The Inconsistency in Scaling Laws for Data Set Size

The loss as a function of data set size was a parallel plus constant offset. And so the L shape you're referring to, the bend in the curve on this log log plot is a result of that parallel approaching what we're calling the irreducible loss. The smaller models will plateau at larger values of the loss because they're constrained by how many parameters they have and then as bigger and bigger models will have lower and lower loss.