Say you have a population of 10,000. You can come to some meaningful statistical conclusions about that group.

If you consciously pick 100 individuals from that original population you also have a group, it's true. But you can't use the conclusions from the first group on the second, much smaller, group.

You can discover some new statically valid facts on the new group -- they may even be similar to the original statistics -- but that is far from a given.

The only thing you can reasonably say is that the new group will fall somewhere in the domain of the original measurements.

Unless, of course, the new group is chosen at random or semi-random in regards to the measured characteristics, in which case you can have some sort of expectation for the new group and even measure (statistically) the amount of randomness.

But when it comes to the question at hand: ability to handle stress, high achievement potential, leadership talent -- these are all things I would assume Google chooses non-randomly from the overall general population pool. So there is no reason to think the relevant characteristics of the small sub set should show any similarity to the general population at all for these characteristics especially when the differences in the general population are relatively small.