Is anyone excited to do ablative testing on it?

With such a high throughput because of sparsity, I'm particulary interested in distilling it into other architectures. I'd like to try a recurrent transformer when I have the time