[chimeracoder]

> I don't disagree with the math, but wouldn't it be the case that it's easier to get a biased sample from 1000/100M people than 1000/1M people?

It's not the population size that matters; it's the variance in the underlying population.

Mathematically, the sample size needed to construct an n% confidence interval for a normal distribution is independent of the population size (the first derivative is zero).