At first it seemed LLM's were perfect assists for coding because they are trained on text and generate text. But code isn't typical text. It's basically a machine that requires a very high degree of precision and accuracy. Seen this way, LLM's are suited for coding only at specific stages -- to generate something like boiler plate, to brainstorm, to evaluate diverse approaches, identify missing tests. Anything that ties LLMs to actual code implementation is asking for trouble in my view.