 The biggest thing that bothers me, as I look around, is that discoveries that don't turn out to hold up. Discoveries that aren't reliable, that aren't replicated when you look at new data. It's easy to find things in the data, but you can torture data until it confesses. And this happens more often than we'd like, especially in extraordinarily important data like medical research. Even the optimists in the medical field believe that half of the results that are published in journals, and again, journals only publish in some cases 5% of what's presented to them. So even cherry picking the results, half of what is published is probably not true. They just don't know which half. And pessimists believe that perhaps 95% of what's published is not true. So there's a lot of money wasted on medical research and we can't trust most of the results that appear. It's very, very disturbing. And half of the problem can be solved by doing data analysis better. And we've developed some techniques to do that. The problems were very well described in a cover article for The Economist magazine a few years ago on how science goes wrong. Very, very well written article. I highly recommend it's available on the web. And half the problems there were scientific in nature. We've developed new methods, pretty much rediscovering old methods, but methods that haven't been in use to solve that half of the problem. And the other half is more the business and the economic incentives of science where people are rewarded for discovering new things and they're not rewarded so much for confirming or verifying old discoveries. And that part of the problem needs more attention by journals and also making people supply the data that backs up their findings so that others can confirm it. This is still not enforced in a lot of places even where it's supposed to be. When I taught at the University of Virginia, which I do occasionally, one of my homework assignments for the graduate students was to find a favorite research paper that they had and ask the authors for the data. And over the semester only one out of my 20 students was able to get the data from a paper that they were found. So it's still not very well enforced for data to be available for other people to verify the work. And this needs to improve. So the medical science is probably one of the most important and one of the most egregious examples where reliability is not there. But in all of our work this reliability issue is paramount because the most important place for the model that you've found you've found some purported relationship between something you can measure and something you want to predict. The most important place for that to work is on new data you haven't seen before. Otherwise you've wasted your time.