How To Use AI to Learn More Effectively

I've always been fascinated with learning. I always tried to learn things on my own, scouring forums, hunting books, binging YouTube videos, which led to my career as a self-taught software engineer.

When AI models started getting good around the Sonnet 3.7 era, I felt I could use these tools to my advantage and learn everything even faster. Talking with Sonnet felt like talking with a stranger on the internet in 2008, both of us barely knowing English, somehow still understanding each other.

I tried various methods throughout the years, applied these methods on my emulator project, what follows is what survived.

How to prepare

First thing you have to realize when working with AI is they are self-explanatory: if you don't know how to begin, whether it's cooking, a video game, or an essay, you just ask, then ask again.

You have to internalize this fact: anything you don't know, you can learn with AI. And not just the thing itself, you can question the learning too: what to ask, where to start, whether you're asking the right question at all. (Here I'm assuming this is not some very niche specialized field.)

Let's expand on this, say you want to learn more about compilers, and let's assume you have very little to no experience when it comes to compilers.

You should simply begin by asking something like this:

Important part here is context, you need to tell the model who you are and what you do, and your question in detail.

If you don't like typing, you should consider voice-dictating: in my opinion it lets you express yourself more smoothly than quickly typing out a prompt.

After this, you should very carefully read what AI said and take note of the words and concepts you don't know, don't skip this part, these notes will matter later.

At every single step you should actively try to create a mental image and connect it to things you already know. The research here is clear: knowledge grows like a tree, new information has to grow out of a branch that's already there, otherwise it falls off.

How to begin

By now you should have a rough idea of what you actually want to learn. It doesn't have to be precise, but it should be more than a blob, something a bit concrete.

Create a folder in your computer, and put everything related you gathered here until now. I suggest making a git repo, it sucks to lose data and work!

Having an end goal is very important, I can't emphasize this enough. Learning for learning's sake is cool, but a finish line puts your learning on a trajectory rather than running in the dark.

Discuss this deeply with AI, get some ideas, take notes, ask other AI models to make your plan stronger. One warning here, most AIs today are very bad at estimating time, in my experience their predictions are generally 2x-3x more than what's required.

Aristotle explains better than me:

— Nicomachean Ethics, Book II

Choose a project where you think you can comfortably do it, it's important to get into groove rather than fighting with your ego and proving that you can finish very complicated projects.

Few AI tactics

There are few known patterns that I tried and still use that I think worth mentioning, and there are few anti-patterns which imo are more valuable for this post.

It's also important to state that AI is unlike anything we humans ever created, so its handling requires different care compared to other tools.

Things you should avoid when using AI

• It's very much okay to feel lost, you are learning new things, if you are not feeling lost or discomfort, it means you are not pushing yourself enough.

  So do the work yourself, never avoid hard work, treat AI as tool, not your brain.

• AIs have a habit of getting sidetracked, don't get sidetracked, this is why you have the project map, sure there could be flaws in the plan factually wrong things, but end goal should be thing you are moving towards
• You can't steer the model out of its base behaviour no matter what kind of prompt you give, you can't make OpenAI models talk like Anthropic models or vice versa.¹
• There are no magic "Skills", "Prompts" or anything like that that will give you huge edge, don't try to optimize a perfect prompt, models change fast one trick that works today might not work tomorrow with newer model, just try to explain yourself better.

Things to try

• Try fast paced, Socratic QA, switch to medium thinking instead of using "max" or "xhigh" for every single request. Flagship models today have very broad knowledge and thanks to web search and other built-in capabilities, most of the time they can reach the things they don't know.

• When you notice AI doing the same specific actions in the same specific order repeatedly, that's a pattern, try to make a skill or script out of it, it will save you a lot of time.

  While building my emulator, the model kept digging through ARM's official ISA XML files to find how an instruction is encoded, so I turned that into a small decode script, one command and the model gets the exact encoding instead of searching for it every time.

• Let AI ask you questions, you think you learned something ? Ask AI to ask you questions about it, you will be able to see your gaps in less than a dozen questions.

• Topic not sticking ? One of my favorite mathematicians, David Bessis, says

  Explain it to me again, but very simply, very slowly. I don’t understand anything about your subject. Assume that I have brain damage and can’t focus my attention for more than a few seconds.

  This works extremely well for challenging concepts that feels unnatural.

• If AI's answer to your question contains more than few concepts/words that are foreign to you, stop, write them down like I said in the beginning, and ask them on another chat window.

  LLMs are heavily steered by their context, so don't put unnecessary tokens into your main chat window, plus this keeps your main chat window clean!

• Make AI do the menial work, would you prefer to see the information on beautifully formatted html, can't bother typing LaTeX ? Just ask it, AI is there to serve you!

• Be ambitious, AI is huge leverage, things that used to be gatekept behind formalities and unnecessary qualifications are open now, want to create a new operating system ? Nothing is stopping you.

  Just remember what I said earlier, be ambitious with the goal, comfortable with the steps, my emulator sounded insane too, but it started as one tiny CPU struct.

Ignore the Noise

There are a lot of online discussions about AI, too many doomer talks, too much hype, people trying to sell something related to AI, unclear and sometimes discouraging talk from top AI labs. People telling you end of world coming, people telling you AI can automate your whole life!

I suggest you just enjoy things as they are, work on things you can change, use AI in a way you want that you feel beneficial to you, you should never feel like you are behind.

Here's a link to Terence Tao talking with ChatGPT on supposed disproving of Jacobian Conjecture, just look at his chat, he is asking one question after another, trying to wrap his head around the problem and its solution, if this isn't inspiring, I don't know what is.

Have fun.