Nathan Lambert has spent the last few years watching AI models get dramatically better at coding and math. Writing an AI textbook made him notice where they haven’t.

In his post “I wrote an AI textbook, how long until AI can do it better?” Lambert argues models have plateaued at long-form, non-fiction explanatory writing, the kind a textbook needs, even while their coding and math scores keep climbing. He writes that “organizing knowledge is a compression. This compression is needed to make insight,” and that current models instead “increase entropy in long-form non-fiction writing,” meaning their long explanations tend to add confusion rather than remove it.

Who Lambert is

Lambert writes Interconnects AI, a widely read newsletter on AI models and research. He previously led post-training work at the Allen Institute for AI (Ai2), shaping models’ behavior after their initial training. He worked on Tülu, a downloadable language-model project, before leaving in June 2026.

What he gets right, and where it’s more speculative

Lambert distinguishes sentence-level work from organizing a whole document. He finds models useful for checking sentences or equations. He says they struggle to combine those pieces into long explanations without contradiction or repetition. That matches a pattern builders using AI for documentation or reports will recognize: individual paragraphs read fine, but the whole document doesn’t hang together as one argument.

Where Lambert moves from diagnosis to prediction is his claim about what this means for AI doing real science. He argues that if models can’t yet organize “the most established science in their area” into a clear explanation, their ability to make genuinely new open-ended discoveries is limited too, predicting that “the progress of LLMs for science will look closer to solving low-hanging fruit” than to major breakthroughs. That’s a reasonable inference from his own writing experience, not something the post proves with a broader study. It’s his forecast, not a settled result.

Why it’s notable

Coding and math benchmarks, standardized model tests, often offer checkable answers. A program’s tests can show whether particular expected behaviors work. Long-document quality is harder to measure. Documentation and research synthesis need facts arranged into an explanation readers can use.

What it means for builders

Treat an AI-drafted report, spec, or document as a draft to review before publishing. Read the whole document for whether it holds together as one argument, not just whether each paragraph reads smoothly on its own. That gap between sentence-level quality and document-level coherence is precisely the weakness Lambert is describing.

It’s also a reason to be skeptical of claims that AI agents can autonomously handle open-ended research or strategy work today. A model’s fluency in short answers does not establish independent judgment on unresolved problems.

End of article