The author of that article admits that the sample size is not large enough to draw meaningful conclusions.

But besides that, I believe LLM code generators can be a useful tool, provided you are willing to go over their output with a fine-tooth comb and assume it is broken until you have proven otherwise, because the hallucination problem is inherent to the technology and they’re never going to completely solve it, and are willing to overlook the myriad ethical issues with all major LLMs in existence today.

Yeah. Even if you run the model entirely on your own 100% solar-powered hardware, there’s still the problem of training data, and there’s still the problem of the myriad adverse effects on the user.

Okay, but they still happen with orders of magnitude less frequency than bugs in AI code. Consider that the time between when rsync first adopted LLM-generated code and users en masse reporting rsync internal protocol errors during a backup was on the order of months.

The person who put “AI is evil” in big block letters is objectively correct. See homes getting claimed under eminent domain to build power lines for datacenters, electricity prices for consumers rising as service quality gets worse, and water usage to the tune of millions of gallons per datacenter per day. See companies buying rare books in bulk, cutting them apart so they can be scanned faster, and not publishing the scans, literally destroying culture so they can have one more scrap of training data guaranteed to be free of slop that their competitors don’t have. See companies operating web scrapers that deliberately evade every block that website owners try to use to stop them from reading their sites, and then scraping so many different pages so frequently that it’s indistinguishable from a DDoS attack, hosting bills skyrocket, servers crash, indie webmasters pay the price for corporate greed. See xAI operating natural gas turbines on the property of their datacenter for extra electricity in flagrant violation of the Clean Air Act, spewing toxic smoke into the air and generating powerful, inaudible infrasound vibrations that have been linked to everything from nausea to irritability to hearing loss to seizures. See Anthropic treating Minority Report as an instruction manual. See how ChatGPT use affects test scores. See AI psychosis. See chatbot-induced suicides. See deskilling. See how little AI companies care.

I’ve been a programmer for two thirds of my life up to this point. I don’t see how the machine that generates code that must be assumed to be broken, and statements about said code that must be assumed to be wrong, until both have been manually verified, because the “hallucination” issue is inherent to the design of LLMs and cannot be solved, is supposed to make my life easier. Especially since a well known problem in computer science is that writing code is more fun than reading it.

If you’re using AI for code generation, I hope you like reading code.

If you’re using AI to learn your way around an existing codebase, I hope you’re prepared to fact check everything it tells you.

If you’re using AI for code review before merging a PR, fucking yikes. What made you think the sycophant machine that once told users a soggy cereal bar was a great business opportunity and cannot even transcribe McDonald’s orders correctly was suitable for that purpose?

If you’re telling yourself it’s okay because there’s a human in the loop, please read some studies about decision fatigue and consider that increasing the number of critical decisions a person must make per day, while giving them a suggested answer that’s right 90% of the time and catastrophically wrong the remaining 10%, and the only way to tell the difference is to read in depth, might not have the desired effect. Especially if the human in question is then put under time pressure.

And all that is before I go on a five paragraph rant about the ethical transgressions of every major AI company.

It’s pretty black and white for me.

Three point landing? Really???

ok well there was your mistake, you have to talk to a crab

Does it have its own repositories? If not, the water gets murky.

Omarchy isn’t even a distro. It’s a set of dotfiles.

WeaponsManufacturerOS is Nix, obviously (they’ve since dropped Anduril as a sponsor but we’re never going to let them live it down), MalwareOS I’m guessing is Arch, but what’s BenevolentDictatorOS?

Agents are better than me at almost all forms of programming.

Literally image admitting you’re that bad at your job.

Alright, fair, probably a lot, but I’d wager much less than half, and in my experience you’ll have an extremely hard time finding people who still don’t like LLMs when they’re not being shoved down their throats outside the Fediverse.

deleted by creator

@zd9@lemmy.world responded to this comment:

Nobody who uses AI treats it like a tool, they treat it like an employee they can claim credit for. And most tools aren’t actively detrimental to the user to the point of proven intelligence drops, lawsuits and environmental destruction. Also, the enemy charges you money to use it.

The only person who benefits from AI is the person selling it. So why are you trying to defend it?

with a reply that amounted to “I’m an AI researcher, and you’re wrong about the technology. There are plenty of legitimate uses.”

Which part of the original comment do either you or zd9 disagree with?

Let’s go through it claim by claim.

Nobody who uses AI treats it like a tool, they treat it like an employee they can claim credit for.

This is subjective, so I’ll let you have it, but it’s true in a lot of cases. Vibe coders and AI artists definitely do, and people who use it as a coding assistant arguably do.

most tools aren’t actively detrimental to the user to the point of proven intelligence drops,

https://arxiv.org/abs/2506.08872

lawsuits

I’m not sure what precisely @susaga@sh.itjust.works meant by this, so until they clarify I’ll tentatively say you can have that one too.

and environmental destruction.

Where to begin? Clean air act violating natural gas turbines on xAI datacenters? Houses being claimed under eminent domain to build beefier power lines because they use that much? Water use?

Also, the enemy charges you money to use it.

Unless you’re using a local LLM, this is obviously true.

Again I ask: which part of the comment zd9 replied to was wrong?

I apologize for presenting an insufficiently nuanced view of your favorite toy. It’s not an excuse, but being subjected (and seeing my friends subjected) to the massive material harms done by LLM companies has a tendency to make me very angry, and I lash out at the obvious target. In retrospect, I realize only about 90% of current LLM deployments and use cases fit the both-unethical-and-harmful-to-the-user model I describe, and the remaining 10% is crucial. We mustn’t throw the baby out with the bathwater.

Please do elucidate what exactly these ethical uses of LLMs are.

You mentioned in another thread you use them for research. Have you figured out how to get them to stop confidently presenting incorrect information because they don’t know the difference between a sequence of tokens that represents a true fact and one that doesn’t hallucinating yet?

First of all, I’m not the person you replied to. Second, which part of what they said was wrong?

Most of them are as staunchly anti-AI as I am, and don’t talk about it unless someone else brings it up. Which if you hang out in programming circles, on or offline, happens approximately every five minutes.

I don’t think LLMs are bad because they’re not useful. If people didn’t think they were useful, ChatGPT alone wouldn’t have 900 million weekly users.

I think LLMs are bad because they buy up all the computer parts and leave none for the rest of us to build datacenters that make their communities worse places to live, and consume so much electricity that houses have to be bulldozed to build bigger power lines and/or have clean air act violating natural gas turbines on the property belching toxins into predominantly black neighborhoods, and suck up so much water that there isn’t enough for residents, and scrape independent websites until their servers crash and evade every roadblock webmasters put in their way in the hopes of one additional scrap of training material, and buying and destroying rare books because they need training data that isn’t polluted by LLM slop like the internet is and cutting the spines off of books lets you scan them faster and more cheaply than if you don’t do that, all so that they can produce a machine that generates answers that are wrong 10% of the time, meaning that you have to check them anyway, and code that its authors do not understand and which is extraordinarily difficult to maintain, and make it the expectation to write three bullet points into a chatbot to turn it into a five paragraph email which the person who receives it will then immediately feed back into the same chatbot to turn it into three (slightly different) bullet points. and measurably hinder children’s learning. and that’s without going into chatbot psychosis and suicides, which we prommy are rare and caused by preexisting conditions.

But they’re also useful for generating meal plans or whatever, so all in it’s even.

If people who didn’t know/didn’t care about all the ethical issues with ChatGPT didn’t find it useful, it wouldn’t have 900 million weekly users.

Obviously they’re useful. They still shouldn’t be used. Both because they do a bad job that is difficult for their users to see as bad, and because they’re 17 different kinds of unethical.

The people that treat it like a tool don’t go around telling everyone about it.

What rock have you been under? None of the people in my social circles who use AI as a tool will shut up about it.