How Might One Address Domain Generality

One obvious way to address the domain generality problem is learning.
Turing mentions it himself.
One popular subfield of AI is Machine Learning (Gates says they're now interchangable).
You give the system some initial knowledge.
It then learns the other knowledge.
This would be particularly important in a conversation, so the system can refer to things it's learned from the conversation.
Unfortunately, modern machine learning tends to concentrate on learning rather simple things.
My pull on this is to build a system that is based on neurons, and exists in an environment.
This way we get grounded symbols, and can take advantage of known neural and psychological problems.
CABot3 is not going to pass the Turing test.