Skip Navigation

Who here has actually used any GPT or Image generator models?

Alternatively, are there any publications engaging with the future potential of AI in a communist society, or in the lead up to a communist society? A lot of forum discussion used to pertain to the cybersyn project in Chile. But I feel like the LLM/ML/AI discussion is currently on a reactionary back foot.

If you've used it, I'm curious as to what potentials do you see from it? Where is it being used effectively where are you encountering it where you go, damn that sucks.

Where it sucks: Every single chatbot and phone tree when calling a support line ends without speaking to a flesh-and-blood agent. It makes me want to throw my phone into the woods.

Where it rules: Fulfilling bullshit work requests like OKRs with AI-generated drivel. Translation work. It tutored me through a Grad-level Statistics class that was all Asynchronous.

21
21 comments
  • I've played around a fair bit with Stable Diffusion running locally on my PC and can confidently say it has an absurd potential as a streamlining tool for art because unlike the simple prompt remote tools you have tons of tools available to control composition, pose, details, etc instead of just giving a vague prompt and hitting "generate" over and over like it's a slot machine and you're addicted to the rush of seeing if the latest pull will be a winner or more garbage.

    That said, if anyone's actually using it well they're keeping it to themselves: the hobbyist AI community itself is compromised of the most vapid, talentless dipshits you'll ever see who want pats on the head for asking the inscrutable machine "pls give good image" while using something like A1111 or Forge because ComfyUI's extremely simple flowchart UI is too hard for them and they want a single text box they can ask for their skinner box treats instead. And don't even get me started on civitai, a site that I think I've actually gotten addicted to raging at, like I'll check it multiple times a day just to kind-vladimir-ilyich about its userbase.

    I've compared it to things like Poser or Daz3d before, which caused floods of absolute dogshit CGI content made with prefab assets that still haven't completely subsided to this day, which is made all the funnier because of how many people seem to want a terrible faux CGI look out of image generators. That, the weird sort of soft-focus glossy style that makes me think of like sketchy airbrush art on cars for some reason, and the creepy grasping for photorealism are all weirdly popular with them and it's all uniformly awful drek.

    Still, I paradoxically think it has a ton of potential for more minimalist aesthetics where it can serve as a labor amplifier for an artist that can feed it a sketch with maybe some crude shading, and then fix up the output in short order, it can create out-of-focus filler for backgrounds, etc. I strongly believe that open source models should be embraced and exploited for that purpose by independent artists and the left in general, because the corporate use of proprietary models is not going to stop and we should adapt to the way that's transforming the landscape and seize upon new capital to compete and survive. The only alternative would be adventure-time no-mouth-must-scream bazinga kind-vladimir-ilyich minecraft en masse, and there's just not a militant anti-AI movement that could do that nor is one going to form because of how abstract and unemotional the problem is.

    • I've been playing with Stable Diffusion too, and the gacha/slot machine comparison is apt.

      1 out of 50-100 images feels worth going through the followup of upscaling, tweaking and making something presentable for my low stakes desire of "gallery of wallpapers matching my particular taste."

      The big deal for me is "low stakes." The catboy twinks have the wrong number of nipples? Your 20-page nobody-actually-reads-it report goes rampant and screams "pork pie!" in the middle of page 12? Nobody dies, no important process is at risk. I don't trust LLMs with code because "plausible nonsense" isn't good enough there. One missed subtle ! or & compketely changes the behaviour.

      I should try ComfyUI; to be blunt, I went with Automatic1111 because there was a decent walkthrough on how to set it up on a Radeon card (using distrobox to isolate the nastiness that is Ubuntu from the rest of my machine)

      • I should try ComfyUI; to be blunt, I went with Automatic1111 because there was a decent walkthrough on how to set it up on a Radeon card

        IIRC setting up comfyui for AMD is the same as with A1111, it just only works on linux because it doesn't have directml support. I just cloned the conda environment I used for A1111, IIRC I uninstalled the pytorch libraries from it, then let comfyui install the versions it wanted. I don't think there was anything more involved than that, just setting up a launch script with the right environment settings to try to mitigate the frequent hard system crashes I get with ROCm which may or may not have been solved by switching from ROCm 5.7 to 6.0, changing a certain kernel value, adding the right environment settings, and underclocking my 6800 even more than it already was - I haven't crashed since then but I haven't stress tested it either.

  • Ive found it mostly useless. It constantly ignores simple commands like (make image a full length portrait). The writing is d tier at best, and it tends to collapse upon itself quickly.

    Also it has no idea of context. Ask it to say..."draw a woman wearing a 1770s dress in a historically accurate manner", and it will instantly draw something that is wrong even to a rank novice who cant tell panniers from a farthingale.

  • The hype is massively overblown (of course) but I’ve recently found ChatGPT incredibly helpful with learning coding and data analysis in Python. It’s been surprisingly good at breaking down scripts of code I don’t understand, as well as giving me the skeleton of a model for me to use. And if you go back and forth with it, you can usually work out any kinks in the code it doesn’t immediately recognize.

    I think if people don’t get any wack ideas in their heads about what they can do, GPTs can be pretty reasonably useful tools in some cases.

  • If you've used it, I'm curious as to what potentials do you see from it?

    AI, in whatever form, is a Force Multiplier. Being that we are massively outnumbered and overpowered by the foot soldiers of capital, you'd think the left would be falling over themselves trying to master tech that wildly boosts our abilities to create propaganda, spread propaganda, and perform uh (illegal-to-say) ethical penetration testing. Because the reality is that this is going to be used by right wingers to do all those things while we ironically yell bazinga or whatever and argue that it's not really intelligence.

    Personally, I'm using ollama locally to run a Mistral llm for coding and computer security related questions. You can bypass most restrictions by asking questions in the context of writing a story. I see a lot of potential in automating much of my security work as the agent/agi frameworks advance. Working with documents is pretty nifty as well and I'm looking forward to when the local models get some of those massive context windows I've been hearing about.

  • I learned how to code (barely) in my 30s and bullshit my way into a dev job for a big corporation. I've been doing it for like 5 years now and I still don't really know what I'm doing a lot of the time. I've gotten really good at using copilot/gpt to check my work. Copilot especially seems to save me a lot of time. I'll write some code and then let copilot look for improvements, simplifications, etc. It's not quite at the level where I trust it to just write all my code from a prompt, but it can use context from the rest of the file (or other files) and really does a good job most of the time, especially if I give it a lot of info in the prompt too.

    It's GREAT at writing tests for code. Like I can just about give it something like "write some tests for this file", but it still makes small mistakes here and there or it'll give really generic names to things.

    The predictive text function, where I can start typing out some code and it'll give me options for the rest of it, I've found to be incredible, but that tech has existed for quite a while now so I'm not as impressed by it

  • I have used both chatGPT and a couple of image generators. They're complete trash, useful for nothing, and only made me appreciate human workers even more. I have since commissioned multiple pieces of artwork from human artists.

  • I've used a self-hosted Llama 3 to answer some questions about css and centering a div that I was having trouble with (I'm not a web dev by profession, nor am I aspiring to be one). You have to prod at it a few times to get it to tell you something useful which it ultimately did.

    That's about as far as I can work with it: asking and re-asking it very common questions that have been discussed and answered 700 times over (but the answer to which is unknown to me, specifically) in the hopes of getting something actually useful. So to that end, of course it can give me an example implementation of common leetcode questions in C, but it cannot reliably do something that requires a bit more originality.

  • If you've used it, I'm curious as to what potentials do you see from it? Where is it being used effectively where are you encountering it where you go, damn that sucks.

    Impressing people with no technological literacy and submitting bullshit work so I can go to sleep

  • I use diffusion models a fair bit for VTT assets for TTRPGs. I've used LLMs a little bit for suggesting solutions for coding problems, and I do want to use one to mass produce customized cover letter drafts for my upcoming job hunt.

    Neither model class is sufficiently competent for any zero-shot task yet, or at least has too high of a failure rate to run without active supervision.

    As for use in a socialist society, even the current version of the technology has some potential for helping with workers' tasks'. Obviously, it would need to be rationed per its actual environmental and energy costs as opposed to the current underwriting by VCs. You'd also want to focus on specialized models for specific tasks, as opposed to less efficient generalized models.

  • I use it all the time. I spent pretty much all of yesterday afternoon using ChatGPT to implement a Reinforcement Learning Algorithm I only half understand into my CCP Milker VR App. [nsfw]A couple months ago, I used it to cludge together the start of a DLNA file browser for Unity. I find it incredibly useful for fleshing out ideas or getting me motivated. It's a lot easier to get working on a project I'm not confident on if I have something basic to start with rather than a blank page. Like, I've been going through my backlogs of Idea Lists that have sat untouched simply because they require too much research to figure out where to start. A quick query and I've got a springboard to jump off from. It's far from perfect, but absolutely helps me get more stuff done.

  • I use GPT for writing inspiration. Nothing it writes is usable - and I wouldn't want it to write for me anyway - but it's quite good at giving me a bunch of ideas and letting me pick the best ones and develop them on my own.

    Image generators are ugly and make me want to cry.

  • I found this article interesting:

    https://cosmonautmag.com/2023/05/artificial-intelligence-universal-machines-and-killing-bourgeois-dreams/

    I've been meaning to read this too: https://www.versobooks.com/en-ca/products/735-the-eye-of-the-master

    If you haven't read it, and while it isn't about AI, Cockshott, Cottrell, and Dapprichs newish book titled Economic Planning in an Age of Climate Crisis, is all about cybernetic planning. It's pretty short. It's on libgen, don't give Cockfuck money.

  • I used midjourney or whatever to make my obligatory images of Drake with Trotsky but it got so boring

    I tried fucking around with AI dungeon too and it's just complete dogshit

    They should be able to find ways to use this tech but trying to make it do everything seems to be all they want to do it's so lazy

    • There are a lot of applications for AI that has nothing to do with art and which the average person doesn't really need to worry about - things like number crushing, boring statistics stuff, etc. It's just that the bazinga brains over at those corporations felt the need to make it into the hot new trend, even though for most people it's at best a novelty that you'll get bored of in a day, and a plagiarism machine at worst.

  • I used it at work to ask questions about various engineering standards, and it often provided good answers without having to slog through forums or blog spam (or, god forbid, actually reading the standard directly). I don't use it so much anymore, because of how much it gave subtly or very wrong answers. You can definately encounter wrong answers from online commenters, but you might at least get some pushback from other commenters, so it's sort of self correcting.

    I also found it weirdly good for asking questions about Marxist theory while reading through Capital. Unlike most people with an opinion about Marxism, ChatGPT actually appears to have read the source material. Still, it can't help but editorialize sometimes.

You've viewed 21 comments.