My learning techniques have always been all over the place, especially in more unorthodox subjects (i.e. those that are not taught in school), because there are just so many options to choose from. As a consequence, I never could settle on learning something by sticking to a concrete method. I stuck to one for a week or so, then switched, either out of boredom or just the hope of landing on something better.
As one can imagine, this approach has been far from ideal, since discipline is key to learning, and I seemed to lacked virtually any in this respect. Thus, as I took up AI, I encountered the same issue I always do. First, I tried watching and coding along to Karpathy’s Zero to Hero series. Pivoted to watching Stanford’s cs231n lectures on Youtube and doing the publically available problem sets. Then, I switched to a textbook. I got bored, so I tried Karpathy again, but, as it turns out, second time’s not quite the charm.
A couple of months in, I had virtually nothing to show for my effort, most of which went into actually picking a learning method and materials rather than actually learning. As a last resort, I decided to try to actually build something and figure out the rest along the way. So, I took up implementing research papers, and, as it turns out, it was a great call.
While it is quite difficult to get started, after getting some inertia and especially after the “study things when you need them” mindset settles in, things get going unexpectedly smoothly. For me, having a clear, attainable goal in sight (i.e. the full implementation) and diving into more complex maths when I actually needed it (thus solidifying my understanding of why, how and when it is useful) kept me motivated and excited about what I was doing.
Bottom line, I learned that sometimes it is okay to not know everything from the start. In life, you can’t plan for every possibility, and I think it’s good practice trying to handle the unexpected and the unknown in a controlled environment, as I have the chance to do while learning.