Mostly holes - part 2: Educating SITA

It takes a whole village to bring up children.
It might take the whole internet to bring up an artificial intelligence.
=> therefore the name Synthetic Intelligence Taught by All.

The goal

The goal in the end should be to have a system, that can utilize a big knowledge base, give correct answers using it and as a glitter on the horizon be creative about it.

... and what you get when you start

Like a small child, the system might just start to repeat repeat the primitive sentences, you have just taught him when it thinks, it might be appropriate.
That is just about what the semantic web has accomplished for now. Links to word definitions in other statements - like in Wikipedia.

It would be nice, when some more knowing responses could be given.

Cheating

It might be OK - for a start - to pass questions, that are too complicated, to real persons working in the background, who answer them in the same manner as SITA would have done.
The system would watch closely and might answer the same or similar question the next time on its own...
And over time less and less human support would be needed... provided an enormous amount of wishful thinking.
Return of invest: You could start out with a (quite slow) system, that answers your questions - and it might be an overwhelming surprise when you get your first answer that no human has laid its hands on... and these cases would increase (hopefully) exponentially.

Showing SITA the world

A first approach might be a tagging of persons and things in pictures of news sites.
The tags would define the area inside the picture and sort the tag-description into SITA’s grammar along with a little sentence that describes the scene and hopefully contains all used tags.
Return of invest: You might get good tagged pictures since the described parts of the picture are pinned and the tags are set in context with synonyms and generalizations.

The Furby effect 

Next to a visual understanding of real life things comes an auditive interaction to make up something that might pass as human-like.
Return of invest: When a user trains the system, it has to his dialect and voice into account.
Over time it would get better at understanding and repeating these dialects so (after long years) a textual chat might be read to you in the voice of the other person, since their voice representation etc. are available in the cloud.