this post was submitted on 13 Dec 2025
569 points (98.0% liked)
Technology
77096 readers
2803 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Not sure, but I hear the Claude Super Duper Extreme Fucking Pro ($200/month) is like the Ferrari of LLM assisted coding
As someone who works in network engineering support and has seen Claude completely fuck up people's networks with bad advice: LOL.
Literally had an idiot just copying and pasting commands from Claude into their equipment and brought down a network of over 1000 people the other day.
It hallucinated entire executables that didn't exist. It asked them to create init scripts for services that already had one. It told them to bypass the software UI, that had the functionality they needed, and start adding routes directly to the system kernel.
Every LLM is the same bullshit guessing machine.
Functions with arguments that don’t do anything… hey Claude why did you do that? Good catch…!
AI is incredibly powerful and incredibly easy to use, which means it's a piece of cake to use AI to do incredibly stupid things. Your guy is just bad with AI, which means he doesn't know how to talk to a computer in his native language
no, AI just sucks ass with any highly customized environment, like network infrastructure, because it has exactly ZERO capacity for on-the-fly learning.
it can somewhat pretend to remember something, but most of the time it doesn't work, and then people are so, so surprised when it spits out the most ridiculous config for a router, because all it did was string together the top answers on stack overflow from a decade ago, stripping out any and all context that makes it make sense, and presents it as a solution that seems plausible, but absolutely isn't.
LLMs are literally design to trick people into thinking what they write makes sense.
they have no concept of actually making sense.
this is not an exception, or an improper use of the tech.
it's an inherent, fundamental flaw.
whenever someone says AI doesn't work they're just saying that they don't know how to get a computer to do their work for them. they can't even do laziness right
yeah, no... that's not at all what i said.
i didn't say "AI doesn't work", i said it works exactly as expected: producing bullshit.
i understand perfectly well how to get it to spit out useful information, because i know what i can and cannot ask it about.
I'd much rather not use it, but it's pretty much unavoidable now, because of how trash search results have become, specifically for technical subjects.
what absolutely doesn't work is asking AI to perform highly specific, production critical configurations on live systems.
you CAN use it to get general answers to general questions.
"what's a common way to do this configuration?" works well enough.
"fix this config file for me!" doesn't work, because it has no concept of what that means in your specific context. and no amount of increasingly specific prompts will ever get you there. ...unless "there" is an utter clusterfuck, see the ~~OP~~ top of chain (should have been more specific here....) for proof...
idk what you're talking about because i'm an amateur, i've never pushed anything to prod. sounds like AI slop is especially bad when you write code for a living, but me.....idk i just dick around on my computer all day so it's like the wild west over here, lmaooooo (i can't work because i had a TBI)
yes, that's exactly the point of everything I've said:
to an inexperienced user/developer/admin the output LLMs produce look perfectly valid, and for relatively trivial tasks they might even work out...but when it gets more specialized it fails spectacularly and it gets extremely obvious just how limited of a system it really is.
which is why there is so much pushback from professionals. actually that's pretty much all professionals, not just in IT.
well i got a bacon number app with django and a .db that was like 2 gigs to work and i vibe coded it. i guess that's the most complicated thing i've put together. i mostly use LLMs for instruction, debugging, and (sadly) writing my code for me when i'm too lazy to read the documentation
sure, and that works at small scales and as long as no change is required.
when either of those two change (large projects where interdependent components become inevitable and frequent updates are necessary) it becomes impossible to use AI for basically anything.
any change you make then has to be carefully considered and weighed against it's consequences, which AIs can't do, because they can't absorb the context of the entire project.
look, I'm not saying you can't use AI, or that AI is entirely useless.
I'm saying that using AI is the same as any other tool; use it deliberately and for the right job at the right time.
the big problem, especially in commercial contexts, is people using AI without realizing these limitations, thinking it's some magical genie that can everything.
As a dev: lol. Do it again, you are good at entertaining
Native language == assembly?
Generative AI has an average error rate of 9-13%. Nobody should trust it wholesale and what it spits out.
It has some excellent use cases. Vibe code/sysadmin/netadmin'ing are not one of those things.
Where does this 9-13% number come from?
There was a study done several months ago. I'll try to find the source again and link it here in a comment edit.
[EDIT]
Couldn't find the study, but here is an infographic that has similar data: https://infogram.com/ai-hellucination-report-2025-1h9j6q7mz7q354g
Hallucinations have improved, but the average for all models is still around 8-9%. Likely the study I read before was from last year, so the higher hallucinations rate is from that.
Even the best model hallucinates an average of around 1 in 100 times and OpenAI has stated publicly that hallucinations are mathematically impossible to eliminate entirely.
I don't trust it wholesale. No one who knows what they're talking about trusts it wholesale. Hallucination rates vary depending on who you ask. And you're wrong about vibe coding, it works great if you're working on some random side project and not working with a team that has to push to production
So expensive, looks great, takes significant capital to maintain, and anyone who has one uses something else when they actually need to do something useful.
it literally doesn't cost as much as a ferrari
What’s with tech people always stating (marketing) things as akin to high end sports cars. The state of AI is more like arguing over which donkey is best, lol.
So...4th in the Constructors and 5th+6th in the Driver's Championships?
unfortunately your code placed last in the driver's so AI would be a HUGE step up for you