Based on the author’s recent work producing a post-training textbook, large language models (LLMs) proved useful for formatting, copyediting and repetitive technical tasks, but remain weak at producing coherent, long-form technical chapters. The book, Reinforcement Learning from Human Feedback (available from Manning and Amazon), was created with substantial LLM support for LaTeX equation formatting, extensive copyediting, and diagram generation (e.g., TikZ and Python).
Where models help
- Bug finding and small fixes: the author reports that GPT 5.5 Pro located deep, surprising minor typos in a near-final PDF manuscript of about 200–300 pages. For verifiable, local tasks, LLMs are particularly effective.
- Editorial suggestions: Anthropic’s Claude models were more useful as editors in the author’s experience — offering better taste, clearer mental models of the task, and more helpful suggestions to get past writer’s block.
- Workflow acceleration: the author had to maintain both Markdown and LaTeX versions and sync a web copy with a Manning editorial fork; automating parts of that process with LLMs significantly reduced the time required.
Why models struggle with long-form technical writing
The author observes that current models increase entropy in long-form non-fiction: they can produce a correct sentence or solve an isolated problem, but assembling whole chapters with consistent organization and accurate conceptual linking remains challenging. Key points:
- Composition problems: models are good at checking or producing individual content units, but they struggle to revisit and cohesively stitch those units together after many successive edits, leading to compounding errors.
- Lack of generalization: although models have progressed rapidly in other domains (coding, mathematics, search and research tasks), their ability to manage the structural demands of extended technical prose has lagged. The author suspects limited targeted training data for these specific writing skills plays a role.
- Underuse of inference-time planning: in typical usage models autoregressively generate text with minimal internal planning. They rarely perform deep, iterative scratchpad-like reasoning or consult judge models before returning a final draft, an approach that could improve long-form coherence.
How the author used models in practice
The author admits that well under 1% of technical explanation sentences in the book came from AI; those were retained because they were judged to be what the reader needed. Typical workflows included:
- Using Claude Code to process editor comments embedded in LaTeX (e.g., with a \editor{} delimiter): the model extracted context, classified whether a comment was a simple typo or a nuanced issue, and suggested text to insert.
- Automating repetitive tasks: background sections, formatting harmonization, and conversions between formats were sped up by assistant tools.
Nevertheless, the author reserved the most interpretive parts of the book — the abstract, introduction, experimental interpretation and conclusions — for human authorship, because those sections convey the story and intent of the research.
How much effort do models save today?
The author gives a conservative estimate that current LLMs can save roughly 10–20% of the effort in creating technical non-fiction: a meaningful but not transformative reduction. The improvement is real, but it does not yet change the fundamental need for expert human craft in producing high-quality reference works.
Outlook
- Short term (2–5 years): the best textbooks will still be heavily crafted by humans; LLMs will assist but will not replace expert organization and editing.
- Capability profile: models excel in two contexts — 1) fully verifiable domains where outputs can be checked, and 2) when provided a lot of context and asked to make a small, specific edit (e.g., finding a bug, solving a precise math step, or giving targeted feedback).
The author remains optimistic that LLMs will become powerful scientific assistants, particularly for formal and repetitive tasks, but cautions that their inability so far to express deeply organized knowledge in underspecified problems is a fundamental limitation if we expect them to autonomously solve broad scientific challenges.
Concluding note
LLMs are already valuable: they speed up workflows, find errors, and help convert knowledge between formats. Yet compressing and organizing knowledge into coherent, long-form technical exposition — the kind of work that creates new, generalizable insight — remains a frontier. Until that gap is closed, models will be revolutionary for local, verifiable tasks but less so for fully autonomous, large-scale scientific authorship.
