I wrote an AI textbook — how long until AI can do it better?
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
Nathan Lambert, author of a new RLHF textbook, reflects on why AI models have stagnated at long-form non-fiction writing even as they've become superhuman at coding and math. He argues models excel at sentence-level edits, typo-finding, and equation formatting but fail to organize and compellingly present established knowledge across a full chapter or book, calling this a lack of generalization. He used Claude Code and GPT models for editing, LaTeX/Markdown syncing, and diagram generation, but kept AI-generated prose to under 1% of the book. He predicts humans will still write the best textbooks for 2-5 years, since AI can currently save only 10-20% of the writing effort.
Table of contents
Why have models stagnated in writing quality?Getting value out of current models as a writerWhere technical writing goes from hereQuestions this post answers
Why are AI models like GPT and Claude still bad at writing full book chapters even though they're great at coding and math?
Long-form non-fiction writing requires organizing and compressing knowledge across an entire chapter, not just getting individual sentences or equations right, and models struggle to string components together coherently once many additions stack on top of each other. They excel at localized edits, typo-finding, and equation formatting but produce muddled organization and conceptual errors over longer stretches, unlike math and code where RLVR has reduced compounding errors. daily.dev surfaces perspectives like this for developers weighing how far to trust AI in technical writing and documentation.
What is Claude Code useful for when writing or editing a technical book?
Claude Code can navigate a manuscript's LaTeX file to locate editor comments marked with a custom delimiter like \editor{}, print surrounding context, flag whether an issue is a simple typo fix or something more nuanced, and offer suggested rewrites for the author to accept or revise. It was also used to sync a book manuscript between Markdown and LaTeX versions, a task that would have taken roughly five times longer manually. developers automating documentation workflows with Claude Code can track these agentic editing patterns on daily.dev.