How could you achieve object permanence this way? Will it 'automatically' appear given more training data or more hidden layers? How is this handled in other approaches?
To achieve object permanence, there needs to be some persistent off-screen data from frame to frame. There's a way to achieve this: train on frame+input+woldstate -> frame+woldstate
Presumably if you were to take that and build a large enough NN to accommodate all the necessary state it needs to carry and all the rules it needs to be able to execute, then after training it on enough game input you'd have a proper world simulation. Of course, as the article rightly notes, then you have just successfully reimplemented Minecraft in a way that is orders of magnitude more computationally expensive...