[robpal]

Nvidia GPU drivers and CUDA installations on Linux. Got better recently but I've spent many nights fixing these installs, wrestling with nouveau etc.

Compilation of deep learning libraries from source on Windows -- don't do it.

Tikz -- LaTeX drawing "language". The images produced can be stuninngly beatiful, but the manual has 1200+ pages and it takes forever to become productive. If only journals accepted hand-drawn images...

[mod]

Can second this.

Also audio libraries on linux. ALSA vs whatever. Had a lot of issues with my particular hardware & Arch.

Also way back when, X itself. Back when it was likely to not work on many systems (~2001). I spent so many hours going back and forth to use my friend's computer & internet to Altavista some help.

[accraze]

lol +1 for audio libraries / drivers on linux. spent many hours on the "battle of configuration".

[k4ch0w]

Absolutely felt your pain too. So many nights trying to build the latest tensorflow package only to find out I couldn't run it on my current Cuda version, so I need to uninstall it, reinstall the new one only to find out I installed 10.0, then 10.1 but I needed 9. It was just so silly I ended up just never updating once it worked.

[tetris11]

To quote Linus: "Fuck Nvidia"

Their proprietary drivers suck and I much prefer using the default Linux ones... but the performance just isn't the same and I always end up relenting and installing their terrible software

[wakkaflokka]

This, this, this. I needed to run some code that was only compatible with an older version of tensorflow. Figuring out how to get the proper Nvidia drivers, CUDA version, etc. was an absolute and utter nightmare.