How I Learn Things

How I Learn Things

A Guide To Studying If You Happen To Be Me

// These are mainly observations I’ve made about myself, for myself, which I found from trying to deliberately study, but it’s echoed (and probably seeped in from) outside research on how to learn things. In between starting and finishing this post, I found Andy Matuschak on Dwarkesh talking about similar ideas.

I’m mainly writing this for myself, so I’m being a bit more firm and solid with advice then I normally would be. Take this all with a grain of salt for your own experience. //

Step 1 - Figure out what I want to learn.

I keep a list of questions and concepts that I want to learn, but haven’t prioritised, and or haven’t found a good resource for yet:

For a while, one of them has been to catch up on technical AI alignment. Let’s select that.

Step 2 - Find the best online course suitable to my level.

I don’t spend a long time on this. I will discover early into the course if it is a good one or not, so only spending an hour or two trying to select a good one is fine. After some googling and asking around, for AI alignment, the answer was ARENA.

Step 3 - Check prerequisites, and repeat from Step 1.

By that I mean, if the prerequisites seem to be “Calculus” and “Linear Algebra”, and I want to make sure I still have a decent understanding, I might try to do a few online quizzes on the pre-requisite. If I don’t feel comfortable with them, then I repeat from Step 1, but this time looking up courses on Calculus. (In this case 3Blue1Brown’s course was the best).

Step 4 - Do the course.

Often this involves watching videos or reading textbooks. When something doesn’t make sense, and it feels like there are missing prerequisites, pause, and go back to step 1 to figure out how to fill that knowledge gap. Ignore the time that course is supposed to take.

_________________________________


Step 4A - Actually Doing The Course

Study slowly.

Ignore the time that course is supposed to take. Don’t tie your ego to it. Expect to take a long time, but make sure you understand it deeply. How?

When you understand something, you should be able to answer all the curious questions that yourself or somebody else can ask about the thing, just by thinking about it and figuring them out.?

Obviously there will be lots of questions you can’t answer like this, that is fine, that just means you don’t understand that thing. Like, if you have studied linear algebra, and you think you understand what a determinant is, but then you think about what it might mean in n dimensions and you can’t figure it out, then you don’t understand that part, and you should try to figure it out and learn it if you want to understand determinants.

This is slow, and difficult, and can be very annoying if you planned to finish in three weeks and now you’re spending all this time on n dimensional determinants when you could just stick with what you know and ignore it...

Don’t do that.?

Well, at least, don’t do that because you feel like there is time pressure. It might very well be the case that studying what determinants mean in n dimensions isn’t relevant for you and you can drop it, but don’t make that decision because it seems hard and you want to claim to understand linear algebra now and finish the course quickly.?

If you do drop it, you should do so knowing that you don’t understand it yet, and hopefully after finding out enough about it that you know that it isn’t going to radically change your existing understanding of determinants.

That way, perhaps when you probe the question in another way, and you find that determinants can be thought of as representing the factor by which a space changes in “size” when transformed by a matrix, you can fill in “oh, although I don’t know how to do it in n dimensions, I get what that would mean now” and be satisfied.

Do quizzes and explore questions.

You are going to need to do examples, answer questions, review your work, and generally actually try to learn. The feeling of being at the edge of your understanding is good, don’t attach your ego to being smart and understanding everything right away, that’s a good way to stop exploring and stop learning.

I generally try to find quizzes and questions online to do to check if I understand something, and also give me practice actually using the knowledge. The other way is via spaced repetition study review.

Spaced Repetition Reviews

When I am learning something, I write about it, trying to categorise things into concepts like “Sigmoid Functions”, “Diagonalisation”, “Macros (C++)”, etc, and I write up my understanding of them as I go.

I find that by writing it up, I often find the edges of my understanding quickly. I get a little stuck, so to speak, and I need to stop and figure it out, or find material to answer the question, or ask ChatGPT to explain. When I am finished on a particular concept, I copy and paste my explanation into ChatGPT and ask it to review it for me. It often comes up with a slightly different explanation, or catches some mistakes, which helps quite a lot.

Then, when I am done for the day, I copy and paste just the concept headings into sections to review after a one day gap, two day gap, four day gap, eight day gap, and so on.

Then I link the whole document into a spreadsheet like this, and the date I need to review is automatically filled in.

When that date comes, I open up the document again, and try to explain the concepts from memory, based just on the concept headings. I check it against the concept explanations above, and also copy and paste the new explanations into ChatGPT to see if it finds anything wrong with them, or can give extra context I have missed.

Then I mark that date as green (and update the date I reviewed the work on if I was late doing it), and go on my merry way. Some days I also review earlier quizzes that I found helpful.

If I get stuck, I look at my earlier answers, and also copy that particular concept into an “Extra Review” doc.

The ‘Extra Review’ doc is checked every day, and I try to explain all of the concepts there. If I succeed without too much difficulty, I can just remove that concept from the ‘Extra Review’ document.

And, that’s it. I’ve been doing this on and off for about five years, often with very big gaps in between (I seem to run out of steam or get distracted by something else after around 6-12 months of heavy study like this, and then come back when I have more time again).

It seems to work pretty well. Recently I tried to apply the system to learning names at a new office, and it worked extremely well -- I’m usually awful at names, but with a little bit of deliberate effort (rather than just thinking “John! John! I should remember the name John!”) I was able to do quite well with it.

要查看或添加评论,请登录

Hunter Jay的更多文章

  • Superintelligent AI is possible in the 2020s

    Superintelligent AI is possible in the 2020s

    Back in June 2023, Soroush Pour and I discussed AI timelines on his podcast, The AGI Show. The biggest difference…

    4 条评论
  • What happens next? Predictions for the 2020s

    What happens next? Predictions for the 2020s

    I had a go writing up some rough guesses and intuitions I have about the next few years of AI. Predicting the future is…

    2 条评论

社区洞察

其他会员也浏览了