There is a stage in learning AI that almost nobody teaches. You have done the courses, you can train a model, you understand the maths well enough, and then you sit down to do something original and discover you have no idea how. Which paper do you read first. How do you know if an idea is any good before you spend three weeks on it.
How do you keep sixty experiment runs from becoming an undifferentiated mess of folders named final and final2. This is normally learned by sitting next to someone who already knows, which is fine if you have that person and leaves you stranded if you do not. Rajpurkar wrote the course for people who do not, and then gave the notes away, and that is a real contribution. The strongest chapter is the one on reading research papers.
He walks through his own reading of actual papers, showing which sections he reads in what order, what he skips, what he writes down and why. The insight, which is obvious once said and rarely said, is that experienced researchers do not read papers front to back. They read the abstract, look at the figures, decide whether the claim is interesting, and only then read the method. Learning that changes how much literature you can process by an order of magnitude.
The idea-generation material is similarly good, framing research taste as a set of habits rather than a gift. The tooling half is competent and the part that ages. PyTorch Lightning, Hugging Face, Colab, AWS, Weights and Biases were the right stack when he wrote it, and several have shifted since. The specifics of a Lightning module or a W and B sweep config are not exactly what you would write today.
That does not really damage the value, because the point of those chapters is the workflow shape rather than the API surface, and you can update the API by reading current docs. Just do not expect a copy-paste tutorial. Two honest limits. First, this is not a machine learning course and does not pretend to be.
You need to arrive already able to train a model. Second, the flavour is academic and medical, since that is Rajpurkar's field. The examples lean towards radiology and clinical datasets, and the implicit career shape is publishing papers rather than shipping products. A lot of the craft transfers directly, because experiment hygiene and literature triage matter the same either way, and some of the emphasis on publication norms will not be what an industry engineer needs.
On the distribution, the course book lives in Google Docs and reads like it. Two hundred and fifty pages of Google Docs on a phone is a poor experience, and it is free, so I will stop complaining. Four point three. Unusual, useful, genuinely free, and the closest thing to a written-down version of what you would otherwise have to get from a good supervisor.