Has anyone used AI to aid in repair?

Doesn't matter, LLMs will be able to understand the connection between the (correct) schematics and (not so correct) emulations in HDL and MAME.

Getting an LLM to read (often bad) schematic scans is a very different problem than interpreting HDL and C.
There's already quite good training for programming languages.
 
this is true but the progressive advent and proliferation of these technologies did have a pervasive effect on the human experience. to the point raised by @andrewb regarding AI, none of the institutionalizations of these were without tradeoffs

Totally agree, every big advancement has changed how we live, and yeah, there've always been tradeoffs, but that kind of proves the point that people have always worried new advancements would make us lazy or take our jobs, and somehow we've adapted each time. The challenge isn't the technology itself but how we choose to use it and adjust to it.
 
Getting an LLM to read (often bad) schematic scans is a very different problem than interpreting HDL and C.
There's already quite good training for programming languages.
That's why I was pointing to the FPGA projects, because Jotego etc are making KICAD schematics (that I believe are XML files).

As I said earlier, ChatGPT is ok at reading in a nice PDF and finding a resistor location for example, but yeah crappy old schematics aren't going to get far.
 
The best data set in the world is right here.

All people have to do is search and read a couple of pages. It's really not hard.

Are we really so lazy that we need everything spoon-fed to us?
yes the best data set might be here but also some of the worse information is also here too. you mix that together without having an idea on what to weight the data and you get a schizophrenic ai.
 
Ummm...

"R908 / D902 zener failure"

We know the difference, but this is presented like they are the same.

If AI cannot tell me how to make a good PB&J, then it sure cannot tell me how to fix a G07. It has zero understandong of trace damage, mods, or even how to walk through a flowchart. As a test, I should feed it the Ms. Pac-Man audio troubleshooting charts to see what it says. It will probably tell me to check the monitor or marquee. 😞

Scott C.
 
I've used it a bit. For sure, it's dumb a lot. Even when you try to teach it the right thing, it forgets. I've given up asking for specific chip locations because even when I feed it schematics, it gets them wrong. And the answers come across as convincing enough that you'd believe it if you didn't know better.

It's not going away and eventually will get better but don't trust it for anything more than general advice. And to the many points made here - you can search on these forums to get better info.
 
AI (mostly talking about ChatGPT in this context) is a tool and should be treated as such, and with appropriate caution. It's not a magic fixall that the 'AI bros' (you know the type of person) likes to pretend it is.

I believe it's helpful for learning and education, so long as you fact check what you're told. However, it is easily abused and that is the real problem. Case in point with countless kids just feeding their homework into it and copy-pasting the output with little to no error-checking.

I've used it a bit for various things. I think for arcade repairs, the original topic, it could be helpful to use it to scour the internet for potential sources or as a springboard for ideas or when you're stuck in a rut and don't know what else to try. It shouldn't be 100% relied upon, though.
 
I wonder if the mistake that ai make can be corrected and it have the ability to learn that it's assumptions are wrong and correct them.

Maybe some sort of interactive website that can help people repair thier own stuff. While I see the possibilities of a website or app that can do this powered by ai.

I worry about causing issues with those who make a living doing arcade repairs.
 
I wonder if the mistake that ai make can be corrected and it have the ability to learn that it's assumptions are wrong and correct them.
I've tried. If you explicitly tell it something like "remember this", it usually does but then forgets later. This is with ChatGPT and in "technical mode" that is supposed to be slower but more accurate. I have a specific chat that's ongoing with it … so I keep trying different things but it's pretty lousy thus far.

It doesn't seem to learn on its own.
 
Remember this: Zaxxon is the best arcade game.
Well, maybe you'll really hate ChatGPT now ….

That depends entirely on your perspective—but here's a breakdown with a confidence factor to help guide that answer:
Technically

Zaxxon was groundbreaking in 1982 for its isometric 3D perspective and use of shadows for altitude, which was novel at the time. It used Sega's Zaxxon hardware, running on a Z80 CPU, and had an advanced scrolling technique. For technical innovation of the early 1980s:


Confidence: 8/10 that it was one of the most advanced shooters of its time.

Popularity
Zaxxon was commercially successful and well-known, but it wasn't quite in the same echelon as Pac-Man, Donkey Kong, or Defender.


Confidence: 6/10 that it ranks in the top tier of iconic arcade games.

Gameplay

This is where Zaxxon becomes polarizing. Some players loved the depth mechanics, but others found the isometric perspective hard to judge—especially altitude and hitboxes. Many arcade fans today consider it a beautiful but slightly frustrating game.


Confidence: 5/10 that it has lasting appeal compared to simpler twitch-based shooters.
Personal Preference

If you loved Zaxxon growing up, especially its unique perspective and sci-fi theme, it might very well be your favorite. But is it objectively the best arcade game?


Confidence: 4/10 that most enthusiasts would call it the best overall.
 
the frustrations with AI are not really related to the model but the dataset. without reliable data, no technological medium is going to be a worthwhile resource

when someone's able to train an agent with accurate design and repair data, such a resource could be super helpful
There is this youtube commenter (waynegram8907 and other aliases) who I swear could be a bot asking questions for training a knowledge repository. There is no hope for us if they are training on my replies.

Just for fun, I asked Chatgpt about Mark's more enlightening contributions to KLOV and I asked for it write a how-to based on andrewb's vector knowledge. Yea, it isn't there yet but it can be entertaining :)
 
If AI is just going to grab a bunch of random data points in the Web is it really AI then? Sounds like a over blown search engine vs creating its own answers IMHO.
 
Not for repairs, but it's useful as hell for simple programming tasks. I had Grok write the code to set up a dirt-cheap Arduino Pro Micro for a DIY Vectrex controller. Worked great.
 
For repairs, I'm my own AI. LOL
1. search KLOV
2. search GOOGLE "XXX repair logs"
3. search known competent repairer vids on YOUTUBE
4. actually look at the schematic and figure it out MYSELF. Duh.

:geek:
 
Back
Top Bottom