That's the one. Taleb essentially recommends that we use real physics-based statistical models while "avoiding econometrics and Gaussian-style non-scalable statistics like the plague."

This is completely a fresh way of looking at the problem, and one that I had been developing the past few years. I knew kind of what was in the Black Swan book, but it wasn't until I finished reading it in the last couple of days that I completely understood Taleb's mind-set. Most everyone that has referenced the book have drawn different conclusions about its deeper meaning. This is partly because of the way the book is written -- very personal, blunt, and also allegorical. This book is so blunt in fact, that if it was written as a series of blog posts on TOD, they might actually get deleted. No one is immune to the criticism of NNT :)

Mucho thanks KLR. Not sure if what I picked up was Web's point but what caught my eye was his position about folks using models as predictors and not just a descriptive mechanism. I've debated others here in that regards. To me all models are correct (as long as the construction is proper). For me a model is just a formula with all the assumptions filled in. One can argue if the assumptions are realistic. Regardless the model does provide an "correct" answer if designed properly. That doesn't mean that correct answer is a prediction. Just that, for this set of assumptions, the model is correct. Where I see folks make huge mistakes is offering such models as predictions. For me the value of a model is its ability to show a system's sensitivity to the various components. A model that shows the great sensitivity of oil prices to global economic activity is one example. Another would be the relatively low sensitivity oil prices might have relative to new production from the Bakken Shale. IMHO where folks stub their toes is when they make predictions of the future. My argument might sound subtle to some but it explains why I don't care to debate the validity of predictive models. For the most part they are not close to being reliable even when I think the model is constructed properly. And then, as Web points out, to make matters worse folks will use the conversation of the model as spin to support their proposed accuracy of their prediction. And the spin suddenly offers more credibility then the "model" deserves.