If your process can be "make a tonne of low-precision instances then screen for accuracy" you can get surprisingly good results in general, as long as the accur

If your process can be "make a tonne of low-precision instances then screen for accuracy" you can get surprisingly good results in general, as long as the accuracy of each instance is relatively uncorrelated. It's a really handy pattern to have in your back pocket.