Also in my model intelligence is a property of layers, defined by how well they can simulate other layers using an API available to them (think about brain simulating physical interaction - lower level or simulating how the society might develop in the future - higher level). The more intelligent, the better the understanding of the surrounding reality. (I use predictive power as a measure of effectiveness of a theory / software that is instantiated on the higher level API). But if we build up artificial layers starting at an arbitrarily selected level - creating a different (unnatural) kind of feedback loops (species), can they simulate as broad range of layers as humans can (be as intelligent / think as much out of the box)? At which level do we need to start, to get close to what natural species are capable of? How much more complicated does the (AI) software needs to be to have equivalent predictive power considering it runs on hardware that provides a very restrictive version of the set of instructions natural species utilize?

(hardware refers to to the lower API a layer uses, software are feedback loops of the layer)