you get a club together for the purpose of sneering at religious people, it doesn’t matter that God doesn’t actually exist, the club attracts psychologically f’d-up people. Bullies, in a word, people who are powerfully reinforced by getting in what feels like good hits on Designated Targets, in the company of others doing the same and congratulating each other on it.
@froztbyte @Natanox
In that moment, the novice was enlightened
Yea it’s a problem already for security bugs, llms just waste maintainers time and make them angry.
They are useless and make more work for programmers, even on python and js codebases that they are trained on the most and are the “easiest”.
People have done it, there’s a bunch of services that do it. But they’re paid.
ie https://devin.ai/
Hey, Devin! Really impressive that the product best known for literally lying about all of its functionality in its release video still somehow exists and you can pay it money. Isn’t the free market great.
fuck off with the unrequested advertising kthx
Because it’s a upscaled translation tech maybe?
These views on LLMs are simplistic. As a wise man once said, “check yoself befo yo wreck yoself”, I recommend more education thus
LLM structures arw over hyped, but they’re also not that simple