Hacker News Reader

Comment by thinkloop | Hacker News Reader

AI fixes Linux on the desktop. Whatever obscure issues you’re facing, you’re a quick prompt away from the solution.

Was not the case when I wanted to use a GameCube controller via Wii U adaptor without a lot of extra lag. Yeah that's a niche, but it works fine in Windows and even Mac.

I believe most of those work with controller drivers in the application (Dolphin Emulator or Steam/SDL) rather than the OS level. That's why the Windows solution requires Zadig to replace the HID driver.

On Linux instead of replacing the driver, you have to add an udev rule that allows applications to communicate with the USB device directly: https://github.com/ValveSoftware/steam-devices/blob/master/6... And you can see in this list, it's not the only controller with that requirement.

SteamOS includes this by default.

I did add that rule, which if I remember correctly was needed for it to work at all. But then the problem is input lag. Many recommend https://github.com/hannesmann/gcadapter-oc-kmod which the author only tested in Arch, and for whatever reason it didn't work on Ubuntu. Was printing that "Failed to acquire lock for USB device" line and then I think "Could not reset device", don't remember. Anyway seems like a hack.

This was for Slippi Melee, so even though I'm not super good at the game, the lag was too annoying.

Whilst I tend to agree, I also don't recommend just pasting every command from ChatGPT into your machine without having some understanding/validating process.

Best flow is

Llm -> manual research -> apply

But no one has the time to validate everything llm writes via manually researching it unfortunately /s

LLMs regularly give wrong advice for debugging and fixing Linux stacks, especially desktop stacks.

Which LLMs? I've had good success with Claude Opus >=4.5

Older/smaller models were far worse

In my experience AI is unreliable more often than not. It is conflating topics, uses outdated information or straight out hallucinates. It can be good if you already know enough to call it out on its bullshit.

The paid plans which enable "thinking" and internet research are far better than the free tiers and open weight models

e.x. ChatGPT-5.4 Thinking, Claude Opus 4.6

Completely false. The models are nothing like they were a few months ago. Try codex 5.4

Assuming you actually take the time to retroactively understand what you just did and don't become a slop-ministrator.

Reply on news.ycombinator.com