A bias in the input may exist because of a bias -- perceived or actual -- in the filter.
Again, I think one of the points of the article is that post-input selection problems can cause selection problems in the input.
As a thought experiment, if we imagine a situation in which a particular minority were to believe that, even if they followed the rules of their majority peers, they would still be selected against, then we might also imagine that as a consequence they would actually select against themselves.
(For example: I'm terrible at casual party-like get-togethers, so I tend not to go to them, which in turn prevents me from getting any better at them; I'm selecting against myself in this situation because of an expected problem. Likewise, if I were a woman and an entrepreneur, I might not attend certain events because I believed in advance that I wouldn't do well there.)
I'm not at all arguing that this is actually the case. But, I don't think that limiting ourselves to looking at problems in education is taking a complete enough approach to the overall problem.
Entities should not be multiplied divergently. A sufficient explanation for the fact that Y Combinator winners are mostly white is the fact that YC applicants are mostly white. As shown by the similar Ruby-enthusiast pool, which has clearly been filtered by the same phenomenon (ie, the vast racist conspiracy - known to some as "The Plan.")
Therefore, it is unnecessary to postulate the additional cause: that YC itself is in on "The Plan." Alas, when tinfoil hats go mainstream, Occam himself is powerless.
Yet another related thought: PG once quoted Caesar as not trusting thin men. This idea may have scientific basis: dietary fats stimulate oxytocin production. http://paleohacks.com/questions/56261/dietary-fats-reduce-ca... http://althouse.blogspot.com/2010/01/people-instinctively-kn...
This stuff is related because it's important to ask whether a meritocracy should necessarily discount the single-generation advantages of genes.
Disclaimer: I am skinny and have the AG variant.
(The point of that link, for anyone who didn't get it, is that the lack of diversity Eric Ries perceives in the output of our filter is also present in the input, which implies it's not caused by bias in the filter.)