I keep coming back to CS229 because it answers the question almost every other course skips, which is why any of this actually works. Where a typical bootcamp shows you how to fit a model in three lines of scikit-learn, Ng spends a lecture deriving the cost function, walking through gradient descent by hand, and showing you where the assumptions live. That is the whole point of the course and it is what makes it valuable years after you watch it. The Autumn 2018 recording is the one most people use, it is on YouTube as a full playlist, and the legendary course notes plus the problem sets sit on the official site so you can read along and actually do the work rather than passively watch.
The coverage is broad and classical, supervised learning from linear and logistic regression through generative models and support vector machines, then unsupervised learning, clustering and dimensionality reduction, a real treatment of learning theory with bias and variance, and a closing run at reinforcement learning. I will be honest about who this is and is not for. If your math is rusty you will struggle, this is a graduate course and it does not pretend otherwise, so the matrix calculus and probability are assumed rather than taught. It also predates the current wave, so you will not find transformers or anything about building with large language models here, and for that you should pair it with a newer course.
There is no hand holding, no certificate unless you enrol through Stanford, and no one marking your work. None of that bothers me because what you get instead is the real thing, the actual reasoning behind the tools you use every day, taught by the person who has probably introduced more people to machine learning than anyone alive. Watch it slowly, do the problem sets, and it will pay you back for years.