[newhouseb]

I went to one of the SIGGRAPH talks they did a couple years ago.

The theoretical plan is basically to write an optimizer that could intuit good schedules, recognizing that they actually don't have a good idea of how to do this. I think they can currently run some ML algorithms that churn through a bunch of different schedules and find the one most fit for a problem but it's rather brute force and slow at the moment.

That said, the conceptual distinction of separating the algorithm from the scheduling also allows you to tune scheduling by hand much easier than would be possible otherwise.