There may be a more subtle issue here. When the specification is interpreted by an LLM that is different than it being interpreted by a person. From the LLM you get a kind of average of how a lot of people wrote that "kind" of code. From the person you get a specific interpretation of the spec into code that fits the task. Different people can have different interpretations, but that is not the same as the random variations LLM's produce. To get the same kind of fine tuning a person can do while coding (for example realizing the spec needs to change) from an LLM you need a very precise spec to start with, one that includes a lot of assumptions that are not included in current specs, but which are expected from people. I see further complications with getting an LLM to generate code where the spec changes, like say now you want to port the same spec to generate code on new computer architectures. So now specs need architecture dependent specifications? Some backwards compatibility needs to be maintained also, if the LLM regenerates ALL of the code each time, then the testing requirements balloon.