@dnsprincess Large language models are an interesting technology that deserves academic research.

As a "product" it just doesn't work. All the promise of what these glorified chatbots could do has not materialized - and as study after study shows - it's not getting closer ("newer models tended to perform worse in generalization accuracy than earlier ones." - https://royalsocietypublishing.org/doi/10.1098/rsos.241776 ; other similar claims of RLHF reducing the quality of the outputs abound).

@dnsprincess I hate the dumb, useless work it's making me do.

I've got a couple of junior engineers that I'm leading on the technology front. Multiple times now I've had to review the AI slop they've tried to use as documentation or change controls and reject it because it's wrong, useless, or incomprehensible.

"But why do I have to redo it?"
"Okay, explain how you're gonna do step 5."
"Uhh..."
"Yeah, your change control includes steps for things we don't have."

The root cause is a people problem, but AI slop is a symptom in the same way that food poisoning is the disease but I'm not any happier about the vomit on my shoes.

@dnsprincess Because for most uses it's fundamentally a stopped clock: it might be coincidentally right twice a day, but it's still useless as a clock, because you can't know when it's right, and if you had a way to reliably tell, why are you "asking the AI" in the first place?

It's "confidently incorrect" frequently enough that it's impossible to tell without extensive cross-checking, and the effort to cross-check is frequently as much or more effort than doing it the "old-fashioned" way.

@dnsprincess You know how when you were a kid, when you didn’t know something, you’d as an older sibling/cousin/relative, and they’d just make up some bullshit that you didn’t know was bullshit but sounded plausible?

It’s that.

It’s nondeterministic - you don’t get consistent responses, so it’s useless for any task that needs to be performed consistently.

It’s that we lose crucial skills when we do t practice the manual way of doing things at least once in a while https://public.milcyber.org/activities/magazine/articles/2023/draeger-the-autopilot-problem

It doesn’t take feedback the same way a learning human does. If I commission some artwork, the artist will send me an in-progress sketch. I can red circle any areas that need changing, maybe sketch in a suggestion, and the artist can work with that! It takes the work it’s already done and makes those changes and moves forward. GenAI doesn’t do that, it’s generating totally new images every time. So it can’t “tweak”.

@dnsprincess The app SUNO showed me I can't tell the difference between modern country/pop and AI generated ones.

Wait, who are we roasting again?

@dnsprincess This isn’t really what you asked for but I’m a white middle class American male so I feel the need to interject my answer anyway.

I don’t hate AI despite so many of our colleagues hating it.

The environment, data privacy, and content theft (I’m mostly anti-copyright already) is all I really have problems with.

I actually want Fully Automated Luxury Gay Space Communism and I think what we’re doing is steps towards that.

@dnsprincess I asked Microsoft copilot a question about a system event id, knowing that the old event id was depreciated with Windows 8.1. Iit gave me the depreciated event id.

I asked google gemini for help building a chronclie rule. It got the syntax wrong.

This is two examples of using the vendor's AI to answer a question about vendor systems, and both times they were wrong.

AI doesn't give you the right answer, it gives you the popular answer. It's Ask the Audience, but the audience has severe cognitive impairments.

@dnsprincess The end of truth.

https://taggart-tech.com/ai-llms/

@dnsprincess

Because it's World Goth Day, and Fuck AI.

@dnsprincess

AI isn't useful. When it succeeds, it's by accident/coincidence. What is being sold as AI is in fact a ton of half-baked ideas and promises that have yet to see meaningful delivery to anything other than authoritarian systems. It's hype that empowers the worst of our society to dole undue and innumerable harm to everyone else.

It's not designed to do it's job well, if we presume it's job is to perform a task. It's intended to Present as useful, but it never measures up to any degree of scrutiny.

The primary thing that AI is, is:
- A hype engine to capture venture capital
- A tool to provide to the ruling classes to intensively devalue the labor of individuals

It's a bubble with no sellable product, to quote from someone else, 'A gold rush where there hasn't been any gold but a lot of people selling shovels.'

We are sitting in front of the world's wealthiest men who are telling us that their product will dramatically change society for the better, all the while pushing to displace people from their employment and replace them with a bot that can't even complete a children's video game.

- There's no profitable end to this venture, it's a $600bn (so far) money pit that everyone is pouring other people's money into
- The products being released are not delivering on what was promised, with everything packaged as "It will be able to' or 'In later updates' etc, all just a series of big lies about the technology to keep the investor money flowing
- The end outcome is there will be a few small models that can be used for some minor data processing tasks, but the long term doesn't include these companies massive LLMs doing anything but poisoning the well of information on the internet to deliberately drive people into walled gardens to find information, information that will be tightly controlled and delivered by the platform (see: Google's recent AI only search, after spending the past 7+ years making regular search bad to push more ads)

The problem with AI isn't the tech. It's the cult of personalities that have come around it, hailing it as the best thing since sliced bread, all the while pedaling a Racist Linear Algebra Machine.

Not a single promise has been met. No significant changes have happened without immense cuts to the labor force with the replacement resulting in worse service across the board, with the big names actively pushing the government to legislate their platforms into permanent fixtures of our society and operating entirely in contradiction to the will of the Market and will of the End User, because at the end of the day, the Customer for LLMs is either:
- The rapidly fascisticizing governments who want it as a tool of public persuasion and information control
- Nominally Literate investors who are caught up in a pipe dream of promises and not wanting to miss out 'just in case'. Think Pascal's Wager, but for investors not wanting to miss out.

There isn't a single useful thing that LLMs can be reliably used for without massive caveats and massive amounts of human intervention to get something usable in the final outcome. As a technology, it's at best used for *some* translation tasks, and even then it comes with strings attached. When the bubble is popped, it'll result is a massive economic downturn that did not have to happen at all. But thanks to the Silicon Valley culture of over promising and under delivering packed up inside an ideology of "effective altruism" that is currently and actively dealing massive harm to the Kenyan communities and other Majority World workers who have been left with PTSD and often times spontaneously have their jobs taken away with 0 recourse.

We will look back at this as a period of mass hysteria and psychosis and it'll be one of the most embarrassing times in human history.

@dnsprincess This is pretty mild, but I hate how people are using it. I think savvy folks know its limitations but 95% of users will think they're talking to a sentient being, form relationships with it, and trust it to always know what it's talking about.

It's just a technology tool that can be used for good or ill. The problem exists between keyboard and chair.

@dnsprincess Because when I’ve asked it to do the things I wanted it to generate, text or image, it refused because the topics I chose were forbidden.

Free speech much? This iteration of “AI” LLMs are free speech inhibitors.

@dnsprincess so my thing is, I'm a weird mix of both early adopter but somehow still some kind of Luddite, I never remember it exists. I dislike it for the environmental reasons you mention and what's a clear link to fascism, that I've seen a lot of this even in my job, the higher-ups just want to automate everything. They'd love to ditch human workers.

But the reason I dislike it most, is probably just how phony it is. The weird, surreal uncanny valley crap, it just really speaks to something when you see it, that I can hardly describe. But yeah. I guess I just like real stuff, let me see those stretchmarks, etc. I don't want glossed-over, phony shit in my life, I never have. I want real shit. And that's what bugs me most about the 'AI' (can that count as something I hate about it? That we're all calling it 'AI' when it's not artificial INTELLIGENCE at all, really just autocorrect, the LLMs?) but I'm pretty sure the thing that bugs me most is the phoniness.

@dnsprincess I hate because it's essentially a bullshit generator. Like that coworker that has absolutely no idea what they are talking about but cannot admit it, so they start to make up stuff from thin air and say it with utmost confidence.

I double-hate it because the people who are getting obscenely rich from it pretend it does a lot more than it actually does to sell this bullshit to C-level folks.

I triple-hate it because the C-levels actually buy into this shit and act as if the fault that they are not getting their castle in the skies was from the dumb people who don't know how to use AI.

Basically in its current format it's a new type of SaaS: Stupidity as a (dis)Service.

@dnsprincess Once we have AI I'll let you know.

Now, the products called "AI" exist to crush labor. They don't have to be good at all, just perceived to be good enough to allow the rich to force highly skilled people to accept work for a fraction of the money/benefits. The only people who need to be convinced LLMs can replace skilled workers are managers and shareholders.

@dnsprincess Because it's an odious trap.
It is just good enough to make people think that they can use it as a lazy substitute for thinking skills for them to start depending on it and lose their own.

Even if it was in fact intelligent this would be hugely problematic, but as it is this is a recipe for disaster as its thoughtlessness will damage things it's directly used for and the thoughtlessness it induces in its victims will do indirect damage.

@dnsprincess Executives are all in on LLMs because they are a firehouse that sprays unpaid interns at a problem. Cheap, replacable labor that won’t talk back. But it’s an intern who never learns, who won’t set the office on fire, but there’s never an outlier worth hiring. Why bother? More interns!

But CEOs are so far up their own asses that they forget what the most privileged intern is there to do: kiss ass and drink coffee. And AIs can’t drink coffee. You can’t motivate a person who doesn’t exist, they’ll just read what the last kid did, hand in something plausible, and then piss off back to their rich parents’ weekend houses to blast music and do cocaine and wine coolers with their college friends forever