In chapter five of Jaynes, "Queer Uses for Probability Theory," he explains that although a claimed telepath tested 25.8 standard deviations away from chance guessing, that isn't the probability we should assign to the hypothesis that she's actually a telepath, because there are many simpler hypotheses that fit the data (for instance, various forms of cheating).

This example is instructive when using Pascal's Wager to minimax expected utility. Pascal's Wager is a losing bet for a Christian, because even though expecting positive infinity utility with infitesimal probability seems like a good bet, there are many likelier ways of getting negative infinity utility from that choice. Doing what you can to promote a friendly singularity can still be called "Pascal's Wager" because it's betting on a very good outcome with a low probability, but the low probability is so many orders of magnitude better than Christianity's that it's actually a rather good bet.

Obviously, you don't want to let wishful thinking guide your epistemology, but I don't think that's what PI's talking about.