Weird I haven't had a graphics issue since like the 8000 series of cards for like, the normal driving graphics right off the card use case. Optimus is definitely a little fucky with multiple GPUs and sometimes new CUDA versions break libs if you're keeping it bleeding edge updated, but most ML workflows use environment management to get around the clusterfuck that is the python ML ANN ecosystem anyway. Nvidia does a lot of work to make it hard to work with their cards but since GPGPU started taking off in the 2010s this sounds like a blast from the past, as a lot of the need for getting good performance out of cards on compute servers - which inevitably run linux because it has any kind of reasonable headless userspace and is easier to write reliable software for - has driven better driver support in general
In discussions between people who don't use linux, their notions of what the pain points even are seem to lag by sometimes over a decade
I find myself equally ill-informed on the pitfalls of windows ecosystems of late, except the ones that make the news for how insane they sound, but in my defense I also don't go around claiming I intermittently try to use it but just can't make the jump
As microsoft continues to tighten the screws, I just hope all these reluctant users can get out before it's too late