[foothebar]

Just a further note: "AI" is a completely misleading name. No one (sane) in the "AI" business thinks of it as being intelligent. It's just automated statistics (it IS probabilistic) which can find patterns in data. So an An A' pattern won't fit. Also it usually has much less than a single hundred layers.

[peterweyand]

With current machine learning (neural networks as I understand them) I believe that there is no "intelligence". But I think it's possible to create an A^ pattern and my paper provides bounding around this area to make it possible. Whether it is computationally efficient to train an AI from a bottom up approach (np-incompleteness) or if it would return results that would be communicable between the AI and a human if it did I do not know. I need to make A^ more formal and find a specific example, but I think it may be possible to create a model using an image classifier (not a neural network as is currently done with say hand writing OCR).