be smarter about sampling?

[nsajko]

Floating-point numbers are denser for lesser magnitudes. This means that a lot of effort may be wasted by the current implementation, sampling uniformly, for the parts of the chosen interval that are greater in magnitude. Alternatively, the denser parts of the interval might not be covered well enough.

[nsajko]

I suppose a smarter approach here would be especially helpful for narrower types, such as Float16.