The thing I have such a hard time understanding (and, in general, I have a very high-level, layman's understanding of LLMs) is how these LLMs are able to understand structure
inside tokens. That is, I understood that LLMs just predict the next token in the sequence. But for a game like Wordle, it's obviously critically important that the LLM understands the
mapping between the individual letters of, say, "apple", and the output result that the author
defined in the prompt, e.g. "XXXXO" means that only "e" is in the word (but the wrong position).
I'm just completely baffled how anything in the training procedure could allow the LLM to learn information about the structure of tokens. Does the tokenization process not treat every token (which I thought usually maps to a word) as an "opaque blob"?
I'm just completely baffled how anything in the training procedure could allow the LLM to learn information about the structure of tokens. Does the tokenization process not treat every token (which I thought usually maps to a word) as an "opaque blob"?