But then again, I don't see much gain there, even if the rephrasing is removed, it may use the same amount of processing / energy as before for the same result.

Plus or minus the fact that the models are not updated, so it's not better.

Same for going trough everything, but it can be implemented like this as a test :

"if answer match our criteria, show a special toggle to ask user if answer was accurate". => anonymise, save in db. Rince and repeat.

And after a while, criteria => db => rephrase if needed => serve to user.

Kinda, but the main pain point is to say that the content produced by the AI is technically correct and must be said correct multiple times across multiple users.

I agree it's a privacy concern, so it can be used for basic questions with basic answers.

Agreed. But in term of efficiency of generation I don't know if the trade is better between "search what we know you answered well and just say it again in a huge db" or "generate again with all the water usage we saw"

Une guerre avec la guerre ?

What for is a trillion dollar company?

What's the point when nobody can eat ?

Elle a un soucis avec la transparence de la démocratie ?

Yeah ? But with more control over the css

Late