Who writes the textbook when the model can: Lambert's post-training book

The book and the question

Nathan Lambert's textbook, Reinforcement Learning from Human Feedback: Aligning and Post-training LLMs, started shipping from Manning this month. It runs about 300 pages, comes with a 12 hour companion course and a codebase of exercises, and stays free to read at rlhfbook.com. Two days after the launch post he published a second essay with a question in the title: how long until AI can write a better one?

The setup makes the question sharper than it would be from anyone else. Lambert works on post-training, which is the set of techniques that turn a pretrained model into something that follows instructions and holds a conversation. The book documents how that is done. The models he is asking about are products of exactly those techniques. So the subject of the textbook is being asked whether it can replace the textbook's author.

What the models did and did not do

The numbers he reports are small. Less than 1 percent of the book's text came from a model, and that fraction was accepted only late, during exhausted editing passes. What models were good at was local. GPT 5.5 Pro caught typos across a 200 to 300 page manuscript reliably. Claude models were better editors in his account, with more taste and more interesting suggestions. Neither could organise a chapter.

His diagnosis is that models check every unit of content but do not revisit components. A model can verify that a sentence is right and that a paragraph is coherent, and it will still produce a chapter whose sections repeat each other, contradict each other in emphasis, or fail to build. He describes the effect as increasing entropy in long-form non-fiction, the opposite of what a textbook is supposed to do, which is compress a field into a shape someone can hold in their head.

This matches what we see when we ask a model to draft a survey section. The individual paragraphs are fine. The document is a pile of fine paragraphs. The work of deciding which three ideas the reader needs first, and which twenty they do not need at all, is still done by the person who has taught the material and watched where students get lost.

Why the gap matters beyond textbooks

Lambert draws a line from this to autonomous research. The argument is that long-form technical writing exercises the same faculty as scientific work: holding a large structure in mind, noticing that a result in section four undermines a claim in section two, and reorganising accordingly. If models are, in his phrase, genuinely horrible at this, then claims that they will soon run a research program on their own are running ahead of the evidence.

We find the argument partly convincing. The failure mode is real and we can reproduce it on demand. Where we are less sure is whether writing is a fair proxy. A research agent can offload structure to external tools, to files, to explicit plans that get reread, in a way a single autoregressive draft cannot. The textbook test measures the model's unassisted working memory for structure, and that may not be the bottleneck once the model is embedded in a system that does bookkeeping for it.

What expertise is doing here

The part of the essay we keep coming back to is what it implies about what an expert contributes. The facts in Lambert's book are in the papers, and the models have read the papers. The value of the book is the ordering, the choice of which historical era to explain before which algorithm, the decision that distillation deserves a chapter and that a particular regularisation trick deserves a footnote. Those are judgments about what a reader needs, and they come from having watched readers.

That is also what teaching is. A lecture is mostly a sequence of decisions about what to leave out. A model trained to be helpful on every question has been pushed toward completeness, and completeness is the enemy of a good textbook. We suspect the cause is an objective mismatch rather than a missing capability, and we would like to see someone try post-training a model specifically against the taste of experienced textbook editors to test that.

His forecast, and mine

Lambert's prediction is that in two to five years the best textbooks will still be mostly written by people. He notes that this cuts against his own earlier expectations. Our forecast is close to his with one amendment. We expect the second-best textbook in most subfields to be model-drafted well before then, with a human doing the ordering and the cutting, and we expect that arrangement to be good enough that the market for the third-best textbook disappears.

The experiment we would want to run is simple. Take the free online version of his book, give a model the chapter list and the papers, and have it draft the whole thing. Then have a few people who teach this material grade both versions blind on structure alone, ignoring prose. If the model loses badly, his argument holds. If it loses narrowly, the gap is closer to closing than either of us thinks.