Working on llm codegen feels a bit like trying to build a faster horse