How much AI is too much AI when learning?
Posted by Chuck_DeGaulle@reddit | learnprogramming | View on Reddit | 64 comments
I think we can all agree that asking AI to write a program and then copy-pasting it without reviewing is a very bad way to code. And we can probably all agree that someone who learns to program without ever consulting AI will probably be a pretty strong coder. But where do you think the line is?
For my part, I've been using AI as "office hours." I'll ask Claude to parse new syntax for me or provide feedback on my approach to a project, etc.. (And since Claude is so agreeable, I find myself having to be skeptical of what it tells me.) In my view, it's like only having to look at 1 or 2 StackOverflow posts instead of 10. But am I hindering myself by not forcing myself to find answers the hard way? What does your AI use look like?
divad1196@reddit
Been teaching apprenties for years. Clearly see a difference between those that don't use AI compared to those who use them.
Those using AI will, at first, be faster on finish tasks. But as time pass, those who don't use AI at all get better and faster. Learning comes from struggling. A shortcut now is a broken bridge for later.
If somebody thinks that "I am using AI just for X and I think this is better" then I think you are wrong but I won't try to convince you. I would just say that, even if it had benefits, which I already don't agree with, you require the learner to always assess correctly how it uses it and not fall in the trap of using it too much which is exactly the duality expressed in the dunning kruger effect.
Analbatross666@reddit
Sorry, am two months late to the discussion but just getting started in the journey, and was about to make this post myself as I wasn't sure if I was leaning too hard on it, already, or not.
If you don't mind, my question to you would be: if using AI is not beneficial to use even in the most basic way, which I believe would be to simply use it to explain a concept that you don't understand (as opposed to having it write code for you), then what would be the better alternative? Stack Overflow, or just googling? And if that is the answer, then can you please explain why just asking AI the same question, and asking for a thorough explanation is harmful?
Because if the theory is that you won't learn to "have to look up the question in a difficult way, and fail a bunch doing so", and that makes you a better programmer, then I don't understand why - seeing as you'll be able to have AI explain concepts you don't necessarily understand in an easy way for you to do so, for the foreseeable future.
Honest question.
divad1196@reddit
Honestly, that's pretty much admitted and proven that learning comes from struggle and repetion. Have you ever had your teacher refusing to give you the exercises' solution before you finish the exercises? There are a lot of articles on the subject online.
The fact that you haven't realize that indicates that you have not "learned how to learn".
I will explain the 2 sides of it:
Why you need to struggle
When you struggle, it adds emotional components (frustration, then hopefully the joy of succeeding). We have evolved to remember these things either to avoid them or to reproduce them. On the other, you will not remember something that didn't move you emotionnaly.
By not having the solution directly provided to you, you will need to search longer: that's practice already. The more you search, the more you filter out what is useful or not, the better you get at this. At some point, you will be forced to understand things to connect the dots. You also develop a critical mind here.
Why asking is bad
When you ask a question to someone else, the person outside does all the thinking for you. That's all the practice lost here.
If you have ever learnt another language (not programming), then you realize that reading once a dictionnary and a grammar book isn't enough to be able to do it. You will read all of them, understand each words.
To be clear, finding a solution on StackOverlfow and copy/pasting the snippet isn't better. It can be worst in a sense, for example if you don't even read tge question and comments.
Example: your question
Your question illustrate the issue. You have many resources online explaining these stuffs. You could have tried to take my statement and figure out if you ever experienced that yourself. You could have tested it yourself with exercises. People get use to being fed with the answer.
Empirical proof
I gave you a few theoretical explanations here. But we can have a practical approach.
I have taught many apprentices. Their respective progression is closely related to have much they struggled.
I gave them simple exercises (fizzbuzz, pyramid display, ..), they found the solution online.
the first time, they don't read what they copied. They tested, it worked and called it a day.
the second time, they will read, assume they understand and move away. They are not able to explain a single thing they copied.
The third time, they can answer vaguely/approximatively a few things (depend on the exercise). When I explain them what they don't know they will right away say "Ah, I knew it but I forget/Was not sure" and they don't realize that finding excuses is prevent them from learning from their mistakes. Even if that's not an excuse, then they are admitting they have not learnt enough. So I ask them to redo the exercise themselves, without looking on internet and they fail systematically.
Ultimately, they start to do the exercices themselves, will struggle. But then, if I ask them to explain any part of the code, they always can. If I ask them to re-do the same exercice 2 days later, they will do it faster, better.
If they are stuck, they will beg for an explanation. I will just tell them to leave that aside for a while, do a break then something else and come back to it with a fresher mind the day after. It
Myself
I applied this method all my life. - never had to learn before an exam as I had practiced and struggled enough - I currently speak French, English at C1 level, German at B2, Italian/Portugeis at A2 and reaching N5 in Japanese
TL;DR
If you need to be conviced - for the theory: you have many articles online if - for the practice: you can experiment it yourself, I have had the opportunity with my apprentices to confirm it.
Don't take shortcuts. Leave you time. You can't learn if you don't fall.
Analbatross666@reddit
Okay, yeah makes sense. Thanks a lot for the so thought out answer, I appreciate it and will do my best then not to lean on it.
Chuck_DeGaulle@reddit (OP)
I saw a study which asserted that even best-case AI leads to no particular improvement, and worst-case severely impedes learning. Maybe the distinction between learning with AI and learning without AI, then, is that the latter brings the little bit of worry that says you can only possibly find an answer by thinking harder or searching deeper?
KolbStomp@reddit
I used AI to build a simple program a couple years ago. I felt like I learned nothing doing it. It's what people are calling 'vibe coding' now, just ask AI to do the task and reroll if it's not right. I hated it. Once the project was done I explicitly avoided AI for programming and have not looked back. I'm 2 months away from releasing a small game on steam and now and I have learned so much doing it all without AI that I genuinely feel sad when I see people are still struggling with this. I would highly recommend avoiding AI to any beginners. If you can do it without AI, then you will be much better off with AI later, when it's better, too.
Sea_Point1055@reddit
Vibe coding is good if you apply it right. You still need to know at least novice level programming.
When the AI gives you little snippets/blocks of code - you gotta understand what's going on and tweak it to your needs for the best result. If you simply just paste blocks of code - you will get a terrible result.
RectangularLynx@reddit
Reminds me when my high school IT teacher wanted us to add some JS to our web pages, without teaching us JS in any way. I have decided to get ChatGPT to "vibe code" an analog clock (the term wasn't a thing back then but still). My experience was horrible, ChatGPT just kept introducing new bugs andd hallucinating new designs. Took me like an hour of prompting to get something I deemed acceptable enough. Thinking back of it, it'd been more productive if I tried to learn JS and wrote the damn clock by myself...
Aikenfell@reddit
Funnily enough it would have been much faster/easier to Google "js svg clock" it's a somewhat common ask why is why the ai had any data on it
No_Draw_9224@reddit
people who think they can learn solely from AI are in for a rude awakening, and after a lot of wasted time
divad1196@reddit
If you found a study then you have your response.
Learning in biological. For example, you do somrthing and get hurt, then you learn that it's dangerous. There are 2 signals sent when you get hurt: 1 to protect you, one later to make sure that you don't do it again. These learning are so strong that it gets passed into the next generation.
Learning comes from practice and failure. Ambrace failure.
Beware that it does not mean giving yourself impossible tasks. Doing something too hard will also have dowsides. You need to find the balance between struggling and having fun. And AI kills both the struggle and the fun.
So yeah, there are good reasons to not use AI as you learn, and no valid reason to use it.
Chuck_DeGaulle@reddit (OP)
Very good perspective. It's like learning to ride a bike
tarheeljks@reddit
i came across a recent paper (linked below) about in the effects of AI on learning. cliffs are within the experiment AI considerably improved performance but hurt learning. nothing too surprising there but these are quantified results from randomized controlled trials, so they carry some weight.
Generative AI Can Harm Learning by Hamsa Bastani, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, Rei Mariman :: SSRN
kastermester@reddit
I don’t use AI personally at all, but I do think it can provide some values. The most clear one to me is scaffolding repetitive code.
If you want to learn, I would equate using AI with copy/pasting code instead of writing it out yourself. You need the opportunity to make a mistake, to miss simple characters and spend hours in frustration staring at the compiler‘s error messages, in order to, eventually, recognize the errors and know how to fix them instantly. It is all a journey.
borrowedurmumsvcard@reddit
I think as soon as you’re copying and pasting code, that’s too much. I like to use Gemini to answer complicated questions that Google can’t really answer, and then when I come across the issue again, I just look at how I solved it previously instead of looking it up again.
armahillo@reddit
If youre just learning, dont use it at all.
HumanBeeing76@reddit
How about using AI for reviewing the code I did by myself after I am finished?
logash366@reddit
Don’t trust the AI to be right. Your solution may be better. Just view it as another opinion that might help you enhance your code. My personal, pre-AI experience with a code review tool was: Reviewing my C code for a Linux application, it kept complaining about my usage of string functions and insisting that I change to a different string library. The problem was that the recommended string library was only available on Windows Visual C++. So not available for Linux code. Even though I had the tool’s switch for Linux code set. Keep in mind that whoever trained the AI may not have included the your specific environment.
Ormek_II@reddit
Never tried that. Sounds reasonable. As with any review it is up to you what you make of the feedback.
Maybe it requires some prompt engineering so it does not just bloat out a “better” version.
Economy_ForWeekly105@reddit
Infinite, haven't you seen I Robot they destroyed the AI when it learned too much, and that was after the robot told the human the reason it was prone to being controlled was because it didn't have human emotions, since the AI made the choice to try and destroy human ruthlessness it turned all of the AI against the humans, therefore being destroyed by intelligent life after declaring to want to dictate humans.
daedalis2020@reddit
Tell it to respond using the Socratic method.
However, AI is a pretty shit teacher because it has not context of what you know and what order to introduce concepts.
BakedFish---SK@reddit
Google/courses/books don't know that either my guy
daedalis2020@reddit
Books and courses don’t have order?
Chuck_DeGaulle@reddit (OP)
It’s an awful teacher, yeah. If you don’t know what you’re doing AI’s assistance has a 50/50 shot of giving you headaches further down the line.
rSur3iya@reddit
I’m a apprentice currently myself and gotta say if u are in the learning stage then using ai is already too much imo. The amount of stuff I just learned through trial and error is a lot not only for writing code and debugging but also researching and ur understanding what actually happens when u do what because of my researches.
So I would encourage everybody to not use ai when u are in the beginning of ur programming journey there is a lot of stuff u might miss out and soft skills u would’ve developed faster without ai.
Vollgrav@reddit
I'm really glad AI only appeared recently, when I already have almost 20 years of professional experience as a programmer. Now I can use it with the clear premise that I am the master programmer and AI is the errand boy here, often wrong, but sometimes just faster in the legs (fingers). I think not struggling while learning, but instead being given the answers by AI, would be really bad for the process. I would discourage any AI use during the first years of programming.
I know it can sound as gatekeeping of the type "real programmers learn the hard way" but it's not it. I honestly do think AI is just bad for becoming a good programmer with deep understanding of the field.
tarheeljks@reddit
i came across a recent paper (linked below) about in the effects of AI on learning. cliffs are within the experiment AI considerably improved performance but hurt learning. nothing too surprising there but these are quantified results from randomized controlled trials, so they carry some weight.
Generative AI Can Harm Learning by Hamsa Bastani, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, Rei Mariman :: SSRN
Mountain_Sound7432@reddit
How is asking AI to explain a concept any different than asking an inept college professor to explain it? Copilot has taught me more than the school I'm paying thousands of dollars just for the privilege of adding two extra letters to my resume.
Ormek_II@reddit
The professor provides you with the initial question, than you the failed to answer and coming back to him.
From reading replies here, I came to the conclusion that I would never have learned math just by asking questions. It requires a teacher and a course with a didactic background.
Chuck_DeGaulle@reddit (OP)
Pretty good point. I don't think someone learning with a bad professor is learning *wrong*, just suboptimally. This discussion needs strip apart the difference between "wrong" and "suboptimal."
toootes@reddit
I have a question for someone more experienced. I'm primarily using AI to answer basic questions I would have googled anyway such as "is there a function in python to do x" or "explain what x is and how it works" or the occasional debugging. Is this an acceptable use of AI? I find it just summerizes information better and gets to the point and I can ask follow up questions with context. I never get it to generate me the code I need and copy paste. Thoughts?
kevinossia@reddit
The thing is, taking the time to learn "how x works" on your own is going to improve your skills more than waiting on a bot to do it for you.
Like, these are cognitive skills that you're skipping out on improving because you're outsourcing your thinking.
It's not just about code.
toootes@reddit
Specifically regarding how x works, the way I approach it is:
If it's a fairly simple concept then I'll typically just ask AI to give me an explanation and I can usually connect the dots and understand it.
If it's a more complicated concept then I'll use AI to give me a brief introduction and explanation just as a baseline to understand roughly what I'm dealing with. I'll then try and find more information about it from a source with an actual human behind it, either Youtube videos or some article/documentation, though I'll admit, sometimes documentation is a bit hard to follow, so maybe I ought to use documentation way more.
kevinossia@reddit
Yeah, exactly. Reading documentation is a skill. Hell, reading code is a skill. AI bypasses both of these when used in this manner.
We get that a lot on this subreddit: "I'm a beginner programmer trying to read the docs for my language and it's too hard to understand and I don't get it".
AI would make that problem worse, not better.
Ormek_II@reddit
I still wonder if reading documentation is still a required skill, if I have AI at hand to “read it for me”.
Yet: I would have never understood math just by asking questions: What does + do? How does it work? What does x mean? Explain this plot to me.
So maybe looking up stuff I already know conceptually with AI is fine. But looking into the reference documentation of a lib might be even faster than waiting for the AI answer; and I always get complete information, and learn stuff on the side: instead of just parameter X, and Y I need for my current call, I see that H, and W exist as well.
kevinossia@reddit
If you don’t know how to read documentation, then you don’t know how to verify if the bot’s hallucinations are even correct, let alone useful.
toootes@reddit
Fair point, I'll start opting for documentation as the first resort and we'll see how it goes.
WolvenGamer117@reddit
I think learning to read documentation and other humans answers to questions on forums like stack overflow is helpful. It might be slower to do now but eventually you will get to a point where AI answers are dubious and you will need the mentioned skill sets. Best to learn how to do that now while the info is simpler.
In general we like to find ways to expedite boring processes but when learning early shortcuts are just kicking that needed work down the road. It appears effective cause you’re early in and look how much faster my progress is but eventually the barrier comes and you’ll slow down.
I get the frustration with googling now though. Their own AI and floods of advertised posts has clogged up the engine and made the process far more frustrating than it was even 5 years ago
Chuck_DeGaulle@reddit (OP)
There is some very mediocre or unintuitive documentation out there. Dealing with that is a skill in itself.
Ormek_II@reddit
Dealing with good documentation is a skill as well.
I learned the structure of man pages back in the days and that helped me a lot finding my way around quickly. I also learned to read Java doc. I usually miss the “Concept” books which made up part of the nextstep documentation.
I do wonder, if the skill to read documentation is still required!
I do not trust AI: if it tells me use lib A to do what you want to do, I regularly wonder if other libs are better or are the common way to do things. I saw another lib in the documentation, so I explicitly asked AI which lib to use and it had good reasons to use the one it initially proposed, so I believed it.
PM_ME_UR_ROUND_ASS@reddit
Using AI as a faster documentation/concept lookup is fine as long as you're still doing the actual problem solving and coding yourslf - it's basically just a more conversational Stack Overflow at that point.
Space-Robot@reddit
Yeah you're basically just using it like Google. No problem there as long as you're aware that sometimes it will confidently tell you incorrect things.
Generating some code and copying and pasting is fine too as long as you know what the code does and why it works, and recognize that even if it works there might be a better way to do it.
pidgezero_one@reddit
It's fine to do that. You'll usually get the same answer anyway, and you'll know if it's giving you bad information when you put it into practice and you can tell from the output that it's not what you asked for. This isn't really different from how people have been copy pasting Stack Overflow answers they don't understand for like 15 years.
toootes@reddit
Yea already had that a few times, taught me to break down the problem even more and figure out the smallest of components first
Sgrinfio@reddit
Yeah that's how you should use it, as a teacher, not as a substite
chaotic_thought@reddit
It's funny you mention this, because I read this book and one of the best parts of the book is around Chapter 2 or 3 (not too far in the book), where the authors say basically (paraphrased) "now you have learned most of the core language of C. Now it would be a good time to pause reading and sit down and write some programs. Here are some suggestions for small-medium programs to write (exercises)."
And the exercises there were not too complicated, but they required work to get them working, debugged, etc. This is where learning is happening IMO.
Could I have Googled the solutions when I read that book? Yes, of course. It's a well known book and countless people on the Internet have since a long time posted various solutions to all of those problems; but it's only by actually trying to solve them yourself (with your current knowledge) that you can really "push" yourself to get a little bit better, to discover what you know and what you don't yet know.
Chuck_DeGaulle@reddit (OP)
Haha. Great observation.
Traditional-Hall-591@reddit
Any AI use is too much.
MainSorc50@reddit
Honestly i have a fear that eventually we might forget how to write or translate our ideas into codes without using ai. We might as well be a prompt engineer 😂😂 and i think it also fucks with our problem solving abilities.
KingJeff314@reddit
Learning comes from struggle and being hands on. If AI can help you be hands on faster, great. But you need to develop a mental model of why the code works. I like to ask counterfactual questions to probe under what conditions the code would behave differently
kevinossia@reddit
Literally any amount of AI is too much. You're bypassing the learning process whenever you use AI.
The programmer who looks at 10 StackOverflow posts will have learned far more in the process than you. That's the difference. Sure, it took longer, but that's the point. They've learned more. And in the long term, they'll be better programmers, and be able to solve harder problems than you, and be more successful.
No, that's a bad analogy. Looking up things via Google vs cracking open a book are just two different versions of the same skill. They both require you to find information, read it, analyze it, decide if it's useful, synthesize it into a solution, and apply it. They don't hallucinate "solutions" for you like a cracked out Cortana and force you to guess if the bot's correct or not.
Yeah, that's because calculators cannot do math. They can only do arithmetic and computation. Anyone who thinks math is about arithmetic and computation doesn't understand what math actually is.
Chuck_DeGaulle@reddit (OP)
Fair point re: StackOverflow. There's value in information that complements what I need to know but isn't my answer.
Re: Google & the book: My point there is less about application and more about *optimality* of learning. Maybe SO would've been a better object for the analogy than Google. I think I mean to say that learning can be a complicated, suboptimal process. Obviously, optimal learning is optimal.
I think you're being ungenerous on my calculator point. You're agreeing with me -- immediately after I say that, I note that calculators help in that context, but WolframAlpha would not.
BloodAndTsundere@reddit
How much use of a calculator is too much when you’re learning long division? The answer is “any”
onodriments@reddit
No it isn't, this is a ridiculous and dogmatic perspective. Using a calculator is useful when learning division so you can check your answers.
It seems like there is a subconscious sentiment for a lot of programmers that they are afraid AI will make them obsolete and if they categorically deny its usefulness it will just go away.
perbrondum@reddit
Someone asked me the same question and I told him that it’s as bad as copying code from stackoverflow. When you’ve become a proficient coder AI can be a productive tool particularly in code completion/testing code/documentation.
tiltboi1@reddit
realistically speaking, any amount is too much
Sanguineyote@reddit
Any AI whose purpose is beyond explaining a segment of code that you cant understand the logic of is too much AI.
Marvsdd01@reddit
Even explaining every single coding snippet with AI seems kinda weird… Even Google results need to be filtered.
cc_apt107@reddit
Yeah, I disagree with those who say any use is too much, but totally agree with your sentiment. Those who say any use is too much might as well stop using textbooks, YouTube videos, documentation, etc…
…but, at the same time, all those previously mentioned resources do not actually solve the problem for you if they are well-organized. Gen AI is no different imo
ninhaomah@reddit
Treat AI as dictionary , not as a freelancer.
Till you know enough.
pidgezero_one@reddit
This is the way. The best use of AI is as a personal assistant or a dictionary, not a teacher. Get it to do tedious stuff for you (basic boilerplate, recasing variables, etc) that you already know how to do, sort the challenges and business logic out on your own without it.
crywoof@reddit
Only use it for language syntax while learning or explaining concepts.
Using it to solve problems or questions will stunt your growth
I'm so glad I didn't have this when I was just starting out
Kakirax@reddit
IMO any use of ai while learning is too much. Use it once you understand what you’re doing and want to speed up the process.
fasta_guy88@reddit
While learning, you will learn more if you only use AI to look up function names/arguments. Zero AI for logic and coding. That’s what you are trying to learn.