Reliability. We can do pretty much anything… with a 5% success rate. Deep learning can take any input, approximate any function and generate the required output, but it’s only as good as the training set and most of them suck. Or it needs to be so large and complex that it’s not fast enough.