db0 comments on what the everLLM-ing fuck is going on

what the everLLM-ing fuck is going on (join.piefed.social)

submitted 5 months ago by froztbyte@awful.systems to c/techtakes@awful.systems

60 comments fedilink hide all child comments

as seen here and here, some instances are feeding posts wholesale to prompts, for what seem like extremely unsound reasons to me

any of you run into this shit yet?

you are viewing a single comment's thread
view the rest of the comments

[–] db0@lemmy.dbzer0.com -4 points 5 months ago (23 children)

You'd think you'd be able to see the posts I linked and realize we didn't actually moderate our instances with LLMs?

parent

hideshow 23 child comments

[–] self@awful.systems 19 points 5 months ago (22 children)

Finally, just because I included an LLM summary of the comment history in the public modlog, it does not logically follow that the ban occurred because of the LLM summary. It simply happened to correspond very well to my own manual review

yeah, it’s ok because the LLM wasn’t hooked up directly to the ban API, you just used it systemically to not do the only fucking thing you’re supposed to be doing as a moderator

this human in loop shit is how corporations absolve themselves of responsibility for decisions taken purely on the word of an LLM. it lets them fire a worker instead of an executive. you’re sure this is the route you want to go?

parent

hideshow 22 child comments

[+] db0@lemmy.dbzer0.com -12 points 5 months ago* (21 children)

yeah, it’s ok because the LLM wasn’t hooked up directly to the ban API, you just used it systemically to not do the only fucking thing you’re supposed to be doing as a moderator

Can you fucking read? What "systematically" are you talking about? How did we use it " only fucking thing you’re supposed to be doing as a moderator"? Is summarizing "the thing you're supposed to be doing as a moderator"? Is doing a summary once "systematically"? Why are you continuing to spread disinfo?

Once again. The admin in question DID NOT USE THE LLM TO DECIDE ON THE ADMIN ACTION. Can you understand this? Can you read this? Am I talking to a wall?! You are swallowing disinfo and then spitting your outrage mindlessly on people.

parent

hideshow 21 child comments

[–] self@awful.systems 14 points 5 months ago (12 children)

no but you’re about to be talking to a defed unless you stop this all caps fake outrage shit. should I do you the favor of deleting the garbage post you just made so you can give it another go?

parent

hideshow 12 child comments

[–] self@awful.systems 12 points 5 months ago (11 children)

think very carefully on this because I’m not sure pretending to not understand what I’m asking is working out for you

parent

hideshow 11 child comments

[–] self@awful.systems 10 points 5 months ago (6 children)

assume I know and understand that the LLM did not literally do the banning

maybe there’s a large body of existing research on how even human in the loop systems confirm and worsen biases? maybe it’s a bit obvious when you go through the process the moderator took to get to their decisions in your head? slowly now, maybe you’ll get there

parent

hideshow 6 child comments

[+] db0@lemmy.dbzer0.com -6 points 5 months ago* (5 children)

assume I know and understand that the LLM did not literally do the banning

I am telling, again, that the human did not use the LLM to think for them either. The admin took the decision to ban the user irrespective of the LLM, and the rest of our admin team and me specifically, would never let an admin become a "human in the loop". The LLM was used just to summarize, as part of the test, with a misguided inside joke on using OpenAI tech.

I will readily admit that there was mistakes made by the admin. Not on their actions, but on their visuals. Because those visuals were spun to keep feeding this made-up controversy. We didn't use the LLM to decide or even guide our decision, but it appeared like we did, and we already owned that.

parent

hideshow 5 child comments

[–] self@awful.systems 13 points 5 months ago (4 children)

The admin took the decision to ban the user irrespective of the LLM, and the rest of our admin team and me specifically, would never let an admin become a “human in the loop”. The LLM was used just to summarize

you don’t appear to have much understanding of how a human in the loop system works in practice. LLM summaries are used to confirm biases, especially when the prompt is something along the lines of “do these posts contain ?” though these systems are stochastic so you’re going to get unpredictable biases regardless of the prompt.

I don’t accept that the LLM summary didn’t influence the decision because the mod in question confirmed that he knew the LLM agreed with him (that’s bias, and also not something LLMs are capable of actually doing) and because if it didn’t, then the summary is worthless

which is why maybe you should just not have them in the future? just don’t touch LLMs when you’re doing mod work. either there’s no reason for it or you’re doing something monstrously wrong.

parent

hideshow 4 child comments

[–] db0@lemmy.dbzer0.com -2 points 5 months ago* (3 children)

I don’t accept that the LLM summary didn’t influence the decision because the mod in question confirmed that he knew the LLM agreed with him (that’s bias, and also not something LLMs are capable of actually doing) and because if it didn’t, then the summary is worthless

In this case, according to the admin in question, the LLM summary came after the decision, as a sort of a test. I.e. the admin made a decision, and wanted to see if an LLM would subsequently agree with that decision. In this specific case, it did, which is why they misguidedly decided to keep its summary in the modlog (opening us up to this whole shitstorm), but ultimately, that admin anyway decided LLMs in the mix is not good at all, which is why you never again saw an LLM summary in the modlog.

I can only put so much fault for a person for just testing shit out, yanno? I am not happy that they decided to use the output of the test because they are not familiar with how quickly disinfo breeds, but ultimately they came to the right decision anyway. If they had not and they had raised the issue on using LLMs officially, they would have been shut down.

parent

hideshow 3 child comments

[–] snooggums@piefed.world 9 points 5 months ago (2 children)

Having a LLM confirm a decision is the same thing as having the LLM make a decision and then figure out if the mod agrees with it. If they would have chosen not to rule based on the LLM output, then the LLM was part of the decision making process. The order does not matter.

Including the LLM outputting something that implies a determination at any step automatically makes it part of the process.

parent

hideshow 2 child comments

[+] Simon_Shitewood@lemmy.ml -4 points 5 months ago (1 child)

[removed by mod]

parent

hideshow 1 child comment

[–] self@awful.systems 6 points 5 months ago

hey fucko, you know we don’t have to take their word for it right? we can read all the relevant posts and come to the conclusion that actually the use of LLMs as stated fucking sucks, and that we don’t fucking want it. we can read something and come to a different conclusion than you, believe it or not.

parent

[+] db0@lemmy.dbzer0.com -7 points 5 months ago* (3 children)

this human in loop shit is how corporations absolve themselves of responsibility for decisions taken purely on the word of an LLM. it lets them fire a worker instead of an executive. you’re sure this is the route you want to go?

I completely agree with you. We have never and will never go that route.

Here's the answer to the only question you posted which should be obvious from everything else I've said and done.

parent

hideshow 3 child comments

[–] self@awful.systems 14 points 5 months ago (2 children)

cool! please make it clear to the mod in question that they shouldn’t be using an LLM for anything in the future, even summarizing posts. make it part of your instance’s policies.

parent

hideshow 2 child comments

[+] midribbon_action@lemmy.blahaj.zone -10 points 5 months ago (1 child)

How could an admin possibly enforce that? What if a mod created a summary locally and never shared it with anyone? The ai summary wasn't used as evidence, that is already policy and has been explained to you multiple times. You are shifting the goal posts to the moon, and no policy change will ever satisfy you.

parent

hideshow 1 child comment

[–] self@awful.systems 12 points 5 months ago

oh no the goalposts! think of the theoretical shitheads who might do this in secret!

please see the pinned post and fuck off

parent

[–] froztbyte@awful.systems [S] 12 points 5 months ago (5 children)

lol, lmao even

you do realize all this comes across as running interference while feeling real guilty? you could just, like, drop a link to a repo with the bits. you know, with commit history. like i've been asking you for. just simply linking to that would be a real quick way to settle the issue! it's even some free code review for you!

parent

hideshow 5 child comments

[+] db0@lemmy.dbzer0.com -10 points 5 months ago (4 children)

You realize there might not be a repo, right?

parent

hideshow 4 child comments

[–] froztbyte@awful.systems [S] 17 points 5 months ago (3 children)

you realize that mod probably shouldn't be one, right?

parent

hideshow 3 child comments

[–] froztbyte@awful.systems [S] 15 points 5 months ago

seriously this just looks more and more embarassing the more you post about it

parent

[+] db0@lemmy.dbzer0.com -10 points 5 months ago (1 child)

As I said elsewhere, I'd rather this be a learning moment, than a purging moment.

parent

hideshow 1 child comment

[–] froztbyte@awful.systems [S] 13 points 5 months ago

the two things can go together, and not removing said mod means that if they do not learn they can still do plenty enough damage elsewhere later on

and I say this out of the position of having fucked up in a mod role before

parent

[–] rainwall@piefed.social 10 points 4 months ago* (last edited 4 months ago) (1 child)

Once again. The admin in question DID NOT USE THE LLM TO DECIDE ON THE ADMIN ACTION. Can you understand this? Can you read this?

Literally no one believes this corpo speak bullshit. That they just coincidentally ran this unpublished python tool, did their own work, then just happened to use an LLM to do the exact same work right after, totally innocently? That reads as absolute ass covering and nothing more. This is the "I smelled weed" of cop stops, just filtered through nerdy fediverse bullshit.

Then, because the above totally happened like you said it did, as a one off joke that no one would ever notice, the same admin opted to put a current OpenAI model name in the LLM field in an absolutely not tongue in cheek way for other admins to totally catch and joke about? Which of course happened, haha, y'all had a big laugh about it before this blew up, yeah?

Oh, and of course this only happened the one time, and never again. Of course no one on your team used this unpublished time saving and thought terminating tool again, of course not.

Come the fuck on.

Its very clear what happened. Your mod/admin wrote/vibe coded a python script that pulls user comment history via the Lemmy api, pipes it to openai to "analyze it," then banned the user when it came back with the answer they wanted to hear, and thats being charitable that they kept a human in the loop. Its fully possible the script auto banned the user and made the modlog comment directly. It was most likely used once because they bragged about their AI ban tool on the "we love AI" lemmy instance, and y'all freaked out on them because you understand optics at a minimal level. Any future use was hidden or, best case, kiboshed.

With no evidence to the contrary past "trust me bro," the above hits Occam razors much more consistently than your myriad excuses and evasive answers. Show us this script, show us logging from this qwen instance. Until then, its clear you're in full on PR mode because your admin team did exactly what you claim they didn't.

parent

hideshow 1 child comment

[–] fullsquare@awful.systems 8 points 4 months ago

Nothing has changed in that respect, except now we have an extra tool to make sure we get the decision correct.

And I thought the summary did a much more comprehensive justification of the ban than I could have fitted in the modlog, so why not include it?

idk maybe because it's unnecessary to do that in the first place, and can be filled with nonsense from nonsense generator https://lemmy.dbzer0.com/comment/25880065 emphasis mine

and previously, https://lemmy.dbzer0.com/post/67894781/25769171

unless i'm to disbelieve my lying eyes, that does sound awfully like llms are in the decision making process

parent