no

LLMs can produce correct information about many different topics, including retro-computing. But at the same time, LLMs can also produce wrong information about many topics (again including retro-computing), that sounds convincing to people not well-versed in that subject. This wrong information generated by LLMs is usually called "hallucination", and the possiblity of getting hallucinations from an LLM makes it dangerous to use an LLM as learning tool.

Without the intent to blame and shame anyone, just to point out the danger, I am posting two links to VOGONs posts that contain AI hallucinations:

• Seeking for the "nocga" utility or other ways to disable integrated CGA (a hypothetical NOCGA utility that doesn't exist)
• Re: Are windows before XP in virtual machines less efficient due to lack of HLT instruction support? (an LLM instructing to modify a registry setting introduced with Windows Vista to disable "fast boot" as a means to enable HLT usage on Windows 2000; while this post does not point out that the "knowledge" in it is AI generated, the presentation of that information with a lot of correct side-tracks around a completely wrong main point are typical for hallucinations of LLMs)

I was disappointed when the coworker at my workplace asked AI about specific solution to minor dilemma. I thought about making minor change to the part of code that retrieves path to user's Downloads folder. Despite Windows Vista being almost 20 years old, .NET still doesn't have the normal means to retrieve path to that folder, Environment.SpecialFolder enum is still stuck in Windows XP era. Unless something changed in .NET 10? Not sure, the relevant issue is still open on GitHub.

AI generated pretty much the same solution how it was done in current code, so get user's profile directory through the usual means, then append hardcoded "Downloads" to path, so it would end being used / created, even if user moved Downloads somewhere else. AI also added the mysterious "using" referencing "Shell32" namespace. It didn't explain where the namespace came from.

That coworker got all mad when I mentioned using WIN32 API to get it, due to supposed plans for Linux builds. As if you'll do updates the Windows way. Well, who knows, but even so, few extra lines of code won't kill you if you'll have to overhaul that whole application anyway.

Another coworker also seems to love using AI, sometimes even remarks it in commit messages.

Programmers are weird. Perhaps staring at the code for thousands of hours fried their brain long time ago.

Sometimes AI can point you in the right direction, but generally, it's a mixed bag. TBH, I don't know the right approach for learning programming in general. There's so many specifics and conventions to do things. How do you know if a part of the task can be accomplished by your runtime library and where do you find the right method? Do you just start by studying some fat book / documentation?

The only book on programming I remember starting and finishing was The C Programming Language. I was still pretty clueless at the end.

I'm also not sure whether AI is making us dumber. Is it much worse than spending lots of time researching on your own and moving at snail pace?

Never gotten an answer from an LLM about an in-depth technical subject that I have intimate familiarity with that isn't confidently wrong in some way.

BEEN_Nath_58 wrote on 2025-10-05, 18:50:

I have been pretty late to leverage the use of LLMs for learning purposes; I have just increased its use with the starting of my […]
Show full quote

I have been pretty late to leverage the use of LLMs for learning purposes; I have just increased its use with the starting of my university curriculum.

In 2025, I feel that most of the questions and doubts I have asked on this forum, or other forums, are getting systematically and well explained by LLMs.

I have asked it on runtime environments, compiler differences, Windows audio structural differences, etc and it presents acceptable answers in a well presented form.

I don't know how much you would use an AI tool for the context of topics on Vogons, but it would be interesting to know how much these LLMs can research and produce correct information on its own.

in my experience recent LLMs do well with things they have had lots of 'training' on. technical topics oftend have not only many articles, tutorials and books as well as official documentation but also lots of online activity in forums.

It's no surprise an LLM can act like a 'super' search engine in many areas.

It can also make subtle errors and these can catch a user out, this is especially true for niche or highly specific questions that simply don't have as much text about, like the kind of sometimes very specific and unusual stuff asked on vogons

so, while i agree an LLM can be a good aid to learning, it can also mislead - as those still learning may not be able to discern the good from bad information

It's actually a great exercise in the difference between "education" and intelligence.

In my opinion you shouldn't use any LLM that doesn't cite its sources which shouldn't be very hard at all but in most cases don't likely due to copyright issues.

Also humans learn by doing, if something or someone else is doing the work then you aren't. Depending on what you are doing or may need done that may be fine or not.

For vogons posts people should only be posting what they have researched, tested or experienced as "true" but as we all know human laziness, inferiority complex, greed get in the way.

There is also the issue of if you aren't asking those questions here then that discourse never occurred. You received an answer which may or may not or is somewhat correct are happy with it and never received any conflicting information and didn't contribute to vogons. Lots of sites are complaining about that.

A lot of this LLM stuff seems like a solution in search of a problem, I don't understand why they are trying to push it as the next industrial revolution instead of focusing on the specific scenarios where it could actually be useful. The underlying technology is indeed very interesting, and its use as a massive knowledge repository is obviously valuable, assuming you can filter out the bias and outright false information. But as others said, the way humans learn is by experience, imagine trying to learn how to drive a car by riding a taxi without even looking at what the driver is doing!

I even have serious doubts about its use on productive work: the fact that many executives are mandating the use of AI tools (and in some cases tying it to the performance evaluations) is a huge red flag to me, why do they need to do that if this technology is "as smart as a junior developer"? Maybe I'm just cynical, but the whole "I increased my productivity by 10x using LLMs!" I see parroted everywhere makes me question if those developers were doing anything useful in the first place.

TheMobRules wrote on 2025-10-07, 04:13:

I even have serious doubts about its use on productive work: the fact that many executives are mandating the use of AI tools (and in some cases tying it to the performance evaluations) is a huge red flag to me, why do they need to do that if this technology is "as smart as a junior developer"? Maybe I'm just cynical, but the whole "I increased my productivity by 10x using LLMs!" I see parroted everywhere makes me question if those developers were doing anything useful in the first place.

I don't see there being a single developer with the level of knowledge of an LLM who could be called junior for it, it'd be an insult. Obviously the more experienced you are the more defined your weak spots, if you can lift yourself in those spots with AI you'll be more rounded already.

gerry wrote on 2025-10-06, 16:21:

so, while i agree an LLM can be a good aid to learning, it can also mislead - as those still learning may not be able to discern the good from bad information

keenmaster486 wrote on 2025-10-06, 16:26:

It's actually a great exercise in the difference between "education" and intelligence.

DosFreak wrote on 2025-10-06, 17:45:

Also humans learn by doing, if something or someone else is doing the work then you aren't. Depending on what you are doing or may need done that may be fine or not.

For vogons posts people should only be posting what they have researched, tested or experienced as "true" but as we all know human laziness, inferiority complex, greed get in the way.

There is also the issue of if you aren't asking those questions here then that discourse never occurred. You received an answer which may or may not or is somewhat correct are happy with it and never received any conflicting information and didn't contribute to vogons. Lots of sites are complaining about that.

I see that the usage of an LLM can be optimised significantly by having prior information about the subject, or better with prior subject. Whether the output from the AI is correct or appropriate, they can only be subjected to test with personal experiences of working with something.

I have never used the LLMs for things I have zero information about, and I can't trust them anytime soon to have them always guide them in the right direction. Maybe a perfect mix is to take 1 step with the AI, take 5 steps yourself however rough it is, and then use the LLM to smoothen and correct things on either ends and then repeat the process.

As for Vogons I have had a lot of technical information answered by citing the sources from here itself, even my own answers (that wouldn't simply come up in search). It helped tabulate and deepen the answers and followups to my questions, things scattered around which would've taken a long time to find on my own.

As for things that have never been answered or developed, I wouldn't have an AI touch it. I remember using GPT in 2023 for Wing Commander issues only for it to suggest running the game with Windows XP compatibility mode 🤣

UCyborg wrote on 2025-10-05, 20:02:

I was disappointed when the coworker at my workplace asked AI about specific solution to minor dilemma. I thought about making m […]
Show full quote

I was disappointed when the coworker at my workplace asked AI about specific solution to minor dilemma. I thought about making minor change to the part of code that retrieves path to user's Downloads folder. Despite Windows Vista being almost 20 years old, .NET still doesn't have the normal means to retrieve path to that folder, Environment.SpecialFolder enum is still stuck in Windows XP era. Unless something changed in .NET 10? Not sure, the relevant issue is still open on GitHub.

AI generated pretty much the same solution how it was done in current code, so get user's profile directory through the usual means, then append hardcoded "Downloads" to path, so it would end being used / created, even if user moved Downloads somewhere else. AI also added the mysterious "using" referencing "Shell32" namespace. It didn't explain where the namespace came from.

That coworker got all mad when I mentioned using WIN32 API to get it, due to supposed plans for Linux builds. As if you'll do updates the Windows way. Well, who knows, but even so, few extra lines of code won't kill you if you'll have to overhaul that whole application anyway.

Another coworker also seems to love using AI, sometimes even remarks it in commit messages.

Programmers are weird. Perhaps staring at the code for thousands of hours fried their brain long time ago.

Sometimes AI can point you in the right direction, but generally, it's a mixed bag. TBH, I don't know the right approach for learning programming in general. There's so many specifics and conventions to do things. How do you know if a part of the task can be accomplished by your runtime library and where do you find the right method? Do you just start by studying some fat book / documentation?

The only book on programming I remember starting and finishing was The C Programming Language. I was still pretty clueless at the end.

I'm also not sure whether AI is making us dumber. Is it much worse than spending lots of time researching on your own and moving at snail pace?

I decided to use some "free AI browser" for a project in my course for improving a site's frontend, not to mention it failed terribly, not just to improve the site but degrading the site, but reducing the functionality of the site.

It can be an addition to solve the problem, but I still back that personal skills and web searches and probably stackoverflow can help more

Recently on Vogons I asked for the most efficient graphics card under XP for DX6-9, then did my own testing. What I found wasn't in the answers and the community consensus I believe was an Nvidia Maxwell, even though they work poorly or not at all with DX6-7. Would've been bad for a new user's build for example. For comparison I asked the same of a few AI now and they simply used Vogons among others as cited sources to repeat the consensus, including from my thread. In this case you can tell the hallucination came from humans and not the AI because you have the sources, but it would've been better for the AI to be critical of them. Amplifying a consensus would be one way to compress human knowledge if you're worried about that.

And human consensus is only reliably correct in niche technical topics where there are no political or personal motivations to be wrong.

For non niche topics it is quite certain if one has to stop at a consensus then they are doing it wrong and will encounter an exception one way or the other later. Or a mix of both rights and wrongs, the LLM might is not a personal testing field so it wouldn't be able to argue much for things that have a variable viewpoint

Studying for taking the CCNA exam, I sometimes ask googles AI to confirm what I know. Was shocked I was wrong about some basic things, looked into it again and found the ai was blatantly wrong about simple networking facts. Don't use AI to learn unless you know enough to be suspicious.

Shagittarius wrote on 2025-10-07, 21:00:

Studying for taking the CCNA exam, I sometimes ask googles AI to confirm what I know. Was shocked I was wrong about some basic things, looked into it again and found the ai was blatantly wrong about simple networking facts. Don't use AI to learn unless you know enough to be suspicious.

There's no single Google AI so there's no single outcome in using it. You can be shocked while someone else seems to get valid answers. Those who have naturally superior learning strategies will get even more ahead this way since at some point down the ability curve the complexity of the system is going to start pushing the learner back. Probably they were saying the same thing about the internet back then, I don't remember.