Interesting read - along a similar vein of thought is the stuff that Taleb looks at in the technical counterpart to his 'Incerto' (it's available online as a free pdf) - in particular how finite variance distributions don't quite cut the mustard when being fit to financial time series.

One more thing - when comparing the MLE estimate of the mean for normally distributed data you mention that it's 'more efficient' than the sample mean. Is this the case? I thought this is a classic case where the MLE *is* the sample mean - so they're equally efficient.