The Counter-Intuitive Path to Machine Learning
you do not need mathematics. if you know basic matrices and calculus you are good to go
don’t waste your time learning all classical ML algorithms and how they work
learn how linear regression, logistic regression, softmax and log loss.
know that all the other algorithms (decision trees, xgboost, naive byes, SVM, etc) are just for these two tasks classification and regression and you don’t need to know anything about them apart from the fact that they exist.
I wasted a lot of time understanding the proof behind SVM, shouldn’t have.
learn about over-fitting and under-fitting of a model.
whenever you have a classification or regression task at hand, show the LLM ( I mostly use Gemini 2.5 pro in AI Studio ) your data and what you wanna do.
he would suggest you the algo and more or less it would be the best. ask him to write code for it too.
bonus tip :- learn about multi-processing, multi-threading and GIL in python. really helps when the data set is too large.
learning how to use basic AWS helps. sometimes the model would require a lot of RAM and your machine would not have it.
Deep Learning
after doing the above you can directly move to neural networks.
lecture 11 and 12 of cs 229 are good.
after these watch this karpathy video to write a neural net from scratch.
after learning about neural networks you can skip everything in between like (RNN, CNN, LSTM) etc and move to LLM’s (transformers basically)
( CNN’s are still used, so anytime you want to work with image classification or regression and you need the model to be light, fast and cheap checkout CNN’s)
unfortunately there are no good theory resources for LLM’s tbh, so I wrote this blog.
(the per-requisite to this blog is neural nets and Tokenizer)
after reading this you can watch karpathy writing a transformer from scratch.
What Now ?
now you are set for learning some advanced things which will differentiate you from the crowd.
first of all ditch linkedin and come to twitter. follow good people, if you do not know who to follow, follow people followed by smart people. (too many follows in a sentence lol). here’s mine
learn about RAG/Graph-RAG and Fine tuning of LLM’s
Learn about multi-model LLM’s (sadly there is not one good guide on these, I am thinking to write a blog when I get time)
Learn about pre-training optimizations, here’s the blog.
LLM’s are trained in two stages
Pre-Training (all of what you learned above)
Post-Training (which you have no clue about)
to learn about Post-Training, you need to know how RL works.
RL is hard. because there is no good simple resource. [ I want to create one but no time :( ]
still David silvers google deepmind lectures are good ( this is from where i studied, though had to use a lot of Gemini to really understand concepts )
after going through these lectures you can read this blogs for LLM post-training.
WHAT NOW ?
after all of this, spend time on twitter. talk to smart people there and you will eventually figure out what now.
and remember “if you want to work on rockets, go work on rockets. instead of waiting for that 100 million crypto bonus first”