Brett Victor's Realtalk is proposing a vision which is opposed to the one put forward by the "current trend of deep-learning models trained on large datasets". I found this section of the FAQ fascinating, and the whole thing is very good, in general. A fascinating project.