I think his point is that it's impossible to normalise such a sample. For example if I sampled only toddlers about their attitudes to ice cream then there's no way to extrapolate from that sample to the general population, and it's the same thing for "people who are willing to fill in a survey for $2".