Is There a Synergistic Relationship Between Application Development and Hardware Development?

I think one of the reasons is that the algorithms themselves are pretty, you know, are pretty complicated. So my intuition is that I think designing custom chips for doing some of those things is more difficult in some ways than tensor operations. We're going to work with the PyTorch community to make sure there's a well-defined interface between the hardware and thePyTorch runtime so we can expose our hardware to PyTorch users without it being intrusive to them. That vision has been significantly more realized in the machine learning community, but it's not yet something that's possible or widely available.