Study log 1

Summary of the day

I completed exercises 0-33 of the 100 Numpy exercises. Overall I was unhappy with my progress, but I didn’t sleep well and had some other stuff to do that day. Overall motivation is good though, e.g. I am writing this at 1:35am after having worked on this a good amount of time during the day :)

I also decided to keep a diary/log each day to note down what I did, both for myself to keep me accountable, share what I specifically did with others, and give recommendations which material I found useful vs which I didn’t.

In addition to this, I am starting an Obsidian vault for the AI safety concepts I learn. For now it won’t be public, but I may publish it at some point.

100 Numpy exercises

So far, very useful. I haven’t used numpy much before, torch tensors are built on top of numpy arrays so it’s highly relevant to be able to understand and review ML code, and the notation has not always been clear to me in the past.

It being exercises also works really well for me - I’m an experienced python developer, so I can read code generally and know how python behaves generally (though today had some details on math.nan I was not aware of :D), so this kind of active learning suits me very well. I usually view learning as occuring on different levels (1 - memorization, read the content for the first time and try to store new knowledge, 2 - application, try to use this new knowledge to solve problems, 3 - transfer, try to use this knowledge in ways not mentioned in the source material, possibly slightly adapting it. Teaching is also a great way of learning at level 3, I just haven’t found a good way to incorporate it into this simple view), these exercises sit on levels 2 and 3, so they are great for me.

Note: Solving them will require many different numpy methods, you probably won’t know them all beforehand. The author provides a hint for all of the questions, sometimes that just mentions the numpy method you need to use. IMO it is completely fine to need the hint, then look up the documentation, play around with the method a bit, and solve the task that way. Having to memorize the hundreds of possible numpy methods seems like overkill. However, always try to solve an exercise first without looking at the hint, or you won’t learn as much.

Knowledge vault

We’ll see how useful this is later. I think writing stuff down like this in my own words improves my learning on level 1 a bit, as I need to know the new material well enough to paraphrase it. It also sometimes shows gaps in my understanding that allow me to reach a better understanding after researching it more. LLMs are basically a free 1:1 pocket tutor for this kind of stuff, which is truly amazing.