Hacker News Reader

learning by trial and error will never generalize well

What is the alternative?

Expert systems with a myriad of rules that fail the moment they get an input that was not considered?

ehem, evolution