Why tokenizers are the hidden magic in language models