31 An ML Curriculum

My interest in ML is recent. A friend tried to get me interested in it in the early 2010s but I did not see the point of it as far as my own work and interests were concerned. You are exposed to the world of heuristics and nonlinearity and probability with Bayes, Monte Carlo, genetic algorithms and similar stuff when you dabble with things. You can even make some nifty stuff with it: board/card game engines, sentiment analysis and so on. But man-or-ape, coffee-or-wine and other probabilistic stuff was not very interesting to me when doing business process engines or hobbyist compiler/vm development where everything must be definite. I have to admit that cellular automata like Conway's Game of Life and other deterministic/probabilistic generators that produce emergent behavior leave you with an eerie feeling sometimes. The output often feels ... alive.

LLMs dragged me out of the slumber. A little bit, at first. And then completely, when they started to become actually useful. This was mid 2024. It is not that LLMs stopped being probabilistic. It is that they became useful for my workflows that needed "definite."

I started using them for Q&A, learning stuff, plotting stories and a whole lot of other things. Mid-to-late 2025 was when they became useful to me as programming assistants. And, today, I write close to 0% of my own code.

I picked up a whole lot of theory along the way, enough to understand how they work and even built a couple of tiny but otherwise useless overfitted things. But LLMs progressed so much in 2026, and subscriptions provided so many tokens for so little, that I spent most of the year working on a wide range of interesting-to-me tools:

ocr engines
compiler/vm/runtimes
conlangs
document markup languages
static site generators
newsletter mailers
sandboxes
LLM harnesses
and so many other things

However, sometime last month, I decided that I must expand my knowledge of how these things work. It is fine to know just about enough to get work done, but not knowing the entire thing from start to end leaves me unsatisfied. So I have been using LLMs and some other material to teach myself AI/ML/DL. And I have been developing a curriculum around it as well so that others too can make use of it.

The first eight chapters of the curriculum are mostly done, and take you from the simplest model that can learn something to a basic 0/1 classifier to something that can learn how xor works. I want to end it with an actual LLM. But, on the way, I would like to visit all the little things that I was not previously exposed to.

ML has had a long history and the terminology can be confusing:

Artificial Intelligence (AI)
Machine Learning (ML)
Neural Networks (NN)
Deep Learning (DL)

I hope to fix that confusion as well. So that I don't leave you hanging, AI completely subsumes ML which in turn completely subsumes NN. And DL is NN, but deeply layered.

I will provide another update on this, and discuss it in some detail, once I make the relevant GH repo public.

Some reference material that I am consulting: