In poker, if a pot is very large, calling a bet (ie investing some money) can have positive expectation even if the probability of winning the pot is quite low.
Similarly, YC's filtering ability only needs to be "good enough" to generate positive expectation, which can happen even if their predictive ability for selecting winners is quite low. Picking winners at a slightly better rate than the competition gives them an edge, even if they're not actually good at it. ie being "barely competent" at something extraordinarily difficult can often be enough to achieve considerable success.
Of course, I don't see any reason to believe that they aren't actually quite good at choosing winning startup teams -- it would certainly be interesting to be able to measure that. It will also be interesting to see if their hit-rate goes up over time.
The only way to measure YC's ability in picking winners is to to have PG & co rate a bunch of non-YC founding teams at the early stage of their startups, wait a few years, and see how well their predictions held up against more or less random predictions. The experiment must involve non YC startups is that YC enhances their startups' chances of success, which can create the likely false impression that YC excels at picking successful teams rather than at guiding the teams it happens to pick to success.