That paper does not imply what you think it implies: that editing all these variants in an individual embryo would have a small effect.
An individual with the top 200 IQ-enhancing variants edited in (and an otherwise normal genome) would be far, far smarter than than than you seem to think, even though the variants edited only explain 5.2% of the differences within the population.
Since they have an average effect size of ~0.05-0.1 IQ points or so, a few hundred of those would in fact be many SDs. You can download the PGS, grab the top few hundred or thousand betas, and sum them if you don't believe me. (I show this with Benyamin et al 2014 in that section above, but no reason one couldn't redo it with a better PGS; the results won't change much.)
And yes, Hsu really does believe this, he is quite explicit in that 2014 paper that that is how it would work (in addition to countless places on his blog & Twitter) and I've discussed it with him in the past.
(As other comments point out, the real problem with this logic is that the hits are usually not causal but tag the causal variant, even if you had some magic CRISPR which could make hundreds of arbitrary edits. That's soluble, but not easily. It is, however, not a problem for selection procedures.)