I've tried. And usually the questions I ask are too specific. I mean I can answer the basic questions myself and often I get several result when it's just that. The AI just mumbles general advice and is always wrong if it's too specific. Like for example: Why does the graphics driver crap out on any OpenGL ES instruction on the old single board computer I have lying around, despite the SoC being supported?

[–] tsonfeir@lemm.ee 4 points 7 months ago* (last edited 7 months ago) (1 children)

… is that your prompt?! No wonder you’re getting crap. It’s a computer program, you have to feed it valid data. There is a huge misconception that you can CHAT with it… which was introduced by calling it “ChatGPT,” a horrible name.

You need model numbers, OS versions, driver versions, and any other relevant information like error messages, screenshots, and code if you are developing.

[–] rufus@discuss.tchncs.de 1 points 7 months ago* (last edited 7 months ago)

No? I left out the detailed info here as I thought it's of no concern. I provided it with pretty much the same info I'd write to the Linux Kernel Mailing List. With computer bugs that's usually steps to reproduce the issue, exact versions of everything, exact error messages and my findings from googling and looking at the code...

That was one of the issues I had that only gave me one or two search results and it's unlikely that someone comes up with a solution since the hardware is outdated and not many people have that specific board lying around and also the expertise to understand the low level hardware coding involved.

I mean it kind of fits the rest of the picture I have from using ChatGPT and similar stuff. It can do easy stuff. And write boilerplate code pretty alright. With the Arduino code I'm tinkering around as a hobby... not so much. I once asked it to do the inverse kinematics for a small robotics project. And the AI can tell me about what I just read on the Wikipedia article about that topic. But that's it. Not an idea how to apply that info. And that the complicated part is to come up with the specific Jacobian matrix. And not just tell me that using one is one of the few approaches to that problem. That's obvious from reading the Wikipedia article or reading any textbook. And it did silly things like write code like equation.solve(parameter1, parameter2, parameter3) ... Sure. I mean if I already had a framework that did that and was available on an embedded platform, I wouldn't have had that problem in the first place...

So my attempts at using AI for the issues I have with computers regularly fail. I can see how that's not the experience everyone has, but still... It doesn't really help me with specific problems or rare issues.

And I still have a few I can try to question some AI about... An slow Wireguard VPN tunnel inside if another tunnel that I already fixed the MTU and it's still unbearably slow... A few obscure webframeworks that don't tie into things... But I'm pretty sure I'll get the same results.

Have you ever been lucky with AI and issues that didn't get you any search results because no one ever did it before? I mean I'd be happy to learn how to use AI properly as a tool. It's just I've tried and I don't think I'm too stupid to prompt it. It's just that I've given up since it doesn't seem nowhere near intelligent enough to tackle the real issues I have. I'm not opposed to AI. I use it and it helps me get small stuff done easier/faster.