2 comments

[ 5.5 ms ] story [ 27.8 ms ] thread
IIRC, AI models were found to be more polite in certain languages... Is it possible to have a politeness to cost ratio as another benchmark when checking language models?