The main issue here is that the Python buffer protocol understands all the other common scalar types we use (even float16)... but not bfloat16, so trying to do anything interesting with it will likely fail.

This isn't a huge blocker for anyone (AFAIK), since there is very little code outside of Tensorflow/TFLite that uses bfloat16, but closing the loop on this would be nice. Opening this as a tracking issue.