I don’t buy the “LLMs are the new compilers” claim.

Compilers take an agreed upon specific intention “implement X via Y code in order to achieve Z” and make it executable. LLMs in the hyped methodology just take “Z”, and decide on an X and Y.

They’re doing things a compiler doesn’t: picking an approach/behavior, and picking a way to perform that approach in code. Those are subjective judgment calls, not mechanical processes that could be encoded in a classical program. That saves time, but given their reward system (“what the user requests is always the desirable goal”), they add considerable risks to the process of software development as well. No amount of guardrails and code quality agents can sidestep the fact that the decisions they make are substantially unpredictable and only as good as the prompter knows to ask for.