Ever since Mendel, genotypetophenotype (GP) mapping has been the defining feature of genetics. The complete GPmap for a trait provides the expected phenotype (genotype value) for all possible combinations of alleles across all genes affecting it. Thus, instead of looking at the effect of every allele averaged across all genetic backgrounds (the marginal effect), the GP-map provides the phenotypic effect of each unique allele combination. If the joint effect of two or more loci departs from simply adding up the marginal effects at each locus, a simple additive model will fall short in predicting the phenotypic effects revealed in the GP-map. Geneticists have for many years debated whether such non-additive patterns are of importance, or if additive models are enough to describe the genetic architecture of a studied trait. The perhaps most important piece of information needed to resolve this debate, i.e. what the true multilocus GPmaps that are modeled actually look like, is however largely missing. Here, we use a large experimental yeast population to perform an extensive, empirical estimation of highorder GPmaps affecting a large number of quantitative traits. Using these as a basis, we illustrate how the estimates obtained from statistical quantitative genetic models will depend on various features of the underlying GPmaps. Specifically, we show that a large additive genetic variance does not necessarily imply that genetic interactions is of little importance, thereby illustrating how variance component analyses can be missleading when making inferences about the genetic architecture of complex traits. We also show how additiveonly genetic models can lead to poor predictions of individual phenotypes.