This is certainly the way things are headed, and they're smart to clarify that vision, but I feel like XLA needs better support for sparsity in order for this t

This is certainly the way things are headed, and they're smart to clarify that vision, but I feel like XLA needs better support for sparsity in order for this to really take off. And I reckon Google knows it, too. Everybody talks a big game about more efficient models but the tooling around sparse NNs is way behind that for dense ones