I took a graduate level class in the 1990s from some SFI luminaries. It was a great class but the dismal conclusion was "this stuff is kind of neat but not very practical, traditional optimization techniques usually work better". None of us guessed if you could scale the networks up 1 million X or more they'd become magic.