they don't see words, only tokens and even with tokens they don't know how to count them at the LLM completion layer they have to be trained with something li

they don't see words, only tokens and even with tokens they don't know how to count them at the LLM completion layer they have to be trained with something like RLHF about word counting at the question answering / instruction following layers or at the application layer (so called "agentic workflows"), e.g. writing a Python code to count words, or calling a function or a CLI tool like "wc"