Over the past 100 years social science has generated a tremendous number of theories of individual and collective human behavior. In general, however, it has not produced anything resembling a "core" of cohesive and consistent theoretical knowledge that is both generally agreed upon and also empirically validated. There are many reasons for this state of affairs, including that social phenomena are almost always complex and multiscale; that social data have historically been extremely hard to collect at scale; and that social experiments have been difficult or impossible to conduct. Recently the combination of vastly increased computing power, "big data," and "virtual labs"---aka computational social science---has generated considerable excitement on the grounds that the lifting of these historical barriers may lead to a revolution in social science. In this talk, however, I claim that social science is also hampered by its traditional emphasis on advancing theoretical (or methodological) frameworks rather than solving practical problems. To be truly revolutionary, I argue, computational social science should reverse this traditional importance ranking, starting first with a problem and then asking what theories (and methods) must be brought to bear to solve it. I then offer a few suggestions regarding which problems we should pick and how we should organize ourselves to solve them.