Machines got more powerful and switched from CPUs (central processing units) to GPUs (graphics processing units)/ Concurrently the different types of software used in AI applications (dozens, from chaos-theory applications to graphics processors to Net search-engine algorithms for selecting preferences, to exotic programming nobody outside the AI field would recognize) were eclectically able to unite the sum of the parts into a whole, a functional digital neural-network. 

This is an achievement in architecture as well as design of hardware or software because they had to create a large body of synthetic neurons by using a separate computer for each neuron - so the network has an unbelievable number of CPUs/GPUs utilizing a vast architecture of elements and controls both hardware and software in nature. 

The precise structure of the system may soon be impossible to chart in exact detail (or may already be) because self-design/self-reproduction are part of AI program design, especially in more-powerful, more-capable computers. By the time technicians were able to graphically represent the structure, it would likely have already changed. 

The danger in this alone is that it might invoke a Technological Singularity in which the machines increase the improvement-rate until the systems become so complex they would be literally incomprehensible to human technicians/analysts. 

At that point, machines might begin to disregard goals set by humans altogether and have motivations we could not predict. They might be capable of helping to solve a lot of human problems but have no interest in doing so.  

The machines might become incapable of explaining their own motives in human terms. This in similar case if human beings tried to explain our motivations to trout in a pond. 

They might be so technically capable they could open new avenues of physics but have no interest in doing so or do so but be unable (or unwilling) to translate that understanding into something Humanity could utilize.