I wonder if the common gap here is the inability to reify specification.

LLMs seem to be a significant step forward in converting language to code, but they don't seem to engage self-directed abstraction.