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

Current artificial intelligence models struggle significantly with long-form, technical, and non-fiction writing, despite rapid advancements in coding and mathematics. While large language models excel at handling isolated units of content like checking individual sentences, fixing typos, formatting equations, or acting as editorial sounding boards, they fail to organize, structure, and compellingly present established science across a full-length book or chapter. When generating open-ended prose, models tend to make random conceptual errors and suffer from compounding organizational flaws.

The author utilized AI assistants as tactical tools while writing a textbook on reinforcement learning to handle tasks such as formatting LaTeX equations, syncing manuscript versions, and suggesting minor edits, but noted that less than one percent of the final text actually originated from AI. The models are fundamentally limited by their inability to reason through the structural compression required to produce genuine insight, typically generating text autoregressively without leveraging deep inference-time scaling.

This stagnation in non-fiction writing has broader implications for artificial intelligence development, suggesting that models are still far from autonomously solving complex, open-ended scientific problems. Because writing well requires organizing vast amounts of grounded knowledge, the failure of current systems to master long-form technical prose indicates that human expertise, taste, and intuition will remain essential for creating high-quality educational and scientific reference materials for the foreseeable future.