Fast and Easy Infinitely Wide Networks with Neural Tangents

Can someone explain the implications of this for performance, or what we can do now that we couldn’t do before?

It really looks more and more like JAX is the internal winner after the tf 2.0 fiasco