For me, Wolfram Alpha has always just worked on the examples given by them, but when I try my own inputs, it fails on almost all of them. Even though my inputs are just variants of the same queries.
Yeah, that's part of the reason why I said that "we'll be able to solve a greater number of problems by 'defin[ing] your goal, and that's about it'". Progress will happen at a steady rate, but the problem space will expand even faster.