Creating a motivation system for a seed AI that remains reliably safe and beneficial under recursive self-improvement even as the system grows into a mature superintelligence is a tall order, especially if we must get the solution right on the first attempt.
