Most public NLP material now falls into two camps. There are practical courses teaching you to call model APIs, and there are older academic courses still built around parsing, tagging and pipeline architectures that the field has largely left behind. CMU's Advanced NLP sits in neither camp, and that is why it is worth your time. Neubig rebuilt this course around what the field actually became.
The sequence works up from language modelling fundamentals and neural network basics through attention and transformers, then into pre training, fine tuning, instruction following, prompting, retrieval augmentation, multilingual and multimodal work, agents, and the persistent problem of evaluation. Interpretability, efficiency and ethics get proper slots rather than a closing lecture. The reading lists deserve particular mention, because they are one of the best curated paths into NLP literature I have seen. Following them will teach you how to read a paper critically, which is a skill with a longer half life than any specific technique.
The lectures are recorded and posted, and they are taught at a genuine graduate pace. Neubig has a habit I appreciate, which is naming the open questions and disagreements instead of presenting a settled story. When he says the evidence for something is thin, you learn more from that sentence than from a chapter of confident summary elsewhere. The assignments are built to make you implement and measure things, which is the correct emphasis, since almost nobody's understanding of these systems survives their first attempt at evaluating one honestly.
Practical caveats. This material exists across multiple offerings, and different semesters have been taught by different instructors, with later versions adapted from Neubig's original design. That means you should spend a few minutes deciding which offering to follow rather than assuming the top search result is the best one, and it also means delivery quality is less uniform than a single authored course. For outside learners the usual gaps apply.
No grading, no feedback, no cohort, and no certificate, so you need enough discipline to do the assignments when nobody is checking. The prerequisites are real. Machine learning background, comfortable PyTorch, and enough mathematical maturity to read a paper without stalling on notation. Someone arriving from a prompting course will find the third week hard.
There is also a matter of emphasis. This is a research course. It will make you a better reader and a better thinker about language models, and it will not teach you much about latency budgets, cost control, or the tedious realities of running an NLP system in production. Those are learnable elsewhere, and it is worth knowing which gap you are filling.
My 4.4 reflects strong, current, intellectually honest graduate material available for free, marked down slightly for the fragmentation across offerings and the total absence of support for anyone learning from outside. If you can only take one academic NLP course and you want to understand the field rather than the tooling, take this one.