So far, end to end seems to be the only way to train complex AI systems that actually works.

Every time you pit the sheer violent force of end to end backpropagation against compartmentalization and lines drawn by humans, at a sufficient scale, backpropagation gets its win.