Murphy's books occupy a specific slot that is worth being clear about, because people acquire them for the wrong reason and then feel bad. These are references, not courses. Volume one, An Introduction, is over a thousand pages and moves from probability and decision theory through linear models, deep learning, and on into structure discovery. Volume two, Advanced Topics, is longer and goes deep into inference, graphical models, generative modelling, decision making under uncertainty and causality.
Nobody reads these front to back, including, I suspect, most people who cite them. What they are extremely good for is the moment you hit something you half understand. You know variational inference is doing something with a bound, you have used a library that implements it, and you would like the actual derivation laid out by someone careful. You go to the chapter, you get the treatment, and the notation is consistent with the rest of the book so you can follow the references backwards until you hit ground you are solid on.
That second property is underrated. A great deal of the pain in learning this material comes from stitching together sources with incompatible notation, and having one coherent two-thousand-page system to work within removes a whole class of confusion. The probabilistic framing is the organising idea and it earns its place. Presenting regression, classification, clustering, dimensionality reduction and generative modelling as instances of the same inferential pattern is genuinely clarifying, and once you have it you stop seeing machine learning as a catalogue of tricks.
It also gives you the right instincts about uncertainty, which is the thing practitioners most reliably get wrong. A model that is confidently incorrect is worse than one that says it does not know, and the probabilistic view makes that a first-class concern rather than an afterthought. The code situation deserves specific praise. Nearly every figure in these books links to a Colab notebook that regenerates it, in Python with JAX and TensorFlow.
This is a large amount of unglamorous work and it changes what the book is, because you can take a figure that has not quite clicked, open the code, change a parameter and watch what moves. That loop teaches faster than any amount of staring at the page. Reproducibility as a default in a textbook is still rare and it should not be. On the free PDFs: they are genuine, author-posted drafts under a Creative Commons licence, and MIT Press sells the finished versions.
They are drafts, so you will find typos and the occasional passage that has not settled, and they are entirely usable. If you work from these regularly, buying a copy is the decent thing to do. The main caution is prerequisites. Volume one wants real linear algebra, multivariate calculus and probability, and is not the book to learn machine learning from if this is your first contact with it.
Volume two assumes considerably more and will hurt without a proper foundation. Four point two. Outstanding as a reference, generous in its free distribution, and the wrong tool entirely if what you need is a course with a path through it.