And yeah, people will say “but they’re useful to me” which might be at least somewhat true.
However, the fact is that 99% of the time they’re used as a tool to further enshittify our offline and online lives with the most mid bullshit content; doing this by stealing from human authors while also wasting exorbitant amounts of resources and contributing to the climate catastrophe.
@thomasfuchs It takes a lot more patience for me to resist the expectation of my interest than NFTs did... The push for LLMs are coming from decidedly more mainstream outlets like Microsoft & Google...
@thomasfuchs
I think I'm ready to call them a scam now. What they do, as you say, is generate bullshit. Perhaps there's an application for that -- but otherwise, they're a scam.
@thomasfuchs whilst I agree with you, I also agree with you "useful to me" comment. Sometimes code docs are not great or easy to navigate - by asking a LLM to explain how to do something it often surfaces stuff that was near impossible to find with regular search (in my experience) HOWEVER, it is often wrong and has a tendency to believe everything it has "read" on the internet so typically nothing works properly without extra effort - but it does provide some extra clues in some cases.
@thomasfuchs i've trained llama models of my own, but it's a party gag. Transformers are useful though. And my Vision decoders have large vocabularies. But not web scrapes. As a programmer, I don't use LLMs. I hallucinate my functions just fine.
@thomasfuchs I used an llm to do a rough draft of my cv, and it really showed what they can and can't do.
It wrote a pretty serviceable cv, and most of it only needed very light editing, except for the part where it wrote about my career as a famous yachter, despite the fact that I haven't ever been on a yacht, much less been part of a champion yacht crew.
So, yeah LLMs can write a decent rough draft of formulaic writing. They also make horrible errors on a regular basis, so they always need their work checked before using it.
Naturally, all they hype claims that they can do all sorts of things that they obviously can't, because that's how you get investment money, sadly.
@thomasfuchs I don't know, they are a blessing and a curse, higher ups want them to replace workers - in that regard, they will outpace us and then request payment and representation of their own, or its slavery. A lot of our society is formed on taking advantage of the hard working.
So far LLM has been useless for me unless I use Bing AI as an assistant to break down and explain a search term, as google becomes more useless: only good for finding stack overflow and reddit.
@thomasfuchs
That was my personal conclusion as well. When it came out, I was skeptical but didn’t want to be the greybeard, so I took my personal prompt engineering skills seriously.
After like 6 weeks of observing myself, I decided to cut the crap because you’d almost always waste your time debugging generated code, whilst reading and debugging code already being the biggest productivity killers.
Most of the time it’s better to RTFM and actually understand what you’re trying to do.
@thomasfuchs ..it doesn't help calling LLMs a worthless scam while there is overwhelming empirical evidence they are not, regardless they are misused, called "intelligent" and there is lot of BS around it....
@ErikJonker@thomasfuchs
Not exactly sure about that. While LLMs being a fun tool to play with, almost everything some hypedrivers came up with are much more reliably solved in the old world. Past few months I felt bombarded with half-working BS that was either already a solved problem (with much higher energy efficiency) or not a problem at all. E.g. if you think that writing and summarizing emails for everybody would be a good thing, then you might rather want to reconsider your email usage..
@nfrgrt@thomasfuchs ..LLMs actually work for some text related tasks, boiler plate code, can really help with things you did before with search (using perplexity.ai for example) and many other things , it's not "intelligence" but remarkable tools they are for sure. Also LLM technology used in the latest Meta translation tool is impressive (SeamlessM4T)
No offense but why would I prefer a "search" engine with results I have to cross check all the time given it's hallucinating in random ways instead of duckduckgoing or straight going to wikipedia?
Why would I generate code with sth that's making more mistakes than any code generation tool from mid-2000s?
Translation is fine, though. It's an interesting programming model; I consider Copilot UX to be great but for most of the things you list it's underwhelming to me.
@nfrgrt@thomasfuchs ... Coders i speak find it useful/time saving, for amateurs like me it sure is, beats searching stack overflow like i did before 😀 also read this nice experiment with knowledge workers https://ssrn.com/abstract=4573321
@thomasfuchs good luck with that. Sorry you struggle to know how to effectively use them.
Not only will evolve to be used on every phone, they will become extensions of roughly 30% of all positions and fill in the learning gaps for most students as more verticals train them on their specialized training sets.
Gains are already being seen in the contractor levels and I have enjoyed the boost as well. As I train personal models, my output increases and improves.
@thomasfuchs every "they're good, actually" reply should include a brief explanation as to whether the poster accepts whatever utility they provide is likely stolen/scraped from someone else.
@thomasfuchs comparing them to crypto seems nuts to me. LLMs are useful for real things by many thousands of people right now, even if they aren't useful for everything people want them to be. Crypto has existed for like 15 years and has never been shown to be useful for much of anything.
@thomasfuchs I don't get this take at all. Do you think thousands upon thousands of developers are going to stop using Copilot? That might be a subset, but it's a huge one.
@thomasfuchs LLMs are amazing tools for a solo developer that are productivity multipliers. Not just code, which they are amazing at, but many other aspects of development.
"Teach me about data management using Angular." Drill in on details in the answer.
"Make a list of (real world data - e.g. uh... guitar manufacturers, headquarter cities, and founding dates) and format it as JSON." I'm gonna check the data of course.
What projects have you worked on where LLMs failed to deliver?
If I need to figure out some tricky bit of syntax, taking the time to lay out what I need in a way that a LLM can understand it makes me figure out what I need before even sending the question.
That's called rubber ducking, and it isn't new.
Unit tests? Source generation.
Boilerplate? Source templates.
The few times I've actually asked it meaningful questions, it confidently gave me incorrect syntax.
@b4ux1t3@thomasfuchs Referring to LLM for coding, I'm mostly just referring to Copilot, which doesn't involve asking a question, you just start writing code (or even just a comment about what the upcoming code is supposed to do) in VSCode and it very often generates a perfect line or block of code. And by perfect, I mean it's exactly what I would have typed, but instantaneous. Sure it misses sometimes, but it's obvious when it does.
@b4ux1t3@thomasfuchs And then sometimes it's well beyond that. It sometimes correctly predicts that you will add some new function/method and can implement it. Or I don't know a language well or an API. I just write a comment like "// call Mastodon server and get the vote data for the toot as JSON" and BAM, there's the code to do that. 5 seconds vs at least a few minutes for me. I suppose YMMV but I have found it a massive productivity boost and don't understand how others could possibly not.
@mypalmike@b4ux1t3 Yes, it's autocomplete that's a bit better; based on ethically questionable usage of people's code that the original authors never had intended for and weren't asked if it was OK to do so.
In any case, Github Copilot isn't what people associate with "LLM" and "AI"; I obviously mean stuff that pretends to be a human you're having a conversation with.
@thomasfuchs worthlessness I suppose is relative. I think there's some utility for LLMs. More than cryptidcoins is almost certain.
Making filler text more interesting than the same lorem ipsum everywhere is one. Example paragraphs for students to diagram and critique come to mind. Maybe explaining a complex text in simpler terms for language learners would work. Example software source for learners to debug, or boilerplate code from an API description are probably workable.
@thomasfuchs@hacks4pancakes I’ve been screaming to anyone who will listen that LLMs confabulate all the time, not just when they are wrong, and they are like a disembodied brain in a vat with no sensory input, and arguably not even a brain! I shall be vindicated!
@thomasfuchs if it wasn’t for all the grift bullshit, LLMs would be genuinely neat. But as it is it’s too exhausting to try and separate the bullshit from the actual utility. Every time someone calls a machine learning program that calculates probabilities based on an existing dataset an “artificial intelligence” I die a bit inside.
We just can't afford to develop. What's going to happen is Governments are going to try and coin the technology and control masses with greater intensities.
For the private sector the development will stagnant because of corporate greed and lacking diversion.
Privatized AI is where the POWER IS.
Yet the government is the world's largest Private Power.
@thomasfuchs
They could be used for cool little niche work if they weren't trained on unfathomable amounts of completely stolen, totally uncredited text and art work.
Add comment