Ten days ago I had an experience with Gemini 3.8 flash that made me wonder if I was being routed to a different model under test. I was trying to use rocm with llama.cpp on my 128gb Strix Halo but could only get it to run Vulkan. I pasted the error message into agy and it proceeded to attach GDB to my GPU driver, reverse-engineer the kernel queue ioctl interface, and author an LD_PRELOAD C shim to get ROCm llama.cpp working on my Strix Halo. My jaw was hanging open the whole time.
Edit to add the fix: <a href="https://gist.github.com/birep/6f2c8d490c7a29820997d57bd654c351" rel="nofollow">https://gist.github.com/birep/6f2c8d490c7a29820997d57bd654c3...
This is why I think llms are a killer app for Linux desktop. They’ve been trained on Linux very hard, and it cleanly solves the “how do I make it do $thing” problem since everything is open and the llm can manipulate it. For example: Sound not working? Just tell the llm.
Is that specific to Linux? For better or worse, I have felt like it'd fairly good at working through most tech stacks I throw at it.
I have a client app on a very old (for the JS world) version of eleventy using NetlifyCMS (also outdated). Claude has quite easily picked that up to add features to it along the way.
Being able to edit and recompile pretty much any part of the OS and userland (often not even needing to reboot!) is not something that can be said about Windows for sure, or even lots of things on Macs too.
Arguably, it requires the LLM to have access to the source code during training. If it's closed off (like Windows), then the LLM can't train on it.
Even if the source code's old, the fact that it is publicly available makes it much easier to train & improve on than if it were walled off.
Not only training, you can literally clone the source code locally and ask the agent to debug your problem by grepping the source code itself. Even if doesn't "know" the answer it can investigate it on the fly
It's at least specific to open source, in my experience. Ask it about some commercial software: it searches some lacking documentation, social media discussions and make some guesses. Ask it about an open source tool: it searches the documentation, if not it dives into the implementation to figure out the edge case.
The more open the operating system the more LLMs can help you. This is what Arch Linux was waiting for all along.
I used to be afraid of Arch, because I don't want a system that takes work because I'm already busy with work. But now I love it, because the LLM can tweak every knob and fix every issue for me, so it ends up being the OS that takes the least work to use. Get an error message? Tell the LLM and they fix it. Something not working exactly like you like it? Tell the LLM and they tweak it for you. This also works to great effect on Win and Mac, but not to the same extreme degree as it does on Linux and especially Arch.
Hell yes, I also started to tweak my XFCE arch desktop to my liking.
Anything I want different now, I just tell the LLM to do it for me.
(For example I can now close lots of windows of the same type with 1 click, not 3, my whisker search now finds files and folders and I am able to run games that refused before)
I still ocasionally run into the usual linux driver issues, but not for much longer I suppose. I probably could fix some driver bugs now already if I point fable towards it and pay some attention.
In theory I could do all this before myself, but not just like that in some minutes, but in days/weeks/months ..
I’ve been playing with the ITS operating system on a simulated PDP-10. Despite it being from the 1960s, Claude has been able to do sysadmin work on it, including a very similar journey where it started disassembling code to identify the source of an obscure problem. I don’t want to take away from the learning experience (the whole reason I’m interested in this stuff), but it’s often almost impossible to find what I’m looking for online, and the LLM turns out to be a better resource when I want to learn how a particular subsystem works. I ended up making a custom MCP server so it could connect to the simulator more efficiently. <a href="https://github.com/masto/pidp10-mcp" rel="nofollow">https://github.com/masto/pidp10-mcp
Drifted off my point a bit, which I guess was meant to be it’s not necessarily Linux-specific training.
I got Claude to set up 5.1 Surround Streaming from my Linux Desktop to my Steam Deck using Moonlight. It needed some help (or really human ears) to make sure things were coming from the right speakers - but it was pretty much a one-shot.
I used Claude to build an entire runbook for how to build my Linux desktop up from scratch to its current state, including all my installed software, configs, hacks, workarounds, etc. Then I used it to migrate from one laptop to another in an afternoon. Now it keeps itself up to date.
Yes this might be a bit too noob linux but I have the AIs maintaining an env.md + logs / updates on that md whenever they change anything about the env. Even things as simple as getting tmux to be fast, they do an excellent job at.
Drivers, Coding environment setup etc. are great too and it's nice to have everything logged so the next (more powerful) agent can come and improve the thing once in a while.
Nice idea, then they want to make a quick little kernel module, oops, you secure booted, this is not your system, can't load...
Since LLMs have been so successful at finding exploits it's been clearer than ever that the people so obsessed with enshittifying every system with ineffective (other than pissing you off and wasting your time) "secure boot" functionality were really just too ignorant to succeed at actual novel security work, so they focused on this make believe crap. Well, glad that's over.
I used Claude & Gemini extensively to troubleshoot Linux since I switched last year, and it's been mostly great. But then it nearly bricked my install last week and I had to get some human help. I was getting black screens during and after boot and tried many suggestions from multiple AI's. After many hours it turns out I just needed to power cycle my monitor. Shout out to Claude for suggesting that (its memory remembered that I have a dock built in to my primary display).
I am still very happy with my switch to Linux. But if I didn't have the AI help I would say linux is still unacceptable platform for those not willing, able, and excited to get their hands very dirty.
Agreed 100%! I got some Steam games that didn't run well (or at all) to run smoothly thanks to AI, after having spent hours trying to figure it out on my own (months prior), fix some issues with HiDPI and external monitor with my laptop, and now I have a tiny shell script that basically runs `checkupdates` (`checkupdates | claude -p`, roughly) and looks online for potential issues. I've been using Linux for many years but still Claude and ChatGPT helped me a ton.
taylorfinley · · focus · HN ↗
Edit to add the fix: <a href="https://gist.github.com/birep/6f2c8d490c7a29820997d57bd654c351" rel="nofollow">https://gist.github.com/birep/6f2c8d490c7a29820997d57bd654c3...
plasticchris · · focus · HN ↗
_heimdall · · focus · HN ↗
I have a client app on a very old (for the JS world) version of eleventy using NetlifyCMS (also outdated). Claude has quite easily picked that up to add features to it along the way.
jubilanti · · focus · HN ↗
Being able to edit and recompile pretty much any part of the OS and userland (often not even needing to reboot!) is not something that can be said about Windows for sure, or even lots of things on Macs too.
x-complexity · · focus · HN ↗
Even if the source code's old, the fact that it is publicly available makes it much easier to train & improve on than if it were walled off.
augusto-moura · · focus · HN ↗
matsemann · · focus · HN ↗
Sammi · · focus · HN ↗
I used to be afraid of Arch, because I don't want a system that takes work because I'm already busy with work. But now I love it, because the LLM can tweak every knob and fix every issue for me, so it ends up being the OS that takes the least work to use. Get an error message? Tell the LLM and they fix it. Something not working exactly like you like it? Tell the LLM and they tweak it for you. This also works to great effect on Win and Mac, but not to the same extreme degree as it does on Linux and especially Arch.
lukan · · focus · HN ↗
Anything I want different now, I just tell the LLM to do it for me.
(For example I can now close lots of windows of the same type with 1 click, not 3, my whisker search now finds files and folders and I am able to run games that refused before)
I still ocasionally run into the usual linux driver issues, but not for much longer I suppose. I probably could fix some driver bugs now already if I point fable towards it and pay some attention.
In theory I could do all this before myself, but not just like that in some minutes, but in days/weeks/months ..
Sammi · · focus · HN ↗
masto · · focus · HN ↗
Drifted off my point a bit, which I guess was meant to be it’s not necessarily Linux-specific training.
c0n5pir4cy · · focus · HN ↗
chinathrow · · focus · HN ↗
oldandboring · · focus · HN ↗
safog · · focus · HN ↗
Drivers, Coding environment setup etc. are great too and it's nice to have everything logged so the next (more powerful) agent can come and improve the thing once in a while.
UpsideDownRide · · focus · HN ↗
stefan_ · · focus · HN ↗
Since LLMs have been so successful at finding exploits it's been clearer than ever that the people so obsessed with enshittifying every system with ineffective (other than pissing you off and wasting your time) "secure boot" functionality were really just too ignorant to succeed at actual novel security work, so they focused on this make believe crap. Well, glad that's over.
samspot · · focus · HN ↗
I am still very happy with my switch to Linux. But if I didn't have the AI help I would say linux is still unacceptable platform for those not willing, able, and excited to get their hands very dirty.
tarokun-io · · focus · HN ↗