I'm not really talking about companies here for the most part, I'm talking about academic ML researchers (or industry researchers whose role is primarily academ

I'm not really talking about companies here for the most part, I'm talking about academic ML researchers (or industry researchers whose role is primarily academic-style research). In companies there is more incentive for good software engineering practices. I'm also speaking from personal experience: I once had to hand-write my own CUDA kernels (on official NVIDIA cards, not even this weird translation layer): it was useful and I figured it out, but everything was constantly breaking at first. It was a drag on productivity and more importantly, it made it too difficult for other people to run my code (which means they are less likely to cite my work).