The lack of "depth" from AI systems is not too hard to understand. Consider some ML network that describes some "problem space". Note that this network does linear algebra to transform input data into some output, and the model parameters are then updated via minimizing some error function or performing derivative calculation.<p>This is not dissimilar to taking a Taylor series and moving in the direction of most/least change to find some local extrema of a curve (gradient descent). That is, these models can explore some local problem space well, and find extremas around the initial location. And it seems unlikely that some arbitrary deep learning network will ever be able to solve some high dimensional problem space (you need a ton of dimensions for AGI given the problems in different domains are often described by different parameters) that describes a more complicated problem .