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

Newsletter/Podcast · Listen

• Nathan Lambert argues that despite rapid advancements in other AI capabilities (like coding and math), AI models are surprisingly stagnant and “genuinely horrible” at long-form, non-fiction technical writing, struggling with organization, conceptual coherence, and avoiding “irreducible compounding errors” beyond sentence-level tasks.
• He attributes this stagnation to writing being “orthogonal” to skills AI excels at, a lack of good training data, and models’ inability to effectively string together components and revisit additions, leading to increased “entropy” in long-form text.
• While AI is an “incredible tool” for specific, verifiable tasks like copyediting, formatting, or finding typos (as he used for his textbook, Reinforcement Learning from Human Feedback), relying on it for open-ended, long-form non-fiction can hinder a writer’s understanding and intuition, meaning human-crafted textbooks will remain superior for the foreseeable future.