Long-form writing is where AI progress stalled, says Nathan Lambert

Nathan Lambert just finished writing a textbook on reinforcement learning from human feedback, and the experience left him with a pointed observation: models have gotten much better at code and math, but their long-form non-fiction writing has barely moved. Less than 1% of his book's content came from a model. The tools were genuinely useful for narrow jobs, copyediting, fixing LaTeX, and keeping the markdown and LaTeX versions in sync, where he estimates a 5x time saving. For actual chapters they fell down. His framing is that models increase entropy in long writing instead of compressing knowledge into insight. They reliably catch a typo or check a single sentence, but asked to fold many additions into a coherent whole, they produce muddled organization and confusing wording.

He points out that some of the models still considered best at writing, like GPT-4.5 and Kimi K2, are comparatively old, while newer releases won on coding and reasoning benchmarks without writing keeping pace. His explanation is that models hold vast knowledge but cannot express it well on underspecified problems, the kind with no single correct answer. That gap matters beyond books. If a model cannot organize what is already known into a clear explanation, the claim that it will soon solve open scientific problems on its own looks premature. Lambert expects expert-written textbooks to stay better for the next two to five years. His full argument is on Interconnects.

Why it matters

If your roadmap assumes models will soon handle open-ended research or long synthesis on their own, this is a concrete counter-signal to test against. For now, use them for the narrow, checkable tasks where they save real time, and keep a human in charge of structure and argument.