Stagnating AI Long-form Writing Raises Concerns Over Open-ended Science Capabilities

natolambert · x · 2026-08-12

AI researcher Nathan Lambert reflects on his experience writing an AI textbook, noting that current large language models have not significantly improved in long-form non-fiction and technical writing as previously expected.

He worries that if models struggle to organize and express lengthy thoughts, they will face difficulties in conducting genuine open-ended science. While AI can solve known math problems and make cross-domain connections, building a coherent body of literature without massively increasing entropy is crucial to advancing collective knowledge, a capability models currently lack.

Related event: Researcher Highlights LLM Limitations in Long-form Writing(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →