Ah, I was arguing BLAS et. al. are effectively non-reimplementable (I'm aware of ATLAS and OpenBLAS, but I'd argue they demonstrate that it's a rare thing to be

Ah, I was arguing BLAS et. al. are effectively non-reimplementable (I'm aware of ATLAS and OpenBLAS, but I'd argue they demonstrate that it's a rare thing to be able to do). I would expect if people aren't willing to use existing tools like conda, then they understand they're going to need to put in the hard yards setting up a proper build environment. Presumably if people want to build everything from scratch, then they know how to set up a correct build environment, and know that they need trustworthy libraries which are numerically correct, rather than a library which hasn't been as rigorously vetted which in a easier-to-install language. If not, I'd be curious what they would do with security/cryptographic libraries, would they be happy with someone's custom SHA256 library that they'd knocked up to avoid needing to install openssl for example. I guess I expect people to treat numerical libraries in the same way as security/cryptographic libraries, as there's just as much subtlety in how those libraries need to be implemented.