this post was submitted on 15 Jun 2026
52 points (91.9% liked)
chat
8624 readers
218 users here now
Chat is a text only community for casual conversation, please keep shitposting to the absolute minimum. This is intended to be a separate space from c/chapotraphouse or the daily megathread. Chat does this by being a long-form community where topics will remain from day to day unlike the megathread, and it is distinct from c/chapotraphouse in that we ask you to engage in this community in a genuine way. Please keep shitposting, bits, and irony to a minimum.
As with all communities posts need to abide by the code of conduct, additionally moderators will remove any posts or comments deemed to be inappropriate.
Thank you and happy chatting!
founded 5 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
The maximalist "anything using any kind of neural network math is exactly the same and is equally bad in all cases" position is nonsensical and feels like people just taking out the frustration they feel towards being unable to stop the actually destructive and bad things (the mass rollout of useless and catastrophically wasteful datacenters, companies trying to replace workers with glorified magic 8-balls that make everything worse, vibe coding speedrunning the complete collapse of the systems and infrastructure modern society depends on) by conflating them with and raging against legitimate use cases (bespoke data processing models trained for specific scientific tasks, or making certain extant techs like text-to-speech or machine translation a little less shit than they were before even if they're still not great) or harmless hobbyist stuff that's at worst kind of cringe and/or annoying (people playing with open source local models, who sometimes spam the worthless output in much the same people who just discovered poser or daz spammed the worst CGI renders you've ever seen everywhere starting in the mid 00s).
Big agree and basically what I think every time I read a hexbear comment that makes me think the last time they interacted with an llm was 3 years ago and that they seem to think the quality of its output has remained static since then. people literally calling it "a markov chain" as if that's even accurate and as if having "a markov chain" bot that is able to pull, contextualize, and then present information in a way that is convincingly human (in that yes, it would pass the turing test) is somehow meaningless.
like there's people here who clearly seem to believe it's literally equal to an AOL chatbot from the 90s and they are just so incredibly incorrect
there's also people who seem to forget china greatly mitigated the environmental costs with deepseek and you don't have to burn up a quarter acre of amazon for every token and that, gosh, perhaps a good number of issues with llms and their use stem from something called capitalism in the same way as numerous other technologies rather than being inherent to those things themselves
Burning much less energy for something useless and bad is still bad. Not doing it would still save lots of energy. I also do not care if it's "improved" since 2023, whatever that means, it's a fundamentally useless technology as it is applied to chatbots and whatnot. Every uncritical interaction with LLMs makes the user dumber and less able to think for themselves.
If this was contained to scientific work like helping doctors identify cancer cells, I would be entirely cool with it. That is the falsest equivalency ever though, the stated goals are entirely not that.
lol yeah let's have the useless lying machine take care of highly sensitive and detailed medical work
Pick a lane, ideologue
Blocking you because we don't have a disable inbox replies and in dont want to argue with nerds all day
There is a difference in application you aren’t acknowledging… training a neural network type computer algorithm to detect cancer on a computed tomography scan is different than a black box that tells you to drink antifreeze - the former isn’t even an LLM… honestly the terminology is so poisoned by the AI bubble marketing, I didn’t even know if you were talking about generative AI or general artificial intelligence at first… not really even sure generative is anything but marketing a specific set of algorithms in a way that implies a level of humanity so that companies can try to sell it as replacing human workers
Do i need to acknowledge anything at all whatsoever when I'm arguing with people who think it still does this?
It's valid to suggest someone pick a lane when they're saying it's so useless it can't give you a recipe but still somehow potentially useful enough to somehow be used for medical purposes. Guess what, I'm a chef and i literally have not once in the last year at least googled a food item or recipe and it not given something correct (even if it might be a list of steps I wouldn't do) and it also frequently exposes me to new culinary things like haitian pikliz when I was looking up haitian black bean shit. That is legitimately useful
Nobody needs to tell me the cost of data centers isn't worth the suggestion of "what if you made haitian pickles" but the point is that if the argument is "it's mechanically useless in its current state" you're just wrong
Gemini told a guy to kill himself last August, less than a year ago. They may do it less and it may be less explicit than "here's how you can make your microwave explode", but they have absolutely not stopped being bullshit machines when pushed in certain scenarios. I'm not a anti-AI maximalist either, though do have a general disdain, but what you are saying is not true.
neural networks have a use case that they genuinely excel at and that's picking up on subtle patterns in immense datasets, and reinforcing those patterns until they become visible to humans. that's basically what neural networks fundamentally are is self-reinforcing pattern-finding programs. biology and chemistry in particular have the kind of near-infinite variables and chaotic systems that really benefit from having a machine that can draw out these patterns for humans to follow up on
advocates for these uses aren't saying the computer should be allowed to directly diagnose people or anything like that, that kind of stupid hubris is a recent thing from the genai people
They specifically said "as applied to chatbots" so it seems to me like theyre talking about the broader field of machine learning as a whole, which is great as smaller models used in specialized applications and isn't limited to LLMs, i.e. their mention of cancer research, vs literally anything that has ever been posted on the internet and can be spoken of using human language being used to amalgamate bullshit for any user that seeks it.
A subjective and ideologically driven opinion where you're just straight up incorrect but ok!
We don't think it has remained static, we realize the "AI companies" have been implementing bandaid after bandaid to try and compensate for fundamental lacks in the technology itself, and that as a consequence results are better now (for like 1000x the cost but who's counting - well we all will, soon, but you get the point).
But this is still a credible text generator, ultimately close to a larger, N-dimensional markov chain. It still "hallucinates". It's still shit. It's still almost entirely useless for coding (you know, if you care about long or even medium-term maintainability or security), never mind the other use cases they're trying to sell it for. It does not work.
And if it did (it doesn't !), the aforementioned bandaids would have made the cost impossible to justify anyway.
You could argue in the same way that car safety technologies are "bandaid after bandaid to try and compensate for the fact that cars are fundamentally huge hunks of metal going really fast".
That's just not true at all, you still need to do quite a bit of handholding and double checking it gets stuff right but it definitely can get medium-high complexity stuff right and overall it is a productivity boost.
The parallel doesn't really work (there is nothing fundamentally incompatible between the purpose of a car and the basic design idea); but to try and answer, this would be a viable answer if the huge "AI labs" admitted the tech doesn't work and were collectively working on entirely new architectures (likely ones that include actual symbolic representation of the world, not just statistical links between words).
But they're not, they're using the same fundamental approach and doing ridiculous shit around that approach - ultimately a genuine, and interesting, NLP breakthrough - to try and sell it as what it's not and will never be.
See, that's one part I disagree about. First as a direct productivity boost; I realize you probably think it's making you more productive; but that's something you have to demonstrate, at scale. And the few genuine studies I've seen suggest the effect, if it exists, is marginal at best; at least two I remember suggest it decreases productivity (and at least one, possibly both, can't remember, said self-reporting from devs was that it increased it despite the effect).
More importantly, though: very little of the code you produce this way will be maintainable. Very little will also be secure, too (both "direct" vulns - they're not actually good at finding them, FYI - and more fundamental logic vulnerabilities - they'll never find these at all outside of trivial cases). If you want a concrete example of that, just look at the leaked Claude harness code: it's genuinely pathetic. They're vibe coding it all and it shows.
Also, I notice you say "you still need to..."; do you think this will eventually get fixed ? it's been nine years. There's a point where "it'll get better" is starting to become more of a meme than an answer. It was getting better, then a lot less, then lately almost not at all, despite greater and greater cost increases.
Shit, lately I've even started to wonder if they're even training new models anymore; I've been wondering if they're just not shipping the same with a different harness / different bullshit surrounding it. It's not like they can improve on the models themselves anyway (not without at least one, likely several, genuine fundamental breakthrough) - they're all out of human-produced data by now.
Yes it is, just yesterday I was having trouble with a script, so I googled it, the Google ai told me to use a command that doesn't exist. It just made it all up. Completely worthless.
China isn't counting the same numbers i tell you that
It's incredibly useful for letting humans interact with a computer system in natural language and maybe you shouldn't take from that "hey we should have it do highly sensitive stuff where the slightest error could have great consequences"
Here's a use for it. You nerds hate Reddit, right? Well hey, it does a great job of summarizing reddit threads, since that's where half of its fucking training comes from. No more ever having to go back to reddit to browse dozens of inane responses hoping to find the one comment with relevant information. Congratulations, you have a machine that made reddit obsolete
That's true; the very quality of results you can obtain with some recent models locally suggests the possible optimizations are huge. But it's also diminishing returns: the moat between state of the art models and one I run on a P40 locally is really small these days; the amounts they need to sink to get even slightly better results at this stage are more and more. They're text generators; yes, if you run ten of them in parallel, vote; have them cross-check their results, introduce harnesses at every steps, etc. (all examples of actual bandaids I evoked above) you can improve results. But all of this is trying to make a tech that fundamentally doesn't answer the problem answer it nevertheless. And all of it costs a lot.
And I dearly hope China is not spending too much effort on backing these domestic initiatives, because again: outside of a few limited use cases (easily identifiable, I've listed a few in another comment in this very thread: those where having no relation to the truth in a portion of the generated text is acceptable), the tech doesn't work.
There are use cases where having the computer completely invent actions or do shit randomly isn't that bad, I guess; games come to mind. But ultimately and more generally I disagree, it's shit for that too. Try one as a daily runner, just for laughs. Or just try an entire shell session where instead of typing the commands, you complete a description of what you want to do through one and then do it. It goes bad very fast, let me tell you.
And I get it: what you describe would be awesome - a SF dream. I like tech and I wish all of this would work; and like many initially I genuinely wondered as well if scaling/attention was all you needed, and had some measure of hope; quickly dashed, though.
Again, no it does not. You have no guarantee it won't pull shit out of its virtual ass; and just as crucially (and even more likely), no guarantee it won't ignore significant parts of the source material. You'll get a result that seem like it summarizes the thread, with no confidence level, and no guarantee of its reliability.
I've seen what it does and how often and the various ways it tends to fuck up, at least Google's search result llm because I'm not going out of my way to use this shit outside of trying to look up shit i was going to look up anyway (mentioning this because as a result i don't know if the model they're using is better but more expensive), and i really don't consider it a different experience than the last 6-10 years of having to look up a reddit thread to parse comments for information. I still have to figure out who knows what and how much they know and figure out how much I trust the information.
With the llm it's the same shit to me, I have to figure out if it's pulling from a source that knows anything, if the information is actually even present in the source it thinks it's from, if it's miscontextualizing the information, and if it's just trying to glaze me or frame its responses in relation to my prompt
The end result is i google shit and tend to find the information i'm looking for faster and without having to be exposed to 100s of tips m'fedora redditisms directly. In this way you could consider the llms use to be like those boxes you use to view an eclipse without hurting your eyes
Fair enough, if it works for you and you're aware of the downsides (which you seem to be, contrary to most users); two things I haven't even mentioned though are the power/environmental costs (obviously) but also the potential cognitive impact.
You say it doesn't change much compared to when you were parsing the reddit comments yourself; doesn't it ? how sure are you "wasting" time parsing these comments wasn't exercising an important muscle mentally for you (getting the jist of a text rapidly - excluding braindead content quickly, etc.) ?
Here the example is pretty ridiculous - I doubt your mental faculties depend much on parsing reddit comments - but you get the idea. It's very early to tell but several papers now suggest the negative cognitive impact is very real and potentially very fast.
Idk, pretty sure, since i feel the "is the lying machine lying to me or is it accurate" appraisal is about the same exercise and there's also the fact that an absolutely disgustingly large portion of reddit comments are themselves being made by llms now (as every single anti bot measure reddit ever uses only impacts real people, apparently)
Edit: Thank you for the concern though btw but if you ever do notice me exhibiting cognitive decline it is almost definitely gonna be a combination of the alcoholism, drug use, contraindicated drug use, and general medical decline
Does that position exist outside of twitter, though ?
When GenAI-critical people say "AI", they almost universally mean LLMs (and sometimes also diffusion models / multimodal ones). And in fact they usually mean "LLMs as sold by the VCs". LLMs themselves do have a few legitimate use cases (for perhaps 0.1% of the target population they're currently being sold to, but still).
They do not mean machine learning in general - which is awesome, extremely useful, and has been in widespread use for decades now. Including traditional neural-network-based classifiers.
It's the standard reddit position and you can even see it in this thread. Like some people genuinely do not understand that there are categorical differences between the shitty chatbots making everything worse at scale just to make things worse and some dork that's just being annoying by posting the weird dream his computer had as if anyone gives a shit.
In my day to day job as an engineer I use LLMs to help me do my work. They are tools just like anything else and enable me to work faster. I've even piloted an agent flow that takes our existing site (a huge behemoth of legacy and new services) and automatically performs accessibility testing across a variety of tools then gives us reports. We have actioned these to actively improve the experience of people using our product with screenreaders for instance (and have confirmed this from these very same people giving us feedback on the changes). The view that they are useless runs counter to reality and in my experience is usually from people who have deep emotional hate of technology. LLMs have incredible potential but as with anything under Capitalism they are used for horrors and their rollout is done with complete distain for human life and the environment. That is not special to AI.