'Superintelligence,' Ten Years On

A "superintelligence" will develop by itself (and this is a gigantic project): inference and machine learning will have to be merged into a realtime model with realtime inputs and outputs which should be anchored in some way with our physical world.

If a sentient being does emerge, it will be probably after a really long time of interactions (years?), and better have a neural net with some induced "cleanup/stabilizing" regimes for weights (aka sleep).

Ok, let's presume you get your sentient being? Then what? Don't forget switching it off will be murder and fooling around with its neural net weights may be even worse.