“Truly important and significant hypotheses will be found to have “assumptions” that are wildly inaccurate descriptive representations of reality, and, in general, the more significant the theory, the more unrealistic the assumptions (in this sense). The reason is simple. A hypothesis is important if it “explains” much by little, that is, if it abstracts the common and critical elements from the mass of complex and detailed circumstances surrounding the phenomena to be explained and permits valid predictions on the basis of them alone. To be important, therefore, a hypothesis must be descriptively false in its assumptions; it takes account of, and accounts for, none of the many other attendant circumstances, since its very success shows them to be irrelevant for the phenomena to be explained.”
I imagine I'd send this as reply to a referee report which criticizes my work on the grounds that I might have found a great dark matter candidate, but not only have I assumed unbroken supersymmetry, my model also has four neutrino generations and, oh, only 2 spatial dimensions. Assumptions that indeed qualify as unrealistic and wildly inaccurate. For not to say, bluntly wrong.
But maybe I am being unfair.
Let me guess what Friedman might have wanted to say. The more parsimonious a model, the easier it is to extract relevant features and get an understanding of its behaviour. That does not mean however, it makes for a better model the fewer and more unrealistic assumptions you have. Certainly, the standard model of particle physics would be nicer if all fermions were massless and chiral symmetry was unbroken. Unfortunately, it doesn't describe Nature then. That's why we distinguish between models of the real world, and 'toy models' meant as testing ground to increase our understanding of the general features (see earlier post on Models and Theories).
But maybe also this is unfair.
He might have meant to say that a simplifying assumption does not have to be shown appropriate for a certain range of validity, the range in which predictions derived from the assumption then can be made. Instead, one can just see whether the model works and such justify the assumption a posteriori. Unfortunately, that too is nonsense. If you don't specify the range of validity of your assumptions (typically by showing that the effect of deviations from the assumptions is negligible for the result) your model is not falsifiable and thus not scientific. If you test it and the outcome does not match your predictions, you can just go and say, well, the assumptions were not fulfilled.
Thus, I am afraid unless you want to redefine what you mean with a scientific theory, this is not a good starting point. One wonders why he felt the need to put “explain” in quotation marks.
See also: Shut up and calculate