Skip Navigation
75 comments

  • I wish people stopped treating these fucking things as a knowledge source, let alone a reliable one. By definition they cannot distinguish facts, only spit out statistically correct-sounding text.

    Are they of help to your particular task? Cool, hope the model you're using hasn't been trained on stolen art, or doesn't rely on traumatizing workers on the global south (who are paid pennies btw) to function.

    Also, y'know, don't throw gasoline to an already burning planet if possible. You might think you need to use a GPT for a particular task or funny meme, but chances are you actually don't.

    That's about it for me I think.

    edit: when i say "you" in this post i don't mean actually you OP, i mean in general. sorry if this seems rambly im sleep deprived as fuckj woooooo

    • peeps who use these models for facts are obv not aware what the models are doing. they don't know that these models are just guessing facts.

      also yes, big sad about peeps in the south being paid very poorly.

      can totally see your point, thank you for commenting! <3

  • I honestly am skeptical about the medical stuff. Machine learning can't even do the stuff it should be good at reliably, specifically identifying mushrooms/mycology in general.

    • that is interesting. i know that there are plenty of plant recognition onces, and recently there have been some classifiers specifically trained on human skin to see if it's a tumor or not. that one is better than a good human doctor in his field, so i wonder what happened to that mushroom classifier. Maybe it is too small to generalize or has been train in a specific environment.

    • Having worked with ML in manufacturing, if your task is precise enough and your input normalized enough, it can detect very impressive things. Identifying mushrooms as a whole is already too grand a task, especially as it as to deal with different camera angles, lighting ... But ask it to differentiate between a few species, and always offer pictures using similar angles, lighting and background, and the results will most likely be stellar.

      • Like I said, I'm just skeptical. I know it can do impressive things, but unless we get a giant leap forward, it will always need extensive human review when it comes to medicine (like my mycology example). In my opinion, it is a tool for quick and dirty analysis in the medical field which may speed things up for human review.

    • From what little I know if it, it's sorta twofold what it does:

      1. It looks through documentation across a patient record to look for patterns a doctor might miss. For example, a patient comes in complaining of persistent headaches/fatigue. A doctor might look at that in isolation and just try to treat the symptoms, but an AI might see some potentially relevant lab results in their histories and recommend more testing to rule out a cancer diagnosis that the doctor might have thought unlikely without awareness of that earlier data.
      2. Doctors have to do a lot of busywork in their record keeping that AIs can help streamline. A lot of routine documentation, attestations, statements, etc. Since so much of it is very template-heavy already, an AI might be able to streamline the process as well as tailor it better to the patient. E.g. the record indicates "assigned male at birth" and an ER doctor defaults to he/him pronouns looking only at the medical birth sex marker, but the patient is also being seen by a gender clinic at which she is receiving gender affirming treatment as a trans woman and brings up that earlier data to correct the documentation and make it more accurate and personalized for the patient.

      In reality, I am sure that practices and hospital systems are just going to use this as an excuse to say "You don't need to spend as much time on documentation and chart review now so you can see more patients, right?" It's the cotton gin issue.

  • My biggest problem with AI is how it was pushed and marketed to us in ways that don't make sense / are unethical. Even the environmental concerns would be ameliorated if AI weren't being pushed into everything. (Using "AI" here to refer to things like LM, image, and art generators,etc.)

    • yes, i completely agree.

      having some LM generate "comment suggestions" for content creators on youtube is such a genuine waste of compute and the environment. (yes this is a real thing)

      it was marketed as this "smart machine" which ends up being too dum for most people wanting to use it.

    • This not relevant to your comment, really, but I would like to applaud you for word usage.

  • I used to think image generation was cool back when it was still in the "generating 64x64 pictures of cats" stage. I still think it's really cool, but I do struggle to see it being a net positive for society. So far it has seemed to replace the use of royalty free stock images from google more than it has replaced actual artists, but this could definitely change in the future.

    There are some nicer applications of image generation too, like dlss upscaling or frame generation, but I can't think of all that much else honestly.

  • I think we should avoid simplifying it to VLMs, LMs, Medical AI and AI for disabled people.

    For instance, most automatic text capture ais (optical Character Recognition, or OCR) are powered by the same machine learning algorithms. Many of the finer-capability robot systems also utilize machine learning (Boston Dynamics utilizes machine learning for instance). There's also the ability to ID objects within footage, as well as spot faces and referencing it with a large database in order to find the person with said face.

    All these are Machine Learning AI systems.

    I think it would also be prudent to cease using the term 'AI' when what we actually are discussing is machine learning, which is a much finer subset. Simply saying 'AI' diminishes the term's actual broader meaning and removes the deeper nuance the conversation deserves.

    Here are some terms to use instead

    • Machine Learning = AI systems which increase their capability through automated iterative refinement.
    • Evolutionary Learning = a type of machine learning where many instances of randomly changed AI models (called a 'generation') are run simultaneously, and the most effective is/are used as a baseline for the next 'generation'
    • Neural Network = a type of machine learning system which utilizes very simple nodes called 'neurons' for processing. These are often used for image processing, LMs, and OCR.
    • Convolution Neural Network (CNN) = a Neural network which has an architecture of neuron 'fliters' layered over each other for powerful data processing capabilities.

    This is not exhaustive but hopefully will help in talking about this topic in a more definite and nuanced fashion. Here is also a document related the different types of neural networks

  • I'll just repeat what I've said before, since this seems like a good spot for this conversation.

    I'm an idiot with no marketable skills. I want to write, I want to draw, I want to do a lot of things, but I'm bad at all of them. gpt like ai sounds like a good way for someone like me to get my vision out of my brain and into the real world.

    My current project is a wiki of lore for a fictional setting, for a series of books that I will never actually write. My ideal workflow involves me explaining a subject as best I can to the ai (an alien technology or a kingdom's political landscape, or drama between gods, or whatever), telling the ai to ask me questions about the subject at hand to make me write more stuff, repeat a few times, then have the ai summarize the conversation back to me. I can then refer to that summary as I write an article on the subject. Or, me being lazy, I can just copy-pasta the summary and that's the article.

    As an aside, I really like chatgpt 4o for lore exploration, but I'd prefer to run an ai on my own hardware. Sadly, I do not understand github and my brain glazes over every time I look at that damn site.

    It is way too easy for me to just let the ai do the work for me. I've noticed that when I try to write something without ai help, it's worse now than it was a few years ago. generative ai is a useful tool, but it should be part of a larger workflow, it should not be the entire workflow.

    If I was wealthy, I could just hire or commission some artists and writers to do the things. From my point of view, it's the same as having the ai do the things, except it's slower and real humans benefit from it. I'm not wealthy though, hell, I struggle to pay rent.

    The technology is great, the business surrounding it is horrible. I'm not sure what my point is.

    • I'm sorry, but did you ever think of the option to try? To write a story you have to work on it and get better.

      GPT or llms can't write a story for you, and if you somehow wrangle it to write a story without losing it's thread - then is it even your story?

      look, it's not going to be a good story if you don't write it yourself. There's a reason for why companies want to push it, they don't want writers.

      I'm sure you can write something, but that you have issues which you need to deal with before you can delve into this. I'm not saying it's easy, but it's worth it.

      Also read books. Read books to become a better writer.

      PPS. If you make an llm write it you'll come across issues copyrighting it, at least last I heard.

  • What does "AI for disabled people" entail? A lot of 'good AI' things I see are things I wouldn't consider AI, e.g. VLC's local subtitle generation.

    • true, we kinda move the barrier on what "AI" means all the time. back then TTS and STT surprised everyone by how it worked kinda good. Now we don't even consider it AI, even tho STT is almost always driven by a neural network, and new models like OpenAIs whisper models are still releasing.

      there are also some VLMs which let you get pretty good descriptions of some images, in case none were provided by a human.

      i have heard some people actually being able to benefit off of that.

      • Yeah, the way 'AI' companies have played with term AI is annoying as heck. The fact AGI has been allowed to catch on at all is frankly a failure of the tech press. I do remember reading a good article on how stuff stops being 'AI' when it gains real world use, that I can't find because Google sucks now.

        I don't enough about running AI locally to know if this applies, but I just can't stomach any of it because I can't help but think of what those companies put people in places like Kenya through in order to get the token data to make these models useful. It's probably unfair to taint the whole field like that, like I'm sure there are some models that haven't been trained like this, but I just can't shake the association.

  • A lot of those points boil down to the same thing: "what if the AI is wrong?"

    If it's something that you'll need to check manually anyway, or where a mistake is not a big deal, that's probably fine. But if it's something where a mistake can affect someone's well-being, that is bad.

    Reusing an example from the pic:

    • Predicting 3D structures of proteins, as in the example? OK! Worst hypothesis the researchers will notice that the predicted structure does not match the real one.
    • Predicting if you have some medical problem? Not OK. A false negative can cost a life.

    That's of course for the usage. The creation of those systems is another can of worms, and it involves other ethical concerns.

    • of course using ai stuffs for medical usage is going to have to be monitored by a human with some knowledge. we can't just let it make all the decisions... quite yet.

      in many cases, ai models are already better than expert humans in the field. recognizing cancer being the obvious example, where the pattern recognition works perfectly. or with protein folding, where humans are at about 60% accuracy, while googles alphafold is at 94% or so.

      clearly humans need to oversee AIs output, but we are getting to a point where maybe humans make the wrong decision, and deny an AIs correct generation. so: no additional lives are lost, but many more could be saved

      • I mostly agree with you, I think that we're disagreeing on details. And you're being far, far more level-headed than most people who discuss this topic, who pretend that AI is either e-God or Satanic bytes. (So no, you aren't an evil AI tech sis. Nor a Luddite.)

        That said:

        For clinical usage, just monitoring it isn't enough - because when people know that there's some automated system to catch their mistakes, or that they're just catching the mistakes of that system, they get sloppier. You need really, really good accuracy.

        Like, 95% accuracy might look like a lot, right? If it involves death or life, it means a death for each 20 cases, it's rather high. In the meantime, if AlphaFold got it wrong 60% of the time instead of just 6%, it wouldn't be a big deal.

        Also, note that we're both talking about "AI" as if it was a single thing. Under the hood it's a bunch of completely different things; pattern recognition AI, predictive AI, generative AI, they work so differently from each other that we'd need huge walls of text to decide how good or bad each of them is.

  • What I think is missing from your viewpoint (and from most people's, this is [IMO] a problem at scale) is the distinction between "simple" and broad machine learning, and the very specific things that are Large Language Models.

    For example, there are no small Large Language Models, and I think that the oxymoron speaks for itself. Machine learning is a very good thing, and automated classification is definitely its best use case, but they are not a small version of ChatGPT, the same way that the average Joe is not a smaller version of a billionaire.

    For more details, these small models are trained on a small set of data, how small depending on how specific the task is; as an example, I worked with models that detect manufacturing defects on production lines, and theses need a few hundreds images in order to produce good results, this make it very easy to produce the data ourselves, and it is relatively cheap to train energy-wise.

    Compared to that, Large Language Models, and their audiovisual counterparts, operate on billions of data, and work on a task so general that they provide incredibly bad results. As a little statistical reminder, anything below 95% confidence is a bust, LLMs are way below that.

    It's very important to distinguish the two, because all of the positives you list for AI are not about LLMs, but about simple machine learning. And this confusion is by design, techbros are trying to profit of the successes of other form of artificial intelligence by pretending that AI is this one single thing, instead of an entire class of things.

    Otherwise, I generally agree with the rest of your points.

    • i completely agree. training an actually small model on your specific task almost always results in WAY better output.

      current LLMs might be great at PhD questions, but are still bad at way simpler things, which shows that they have been trained on these questions, rather than generalizing to that level.

      training a "cancer recognizer" will be way more efficient and accurate than a general, much larger VLM trying to do the same thing.

    • wait no, there are small language models! like the one in the phone keyboard, suggesting the next word. sometimes there are rule-based but in many cases, they are real neuronal networks, predicting what you will type. in my case it even trains on what i type (an open source keyboard i got, running locally obv)

      • I'm pretty sure that phone keyboard use heuristics and not Machine Learning. Basically, it does not create a neural network through trial and error, but whenever you type, it saves the context of each word, and when it sees the same context again, it "knows" what the next word is.

        For example, if you type this big brown fox, it might saves something like "{ fox", ["big", "brown"], 1 } (assuming two words of context, and the 1 being the number of times it was encountered). Then when you type my big brown, fox will be suggested.

        Using the technology of LLMs for keyboard suggestions is impractical, as your typing habits would be drowned in the initial training data, and would yield worse performance as well as results compared to the simpler approach.

  • I think generative AI is mainly a tool of deception and tyranny. The use cases for fraud, dehumanization and oppression are plentiful. I think Iris Meredith does a good job of highlighting the threat at hand. I don’t really care about the tech in theory: what matters right now is who builds it and how it is being deployed onto the world.

    • oof this is brutal. but a good analysis.

      at the end of the day it, no matter what good uses people might have for this tech, it's hard to reconcile the fact that it's also being used by the worst possible people, with the worst possible intentions, in the worst possible ways.

  • Honest question, how does AI help disabled people, or which kinds of disabilities?

    One of the few good uses I see for audio AI is translation using the voice of the original person (though that'd deal a significant blow to dubbing studios)

    • fair question. i didn't think that much about what i meant by that, but here's the obvious examples

      • image captioning using VLMs, including detailed multi-turn question answering
      • video subtitles, already present in youtube and VLC apparently

      i really should have thought more about that point.

  • i'm personally not too fond of llms, because they are being pushed everywhere, even when they don't make sense and they need to be absolutely massive to be of any use, meaning you need a data center.

    i'm also hesitant to use the term "ai" at all since it says nothing and encompasses way too much.

    i like using image generators for my own amusement and to "fix" the stuff i make in image editors. i never run any online models for this, i bought extra hardware specifically to experiment. and i live in a city powered basically entirely by hydro power so i'm pretty sure i'm personally carbon neutral. otherwise i wouldn't do it.

    the main things that bother me is partially the scale of operations, partially the philosophy of the people driving this. i've said it before but open ai seem to want to become e/acc tech priests. they release nothing about their models, they hide them away and insinuate that we normal hoomans are unworthy of the information and that we wouldn't understand it anyway. which is why deepseek caused such a market shake, it cracked the pedestal underneath open ai.

    as for the training process, i'm torn. on the one hand it's shitty to scrape people's work without consent, and i hope open ai gets their shit smacked out of them by copyright law. on the other hand i did the math on the final models, specifically on stable diffusion 1.0: it used the LAION 5B scientific dataset of tagged images, which has five billion ish data points as the name suggests. stable diffusion 1.0 is something like 4GB. this means there's on average less than eight bits in the model per image and description combination. given that the images it trained on were 512x512 on average, that gives a shocking 0.00003 bits per pixel. and stable diffusion 1.5 has more than double the training data but is the same size. at that scale there is nothing of the original image in there.

    the environmental effect is obviously bad, but the copying argument? i'm less certain. that doesn't invalidate the people who are worried it will take jobs, because it will. mostly through managers not understanding how their businesses work and firing talented artists to replace with what is basically noise machines.

  • In my experience, the best uses have been less fact-based and more "enhancement" based. For example, if I write an email and I just feel like I'm not hitting the right tone, I can ask it to "rewrite this email with a more inviting tone" and it will do a pretty good job. I might have to tweak it, but it worked. Same goes for image generation. If I already know what I want to make, I can have it output the different elements I need in the appropriate style and piece them together myself. Or I can take a photograph that I took and use it to make small edits that are typically very time consuming. I don't think it's very good or ethical for having it completely make stuff up that you will use 1:1. It should be a tool to aid you, not a tool to do things for you completely.

    • yesyesyes, can see that completely. i might not be the biggest fan of using parts of generated images, but that still seems fine. using LLMs for fact-based stuff is like - the worst usecase. You only get better output if you provide it with the facts, like in a document or a search result, so it's essentially just rephrasing or summarizing the content, which LLMs are good at.

  • I don't see how AI is inherently bad for the environment. I know they use a lot of energy, but if the energy comes from renewable sources, like solar or hydroelectric, then it shouldn't be a problem, right?

    • i kinda agree. currently many places still use oil for engery generation, so that kinda makes sense.

      but if powered by cool solar panels and cool wind turbine things, that would be way better. then it would only be down to the production of GPUs and the housing.

      • Also cooling! Right now each interaction from each person using chatGPT uses roughly a bottle's worth of water per 100 words generated (according to a research study in 2023). This was with GPT-4 so it may be slightly more or slightly less now, but probably more considering their models have actually gotten more expensive for them to host (more energy used -> more heat produced -> more cooling needed).

        Now consider how that scales with the amount of people using ChatGPT every day. Even if energy is clean everything else about AI isn't.

    • The problem is that we only have a finite amount of energy. If all of our clean energy output is going toward AI then yeah it's clean but it means we have to use other less clean sources of energy for things that are objectively more important than AI - powering homes, food production, hospitals, etc.

      Even "clean" energy still has downsides to the environment also like noise pollution (impacts local wildlife), taking up large amounts of space (deforestation), using up large amounts of water for cooling, or having emissions that aren't greenhouse gases, etc. Ultimately we're still using unfathomably large amounts of energy to train and use a corporate chatbot trained on all our personal data, and that energy use still has consequences even if it's "clean"

75 comments