BTW, "filters" as in, "filter assisted decoding" is actually really helpful and AWESOME for fixing some of the problems with ChatGPT at writing poetry or writing lipograms (text with correct english but where you omit a letter systematically). I wrote a whole peer reviewed paper about this actually:
So, when we call this "filters", it's more that it's doing "content filtering", because there doesn't appear to be the kind of token level filtering that I describe in this paper going on with ChatGPT.
https://paperswithcode.com/paper/most-language-models-can-be...
So, when we call this "filters", it's more that it's doing "content filtering", because there doesn't appear to be the kind of token level filtering that I describe in this paper going on with ChatGPT.