I don’t remember clearly when I got interested in psychology, but I’m sure it was related to accidentally discovering this book called Mathematical Biophysics by Nicholas Rashevsky who had published a journal... for many years, and one of the important papers in his journal was the paper by McCulloch and Pitts on logical... how... how logical modules, little ands... devices that computed ands and ors and nots, could compute more complicated functions of stimuli and produce responses and so forth.

So that was a general theory, and general theories are often easier to construct than specific theories, so this said how you could make any machine, but it didn’t give details.  And I got more interested in, how can I make a particular machine that would learn from experience and do some of the things that intelligent animals can do? So I worked out some ideas about how one kind of neural network could generate different possibilities. The simplest idea would be to try things at random, but that’s rather uninteresting in some ways, it’s much harder to experiment with, and another alternative is to systematically generate different possibilities, and then that could be more efficient because you don’t waste time duplicating things, and on the other hand it might be inefficient, because it makes too many meaningless... explores too many meaningless possibilities.