I think of this as the "abstraction level problem". LLMs work at the wrong abstraction level (statistical relationships between individual words) for the tasks

I think of this as the "abstraction level problem". LLMs work at the wrong abstraction level (statistical relationships between individual words) for the tasks people want to use them for (high-level reasoning). Image generators have a similar problem -- they suck at large-scale composition and have problems with details like hands because the abstraction level they're working at (individual pixels and their close neighbors) is much lower than the level that a human artist works at.