Reflections on Self-Learning with AI

Thanks to Longview Labs for sponsoring this article. Check out their work at arweavehub.com/weekly.

It feels like there is a proliferation of posts and studies around these days about how AI is ruining people’s brains, or how people are losing their skills, or talking about how no one knows how to do anything anymore.

In these posts AI tends to get painted as the “big bad”, and we should all exercise caution using it lest we become something less than human. I’m personally not convinced it’s anywhere near as bad as people make it seem.

To some extent, I can understand these comments and the general sentiment towards new technology. But I’m not entirely convinced that it’s as bad as people make it seem.

We do a lot of things considered “bad” for us, like consuming too much social media and frying our dopamine receptors. We’re human. Compared to some of these, I don’t think AI is as negative as people make it out to be (at least for personal growth).

I’m someone who has been learning independently for over five years now, and these tools seem like a positive on how I learn. Maybe this will age like milk and I’ll be proven wrong. However, for now I have to trust my own lived experience.

In this post I want to talk more about why I think they’re useful, how I use them, and where they fall short. And as someone who writes all of their own posts by hand, I’m not advocating for the replacement of human learning resources: rather a supplement.

Let’s start with the first important thing: mindset.

The first and most important thing for self-learning is having the correct mindset. Much writing has been done on the topic of the Growth Mindset and the Fixed Mindset (which I’d recommend you read up on if you’re unfamiliar).

With this mindset in hand, you also need to have the discipline to put your personal growth first over output. Don’t worry about whether you’re shipping fast or adding new features; your first goal is acquiring knowledge and building neural pathways. It’s not about how much output you’re doing. We’re not corporate, bro.

The reason discipline is so important is precisely because AI can do everything for you. Many of us have traded convenience and ease of use for things in our life, and AI is no different.

The same mindset that makes you copy and paste a code snippet from StackOverflow without understanding it is the same one that will lead you to “vibe code” without understanding the output.

The short-term results might seem beneficial, but in the long-run, you’ll be no different anyone else unless you’re the one who bothered to put the work in and build strong domain knowledge to solve problems effectively.

So with that being said the goal is to use LLMs to learn, which requires the correct attitude first and foremost. You need to realise you have agency and build the workflow required to do so.

The meat and potatoes of using LLMs to learn is with a technique called inquiry-based learning.

In short, you should be doing exploratory work and using AI more like you would a teacher or subject “expert” you have access to.

You start working on something which you don’t necessarily have full prerequisite knowledge of, and you start to form questions. Then, you try and collect evidence to back up those questions. Finally, you ask the chatbot whether or not your assumptions are correct.

For example, let’s say you’re learning a new programming language. You want to understand whether or not when you assign an existing list to a new variable name if ti creates a “shallow” or a “deep” copy. So, you test it out, and think that it’s making a deep one.

You present your process to a chatbot with your evidence, which will either confirm your theory or correct you, and give you information about the underlying implementation so you can understand it more deeply.

Notice that in this process the AI is not replacing your thinking; you’re simply using it as a way to extract valuable information that would otherwise have taken you down a Google search rabbit hole, or into forums or Discord servers trying to find an answer.

The learning approach I just mentioned is something you can see as either a good or bad thing: it’s good as you save the effort on menial tasks and having to interact with people who aren’t always as friendly as you’d like. Or, the bad, which is the fact that you are forgoing social interaction to get an answer (which you might have built some friendships or professional connections from).

It doesn’t have to be all or nothing, or one versus the other.

You can ask an LLM questions, and still frequent forums. You can use it to generate code snippets, and still read critically acclaimed technical books. In fact this is exactly what I do: I learn Rust by both reading the Rust book, tinkering with code myself, and then asking a chatbot questions on anything I find unclear.

Something I haven’t been using AI for is what I would consider “creative” tasks such as writing or making art. Maybe some would even consider programming to come under this too, but we all have to draw a line somewhere.

For me, that’s using AI to help me become more knowledgeable, and doing the menial tasks. There is still an argument for how much you want to delegate your thinking to AI, and I believe it likely depends how far along a growth path you are.

For example, if you’re a seasoned writer, the task of going through and finding typos probably isn’t going to have a big impact on you becoming a better writer. It’s just something you have to do in the editing phase. So, I’d probably ask the AI to do it.

However if I was a new programmer, contrary to popular belief I’d probably rely less on these tools until I’ve built up those neural connections and have a good understanding of the patterns I should be looking for when I program.

This is where it gets tricky though. Couldn’t you just use AI from the beginning and get it to teach you?

Maybe. I do believe you need a level of discomfort to learn effectively, so making it too easy is also a risk.

The most important thing in my opinion is keeping an extremely high level of curiosity and a desire to understand things. If you can keep that, then it doesn’t really matter which tool or resource you use because you’re still trying to peel back the layers and improve the way you understand things.

Thanks for reading.

If you want to read more posts on functional programming, PL development, and self-learning, feel free to subscribe.