> they produce incoherent sentences with no relation to the text they are supposed to extrapolate

I think we’re agreeing here?

Language models of the size of SD produce literally useless output. SD can produce kind of fun output that people can make use of sometimes?