Maybe a hot take, but: Even using AI as a learning tool is wrong

looking up stuff is an important intellectual skill too.

It doesn't do you as the user any favours so it's definitely wrong in that sense. It's bad for the environment too. People will still do it because it was pushed on them hard by universities and employers; it'll take more time to unlearn

Not sure why is that a "hot" take. It's obvious.

A learning foundation of a human is crucial. It must be supervised by another human or originate from a human... A human who realizes the effort, empathy, pain, value, experience, love for discoveries, academical or not, and from one who realized the information passed for another human, improving it together in alive, respectful, careful, manner for the authors, artists, people, with names who ask, refer to, quote for, to rely on, to trust and discuss with together realizing it all ineffably magnificent...

ML/LLM/AI "knowledge" is limited, indeterministic, unverified, unstable, and no one is irresponsible for it, and never will be.
This is a reckless, lonely, sorrowful, self-damaging experience... that no one alive must experience.
It is a path of degradation, atrophy, and disrespect towards anyone involved.

The older I get the more I realize most people don’t actually care about how or why anything happens

ice cold take. the point of education is enhanced understanding of world. achieving result of it, be it thesis or working arch install, used to be more of a proof of ability than anything else. that, and friction coming from the process is what allows for actual learning to happen. chatbots substitute for both

I don't get why you say "even using AI as a learning tool is wrong"

Using AI for learning is among the worst ways you can learn things.

AI/LLM don't have the capacity to plan lessons, or to explain things pedagogily, it requiers you to keep ahead of what you want the AI to teach, this to enable you to explain to it what you want it to teach.

Also, unless you know the subject you have no defense against AI halucinations and lies.

The one time I have found AI somewhat useful is when I need specific targeted assistance with something special, say the syntax of a command or something similar.

The least bad use of LLMs is probably "shitty search". Because Google et al have let the internet rot, it's often hard to find things through normal searching. So people use the LLM that's at the top of the page. This is not great.

They tell me to use AI at work. I tried entering in like "tell me about X. Link sources". All the sources were also AI. Horrifying.

From: https://wiki.archlinux.org/title/Help:Reading

Most articles on the ArchWiki do not attempt to provide a holistic introduction to a single topic; they are instead written in adherence to the "Don't Repeat Yourself" principle, under the assumption that the user will seek out and read any supporting material that they do not yet understand. Where possible, such supporting material is indicated in the article via special formatting, see #Formatting.

Because of this organization, it may be necessary to examine several related sources in order to fully understand an ArchWiki article. In particular, users who are new to Arch (or GNU/Linux in general) should expect to end up reading a great number of articles even when solving simple problems. It is especially important to study the supporting material before seeking additional help from other users.

Read the Wiki! It even has a wiki page on how to read the wiki, and yes, it was one of the first things I looked at.

Arch Wiki articles are pretty well structured, and usually the various keywords and jargon mentioned can be found on either Wikipedia or Arch Wiki, and then it will also link to other documentation or resources for the harder to find stuff.

I am growing ever more curmudgeonly to the point where I’m often thinking „[some] gatekeeping is good actually“. Communities should absolutely have a minimum level of engagement and understanding required to be a part of, and I think that’s actually orthogonal to how open the community is to new memebers. As OP points out, the Arch wiki exists and while it’s pretty technically dense (by necessity given Arch is a power user distro even by power user standards), it‘s open to anyone with a web browser.

I think users of LLMs often suffer from a sort of Gell-Mann amnesia phenomenon. If you are well versed on a topic, you can often find the mistakes they make; but, when researching a topic you are unfamiliar with, the responses feel like the "AI" knows what it's talking about.

I'm really on the fence here.

On one hand yes it will make everyone dumber because most won't stay curious but they haven't probably been the most curious anyway and it will make it so they don't understand the how or why.

But against that argument, people said that about all technology. Writing was dumb because you're writing about it not doing it. People said algebra was cheating because you're replacing these huge long form math with equations that made them more approachable.

Unfortunately LLMs are here to stay and it's important we call them that. They're glorified translators and thesauruses. They can synthesize information well and can create diagrams with said information pretty well. They're trash now but so were the first cars and planes.

It's likely most people won't ever learn how computers work, Linux will become easier to understand and LLMs are already in that space to digest that workload to get there.

So I dunno I hear you but overall I don't think it's actually that evil. I think this is what humans do. Work smarter, not harder

I feel AI definitely has a place in more specific fields of learning.

For instance, I've played with a local model and used it as a more intelligent search for an hours-long D&D audio transcript, I could get useful shortcuts to things like "make a list of items the group used today, and include an exact quote when it was used", or "Here's this name, find every word in the log that could be that name spelled or transcribed incorrectly". The second is very handy when there's names like Vlaxor and the transcript includes things like "flax ore" and "blaster".

It really is a decent unpaid intern. It's just unpaid interns need specific instruction and need to be checked, but people treat it like they're experts. Your friend could also get a similar understanding in beginner forums, or, if it's a common issue, even YouTube videos going over it, but posting on a forum isn't as convenient as asking the unpaid intern who cannot think less of your skills as a Linux user.

I agree LLMs are not easy to learn from, though it can happen sometimes. I actually find though that the problem is usually the incredibly verbose but hard to absorb writing style that comes back. My (software dev) workplace has been insisting we use Claude so I've been trying it. Everything it does, it writes 10 paragraphs explaining. I try to read it and my eyes glaze over. Then of course it makes many mistakes and I think that's also part of my issue. I become demotivated in playing the game of "sift through all this bullshit to find how it might be wrong". It's exhausting.

I agree and disagree.

I agree in part because yes, learning requires effort. Subjecting yourself to the instant gratification machine won’t do you any favors when it comes to learning. The only thing you’ll learn is that there’s an instant gratification machine ready to help you anytime.

I disagree in part because I think anything that helps Linux adoption is good in my eyes. Linux doesn’t need any ”filters”. Not everyone has English as their first language for example. Should they be closed off access to the wealth of information in Arch Wiki?

honestly I can probably count the number of times “AI” has helped with learning something on one hand. Maybe with extremely surface level “explain like I’m 5” explanations, but as soon as you expect more detail or point out inaccuracies, it stops being useful and more of like a confidently wrong tutor.

There's been genuinely awesome advances in adjacent topics like vector search, but ultimately when you put an LLM in that layer in between you and whatever you're trying to find out, you're just robbing yourself of the learning experience.

Well, the correct way to tell is to do a study on AI as a learning tool. And turns out we got some. The answer to this question/debate was in the news a few days/weeks ago.

And some things just are complicated. Especially computer stuff like what's in the Arch wiki. There's 3 million ways to skin computer cats, every rule comes with several exceptions. People don't really know what they want. People do widerly different things with their computers... Some have brand new laptops with very different quirks than other people's 15 year old hardware. I mean the authors aren't cruel if they write 10 pages of very detailed info. It's just that it's complex. And sure you can dumb it down... To the point every statement is wrong on several levels.... It's just not a common conversation style in the Linux community.

It is however common in mainstream culture to desire some "simple truths". They're mostly wrong, no matter where they come from.

From first person experience, both with software developers and business users: it's fucking up their brains.

It's making them less competent software designers and basic programmers, it's launching a thousand careers where their only skills are copying and pasting LLM output, and it's making them hyperconfident while actually rotting their fucking brains.

It's a learning tool like an answer key is a learning tool. I'm pretty sure my college textbooks included correct answers for odd numbered problems but never complete solutions.

What do you think about learning languages with LLMs? I use them to help when I want to make a translation sound more natural or understand why a native speaker phrased something a certain way.

You and I both want people to learn. You and I both see AI as an obstacle to learning. We just see things slightly differently.

LLMs are bad for education because they oversimplify.

I actually think this is not the main reason AI is awful for learning. The gravest problem, in my opinion, is the illusion of competence or the illusion of learning that it brings. This phenomenon is very well documented in the learning literature, well before LLMs were a thing. But LLMs amplify it.

Just take a look at PISA results. Students do poorly because they use LLMs and they never properly understood the content.

However, the education literature also paints a clear picture: after someone learns the basics, they're ready for other kinds of activities. In fact, if someone already understood thing and yet they still go over the same thing over and over, they stop learning. This is called the expertise reversal effect. So the alternative is to move on to other stuff.

I'm not arguing for having AI in the classroom. I'm just thinking of the implications of a well-researched theory: cognitive load theory. From that point of view, students that already learned the basics can then look at LLM slop and see its deficiencies. They can now plan an activity with an LLM and see how it nails some stuff and falls apart in other stuff. In that case, and only in that case, LLMs can improve learning in the same way that looking at your friend's essays can improve yours.

But, again, I'm not arguing for AI in the classroom. I think the world would be better off with a blanket AI ban in education. I think this is similar to how nicotine actually has some positive effects, but the bad outweighs the good by so much that it's much better to ban it.

Having a verbose wiki that is incomprehensible to most people is good because it filters out people who don't understand

Maybe.

I guess here I ask: what's the goal? What do we want more of? What do we want less of?

If our only policy is "out wiki should be incomprehensible to most*", I'm not sure we'll have more people in the world who understand the wiki. If, instead, we think "our wiki should be as accessible as possible", then we may find ways of teaching people.

• Edit: better characterize JettuhGlowie's stance

me when I open some source code in one of my 6 text editors, but it isn't one of the ones that has an lsp installed for that language:

Relying purely on the knowledge from an LLMs dataset is never a good idea. Frontier models at least append their data by popular demand. Google used to do this back in the day. If a search used to show no results they would note that and start searching their indexed sites for the unresolved search. Maybe a week or so later you could try again and go Google would finally come up with something. Same thing happens with LMMs now. If a question or prompt is popular enough they scrub their indexed sites with more focus and refine the dataset to give more accurate responses. But more accurate is still not 100% reliable. Open weight models usually have Datasets up until 2020, since filtering out AI slop from current day data is a science in itself and smaller labs and institutes don't have the rescources for that.

But where LLMs shine is in the fact that they can understand natural language and use tools. For example if I where to research a topic I'd go through hours of web searches, clicking all the links, reading all the abstracts etc. Why not just have an LLM with a websearch access do some agentic work. You give it relevant search terms, and it will go through several pages of the Google results, create summaries and provide the sources directly.

Add some MCP servers to that so LLMs can directly interact and navigate websites and it will massively speed up research. Instead of reading through several pages of a paper to figure out it is only vaguely related to your topic turns from 10-20 minutes to 1-2 minutes.

For example Qwen Modells can easily handle 300 page PDF documents without hallucinations.
And once you've got enough material, then you can go back on the sources and properly read through it yourself.

LLMs can very much be helpful for research, if you use them properly for what they where designed to do, namely data analysis, instead of some know-it-all Guru.

But AI is good at suggesting the direction: correct names of the relevant scientific field, names of algorithms, names of theorems, or names of certain classes. Then you can later find and open the right book and the right section. Some times it happens that you don't really understand the topic yourself, and it's hard for you even to formulate the topic correctly and properly to find the right section or topic. This still speeds up the search for the needed infdormation. But I myself try to minimize the use of AI, but sometimes I use local AI exactly in the above-mentioned cases and for the purpose of learning|studying something.

As a tool for learning Linux AI is fantastic. I'm going to have to disagree with you there.

Example: You get an error. You don't understand it or how to fix it. You copy and paste the error into Google: The answer you seek is probably on the 5th page after all the SEO bullshit. You copy and paste the error into Gemini, ChatGPT, Claude, etc and you get immediate details and possibly the exact answer you're looking for. Not only that, but if you still don't understand you can ask it further questions and have it clarify and that actually works really well.

IMHO, you're conflating "AI slop" aka "AI output, generated with as little effort as possible to avoid having to think" with "AI output, generated because you're thinking and want to find or understand something."

The two are miles apart.