Inertia is part of the answer to the first question. In some industrial areas, such as paper making, NN are replacing human production managers. Typically these areas are ones where some old guy with years of experience make the dozens of adjustments to keep the product quality up. In the paper production plant there is the pH, the moisture content at dozens of stations in the production line, the temperature at these stations also, and a dozen other factors I do not know about (perhaps he even periodically tastes the mass.)
Much the same is true of the wine and beer industries. The "master" may die, taking his not even verbalizable knowledge to the grave. Thus many of these complex industries where quality control is an art, are switching to NN networks, but one drawback of the NN is that there is only the data set of the current production runs for the NN to learn on - I.e. it may take more than 10 years with all these factors measured hourly by instruments (and some may not be known to even measure) with information about the quality of the beer, paper, etc. being produced before there exists a "training set of data" which will let the NN even match the judgment of that old master's years of experience.
Human intelligence took more than a million years to develop. The NN will probably take a few hundred to equal it. But then intelligence will rapidly advance as they can significantly improve the next generation - in contrast to humans where thousands of generations were required to make even a one point increase in the average IQ.
Im not argueing that it will be impossible for neural networks to equal humans.
The design of them are very innovative, but are vastly complicated to construct in a simulation let alone in real life. Of course, thats really not a limiting factor because hey, if we were willing to build a machine to take us to the moon, we will be willing to make a somewhat adaptive computer.
But the problem is that it is not equal to a human's general ability to make adaptive decisions, when you think about it, yes a NN would probably be better in several fields that would be occupied by a human, but the problem is, that is all that NN would be good at, put it in the driver seat of a car and it wont even know what is going on.
They would be great in a specialized task, but the second the parameters change, the NN would have to adapt, and the problem is that it adapts based on patterns, if something is out of place once, it adjusts maybe a fraction of lets say a millimeter (lets say it has to do with attaching parts to a....car I suppose), but the distance needed to move is a foot, it could potentially mess up several times before it hits the right place. It's called trial and error.
The problem is, that it won't always get a second chance. ie, in space, if something changes the parameters, messing up once could end it.
And that's if it even knows what is the problem in the first place.
NN is going to be a great invention, but it will still need a problem library, it won't need the answers necessarily, but it will need to know for things like outer space that if hypothetically the space ship keeps rolling at a rapid rate, that it is a problem with the valve of the thruster.
