I wonder how many llm tokens are burned dealing with Unicode issues?