LLMs... Do you use them?

[Monoki]

I haven’t been itching to use LLM’s personally. I have dabbled with Image Generation a couple times. But overall I personally can’t find much of a reason to be bothered to do so.
Can’t say why really. Part of it may be that I’m not really all that impressed, another may be that I prefer figuring things out myself and researching. AI flooding every bit of my search results annoys me.

EDIT: I might add an extra thought here while I lie in bed but I also think AI has created a mini-crisis of humanity of sorts.
I’m not anti-technology. At its best, technology is a tool that can extend our abilities as humans or a toy that can give us unique ways to have fun.
But I fear there is a dark side where technology is slowly consuming and replacing our humanity, and not just in a, “corporate businesses want to replace humans with robots so that they can make more money” kind of way, though that’s undoubtedly one symptom.
But in a more pernicious, misanthropic way. One that demands humans upgrade or perish because humanity as it was conceived of is obsolete.

I think, at a deeper level, this is my aversion to AI and probably why new tech doesn’t wow me like it once did. It’s no longer a tool in my hand that can enable me to create fantastic works of art. But instead it’s there to fundamentally replace me. Or create bigger pressures no human is meant to bear.

[varve]

I don't and won't voluntarily use an LLM, nor an image generator. Whenever I've seen somebody's results the text is bland and generic, and the images have more problems the longer you look at them. Summaries and search results are a thoroughly mixed combination of "true and useful", "true but not useful", and "false" information, and it takes an expert to figure out what's what.

I also try to block the scrapers that feed them from my writing and art; a perpetual arms race, but still worth trying.

[grilledmint]

At this moment I have a strong opinion on it. I'm afraid if I use it I become even more lazy and addicted to easy information(which isn't always even correct). Which, actually, was one of the reasons to get to this part of the internet. I plan to avoid this technology until it's still possible. I prefer to be not enough effective and progressive at the moment but at least I won't get any harm that may be founded later

BUT! Back in 2023-2024 I used it for one specific purpose - roleplay. In my case, mostly about topics that I'm too shy to roleplay with real humans (it's easier to say "oh yeah, I'm so kinky, I'm into this, that and those " then ACTUALLY saying this stuff “out loud”...). And it has one more huge advantage  - it's WAY faster. You don't need to wait hours or days to get an answer. And, even I really love to roleplay with humans...man, how many RP was abandoned because people just lost their interest or got too busy. Why do I stoped then? First: the more genAI around me, the more "I asked ChatGPT..." peeps online, the more I start to be annoyed by technology itself and it stops relaxing me. Second: for some reason answers just got worse. I don't think it's because LLMs are getting more stupid (I wish haha) but maybe they start to hide more poverfull one by paywalls and I don't wanna pay for this/host it. Third: I open Solo TTRPGs for myself and I find out that generating random stuff with my brain, dice and tables is more than enough and fun for me.
Still, I think RP with LLM is one of the harmless and still fun stuff you can do (except ecology stuff of course but I think it's just not a main possible danger)

[NoxidKin]

I've used them quite alot.  I'm learning to code this year and I can't imagine how people got on before them.

I know they're not popular around here, and I definitely am concerned with the many ethical issues with them as well.

I'm also trying to learn 2D art and finding that extremely difficult.  It's tempting to use image generators, but if I hand it over entirely it doesn't seem like my art any more.  But I have used image gen for filters and references.  I probably would use it more if it were better.  Although, again, I'd never have it make something from scratch, then release that as "finished" or "my" work.

I have used LLMs for casual conversation in the past, but I generally don't now.  It doesn't feel healthy.  That being said, I do bounce ideas off it, and share things that I don't think any humans are interested in.  Maybe more than I'd like.  (BTW check out my blog.  :)  )

But, like I said, I definitely recognise alot of problems.  Between the environmental, social and psychological problems, it is hard to justify using it at all.  Although the same could be said for many modern conveninces to be honest.  So I do use them, because they're very useful to me.  But I try to do so minimally and carefully.

[st3phvee]

I'm learning to code this year and I can't imagine how people got on before them.

We learned how to code by asking one another questions on forums/chatrooms, by taking classes, by following guides posted on random websites, by inspecting source code, by trying different things until we figured it out, etc. It was very doable to learn coding without LLMs (and still is -- even more so now than it was in the 90s and early 2000s). Teenage me would have been astounded by all of the excellent coding resources readily available on the internet today.

[lakes]

We learned how to code by asking one another questions on forums/chatrooms, by taking classes, by following guides posted on random websites, by inspecting source code, by trying different things until we figured it out, etc. It was very doable to learn coding without LLMs (and still is -- even more so now than it was in the 90s and early 2000s). Teenage me would have been astounded by all of the excellent coding resources readily available on the internet today.

Yeah, I sometimes wonder if people who rely on AI never learnt how to research things online before. Obviously this doesn't apply to everyone. For some people, they just think it saves time. (I'd argue you should still factcheck AI because it gets a lot of things wrong.) But there are plenty of resources online: W3CSchools, JQuery, forums for other webmasters & web hosting services, random source code repos on places like github and gitlab, even right-clicking a webpage and clicking "Inspect (Q)".

[fairyrune]

It's tempting to use image generators, but if I hand it over entirely it doesn't seem like my art any more.  But I have used image gen for filters and references.

hi, artist here. i would strongly advise against using them for references. the reason this is is because image generation, while it looks fantastic thanks to all the people who have been using it nonstop and feeding it information, still doesn't depict reality.

when you're learning to draw, it's very, very important that you use real references. image gen can make mistakes, won't account for the natural rotation of joints, won't be able to depict realistic facial expressions accurately, doesn't know the subtle differences between different animals and how they move if you're learning to draw those, accounts poorly for things like disability, weight and age... it doesn't know how those things work. similarly this also applies to inanimate objects. image gen can't account for what it can't see - it can't know the overhead structure of a building, and will make impossible corridors, windows, walls, etc. it can't know how an object looks from behind or beneath, so it will create something impossible there too.

as an inexperienced artist you might not notice it at first, but it's best to use references you can verify as real - even ones you take yourself! - to learn how things are physically made up, and then once you've learned the rules, you can learn to break them.

[NoxidKin]

Oh yeah I agree.  Unlike coding, I wouldn't suggest using AI to -learn- art.

And to those pointing out the ways you can learn coding without AI, I also agree, I was exaggerating when I said it seemed impossible.  But it does seem like shooting yourself in the foot.

[lakes]

And to those pointing out the ways you can learn coding without AI, I also agree, I was exaggerating when I said it seemed impossible.  But it does seem like shooting yourself in the foot.

No worries. Sorry if I came off as presumptuous. I'd imagine AI has more accurate coding knowledge than for other fields of study, since the developers knew how to code obviously. But I'd still recommend factchecking the stuff you get from AI if you don't already because I remember hearing of cases where AI trained on code from other AI can get all screwy.

[varve]

The problem with using an LLM to learn to code is that once you get past the super basic stuff (which it has enough samples of to mostly put together something that works) it doesn't have enough samples to put something together for you... and you're stuck not knowing where else you can learn things! That super basic stuff is not only about learning the basics, it's about learning where to look for more.

[stephan_e_perez]

I've used them quite alot.  I'm learning to code this year and I can't imagine how people got on before them.

Before LLMs, people would go search on Google, StackOverflow or software-specific forums to see if someone else had asked the same question and already gotten an answer. Otherwise, you post about it yourself and waited for a response. If no one answered, you could try another forum... or else keep trying to figure it out with the framework's documentation or through trial and error, look for a tutorial on YouTube, or give up. You can still do this, of course. Tutorials in particular are still a great option if you want to learn the basics of a framework.

The problem with using an LLM to learn to code is that once you get past the super basic stuff (which it has enough samples of to mostly put together something that works) it doesn't have enough samples to put something together for you... and you're stuck not knowing where else you can learn things! That super basic stuff is not only about learning the basics, it's about learning where to look for more.

LLMs sometimes make mistakes and a manual web search can occasionally give better results (this is especially preferable for very basic things). However, compared with one year ago, I would say LLMs are making less mistakes and are generally much faster than the above alternatives, ethical concerns notwithstanding. It's good to know where the documentation is for a given framework and to consult it if needed, but I'm skeptical that only using the documentation to look up how to do things will really teach you more... ultimately, it comes down to memorization and you don't memorize things by just reading them and typing them once- whether it's something you found after searching a long time through the documentation or something that was delivered to you near-instantly by an LLM. Either way, you probably won't memorize most method names, etc., so what I (and I'm sure many others) end up doing is either a) referring to similar things I've coded before, b) struggling to remember and relying on the IDE's autocomplete, or c) just looking it up again.

[NoxidKin]

Yeah again, pretty much agree with everything you guys have added.

It's useful as a sort of English > Code translator and dictionary.  But it's pretty terrible at anything beyond that.

It should be one of many sources, not your only source.  But that pretty much goes for any other source too.

I did recently make the mistake of trying to get and answer out of LLMs for 10-15 minutes, before eventualy resorting to Discord, and getting the right answer form a human immediately.  So it can be a bit of a trap in that way.

[Solinus]

Nowadays, I NEVER use them, and I have no need for them. As far as I can remember, these were my experiences with using LLMs:

Mid 2023: Tried to use ChatGPT to generate descriptions of different interior zones for a restaurant project. They were too long winded and felt inauthentic, and so I gave up and written the descriptions themselves.

Late 2023?: Tried to use ChatGPT again to help me out on physics homework, making sure I don't just make it spit out an answer so I can actually learn. Gave up as it spat out nonsense formulas.

2024ish: Comp Sci teacher had me use ChatGPT to make a python chess game with graphics from scratch, despite me not knowing about how graphics are made from code. Guess what, nothing worked.

Off topic but besides some screwing around pre 2024 and making a site parodying nostalgia-based slop (https://solinus.neocities.org/Super-frutiger-scam-slop/) I haven't really generated any AI images. I also remember in school that a friend in a hurry to finish a presentation for next class or so and attempted to use AI to generate a visual. Eventually he gave up and switched to Google where a much better image was found.

[st3phvee]

But it does seem like shooting yourself in the foot.

How so? I think not having all the answers immediately accessible via chatbot is actually better in the long run for learning. If someone just asks an LLM for a code solution without spending the time thinking about it on their own, without doing any trial and error on their own ... are they actually going to retain what they've learned as well as they would if they took the time to try to figure it out with extant non-LLM resources? Would they be able to extrapolate from what they've learned and apply those solutions to similar, but slightly different problems?

Some of my favourite moments in coding over the years have been those "Eureka!" moments, when I've taken some code example from some forum or Stack Overflow post and learned how to modify it to suit my unique needs. All those joyful moments of discovery and satisfying feelings of "shit, I'm smart!" go away if Claude or ChatGPT come up with those solutions.

I can totally understand why someone might want to use an LLM to speed up the tedious aspects of coding that they're already knowledgeable about, but I would strongly encourage novice coders to learn how to code without LLMs. I wouldn't even trust it as a code dictionary... It gets basic things wrong all the time, because at the end of the day, it's just a fancy predictive text generator.

[stephan_e_perez]

I would recommend to anyone of any experience level to make regular use their own brain for problem-solving- if you don't use it, you lose it.

Having said that, many languages and frameworks have bloated librairies and idiosyncratic idioms. Asking an LLM what tools are available within a given framework for performing a specific task can be a huge time-saver without atrophying your brain too much.