LGM Coding as an Art

LGM, the large guessing models, are already able to write code based on our description. But how to understand and control the result if reading the generated code becomes less meaningful?