>>105582530
>Type-1 people read SLOW
Now that's something I haven't heard about, but makes sense in the sense you're replaying a movie.
Ideally all responses you get are perfect, but (inb4 skill issues) at times you feel something is shit or off, or at least need to make a little edit. Are you really expecting everyone to slow read every generation from the start?
When the response is a banger, usually 500+ tokens "story"/story stuff rather than <100 tokens RP chat (might reread the log in the end in this case), I'm more than willing to read the passage fully for the second time.
Since there are less words than tokens, 6 T/s will give users a slight buffer zone over 5 T/s in case they aren't speedreading but aren't slowreading either. When I'm reading a visual novel, I don't want to be at the tail end of the words popping in. I want all text to appear, then I'll just read it without the distraction of the animation of words appearing.
20+ T/s stuff is not for pure creative reading, but to be able to skim outputs and decide, maybe you're seeing how high you can push Temperature before it goes a little weird, maybe a user is interested in the ideas the model will take on swipes, or other purposes you may label as "ADHD and/or writing slop". Biggest thing for me if I'm trying to direct the response; I may need to adjust my inputs a few times and swipe to get something that feels nice.