I don't think text generation is realistic at this point, nor will it be in the near future. It's hard enough for a human to figure out the right thing to say to a child, let alone design an algorithm to do that. When I read Stephenson's book I thought that good speech generation was a much smaller problem to overcome (Indeed, it was dropped for the version of the Primer given to the cast-off girls) than content generation. I remember feeling that, were the Primer built today, the on-demand talent required would be guided improvisational interaction rather than mere acting.
However, I could see a Watson like search engine being a possible alternative. The child could carry around a monitoring device that would seed the search with the events of the child's day. If the child struggles in class, gets beat up by a bully, etc. the engine could then set programming from an existing library of stories, movies, etc. that is immediately relevant to what the child is experiencing. Entertainment for children is often designed to teach lessons. This would just time the lessons to have maximum impact. Teaching children is a bit like training dogs in that timing is massively important. Catch a dog in the act of peeing on the rug and you can have a far greater impact than if you don't notice the mess until after it's dried!
AI cannot presently compete with humans when it comes to content generation, but the above device could be with a child night and day throughout their schooling, which is something few, if any, parents can do.
