My understanding is that currently several letters are grouped together into tokens for efficiency reasons. For instance, (The) might be a token, as might (cast) and (ing), so an LLM might read "The casting" as "(The)( )(cast)(ing)". Thus it would be quite difficult for such an LLM to answer how often the letter t appears in that sentence unless part of the training text indicated how often t was present in each of its tokens. However, if we didn't care about efficiency, we could make each letter its own token. In this case it would be pretty easy for the LLM to reason about how often t appeared in the sentence.