QotNews
Hacker News, Reddit, Lobsters, and Tildes articles rendered in reader mode.

A lukewarm defense of LLMs

Source: herbofmints.neocities.org
20 points by Tygrak a day ago on tildes | 11 comments

precipiceSmall

Intro

So to start of I will try to describe who I am a bit so you can at least sort of understand where I am coming from and so you can be angry at me properly.

I am a programmer, more specifically a graphics programmer. As a kid I was really into videogames (pretty unique right) and I really wanted to make my own games, first messing out with stuff like the Warcraft 3 map editor, later gamemaker. As I got older I got introduced to Python by a friend. I genuinely started enjoying programming, and found that programming truly can be beautiful. I made programs to generate mazes, to render various fractals, I made my own stupid esolang, I started doing game jams (I made over 20 game jam games over the last 10 years), and did so many other different programming projects. From there it spiralled into me actually studying informatics (which is what some European countries call computer science) and later even sadly actually getting a job in game dev. During all my time wanting to learn how to make games I also learned how to do a lot of other creative skills, at least badly -- how to draw, 3d model, play guitar and make music. I am like half of the letters in the whole LGBTQIA+ soup. I consider myself to be a leftist on most issues, in practice when voting I voted for the most leftist party that was available to me every time.

Also I have to mention this blog post by Pixelcluster, discussion of which finally pushed me to finish writing down my thoughts here. I think it's much better written than whatever I will word vomit out here, it's level-headed, and just analyses all this LLM stuff quite accurately. If you don't want to waste your time reading what I wrote here you should read her post instead.

So with that out of the way. My actual point of writing this is that I think LLMs in general are not the devil (or don't necessarily need to be?).

devil

The defense

So here are all the problems I see with LLMs:

  1. Capitalism bad/big tech companies evil
  2. They will probably cause deskilling of literally everyone and prevent people from getting skills in the first place
  3. They are built in ways that encourage wasteful usage and are also addictive
  4. They cause psychosis/human brains aren't ready to have machines talk in natural language to us
  5. They are plagiarism machines
  6. They are built on amoral grounds
  7. Techbros/libertarians like them
  8. They are stealing peoples jobs
  9. They give people too much confidence in areas they don't understand

1. Capitalism bad/big tech companies evil

In my opinion almost a lot of the biggest problems with LLMs are caused by the fact that they are being built and pushed by companies that are close to being unironically cartoonishly evil. The two biggest startups focused on "AI"[1], Anthropic and OpenAI, are really really difficult to not hate. These companies literally stole the whole of internet and more. I myself am not really much of a fan of how copyright works, but there's a complete dissonance in how these companies operate and how the average person can work -- if a normal person tried to steal this much shit they'd get jailed immediately. One of the loopholes Anthropic chose to make stealing the collective knowledge of all of humanity be ok in the eyes of law is to literally buy up millions of books, and then shred them. That's really just... it sounds like a joke, it's completely unbelievable, but sadly it's just completely real. One of these companies also has the word "open" in the name, which is an interesting choice considering almost nothing they ever did has been ever released openly. If you are stealing everything from humanity at least give something back. These companies also have the gall to repeatedly say how other companies distilling from their models is somehow bad.

These companies are also incredibly wasteful and just setting money on fire. The actual models made by these companies are just absurdly gigantic and way too expensive to train and run, which is why these companies have to be subsidizing their usage quite heavily. They need people to use them to build hype, despite it costing them money to run them. Their plans are so megalomaniacal that they single handedly caused the price of many PC components to skyrocket. While I think the environmental problems of LLMs are a bit overblown from reality, its clear that these companies are building way too many data centers, way too fast. Building new data centers inherently isn't bad. But the current build out is absolutely bonkers, there's data centers being built to meet some magical future hypothetical demand that just doesn't currently exist. All this is also propped up by fake monopoly money being basically created out of thin air, with companies borrowing each other absurd sums of money. This is just not sustainable, both in an economic and environmental sense. We are in a bubble. And while it's perfectly fine that these unsustainable companies are gonna explode at some point, the bad part is that when they do it's most likely going to trigger a real recession that will cause a lot of problems for every person who doesn't even have anything to do with "AI". We are all going to eat shit for the sins these garbage companies make. See for example https://paulkrugman.substack.com/p/ai-versus-everything-else from Paul Krugman to read about this more from someone smarter and more qualified than me.

Oh and also it's not like everything is the fault of just those two companies. There's also all the old boring tech companies participating in all this too, which all did their own share of unbelievably evil shit (Facebook[2], Google[3], Elon Musk and so on).

It's not inherently necessarily to use LLMs made by those companies. The model that I am currently using the most is Deepseek. Now I am not saying the company that is making it is perfect. The biggest problem for a lot of people/companies is that it's Chinese, which China is I guess like basically sort of an enemy of the US and the West in general. But also the US is kind of the enemy of the West in general too. So like as a random person I don't really care if I am sending all my data to the USA or to China. In the best case scenario no one would need to send their data anywhere, and some people can run LLMs locally, but most people don't have hardware good enough to run useful models at reasonable speeds. I don't give a shit that Deepseek is "distilling" OpenAI and Anthropic models. There's definitely going to be some other good reasons to dislike this company. China is doing some messed up stuff and Deepseek might be censored in some areas due to that, but I haven't really ran into that much myself, but I also don't really ask it too often about how the Chinese government treats minorities.

To get to my actual point, Deepseek is burning vastly less money than the other companies, and the newest flash model they are providing is dirt cheap, while performing close to the same level of the other models, and that's with it not being subsidized in some crazy way. It is open weights so other providers can and do offer it, and it's dirt cheap everywhere. AFAIK Deepseek is still expanding extremely fast so it is quite possible that the company itself will fall into the same "infinite growth" traps as the other companies later and will start doing evil shit too (or they might even be doing them already and I just don't know it). But I think even just the current 4.1 flash model existing proves that not all LLMs have to be a gargantuan money burning wasteful dumpster fire, while still being very useful.

2. They will probably cause deskilling of literally everyone and prevent people from getting skills in the first place

This is kinda the Pandora's box innit? The problem with this is that there are many tasks where you genuinely don't care about the journey but just about the destination and where it's fine that you won't learn anything. Some people like to liken LLMs to other things, for example calculators. Which I mean sure sort of, but it also doesn't make sense. Calculators can do math for you, but we still teach kids how to do maths, because actually understanding what is happening is just so useful in so many different areas when learning everything else, and even in everyday life. Besides that, a calculator being able to do maths won't help you if you don't know which buttons to even press on it in which situation. I sort of hope that people will calm down a bit and figure out that actually understanding how things work is still extremely useful and will still try to learn. It's pretty clear that at this point people who understand some problem area are more efficient at solving problems in that area than people who don't, even when LLMs are involved, and I don't see that changing any time soon.

I do think this is still a gigantic problem, just the temptation of the ability to skip most kinds of learning instantly is sure to prevent people from learning things even if learning them is still useful. And there's also the issue of students in school being able to cheat homework on unprecedented scales. Most students don't really want to do homework anyway so they will be tempted even more to cheat. I know that I cheated sometimes for subjects I didn't like just by copying my homework from friends. I am sure it will be very difficult for teachers to come up with ways to make students do homework properly without cheating. In a magical perfect world there'd be enough teachers to be able to pay enough attention to every student to make sure they are doing the work, but we are nowhere near that world.

I also want to talk about the benefits a bit more here though. There are plenty of cases where just having some output can be very useful, even if you don't know how it was made. If you ever use proprietary software you most likely won't ever be able to understand the inner workings of that software. Buying Photoshop doesn't teach you about how Photoshop works. So what is the difference between using real Photoshop or a vibeslopcoded Photoshop clone? Now that I mentioned Photoshop clones, people will immediately think of that vibecoded Photoshop clone which is apparently soon going to be at feature parity with real Photoshop. That shit is probably the clearest example of pure slop vibecoding at a gigantic scale currently in the wild -- new pull requests and issues are appearing in that Github repo every few minutes, probably without literally any human seeing any part of what is even being done, and the result is a steaming pile of crap. I didn't try running it myself but from everything I've seen it's missing even super basic features, and the features that it does have are broken in unbelievable ways. So no, I don't encourage using that. It does show something very interesting though. There's a reason why it blew up on the internet despite it being a pile of crap. It is because people really do want a free open source alternative to Photoshop. Even people who normally hate all kinds of "AI" came out and said that something like this could be good. An artist coworker of mine literally said that they can't decide if they hate "AI" or Adobe more. It shows that if used for something that people consider to be a good thing, people can tolerate "AI". Hopefully with time there will be more and more examples of it being used for "good".

Also what I did want to say is that Photoshop has thousands of features, but the average user uses just a small amount of them. It is not unreasonable to prompt up a simple app that can do just that small set of features, that you can yourself use instead of real Photoshop and save yourself the absurd monthly subscription that Adobe asks for. And I am writing this because I did exactly this. After just a couple of weeks of very leisurely sloppy prompting I have an app that does basically everything I used in Photoshop. I don't care if you think LLMs are evil, I think Adobe is evil too, so I don't see a negative with giving them less money. Worst case it's just net neutral. And in this process I didn't learn too much new about how exactly Photoshop works, but I do get to improve my drawing because I get to use a program that I don't hate.

3. They are built in ways that encourage wasteful usage and are also addictive

I've also seen people say that they think LLM usage is basically addictive. Everyone probably saw someone not using "AI" at all, and then suddenly using LLMs for everything a few weeks later after trying them. I think there's two halves of this. One is that LLMs are just genuinely useful for some things, so it's no surprise that people will use them for those things after discovering that. The second is that the big companies are designing their "AI" applications with some pretty obvious dark patterns that sometimes even gacha games wouldn't dare to try. What exactly is the purpose of a 5-hour usage limit? I guess if we are charitable we can say that it's there to prevent using up the whole monthly limit in one bad prompt. In reality it's pretty clearly there to build FOMO, to make people feel like they need to keep coming back over and over, even multiple times each day to make proper use of the money they spent.

Connected to this is how LLMs are setup in ways that encourage being very wasteful. Something I see very often both online and in real life is people talk about how the limits on their sub is resetting soon and therefore they feel like they need to go and use up all the remaining usage. I then see these people start spamming the LLM too do some absolutely pointless things just to use up all the usage just before it will get reset. This always results in the worst of LLM slop -- stupid oneshots created basically just to tell some social media that wow this LLM made this garbage in one prompt!!!

Now I don’t really know what is actually the best way to solve this. Again the obvious part is to get rid of the stupid 5 hour and weekly limits. It's an improvement, but if it is a monthly sub it probably does need to have the usage reset somewhere. And I am guessing the same thing would just happen at the end of the month in an even bigger scale. Best option would be if the subs just didn’t exist and people just paid the actual API prices. But the API prices would need to not be so absurdly high for people to use them, and we already covered how the subs are heavily subsidized just to get people to use them. Again with something like Deepseek, even using the API prices it’s still dirt cheap, so you can easily do that. Maybe the 200 dollar subs shouldn't exist at the very least. 200 dollars per month has to be a pretty substantial amount of money for the majority of people, so when some people are spending that much they really do feel a lot of pressure to actually get something out of it.

4. They cause psychosis/human brains aren't ready to have machines talk in natural language to us

There's been a lot of cases in the news of people getting pushed to go crazy, even pushed to suicide from talking to LLMs. I feel like it stopped showing up in the news a bit, but maybe just because it stopped being interesting, it just became too normal. This was a pretty common topic especially around when Open"AI"s GPT-4o[4] was new. But it's obviously not a problem that's exclusive to it. Despite the company putting out a whole article about how they fixed their models and they aren't sycophants any more, it's completely obvious that their newer models still work basically the same. I tried asking free ChatGPT about the timecube and while it said that it probably isn't a real thing it also said that it's very fascinating and asked me if I myself think it's real... Also, there are other companies that try to optimize for this much worse than OpenAI. No idea which one is currently popular, but I know there are things like Character."AI" that specifically try to make people have LLM friends/lovers, which probably isn't the best for our brains. Especially for socially awkward people, it's genuinely scary to imagine that someone might decide that these things are good enough as friends and that they don't even need to try find actual real friends. I can sadly imagine myself isolating myself due to talking to these things if I was a kid growing up now in these times. Humans are social animals, and getting the only socialization from crazy sycophantic bullshit machines is bound to result in a lot of bad things.

Again, similarly to the addictiveness problem, the solution here is pretty obvious. There's nothing inherent that says that LLMs have to respond to us in natural language at all. The best would be if you tell a LLM to do something it would just do it, and just respond with some success code like other computer programs do. There's no reason they need to output any text at all. Or that the text they output is so "human". If it's a task that requires outputting text the text can at least be as machine like as possible. No one ever thought that Google translate is sentient. The coding models already sort of try do this, but they still do mimic the language people use when talking to them, which would ideally not happen at all to help prevent this.

Obviously in practice I don't think is getting fixed any time soon, considering there are many companies specifically in the business of making people have fake LLM friends. There are even companies literally trying to replace therapists with LLMs.

5. They are plagiarism machines

There's this idea that LLMs can't actually have any novel thoughts. Despite it not being completely true in the strictest sense, it is still pretty accurate. For example there's a specific UI style that all the zero effort vibecoded LLM apps have. They always seem to pick the same technologies if people don't specify what should be used -- every single garbage one-shot vibecoded slop game I see in social media posts is using three.js, because that is what LLMs pick by default if someone tells them to make them a video game.

I didn't originally think that this is where I would talk about this but there's also the whole garbage lingo LLMs have where they absolutely love certain words and symbols. The most memed on one is probably like calling things load-bearing, but there are ones that annoyed me much more. For example that something is landing, some changes, a build or something. I don't think anyone ever says that. And the worst one for me is the word "honest". It's tragic that even such a common, normal word is now tainted for me. There's a billion other examples of "AI" writing smells, em-dashes being maybe the original most notorious one.

Honestly (crying cat emoji) I don't really mind that LLMs just regurgitate already existing things, LLMs are still useful. Stack overflow used to be useful, and that literally by definition couldn't contain novel thoughts since you could always find good answers there only for the most widely used programming languages and technologies. Google/search engines in general are the same, they just find you already existing things, and they are also still even now extremely useful.

I do also have to mention that them just regurgitating the same things that make sense in the current context means that they can have some really bad biases in other ways than just using annoying words. If the training data has a lot of text that says that glue is a great addition to pizza, the LLM will say that. If the training data for a LLM built for agentic coding would most commonly have someone using bash commands that end in deleting every file on your hardwire in some niche specific context, the LLMs will obviously then be able to do that if it gets in that specific context. If the LLM gets into a context that somehow reminds it of online discussions where people bash minorities, the LLM will probably start outputting text like that too. Not only are LLMs plagiarism machines, even more importantly they are a statistical assumption machines. If the input the LLM gets has something bad as a likely output, the LLM will just probably output it, it isn't grounded in reality in any way. RLHF[5] is supposed to try negate the worst parts of that, but it is kind of just in the LLMs nature. If it is trained from the whole of basically all human writing and you want it to produce text, it will reflect how human writing looks.

6. They are built on amoral grounds

I don't have any defense of this. It's obvious that they really are built on amoral grounds and there's no way around it. The only arguments I can come up with are whataboutisms about how there's a lot of other amoral messed up stuff going on in the world. Stuff that in my opinion is much worse than whatever is happening in fake computer land. God I need to try really hard to not write all the whataboutisms here.

7. Techbros/libertarians like them

Ok, I didn't actually see people point this out specifically too much. But I think that it's just a huge red flag. I also don't really have anything more say for this. I guess I'll say that it isn't mandatory to even have LLMs get served by for profit companies. People could setup nonprofits. Or it could even be something that a country's government handles as public infrastructure[6].

8. They are stealing peoples jobs

We live in the real world and real people are losing real jobs over some people in management deciding they no longer need real humans to work for them. I think it’s obvious that in practice almost all the time this is just a bad decision if you want to have quality output. "AI" just does a shit job by itself in almost all areas. Unfortunately the bosses everywhere don't really always care about making quality products. I think this future where some CEO just commands an army of clankers to run their company that some people are imagining is nowhere close to reality. So people being fired due to "AI" is mostly just a shitty cover for some different reason (the real reason usually being short term gain). Or it's just plain stupidity. But it’s still obviously happening, and I can't deny that it really sucks.

With that out of the way. This argument annoys me a bit. To defend this, I saw people genuinely argue about how all different technologies that just clearly save time and are more efficient are evil too. An example being self checkout machines in stores. I never thought self checkout machines are controversial yet I heard this sentiment that they somehow are evil multiple times in different places in the past few weeks. Ok... so the issue is that someone doesn’t get to mindlessly sit all day scanning groceries and mess with a cash register. Is that a loss? I don’t think this is a job that almost anyone wants to do anyways. If people had a choice to not do it, they would surely be able to figure out something much more interesting with their time. Are there people growing up hoping they will be able to become cashiers? I doubt it. So why would this be bad. I think it’s evil to force people to do these mindless bullshit jobs to survive. Of course the real problem is again just that capitalism bad. Despite the fact that I am quite cynical about humanity, I believe that humans are in general good and want the best for each other. So why do we just roll over and say that things can’t get better and that we should still have escalator attendants[7], call center workers, cashiers, people pumping gas at gas stations, people washing dishes...

You might say that the difference is that "AI" is taking just all the cool non bs jobs. And I would say that you are right and that it is sort of unfortunate. But also as I said in reality if things made sense people wouldn't be getting fired because their job can now be done more efficiently. Programming in my opinion really got at least a bit more efficient, even if you are just using LLMs as improved autocomplete and improved context sensitive google search. It doesn't make a programmer useless. You still need a programmer to create complex software. You still need artists if you don't want everything to look like shit. And if it does make some jobs more efficient, maybe in the future that could just mean you spend less hours at your job, less days per week. Humans currently spend way too much time at work. I wish people could spend more time doing things they actually want to do. I wish everyone had more time to explore new hobbies, to spend more time in nature, to do more sports, to spend more time with their friends and loved ones. I believe humanity can do better. Of course this isn't really fully about "AI". If humanity didn't expend so much effort on killing each other and making each other watch as many ads as possible we could probably already be there.

9. They give people too much confidence in areas they don't understand

I guess this is already sort of kind of a good thing, at least partly. It's not bad that people can for example do things that would have normally required a programmer. I feel like many of these points are very connected so I feel like I might be repeating myself a bit already. "AI" enables people to create things without any effort or understanding. That is fine if you keep the outputs to yourself. But if you don't understand something, you won't even be able to tell if the results are even anywhere close to correct. LLMs are really good at making something look correct in the context, but looking correct isn't necessarily the same as actually being correct. In the past making something in an area you don't understand required huge amounts of effort, so you inevitably learned just by making that thing, and spending that effort made you care about the thing actually being correct. Now you can get things that look impressive and look correct without any effort. Effort is the magical important component that allows humans to actually accomplish anything and to make quality stuff. When using LLMs it's easy to skip expending any effort.

This is just how LLMs inherently work, and since the outputs really can look correct, and the LLMs can even write smart sounding soup assuring users that it definitely is correct, it can convince people pretty easily. After that, even well meaning people might feel like sharing that output with others. It is frustrating to get some big wall of LLM output text given to you by a coworker. Just a few days ago I had a coworker send me such output, and even though they said that it is "AI" output and that I should take it with a grain of salt, I still felt like I needed to go through it since if they are sharing it with me, it probably has to be something, right? Sadly it contained a lot of things that were just plain inaccurate and having to go through it was basically just a waste of my time, and I even had to dig up proof that it contained hallucinated inaccurate things and even had to try find out why it might be hallucinating those things.

I don't think there is any solution to this. Currently people just trust the LLM output too much, maybe LLMs could be trained to try have them be less confident about how well they do everything, but I don't think it's actually a solvable problem. Ideally some change would happen socially. People should just not burden others with their LLM outputs if they don't understand them. At the very very very least people should make sure that the outputs are actually correct if they do want to give them to others by somehow testing them to the best of their abilities, so it is at least worth it to go through the walls of LLM text. People should at least rewrite the LLM output in their own words. Please.

Ending thoughts

There's a million good reasons to hate "AI". I can't fault anyone for deciding to just fully not want to use them for anything ever. It really is just the best choice if you want to live as morally as possible. Or even if you just really care about all the parts of your craft, whatever it is. Everything isn't just about the final product. Especially if something is your hobby. You will also just be able to learn much more if you don't use them, and you don't risk having the skills you already have slowly atrophy away by being too comfortable with using them for everything.

At the same time, I do think LLMs are just undeniably useful. If anything, for programming the issue is that they are a little too good, they give pretty good results in zero time, which gives people confidence to just trust them too much. They are too good at implementing simple ideas, they are too good at reverse engineering. And outside of programming, I don't really know, I just think it's not so simple to universally say it's all bad or all good. That's kinda what I am trying to say the whole time here. Look, there's just too many ways LLMs can be used. I think literally everyone is already ok with using LLMs in some forms, since they really in the grand scheme of things are nothing new. I understand that usually the people who are the most against "AI" specifically mean the kinds of "AI" that are bad, probably like specifically all the generative "AI", and specifically the chat bots and coding agents doing big tasks on their own, and they don't really mean anything else. And like yeah those are the ones causing problems the most, so I understand. But in my opinion it's not so easy to draw a line between what LLM usage is ok and what isn't.

I'll mention a comparison that Pixelcluster made in her post about LLMs which I recommended right at the very start. I apologize in advance that I am using this comparison while somewhat defending LLM use. She likens deciding to not use LLMs in any form to digital veganism. I think it's a pretty accurate comparison. The meat industry is a complete nightmare morally and ecologically, virtually everyone can agree on that. I think virtually everyone can also agree on most of the problems "AI" is causing. People might say they just like how meat tastes, but some people will just decide they can't just knowingly ignore all the environmental and ethical problems and they decide to not eat any meat or animal products. It's clear to me these people are just doing the right thing. People like me might say that LLMs are useful, but other people can decide that they can't just ignore all the societal, environmental and ethical problems and decide to not use them at all. It's clear to me these people are just doing the right thing.

I do see one difference between the people from these camps though. This is just my own observation from what I myself see, obviously not everyone is doing this. I myself have been vegetarian for over 7 years, and I do know a lot of other vegetarians and vegans, so I hope I can at least sort of talk about this. I don't think in general vegans and vegetarians hate everyone who eats meat. In general I don't think we immediately shame anyone who for example posts an image of their lunch and it has meat in it. I don't really think it's helping anything to immediately vilify the majority of programmers[8] or anyone else who uses LLMs in some form.

"AI" usage disclaimer

No part of this text was written or edited in any shape or form by any LLM. The images were drawn in the vibecoded Photoshop clone that I described earlier. They were drawn by me using a drawing tablet. I asked Deepseek to make me a web app which can take markdown with footnotes and images in it and output plain HTML in which I wrote this text. I googled some stuff while writing this too.

precipiceEmpty

Footnotes:

  1. I will always write "AI" with the scare quotes because I don't respect it and I don't think this shit is intelligent, it's literally just a big pile of statistics and matrix multiplication. ↩

  2. Here's a 3 hour video outlining some good reasons to dislike Facebook or "Meta". https://www.youtube.com/watch?v=MPyJBJTHyO0 ↩

  3. There's the whole thing with having "Don't be evil" as Google's motto and then later getting rid of it. That is a pretty quick summary imo... I am too lazy to write another footnote for Elon Musk, but I hope I don't need to make one. ↩

  4. Pretty interesting video I watched a year ago showcasing just how crazy the openai models were at this time: https://www.youtube.com/watch?v=VRjgNgJms3Q ↩

  5. Reinforcement learning from human feedback https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedback ↩

  6. South Korea seems to be trying exactly something like this https://www.techradar.com/ai-platforms-assistants/other-governments-will-be-watching-this-closely-south-korea-starts-testing-free-ai-for-the-public ↩

  7. This is a real job. https://www.eurogunzel.com/2015/12/escalator-attendants-former-soviet-union/ ↩

  8. This survey that asked over 15 thousand professional programmers about how they use LLMs for code says that only 27% of them code fully manually https://blog.jetbrains.com/research/2026/08/how-much-code-do-developers-really-let-agents-write/ ↩

end