There is nothing out-of-distribution

2025-11-20

Summary: A sourced summary of an argument that LLM generalization failures are often intelligence limits rather than pure data-boundary limits.

This sourced note presents the author's view that text is a very broad interface because so much of human knowledge, planning, reasoning, and communication can be described in language. From that view, LLMs can potentially generalize to many tasks if the task can be expressed well enough in text.

The author argues that when models fail, the problem may often be a lack of intelligence, reasoning strength, or training quality rather than a strict inability to work outside a narrow dataset. This is an optimistic view of LLM progress and should be read as opinion, not settled science.

Free Basics version: clear descriptions help AI tools perform better, but AI can still fail and must be checked.

Source: Daniil Sedov, Gusarich's thoughts, gusarich.com/blog/there-is-nothing-out-of-distribution