The supposed fix? Switching from 144hz on my DisplayPort to 120hz. Around the same time I also did the latest round of updates so I genuinely have no idea if it's a displayport issue with the AMD driver or some rogue bug in an update that hadn't accounted for a rare hardware/software combo. Either way it fucking works now and I'm not updating anything until I have another issue or there's a major security vulnerability. Back to being a rock solid system with zero issues.
I don't know precisely what hardware you run, but I'm 99% sure it was an issue with the kernel+mesa implementation on Debian. The display stack churns stupidly fast on Linux, such that there can be substantial differences between minor version numbers. For example, Linux kernel version 6.18 on Debian Trixie forcing GCN 1.x cards (re: the HD7970) to use the legacy xf86-video-ati driver vs. Kernel 6.19-onward finally mainlining AMDGPU support for them. Their Mesa implementation (re: the libraries that provide OpenCL, OpenGL, and Vulkan) also falls behind the curve as part of the release cadence.
AMD and Valve work stupidly close with the kernel, Mesa, and WineHQ/Codeweavers (in Valve's case) team. HDMI and DisplayPort 2.x were huge sore spots on Linux until Valve set their heart out on the Steam Deck, the Steam Machines, and all that stuff. So much so that there's actual progress being made despite the standards bodies categorically rejecting Linux support like a decade ago.
It's also completely turned me around on AI as google basically helped troubleshoot my issue by scouring some of the most niche linux forums for information and answers and I finally feel comfortable with messier troubleshooting that's not just "erase everything and start afresh"
Don't play
too fast and loose with the LLM whispering. Using AI like Claude, ChatGPT, Gemini, and so on is cool until you start venturing into the world of troubleshooting something that involves doxing text files full of UUIDs, private keys, and so on. Frontier models like the aforementioned are really fucking good at pointing out "oh hey, it's this problem from 2015 that's fucking with ya. Here's how you solve it." That help comes at a price: your data. All frontier models actively train off the data you feed into it, and you have no discernible guarantee they're honouring the promises they made about user privacy and data security. The models all say more or less say "don't feed confidential information here, you stupid fucking dipshit." You also never know who on the server side is trawling through logs for shits and giggles, and then happens across your LAN IP with your WireGuard keys. The odds are low, but never zero.
There's also the matter of LLMs giving you the answer without the rationale behind the answer. ChatGPT sees something's borked, it tells me "oi m8 ur script broke at line 21. Put in a closing bracket, y dontcha?" but it won't tell me
why the closing bracket is important, let alone the breakdown of the file syntax I'm working with. It's
painfully easy to get complacent when the LLM in question gives you an answer that works
in that moment, but might bite you in the ass 3-4 months down the line. Frontier models also have hard usage limits, and you
never wanna be in a multi-hour troubleshooting session where you're waiting for ChatGPT or Claude to refresh your tokens so you can finally get to the bottom of whatever obscure problem you're facing.
Local LLMs help mitigate that to a certain extent; God only knows how ubiquitous Ollama is nowadays. Yet Qwen2.5/3, Llama3.1, Gemma4, DeepSeek-R1, and so on are
not panaceas unto themselves. You're constrained by parameter count, not to mention you're also at the mercy of whether or not ROCM will play nicely with your GPU. Llama3.1 can
probably help you troubleshoot some obscure system failure... but you need to consistently interrogate Llama (or any other self-hosted model) to make
damn sure that they ain't goofing on you (which they will... far more often than the documentation will ever admit). You'd think that interrogating an LLM is easy, and it "logically" is, but we've been collectively spoiled by 3-4 years of ChatGPT and its consequences. There's a huge psychological incentive to just take the answer at face value without pressing any further. That's why people like You Suck at Programming say that LLMs are basically the latest iteration of the "for you" page problem: you get the answer from the LLM, maybe have search enabled and click through one or two links, but that's where your research ends.
