1. The authors only show an association between phenotype (autism) and a rare genetic variant. They do not prove any kind of adaptive function to that variant, or use any method to infer adaptive function of that variant. Thus, this isn't finding positive selection, it is finding an association. Associationed locis effect on phenotype can be positive, negative or completely null. All an association is, is a sign post pointing the way to something interesting. It doesn't mean the associated SNP itself is functional.
2. The PGC ASD dataset does not have the power to detect associations at genome-wide base-pair resolution. They had 16,000 individuals, but used a genotyping array. This is great work by PGC and an extremely important dataset, but it is best suited to finding larger associations with disease (spanning thousands of bases), not individual associated causitive nucleotides. It is for this reason that I don't trust claims of an adaptive single-nucleotide loci in the absence of any functional studies or adaptive inference.
3. Nobody in genetics takes GO Term enrichment at face value. It is an extremely noisy analysis that often turns up false positives. It sometimes points you in the right direction, but it really isn't a good analysis to hinge your paper on. This is usually interpreted among colleagues as grasping blindly for a meaningful result.
4. The autism-SNPs->intelligence link is just wholly speculative (edited for clarity).