Comment by rapatel0
1 year ago
Anyone else find it funny that the entire concept of prompt engineering is probably very close to good management and communication principles?
It's kinda hilarious when you think about it.
My mental model for LLM chatbots is to treat them like a junior intern that has access to google. Sure they can get things right but realistically I have to check their work to prevent any show stopping issues.
Kinda, but I worry that actually overstates how practical/tractable the situation is. (As opposed to being just a very accessible form of endless statistical whack-a-mole.)
We're just trying to guide the the chaotic fever-dream of a text-predictor.
I run our postmortem/incident review process at work, and there is one area on our incident report that is intended to capture any investigation threads that are still ongoing at the point the report was written, so as to note that there is more information to come but that the authors didn't think it necessary to hold up the whole process for that thread.
Getting people to not put action items or proposals in that section (i.e. propose investigation threads that are NOT in process yet) has been... challenging. But every time I change the description of that field on the report to try to better convey what it is for, I think about prompt engineering.
exactly this. All of the well documented specs I wrote years ago are now fully coded apps after a copy and paste. What a time.