He makes a lot of points here, but the most important IMO is that many measurements are noisy, and the correlation between X and Y+noise is less than the correlation between X and Y.