I think the quote is good, regardless of the actual source. It is some of a wordplay, because, as you say, simpler than possible is impossible, whereas trying to make an explanation even simpler when it is already the simplest possible is indeed possible. And that is the point, I think. One shall not sacrifice accuracy for simplicity. The best explanation is the simplest one among the most accurate ones.
My last formulations above are a bit approximate. As a Machine Learning student, I learn that model overfitting is a major concern, that is fitting a model very well to known observations, while losing performance on unseen data. ML models usually have a lot of free parameters. They are needed for representational power, and are also very useful for finding paths toward acceptable local optima. In ML research (and statistics in general), one will typically hold out subsets of the data for validation, to monitor overfitting and stop training before severe overfitting harms generalization.
The point is that sacrificing some accuracy for simplicity may be right sometimes, e.g. when the amount of data is insufficient for benefitting from hold-out validation.
Update: Here is a recommended link in response to the comment from Frank Hubeny: https://en.wikipedia.org/wiki/Overfitting
And here is an introduction to a basic validation scheme: https://en.wikipedia.org/wiki/Training,_validation,_and_test_sets