[hansvm]

Steelmanning their argument, an ideal output would consist of every possible float in [0,1] occurring with an appropriate probability -- some of those vanishingly unlikely. Kind of like how randint() should be capable of returning every possible integer rather than, say, every multiple of 64.

It's fine to counter that most of the time the difference wouldn't matter or that it might be problematic to compute for some reason or another, but I don't think it's reasonable to critique the idea on the grounds of the result not being uniform when they didn't actually ask for each float to have equal probability.