In the q&a with Guy Steele at the end he says one thing he would do differently is integrate transducers more deeply into the language from the beginning, but the concept emerged around the time of core async when they were wary of rewriting map filter reduce etc again for channels (rough paraphrase from memory, watched this a few days ago).
I wonder what that would look like, what core functions would use transducers that do not, and how.
