I was reading an article on massive open online courses (MOOCs) ("MOOCs Assessed, Modestly"), and struck by the final quote:
“In a regular Stanford class, if 2 of 100 students got something like that wrong, we wouldn’t even notice it,” [Andrew] Ng said. “But when 2,000 out of 100,000” do, it’s immediately evident. “It’s ironic that in order to achieve personalization at the level of telling students exactly what their misconception is, what was needed was to teach massive amounts of students.”
It's not ironic, it's exactly why we're expanding genetic studies to include hundreds of thousands of subjects. A complex system can fail in many ways, most of which will be rare. Finding rare causes requires giant samples. But what I love most about this Coursera example is that they figured out a way to flag the error as students make it, so that they can learn at the moment when they might make the mistake. Following students through the system, on a massive scale, gives a new way to improve learning.