---

One discovery I've had about myself in 2026 is that when it comes to computer science I'm more adaptive than I initially thought (and less adaptive than I needed to be, obviously, or else I'd probably be able to stay in Ireland…). I joked about learning deep learning in early 2019 but never actually did, because just like blockchain and cryptocurrency for whatever reasons I have always had this aversion towards ML/DL; and now the situation has worsen so much it finally no longer allows me to avoid it anymore. It's all very calculated and strategic; but seriously, when hasn't my choice been a calculated and strategic choice when I'm not relapsing and becoming a slacker who fucks everything up?

There is a hypothetical question where one would naturally ask seeing how things have become what they are right now, "if LLMs are only going to get better (which it seems to be for the last 3 years) and surpass living humans eventually (since it certainly has surpassed many living humans now), why bother learning anything at all? Why not just delegate everything to LLMs?" To this question I have no answer (not for now) other than some transcendental takes like that one mountain climbing guy saying that he climbs the mountain simply because it's there. I'm choosing to learn this, because I believe it's good investment, and I cannot stand the sight of me sitting around doing nothing while the LLM tsunami engulf us all.

And with that mentality I began. I (even up to now) still believe Yann LeCun is right about world models, so my goal is to be able to do that kind of research eventually; but obviously I cannot afford to spend months re-learning the maths I've forgotten over the years, so I've decided the most practical strategy is to use LLM to come up with a learning plan that starts writing code in Python from day 1 (first with JAX, later switched to PyTorch because I have an AMD Ryzen machine and couldn't get JAX working on it…) and learn the theory only when needed. It's far from my favourite way of learning and it would give me a flimsy and crippled basis of understanding since I've decided to trade that for speed, but again with the pace of things I only have so much choices before I could no longer make any.

And you know what? Despite me saying on more than one occasions that I was seriously done with computer science and only using this as a way to earn money, I did feel a little bit of joy during the learning, especially when you clearly know the code contains no special algorithms but somehow still managed to do special tasks (e.g. digit recognition via training on MNIST). I suppose in this era of "you don't need to learn because LLMs will take care of it for you" (or in any era, probably), learning itself is recreational and should be seen as a recreational activity outside of its supposed functions. I knew a certain guy who's originally a PLT guy but later turned to deep learning research around 2016. It's not like we talk with each other anymore these days, but there is a part of me that wonders if he himself had felt the same way I do.

---

Back

Last update: 2026.8.6