Does AI help people learn faster, or just help them skip learning?

You can't learn anything by getting anybody or anything else to do your work for you.

Having been self-taught in programming and seeing how LLMs behave I will tell you my hands on experience. As a disclaimer, I am sure you already know, but I'll tell you anyway, that LLMs don't know how to read, speak, understand or get nuances (such as irony, sarcasm, under-handed compliments, etc., not can they be creative OOTB); they base all they know on the (possibly coached) learning they had and existing texts that they were trained on.

Moving on-topic, let's address the elephant in the room:

Many people say AI helps them learn faster by providing instant explanations and examples.

This is false. It is just an impression, a way that the person perceived the experience of that conversation. Consider for reference the classic delusional scenario of entitled people that I think all of us have witnessed or experienced at some point and I will use a bit of a niche topic just to emphasize it: some random person with absolutely no experience in a domain will read some information on Wikipedia (not even the full articles, just random paragraphs that they are interested in), let's say Quantum Physics. Then they go to post a rant on an experienced person's Twitter (I refuse to call it "X") feed on how they think the experienced person that does what they do for 40+ years, wrote countless papers, participated in even more countless conferences, wrote books, taught courses etc., should do some random thing or that they might be wrong, all based on a few paragraphs they read and some "opinion", not fact.

While extreme, that experience was surely felt by a lot of us when some business person said that "anything is possible in IT, and how hard is it to do X or Y? it should be a couple of lines, right?".

This is no different. When you don't know a topic and a LLM summarizes it in a(n approximate) ELI18 you feel like you have a good grasp on the topic, but let's be serious, there are tons of nuances that you can only learn from experience and a lot of digging, rather than just "how do I do X or Y?".

OTOH, if you have some experience, a LLM might help you generate proper code but this is entirely dependent on how well you write the prompt (and this is where experience comes in) and its "mood" (because it can be on topic, off-topic, embellish code where it doesn't need to, suggest things that are out-of-scope, etc.). Also note that whatever code it generates is great for a PoC, but it should not be deployed as production code, at least not without review. Depending on the prompt, once the security team runs pentests, you might soon see that all that code generated in seconds needs a serious rewrite that would still take some time (maybe not as long as originally, but still enough). That is of course if it respects your constraints.

Others argue that AI makes it easier to skip the learning process and rely on generated answers.

This is very dangerous. There are not only times when the code that it generates is not enough to address the problem or the answer might even be wrong, but even if it is correct, by some miracle, relying on some text without understanding it is just plain stupid.

TLDR: LLMs are great to give you some starting points on your queries, some references, some information, but you need to research it yourself. Relying on the LLM too much might result in incorrect results or worse. This is not even taking into account the brain rot effect. So use it wisely!

"Manual" learning is still the best way, especially since that is what was used to train the LLMs. That alone should tell you everything.

In your experience ...

I don't think that it is useful to ask for programmers' personal experience / anecdotal evidence on how to learn. You are better off reading some papers / articles written by experts; i.e. educators and psychologists. For example:

• How to help your students use AI without losing the learning
• Does generative AI help in learning programming: Students’ perceptions, reported use and relation to performance
• Should You Use AI to Learn to Code?

The last one includes this pithy summary:

In short, the more of a beginner [programmer] you are, the less you should use AI to generate code for you.

... followed by a list of reasons why ...

In my own experience in learning to code via courses and self learning by doing and failing I must say that I do have to find some answers to problems using LLMs but they often don't give a helpful explanation on how somethings work when it gives me a answer.
so I will still have to search it up on why it would work either way.

So for me, Yes it can help with problem solving but if it solves your problem and you learn nothing from the problem it solved for you, is it really learning to begin with.

Depends on the person—you’d need a longitudinal study and not sure how you’d be able to quantify it anyway.

Neither of them.

LLM (AI is very broad term) tools do what people usually do in the context. They are called a stochastic parrot because they have a compressed model of probabilities (statistics are analyzing data, probabilities are rules inference from the analyzis, and stochastic means that an algorithm considers the probabilities)

It's the exact same thing when your company has a list of 5 colors they consider the clients are the only ones they want. It's a simplification of reality, to a point a continuum becomes discrete. Logic allows us to work on continuum, but algorithmic requires discrete (yes, even double-precision have a finite amount of numbers they can represent)

So. LLMs only work when what you ask is already among the things that were deemed acceptable when the data was analyzed. Or very close to. If it's not close enough, the LLM will still deduce things, but the result is far from what you expect ( "hallucinate" ) . Selecting what is deemed acceptable is part of the training, after curating and annotating data, it's the neet part of the LLM.

This does not mean that LLMs are bad. It means, they are not a generic cure-all. They are actually good at doing simple tasks that are still too complex for human people to spend time "patternizing", that is formely deducing a pattern for. For example, summarizing mails :)

However, you also need to know when you are asking something that requires a specific human validation. This means, until you understand correctly the topic and how the LLMs addresses it, you should not trust the LLM at all .

But this does not mean you should not use LLMs to learn. Actually, asking an LLM can give you a lot of information on what is possible to do - just don't trust it on its explanations.
So, ask your LLM ; but always read the whole result and check anything that you are not already aware of. If your LLM uses a new library : what is that library ? Is it still maintained, is the code used the most recent one ?

LLMs are just tools, they can be used however you want. Their promotion as a "dev replacement" is plain BS but that dishonest advertising should not make you simplify this tool to a "good/bad" false dichotomy.

I don't like the over promotion of LLMs, yet I strongly disagree with you saying

This is false. It is just an impression, a way that the person perceived the experience

Sometimes you have very complex problems, and you can't even find where to start to solve it.
Having something that gives initial starting code, even when completely wrong and hallucinated, can be a good motivation.

LLMs, by searching what exists in similar cases, can help you get one foot in the stirrup.

Just like a good Google search algorithm if you provide the right keywords. And when you find too little breadcrumbs, there's StackOverflow. That's how I did it before LLMs and when searching was not enough, I'd ask questions (fortunately in my case, most of what I wanted to know was already answered on SO).

even when completely wrong and hallucinated

When it nonchalantly presents made up things as facts due to a hallucination, you can hardly call that a motivation (since you're wasting time on a false trail).

Now I know what you generally mean, but until it will reach the point where it can at least learn to read and understand (even without nuances, irony, sarcasm, jokes, etc., just literally), it is not something to be considered reliable. Having said that, I do admit I agree with your point that when it DOES manage to give you something on-point, it does help to get you started at times as its context dissection is a lot faster than what it would take you to find and extract the needed information if you had the same links/books as sources.

But to me, Perplexity is the closest to what LLMs with browsing features can do: being very good information aggregators. Still, I will give kudos to ChatGPT and Claude when it comes to coding. I'd lie if I said I wasn't impressed with just the concept of text-to-code even if the quality is not amazing.

There’s literally search engines explicitly advertised as non-ai. There’s a reason duck duck usage went up 30% when avoiding Google ai results weren’t configurable anymore

But learning also entails work it isn’t necessary to do by hand, just like I use a calculator. Depending on your goals a reduction in mechanical effort is still advantageous.

Sorry, I know I have a tendency for ... "long" (understatement) responses, but sometimes it can't be helped. Don't laugh at me :)

If you just copy-paste whatever code it gives you then copy-paste new code when the first one doesn't work right you're not going to learn much.

If however you read the output, apply your own knowledge to the offered solution and then formulate good questions when the first version doesn't work you will learn a lot.

I use a work version of Microsoft Copilot and it can read our internal docs, surface forgotton-about references, and pull in top answers faster than I can do a manual search; I feel even as a very experienced engineer I learn a good bit every day, or at least get riminded about something I've forgotten!

You learn by try and fail, then reiterate and do something different. When you ran out of options and cannot think of something else, talking to others helps give new insight or inspiration, or simply share knowledge.

AI has evolved so it can talk like a human on a topic. It can listen to your input and answer with different insight or new aspects. There is no guarantee of hallucinations, outdated facts or simply being wrong. Yet talking to a human being would neither guarantee all that. But it may give you new insight, inspiration or transport the knowledge from someone else.

@TheNomad do you usually write? like blog or something. If yes please share link.

I would say it induces I'll learn this later behavior more than learn it faster, learn it now. Usually, the projects demands something to be delivered faster, causing 'i'll check this later'. World is exploding with information, so its hard to look back what you bookmarked.

It depends on what you want to use AI for. If you want to use it for learning, then it can be a powerful learning assistant. If you want to achieve a certain result quickly, such as vibe coding, it can allow you to skip the tedious and systematic learning process

Yes, I think AI has both advantages and disadvantages for us, but so far it seems to be more advantageous.

My experience comes mostly from training an LLM, rather than using it as a tool.¹ And when training an LLM, the experience of "learning" anything doesn't really happen. At least it didn't to me. Your miles may vary.

¹ I do use LLMs as part of a search engine now and then, but would actually be happier if it didn't summarize bullshit, give wrong answers, and refer to meaningless textual pages. It's much better if I could get the old fashioned lists, but as far as I know, those are completely gone.

It's different for different people. You'll need to specify a little bit more about your own background and past study to get any real feedback on how this pertains to you.

But in general, I've taught first-year programming labs at a university, and it's incredibly obvious who is learning the core concepts while using AI and who is just using it to get across the finish line. AI helps people who would have failed or reconsidered their study get further; but at some point everyone needs to gain experience and judgement and if people are outsourcing that to study 'prompt engineering,' they're in for a bad time.