LLMs... Do you use them?

[Yoylecake420]

I don't use LLMs at all. I'm soooooo tired of seeing ChatGPT-generated writing and AI-generated images all over the internet. I have misophonia (meaning I get irrationally upset when I hear certain sounds, like people talking with their mouths open, and have to leave the room or drown them out) and I think I might have a LLM version of that particular affliction as well. AI-generated stuff is very obvious to me, and it drives me nuts when I see other people congratulating the "author" or "artist" on their "excellent writing" or their "artistic skills." I don't care if the people who use them are upfront about it, but it rubs me the wrong way when I see them taking credit for something they obviously didn't create.

I also have major issues with the companies behind these products. The way they stole basically everything on the internet, the way they exploited low-wage workers in the Global South to "safeguard" the tech for widespread consumption... If anyone's on the fence about it, you should read Karen Hao's excellent book, Empire of AI.

What happened to curating our online experience, blocking shit, or calling out people for AI? Have you considered taking a break or filing complaints or heck, lashing out?

[st3phvee]

What happened to curating our online experience, blocking shit, or calling out people for AI? Have you considered taking a break or filing complaints or heck, lashing out?

Oh, I mute it when I see it. I'm just getting tired of having to do so much muting, y'know? It makes me sad that the internet is fulling up with slop so rapidly. It's one of many reasons why I barely use social media anymore. The few times I've tried calling them out in the past, the authors just blocked me, and their fans piled on saying that it's not AI writing, not AI art. I constantly feel like I'm that scene of Mr. Mugatu from Zoolander ... "Doesn't anyone notice this?! I feel like I'm taking crazy pills!"

People just seem to be really bad overall at realizing when they're looking at AI-generated stuff these days (esp. when it comes to writing), and I wonder if it's because a lot of people just haven't really tested out the different things that are possible with gen AI yet? Or maybe it's because they're so used to skimming stuff online and not engaging with it closely? I tinkered around a little bit with ChatGPT back in 2024 when everyone was raving about GPT-4o (I wasn't impressed), so all those annoying, repetitive stylistic tics you see in AI-generated writing (stuff like overuse of "it's not just [X], it's [Y]") are really obvious to me now.

[star dreams]

In a similar boat as a few people here, I use Chatgpt every once in a blue moon to help with a coding problem. Same experience; good for a trivial line of code you know is possible but can't figure out (and happens to be juuuust too niche for stackoverflow to have an answer), not so helpful beyond that.

Most frustrating would be when they hallucinate a function and I keep bashing my head against a wall trying to get it to work, only to realise it was never possible in the first place. An unfortunate consequence of companies trying to prolong user usage (and prove to investors that their money is not evaporating into nothingness) is that big corporation LLMs like Chatgpt will always try to agree with you or accommodate for you, even if you say something completely wrong. (though it can be kinda funny sometimes). Only recently learned of the idea of self-hosting an LLM, and I've seen people train them to be actually useful, so that's an exciting idea I'd like to try someday!

I used to love it when it was janky and you could tell it was clearly made by a machine

This has always been a fascinating thing to me as well. My personal favourite is that chair excavation video. I think it genuinely blew my mind the first time I saw it? The fact that Sora doesn't understand that a rigid object doesn't change shape or move on its own, that a person shouldn't disappear after walking behind someone, the way gravity is just barely mimicked but not obeyed at all... I was suddenly made aware of the unconscious assumptions of the way the world works by witnessing the image generator's unawareness. As a visual artist it made me realise: holy shit, I've Got to get weirder with my craft. Cuz yknow what, why does a chair need to obey gravity, even in a world of my own making.

At the very least, I find ai has a place in art for that, challenging the way we conceive of the world and its boundaries. Though that video was defo just someone trying to see how "good" Sora's generation capabilities are, with specific intentionality I rlly think Ai could be used as an interesting medium for artistic exploration; but I guess we don't get to be in that timeline.

As a final note, I'm paraphrasing a little bit from my posts in another forum thread on the 32-bit cafe's discourse forum. The thread also has a little bit to do on Ai usage on the small web and its implications on creativity, I'd recommend a read! Lastly, I highly recommend Nicky Case's Ai safety explainer for anyone who's tired of sifting through online discourse about the vague concept of Ai. Specific, critical, and made for a layman like me! They also made a blog post a few months ago about using Ai for some more niche purposes (therapy and even trying to clone themselves) that I find pretty interesting.

[nvalk1]

The popularity of LLMs has the same implications and needed considerations as other popular technologies that have popped up during this century. I personally see a lot of similarities between LLMs and the digital walled gardens colloquially referred to as "social media".

The main similarity I see is that these programs are designed as mirrors—they exist as reflections of the users. A tool is only as good as the person using it, meaning that someone with knowledge on how it works can use it to produce very meaningful and high-quality work because they understand the tool's capabilities and limitations. That's the primary aspect I see with LLMs and other "AI" tools, but it does fall for the same capitalist trappings we've seen with programs like Faceboook, Twitter, Tiktok, etc.

The main trapping I see is the implementations of Bayesian clustering, which in layman's terms, basically is just the idea of all of the program's users being put into distinct "bubbles" and as such the users can only see what's within their bubble. Software designers implement this because the more a user believes that the platform is a reflection of themselves, the more they'll trust it and the more they'll use it.

For me, I think LLMs have been helpful with processing ideas and the ability to receive immediate feedback is much more convenient than discussing whatever I'm thinking about with a person I know. Initially, I used LLMs as a "sparring partner" very frequently and would have multiple chat threads per day, but now that I see the limitations of them for my own use cases, I don't use them that much anymore.

I write and code, so incorporating LLMs into my workflow has been more helpful than not. I don't have LLMs write anything for me, but I do use ChatGPT kind of like one would use Grammarly (which is also "AI"). I'll ask it for help with finding typos, grammatical mistakes, and style suggestions. I don't take every suggestion at face value and more often than not reject its suggestions than accept them. But when I do accept them, it's typically for what I find to be a good reason.

For coding/programming, I have found LLMs to significantly reduce the barrier to entry on that stuff. Things that would've taken me years to learn, I can implement in a matter of weeks. I avoid "vibe coding" by trying to keep my prompting as specific and as technical as I can regarding certain requests. Even then, LLMs are most optimally used for small-scale projects. Doing anything at scale makes it hallucinate too much and break things.

I think for me, the sweet spot with LLMs was around this time last year (2024) and once I can self-host a model with a few hundred billion parameters on a relatively modest hardware setup, I'll stop using commercial models entirely. Model distillation is getting a lot better, so hopefully within the next year or two I'll get it going. Fingers crossed, though.