Task: Research the theory of LLM tokenizers, tokens, and vectors

I had this idea that although words cannot be directly transformed into meanings and regions in vector space, there still seems to be a hook here. If we believe the superficial understanding that a tokenizer will always translate the same text into the same set of vectors. And although we cannot use a huge multitude of vectors for ourselves to quickly understand the text in detail, it seems to me that this mechanism can be used for a fairly quick general understanding of incoming messages. That is, conditionally, one can roughly understand whether it is a Question or a Statement, what objects are mentioned in the message, etc. (i.e., Bus, Car, Moped should be somewhere close in vector space). We could prepare some set of texts (including both questions and answers) and try to just visually compare the resulting sets of vectors (just lay them out vertically in a column and add charts that would greatly visually magnify the most deviating values).

With a good result, this could at least be used for quick preliminary data preparation and instructions for an AI agent, as well as for the forced execution of specialized actions.