And much more information when selecting just as many examples from quadrillions of randomly generated examples.

The information from the selection criteria isn't available to the model, just the chosen samples.