There is a specific kind of person working in AI right now who can build a retrieval pipeline, tune a prompt, evaluate an agent, and cannot tell you what a tokeniser is doing or why perplexity is the number everyone quotes. That gap is filled by exactly one document, and it has been free for twenty years. Jurafsky and Martin have been writing this book since the 1990s. The second edition was the standard NLP text for a generation.
The third edition has never been formally finished, and instead sits on Jurafsky's Stanford page as a continuously revised draft, updated as the field moves. The most recent release in January 2026 was a cleanup pass with refreshed transformer illustrations, which tells you something about the current state of it. The structural argument has stopped being about whether to add LLM content and started being about presentation. What makes it exceptional is not the LLM coverage, which many books now have.
It is the first third. Chapter by chapter you go through regular expressions and edit distance, n-gram language models, naive Bayes and logistic regression for text, vector semantics and embeddings, before you get anywhere near a neural network. Every one of those chapters pays for itself later. Understanding smoothing in n-gram models is what makes the concept of a language model assigning probability to a sequence feel like a mechanism rather than a slogan.
Working through cosine similarity on count vectors is what makes embedding space intuitive when you meet it again inside an attention head. People who skip to the transformer chapter get a description. People who read from the start get an explanation. The modern chapters are good and increasingly comprehensive.
Transformers are built up properly from attention rather than presented as a diagram to be memorised. There is substantial material on pretraining, fine tuning, instruction tuning, alignment and RLHF, and on the evaluation problem, which is treated with more care than almost any other source I have read. There is also serious attention to bias, harm and ethics, integrated into the relevant chapters instead of quarantined into an appendix nobody reads, and it is written by people who study this rather than people covering themselves. Now the honest limitations.
It is a draft and it behaves like one. Chapters get renumbered, sections move, and a reading plan written against last year's version will not survive contact with this year's. If you are studying with other people, agree on one PDF and stick to it. There are exercises but no solutions, so you get no feedback loop, and for a self learner that is the single biggest weakness.
There is also very little code. This is deliberate and defensible, since code rots faster than concepts, but it means the book cannot be your only resource. Run something practical alongside it. The Hugging Face LLM course, Karpathy's neural network series, or your own implementation attempts will each work.
The maths is another honest barrier. It is not intimidating by graduate textbook standards, but it assumes you are comfortable with vectors, matrices, partial derivatives and basic probability, and it does not pause to reassure you. If those are rusty, spend two weeks somewhere else first rather than grinding through this and concluding you are not smart enough, which is a conclusion a lot of people reach for reasons that are entirely about preparation. My 4.7 is close to the top of what I give anything, and the deductions are for the draft instability, the absence of solutions, and a breadth that can drown a first time reader.
Do not read it cover to cover on a first pass. Read chapters two through six properly, with a pen, then jump to the transformer and large language model chapters, then go back and take whatever else you need. Treating a reference work as a linear course is the most common way people fail to finish good books, and this one deserves better than to sit unfinished on a hard drive.