I've been trying to figure out why the term "prompt engineering" bugs me. I believe that @xgranade@wandering.shop on mastodon may have shone some light on it for me with this thread:
https://genomic.social/deck/@xgranade@wandering.shop/113222869786989028
The "engineering" part of it feels like it's suggesting that is this is akin to something like software engineering - also a less than solid use of the term "engineering" but one that I can cope with.
As @xgranade says in the thread, the prompt to an LLM is input and not the program itself. I rather like their definition of programming:
"the careful management of the state of a computer, including both ensuring the initial state of a computation and the mechanisms by which that state evolves."
You can have evolutionary process that explore different spaces. You can have stochastic processes - Markov Chains are one such. The act of "prompt engineering" is that of defining the initial state only. There's nothing in the term prompt engineering suggesting that there is any knowledge or control of how the state evolves. Which makes it all seem a little batshit insane to me. I don't think that I can currently accept that carefully crafting the initial state of a stochastic process, with no control of how that input is transformed, and most of the time without even an abstract knowledge of how the processes that do control how the input is transformed are constructed, can be called engineering.
I could see it as part of an artistic process maybe. Not as part of anything that needs to be reliable or defensible though. As part of an artistic process, the term "engineering" feels out of place. As part of anything else it feels like it's not engineering.