"Composition often allows optimization to be generalizable: If a system can optimize A and optimize B, then it can optimize programs that combine A and B" how true can this statement be in practice? I do understand that this can hold but if a program has A and B tightly combined then I feel like it isn't a given that a good system can optimize for both at the same time. But that probably shouldn't exclude it from being regarded as a good system...

Standout takeway from this slide for me was the last point. Engineers will always find new ways to use a well designed system that was never anticipated, like video detection for tensorFlow