You're correct about their error; however, Hinton views that a sufficiently scaled up autocompletion would be forced, in a loose mathematical sense, to understand things logically and analytically, because the only way approach 0 error rate on the output is to actually learn the problem and not imitate the answer. It's an interesting issue and there are different views on this.