Can We Just Keep Putting in Bigger Data Sets and See Better Performance?

For most of the tasks that we include in this paper, in my opinion, we have basically saturated them. While we definitely did see gains from making the model bigger, it's not clear how meaningful additional gains would be on these tasks. So there's actually a lot of really excellent work in the n l p community right now trying design harder and harder tasks than are difficult for models like this to solve.