Do you expect the ability to reason about geometry would emerge, or would it be more like pattern matching limited by the training data problem space?

Meta: testing in progress