I wrote an AI textbook — how long until AI can do it better?
| Source: Interconnects (Nathan Lambert)
Tags: LLMs, AI writing, scientific AI, Nathan Lambert, AI limitations, long-form reasoning
Interconnects author Nathan Lambert argues LLMs remain surprisingly weak at organizing knowledge for long-form non-fiction — a structural gap, not a capability lag — that limits their autonomous science potential beyond narrow, isolated problems.
Details
Nathan Lambert, AI researcher and author of an AI textbook, offers a grounded critique rooted in direct experience: despite rapid model progress, current LLMs increase entropy in long-form structured non-fiction writing rather than compressing knowledge toward insight. He tested this while writing a textbook and found models could not organize established scientific material into coherent, progressive explanations. His broader argument: the ability to organize existing knowledge is a prerequisite for autonomous scientific progress. If models fail at this for settled science, their contribution to open-ended research will remain narrow — solving isolated problems and bridging distant fields rather than generating systematic breakthroughs. Lambert frames this as a structural issue. Models are optimized to generate plausible tokens, which is the wrong objective for editorial compression and the judgment required in explanatory writing. He contrasts this with domains where AI genuinely excels: mathematics and translation, where success is verifiable and structure is more explicit. He remains bullish on AI as a research assistant and optimistic about progress overall. The distinction he draws is between AI-assisted science (strong) and fully autonomous scientific reasoning (not yet ready) — and he argues the writing gap is a leading indicator of that gap.