Comment by int_19h

I don't think it's hard for the LLM to treat a sequence of two tokens as a semantically meaningful unit, though. They have to handle much more complicated dependencies to parse higher-level syntactic structures of the language.