They're looking through a huge number of genetic variants for genes that each have a small effect on expected intelligence scores. Two obvious problems that the

They're looking through a huge number of genetic variants for genes that each have a small effect on expected intelligence scores. Two obvious problems that they run into: 1. When you test that many genetic variants, some correlations will appear purely by chance. To correct for this, they require very strong evidence for each variant that they declare significant. This means they end up having a lot of false negatives, and it's only through having an enormous sample size that they manage to get results this good. 2. There are a lot of genetic variants that they're not even looking at. (They try to pick a decent subset to measure.) So they end up with some set of extensively-vetted genetic variants that, combined, predict a fairly small percent of variation. This is what you'd expect to see if there were a lot more that they weren't able to detect, which would be consistent with the much larger heritability numbers from things like twin studies.