I wonder if we will reach insane quality when we will be able to allow models to generate not exact same text but something similar (thus allowing for more compression). I think math leans a lot to this type of training as one can super easily verify if generated result is fine or not.
We probably should do similar thing for language when other model judges if generation is ok or not (but doesn't require exact generation). This probably can be done as fine tuning? Anyone know if this has been tested already?