I haven't read the Sandberg, Drexler, & Ord paper, but from what I've read, the most important point is that people should appreciate that averages (expected values) are not always the best piece of information to extract from a probability distribution.

I wish Steve would have explained this on the show-- he made it sound technical, but I think the general concept is accessible and important.

Suppose scientists can agree on a probability distribution for each of the parameters, and each parameter is independent. The usual approach to the Drake equation is to find the expected value for the number of civilizations by multiplying the expected value of each parameter. This is totally legit, but the expected value just tells us how the average number of civilizations if we simulated the universe many, many times.

The average can be relatively large, even if most simulations have zero or one civilizations (e.g., if a few rare cases yield a plethora of civilizations). The probability distribution would be right skewed, with a fat tail representing these rare outcomes of lots of civilizations. I think the central point made by the paper is that we should consider the probability of <=1 civilization, not the expected number of civilizations.

To tweak Mr Beagle's example: It's possible that a lottery has a positive expected value of a pay-out, even if there is a 99.99% chance of losing (of course, nobody would run such a lottery). Even though the expected value tells me I should expect to make money by playing, I don't think anyone would call it a "paradox" if I didn't win money after playing once.