[leiroigh]

Comparability between implementations: Say GPU vs CPU, or between languages, or to pseudocode in old papers.

Typical example where the badness of the floats bites you is if you do something like log(rand()), or 1/x, or more generally you map your uniform (0,1] interval via a cdf to generate a different distribution, and your mapping has a singularity at zero (this is like super standard -- e.g. how you generate exponentially distributed numbers. Or if you generate random vectors in an N-d ball, you're using a singular cdf to compute the magnitude your vector, multiplied with a normalized normal distributed N-d variable to generate the angular component).

Then the bad random floats (i.e. the non-smooth distribution close to zero) introduce real and visible bias after transformation. Afaiu the problem is well-known and serious people correct for it somewhere. If you fix the underlying problem without revisiting the now-obsolete fixes, then your results are biased again.

I'm not arguing for keeping the bad random float generation everywhere. I think it should be fixed, not just in julia but everywhere. I'm just saying that it's not a no-brainer and there is discussion and compat and communication and code audit involved in doing such a momentuous change.

Also, I'm not really up to putting in the work to champion that at the moment (arguing on the internet doesn't count ;)).