Depends on the field kinda. For a really extreme case, something like big dense linear algebra is going to just use BLAS calls anyway. For big enough matrices, all of the flops are coming from the library, the cost of calling from Python vs the cost of calling from C or Fortran is amortized anyway, and probably most people won’t beat a tuned BLAS however much developer time they throw at it.
It makes more sense to tune libraries excessively, we could say it is less of a trade off, more of an allocation of the finite low-level tuning developer resources to high-impact libraries.
Anyway it turns out that the only way to get most people to link to a good BLAS is to use Numpy and distribute it through some giant Anaconda Rube Goldberg machine, so, I dunno, people are weird.
It makes more sense to tune libraries excessively, we could say it is less of a trade off, more of an allocation of the finite low-level tuning developer resources to high-impact libraries.
Anyway it turns out that the only way to get most people to link to a good BLAS is to use Numpy and distribute it through some giant Anaconda Rube Goldberg machine, so, I dunno, people are weird.