yea its not too hard, see fastai, courseraml, deep learning book, talks on yt etc and loads of huggingface models/trainer scripts/pytorch very accessible these days, can get some gains with better data/retraining etc. Can get good at data oriented/parallel programming/cuda which helps.
Math fundamentals generally help, like its easy to brick the differentability with some hard if statements (use soft functions instead) in pytorch so good to understand you are programming for differentiability/autograd too