Can someone explain why this is better than using a larger tokenizer? To me it seems like this would just make the LLM have a harder time understanding the content (when a token might have multiple meanings and isn't full, it can't have a good embedding)