People enjoy solving puzzles. In real life, solving puzzles is useful, for instance recognising a camouflaged predator. But this is not a sufficient reason to be rewarded (emotionally) for solving the puzzle, rather than just for the outcome.
In a complex world, with a favourable outcome depending on many factors (including chance), learning only from ultimate successes and failures is limited. It is better to reward actions that lead to success – solving puzzles, hard work, learning, etc. It is even better to develop ones judgement and taste, i.e., to know what to reward by how much.
While computers today have many wonderful reasoning powers – being the best chess players and mastermind quizzers, they seem to lack the taste and judgement that will lead to deeper discovery. But one can argue that this is because we forgot to put these qualities in them. While automated reasoning functions do have reward functions, and even compete for survival, computers are not set up to decide, or even fine tune, what deserves a reward. A good way to introduce such taste may be to mimic the interplay between our own cognitive and emotional systems, and their relation to the world outside. Many computers may simply become tetris addicts, but perhaps a few deep thinkers will emerge.