Can We Get Superhuman Performance Out of Smaller Language Models?

Chinchilla out performs the human base line on, i think it was like 80 or 90 % of the tasks that it was given. You don't actually even have to give the answer correctly once you're hen, you're doing the tromting. And then when it answers, it's much more likely to give you an answer that's kind of in the right form. Thats a fun result that came up a couple of months ago. It didn't matter that you, youe said the wrong thing. I believe if i rememer in corect ly had definitely one shot, zero shot. All these things are garbage. There's no what are those two terms? What