So, I’ve heard that ML manipulates tokens and specifically for the English corpora they take place of words. If we want model to be polite and not to speak uncomfortable language we can remove certain words from the internal array where all tokens and their associative data are stored, for example “fuck”.
I believe that the 3rd or nth, word is because it sounds more human. The statistically first correct word ends up sounding very robotic and forced, where the 3rd is still very likely correct, but leads to variation in responses
This is all from what I remember reading a mini-paper about it, so I could be wrong