Photometric redshifts have been with us for two thirds of a century, although they have become scientifically useful only in the last ~15 years. But what drives their precision and reliability and is the state-of-the-art good enough for the surveys of the future? We present some photo-z theory, present a bias-free, empirical KDE approach, discuss their application in large surveys such as RCS2 or DES, requirements for the future and where the OzDES-deep survey could help.