Start with validation, not models
The single most common failure in beginner machine learning is a model that looks excellent and generalises terribly. Kaggle's Intro to Machine Learning is short and unglamorous, and it spends most of its length on exactly the right thing: holdout validation, underfitting and overfitting.
Once those are second nature, adding neural networks is a matter of weeks rather than months.
Take one theory course
Applied courses tell you which function to call. One theory course tells you why it works, and that is what lets you debug a model that misbehaves. MIT's 6.034 lectures remain the clearest free option, and cost nothing but attention.
Generative AI courses are worth taking last, not first — they make far more sense once you know what training and evaluation mean.