LLMs work as text prediction models, and the way tokens, words, and letters are structured across a context window in the right order affects how the model predicts the answer at the end.