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

Nathan Lambert, who recently finished a post-training textbook titled Reinforcement Learning from Human Feedback, published reflections on AI writing ability and how AI models get more capable. He wrote that he used LLMs throughout the project, including for LaTeX equation formatting, extensive copyediting, and creating diagrams in TikZ and Python.

Lambert said he expected far more progress on non-fiction writing from models. He noted that some of the most famous models for writing ability are relatively old, citing OpenAI's GPT 4.5 and Moonshot's Kimi K2, while models have gone from okay to superhuman at tasks like coding and mathematics around those releases. He said writing well feels orthogonal to most other skills and that it is challenging and lacks good training data to specifically intervene on.

He described current models as genuinely horrible at long-form technical writing, saying they can get a sentence right but produce chapters with confusing wording, muddled organization, and random conceptual errors. He added that GPT models have long been strong at finding typos, and that GPT 5.5 Pro located minor typos across his 200-300 page manuscript, while Claude models were more useful as an editor, with more taste and better suggestions.

Lambert wrote that fewer than 1% of sentences in his book came from an AI model, included because he loved them. He estimated AI can save only 10-20% of effort today and said he does not see that percentage becoming the majority anytime soon. He said he expects the best textbooks to be heavily crafted by human hands in 2-5 years, and is unsure after that.

He argued that models being stagnant in long-form non-fiction should be alarming to those reliant on models autonomously solving grand, open science problems, since models today struggle to organize and compellingly present established science. He mentioned that Anthropic published a blog post on Claude making some progress on the Riemann Hypothesis on the same day. He said models are strong in truly verifiable domains and when given a lot of context to make a small edit, but not at open-ended prose generation.

In one example, he asked Claude Fable 5 in Claude Code to write a poem about a goldfish; the model said it made a minimal plan in its reasoning tokens and then autoregressively generated the poem.