Stanford's CS224N gets most of the attention in this space and deservedly so, but it is not the only good NLP course online and it is not always the right one. CS685 is pitched a level higher and aimed at a narrower audience, graduate students who already know the basics and need to get to the research frontier, and if that describes you then the fit is better here. The defining feature is that Mohit Iyyer teaches from papers. Each topic is anchored in specific work, and rather than presenting a smoothed-over textbook consensus, the lectures walk through what a particular group tried, what they found, and what was still wrong with it afterwards.
That is a different skill from understanding a finished idea and it is the skill that actually matters if you intend to work in the area, because the literature does not arrive pre-synthesised. Crucially this stops short of being a reading group, which is the failure mode of paper-driven courses, where students present unevenly and nobody ties anything together. Iyyer does the synthesis himself in lecture and the papers support it rather than replacing it. The course also handled the arrival of large language models better than most.
Plenty of established NLP syllabi responded by leaving the old structure intact and appending a lecture or two about transformers at the end, which produces the strange experience of spending eight weeks on methods the field has largely abandoned before being told, briefly, what everyone actually uses now. CS685 restructured. Recent offerings put language models near the centre and treat the earlier material as background that explains why things are done the way they are, which is the right call and a slightly uncomfortable one for an instructor whose earlier lectures were about the displaced methods. Comparing offerings across years is itself instructive.
You can watch the emphasis move, watch topics that once took a full lecture compress into a slide, and get a sense of the pace of change that no single snapshot conveys. I would genuinely recommend sampling the same topic from two different years as an exercise. The production is the weak point and prospective learners should set expectations. These are real lectures recorded in a real room.
Audio quality varies, questions from students are sometimes inaudible, the camera does what classroom cameras do, and there is none of the polish that Stanford's online operation brings. None of this bothered me much but it bothers some people a great deal, and it is a fair reason to prefer a slicker alternative if polish affects whether you keep going. Coverage is uneven in the way that any course taught by an active researcher is uneven. Areas close to Iyyer's own work get depth and care.
Others get a lecture that does the job without much enthusiasm. That is not a flaw exactly, a course taught with genuine interest in most of it beats a flat comprehensive survey, but you should not treat this as complete coverage of NLP. Pair it with a broader course or with the standard textbook if you need the full map. The currency issue applies here as it does to most academic courses.
The most recent public offering is a couple of years old and the field has not stood still. The methodological content holds up well, since how to read a paper critically and how to think about evaluation do not expire, but specific model discussions have dated. Four point three. Strong teaching, useful pitch, let down slightly by production and by the fact that keeping an online course current is work that almost nobody is paid to do.