I'm not sure it moves the needle on NLU/classification tasks very much, compared to models with many fewer parameters. But it does seem to make the NLG better,

I'm not sure it moves the needle on NLU/classification tasks very much, compared to models with many fewer parameters. But it does seem to make the NLG better, which is what Microsoft seems obsessed with lately.