And LLMs can’t leverage these abstractions nearly as well as humans…so far.