all genes on X is a meaningless statement unless all genes are involved in X... if you read the paper the authors seemingly put quite a bit of effort into the rigor of their model.

they used multiple validation cohorts, they controlled for environmental factors, they controlled for non-autosomal factors, applied multiple methods to boost the predictive performance of their model. adding more snps typically leads to an overfit and poor performing model on validation cohorts, which is why they use so much rigor and end up with usually 100-400 snps for each of the phenotypes.

in anycase they found a lot more than 10 SNPs, and they did more than one analysis in the paper.

Parent educational status is a significantly better predictor than the polygenic risk score for educational attainment. heritability (also known as r^2) for cognitive performance they estimated at an even lower ~7-10%.