Skip to content

Back to the catalog

I wrote an AI textbook — how long until AI can do it better?
interconnects.aiarticle

I wrote an AI textbook — how long until AI can do it better?

Nathan Lambert-style essay (Interconnects) arguing that LLMs have plateaued on long-form, non-fiction/technical writing even as they became superhuman at coding and math, drawing on his experience writing a post-training textbook (Reinforcement Learning from Human Feedback) with AI assistance for LaTeX formatting, copyediting, and diagram generation. He contends models excel at checking or generating individual sentences/equations but fail to compound edits coherently across a document — an "irreducible compounding error" problem RLVR solved for math/code but not prose — and argues this reveals a broader generalization gap relevant to claims about AI autonomously solving open scientific problems. He predicts human-crafted textbooks will remain dominant for 2–5 years and that current models are best used as narrow editors (Claude) or typo-finders (GPT) rather than autonomous long-form writers.

TL;DR An AI researcher argues LLMs have stagnated at long-form non-fiction writing — despite superhuman coding/math gains — based on his experience writing an RLHF textbook with AI as editor rather than author.
Open original ↗
Document

What it claims

The useful part for an engineer is the author's concrete accounting of what AI did and didn't do while writing a 200–300 page technical book (Reinforcement Learning from Human Feedback, Manning):

  • AI-authored prose: well under 1% of the text, all of it individual technical-explanation sentences the author chose to keep.
  • Effort saved: about 10–20% of total writing effort, and the author does not expect that to reach a majority soon.
  • Where AI paid off most: mechanical, high-volume work, not drafting.

How it works

The page describes three workflows worth copying:

  1. Editor-comment triage. Editor questions were embedded in the LaTeX source behind a delimiter, \editor{}. Claude Code jumped to each one, printed the surrounding context, and classified it as an easy typo fix or something more nuanced. The author then wrote the replacement text or asked Claude for suggestions before applying the fix.
  2. Dual-format sync. The book existed as Markdown (the web version, which took reader feedback) and as a forked LaTeX copy (for the Manning editors). AI agents kept the two in sync. The task still took tens of hours, and the author estimates it would have taken about five times longer without agents.
  3. Whole-manuscript proofreading. A near-final PDF went to GPT 5.5 Pro, which found subtle typos across the full manuscript.

Other uses mentioned:

  • LaTeX equation formatting, described as "trivial" with AI.
  • TikZ and Python diagram generation.
  • Refactoring the repository and porting between languages.

Model split:

  • GPT models are best for finding typos and minor errors.
  • Claude models are the better editor: more taste, a better grasp of the task, and more useful help getting past writer's block.

For scientific papers: use AI for related-work or background sections you already know by heart. Do not use it for the abstract, introduction, experiments or conclusion.

Limits

Failure mode. Models check or produce single units (a sentence, equation or figure) well. Across many stacked edits they lose coherence, which the author calls "irreducible compounding errors." Whole chapters come out with confusing wording, muddled organization, needless cuteness and random conceptual errors.

Prompting caveat. The author argues that much of the weak output comes from how models are used. A plain request like "write me a poem" gets a minimal plan followed by straight autoregressive generation, with no iterative scratchpad. The suggested fix:

  • prompt heavily,
  • make the model work extensively before answering,
  • have it consult other judge models before returning text.

The author proposes this but did not test it.

Expertise requirement. You need to be an expert already to catch the model's flaws. Leaning on AI also slows your own build-up of intuition and taste.

Evidence quality. All claims come from one author's experience on one book. There are no benchmarks or controlled comparisons. The forecast that human-crafted textbooks will stay best for 2–5 years is opinion.

Organisations named here

  • OpenAI
  • Moonshot
  • Anthropic
  • Manning
  • Jasmine Sun

Technologies named here

  • Python
  • Markdown
  • LaTeX
  • GPT-4.5
  • Claude Code
  • TikZ
  • Kimi K2
  • GPT 5.5 Pro
  • Claude
  • Claude Fable 5

Techniques named here

  • RLVR
  • Inference-time scaling for writing

Concepts named here

  • Riemann hypothesis

Related in ai-policy