How many llm tokens are wasted everyday resolving utf issues?