You are right, there has to be limits to the cases it can learn, but imho it doesn't need to learn anything exhaustively ( probably impossible anyway), it only needs to learn enough to convince us that it has developed at least as much common sense as us. The process and the object of learning is surely to progressively remove what we don't know, and what we don't know is marked by uncertainty, so the object of learning is to eliminate uncertainty, and it's power is in it's immense scaleability.

Deliberately putting uncertainty in, so as to cut down on resources taken up by learning, seems to me to be a backward step. The power of Ai is surely that we don't care so much about how efficiently it learns, we are much more interested in the accuracy of its learning.

From an information theory point of view, we can't get around there being a fundamental minimum amount of resources needed, to learn something completely, nothing taken out. The technique of reducing system complexity by reducing fidelity of information looks a little like the lossy compression techniques used to compress musical data in ways we can't hear the difference. It works for music, but I don't think it would be good for learning :)

If the concern has anything to do with energy consumption, folk should forget this, given that in a scenario of 100% solar power, it doesnt matter how much solar we use, we can't use enough to "Steal" a significant amount from nature, all we can use of it is beneficial, the more the better, because our use of it alone has the effect of lowering temperature thermodynamically. This is the opposite of what happened with energy extracted from Earth.

So why not max out the resources available to Ai, pour it on I say, ensure it is all solar powered,, and whilst we are at it, sack all the presidents of every country, and put Ai in charge of all. It will soon learn the ropes, I am sure :)