Deutsch's point (made in greater length in the book) is that predictions are lower level than the true target of science- explanations- not that they aren't valuable. One of the main ways to test explanations is to get predictions from them, and then check out the predictions, and getting too many predictions wrong is fatal for an explanation. Your example of "interesting" predictions highlights his point: the explanation of how the trick work can readily generate a prediction of what would happen if the cups were transparent, but the prediction that the cups would later be empty does not readily generate a prediction of what would happen if the cups were transparent. By focusing directly on explanations, he makes it obvious which predictions are the interesting ones. Indeed, I'd even speculate that someone who didn't have and couldn't acquire the concept of explanations would have trouble grasping the idea that some predictions are more 'interesting' than others and that there's a reliable way to determine which predictions those are.