To start, this was a great read! The more I understand the space, I feel we are still far away from super intelligence. AI is yet to get even pronouns right — http://whatsnext.nuance.com/in-the-labs/winograd-schema-challenge-2016-results/.
If I were to guess I would predict a general intelligence may actually appear much earlier than we anticipate. This general intelligence can apply itself to a huge variety of problems and may or may not be better than humans. This is kind of an extension of Deepmind’s approach to using RL for playing a variety of Atari games. I believe this should only be a matter of throwing an extensive neural net (several magnitudes larger than what we have seen yet) and encoding the right kind of redundancy structures (like Dropout). This would be a fundamental departure from today’s focus of developing ML solutions for each application.
For this general intelligence to get to super intelligence, it needs as many or more opportunities to learn as humans get in the real world. For some problems, the internet provides an avenue, for others there may be no obvious methods. Which is why I feel it will take a good load of time for AI to be qualitatively better than humans at every aspect.