Much of quantitative genetics is based on the "infinitesimal model", under which selection has a negligible effect on the genetic variance.  This is typically justified by assuming a very large number of loci with additive effects.  However, it applies even when genes interact, provided that the number of loci is large enough that selection on each of them is weak relative to random drift.  In the long term, directional selection will change allele frequencies, but even then, the effects of epistasis on the ultimate change in trait mean due to selection may be modest.  Stabilising selection can maintain many traits close to their optima, even when the underlying alleles are weakly selected. However, the number of traits that can be optimised is apparently limited to ~4Ne by the "drift load", which is hard to reconcile with the apparent complexity of many organisms, which arguably implies that selection maintains a very large number of traits. Just as for the mutation load, this limit can be evaded by a particular form of negative epistasis. A more robust limit is set by the variance in reproductive success. This suggests that selection accumulates information most efficiently in the infinitesimal regime, when selection on individual alleles is weak, and comparable with random drift. A review of evidence on selection strength suggests that while most variance in fitness may be due to alleles with large Subscript[N, e]s, substantial amounts of adaptation may be due to alleles in the infinitesimal regime, in which epistasis has modest effects.