The closest thing may be the work going on in "active learning" where the learning process is self directed, the agent adapts to limitations inherent in the data, and employs multiple techniques as needed, especially interactive Q&A to resolve unlabeled / ambiguous exemplars.

Until an agent takes some sort of initiative in directing its learning process, its behavior will remain scripted and passive, never amounting to more than a recognizer of patterns -- a rather low bar in the quest for autonomous intelligence, IMHO.

I'd be happy with a framework which could decide: this is a point in the free-form conversation to answer a question or interject some information.

I think when we finally start building hierarchies, when NNs "go meta", that's when we'll start seeing truly mind-boggling results.

There can be no AGI without hierarchical, "meta" NNs.

By most definitions, people have been building hierarchical NNs of many different kinds for years.

this is a great idea. then maybe we can put an executive-executive layer above the executive layer. then we would be getting somewhere