When integrating LLMs into any system (e.g., using OpenAI's Python API), you inject unpredictability into the system - even if all other system components are perfectly predictable.
As exponential coders, you and I will happily pay these costs given that LLMs are powerful "Alien Technology" (Simon Willison).
However, we must accept that the paradigm shift of "nets, baby, all the way down" (Elon Musk) will make many rules of traditional software engineering obsolete.