ChatGPT useful in designing ?

GroupDIY Audio Forum

Help Support GroupDIY Audio Forum:

This site may earn a commission from merchant affiliate links, including eBay, Amazon, and others.

Script

Well-known member
Joined
Nov 27, 2008
Messages
2,273
Location
Japan
I have been wondering how this language-based AI engine could be useful in designing anything.

Today I asked ChatGPT "How to troubleshoot my GSSL" and the answer was a quite comprehensive synopsis of basic troubleshooting -- overall good advise, although rather general and basic . But certainly helpful to beginners.

Also, it knew in answering a second question that the "Turbo board is an "add-on"(!) that needs establishing several wire connections.

And it explained how a 'tilt' filter in the sidechain can be used to alter the sound.

So obviously, I thought, ChatGPT is reading this forum.

However, since design decisions are often a trade-off between different parameters or goals (e.g. op amp gain factor versus noise etc etc), I wonder how far or deep this can go.

Of course, those chat answers are better when the questions are precise. Little-knowledge questions generate little-content answers. But what about high-knowledge questions? Does it generate anything useful?

Anyone around who tried it and is willing to share opinion or experience?
 

a guy I know on facebook has been asking chatgpt physics questions. I don't know if he got anything useful, I think it kept him amused for a while.

JR
======

Tony Boutwell

Today, I created a new blog/website to delve into an intriguing quantum theory that I've been exploring with the help of AI. While there's just one post so far, every journey starts with a single step.
🙂
#ai #quantumphysics #physics #chatgpt
 
For those of you who use or at least are familiarized with stack overflow (and stack exchange), it seems that it is getting abandoned due to ChatGPT. More info here.

1690287146906.jpeg
 
I guess the biggest problem is that ChittyChatGPT can't handle schematics, and it does not have access to a circuit simulator.

Also, I just tried it, it can't calculate properly, doh! ... and then apologized for the wrong result, LOL.

Still a long way...
 
I used chatGPT to select matching pairs in 30+ FETs I had measured. It took me quite a while to type all the data but after that it helped me in finding matching pairs. There was more than 4 parameters per FETs so my brain was melting. It did the matching in a few seconds. Very happy with the result. I guess a software could have done this, but it felt very convienent just to « speak » to the machine with common words and sentences. I had to explain a thing or two but it did well.
And I have tried and it’s able to « draw » some basics schematics. I’ve had it successfully found out pad values.
Obv when it comes to more complicated schematics it is not working (yet?)
But data sorting : perfect.
 
Last edited:
The biggest problem with chatgpt is that it will give you the wrong answer with confidence. It would be a much more useful tool if it expressed indecision, or confidence percentages etc. You sort of have to have the knowledge to tell it when it's wrong for anything that isn't searchable. As a search engine it's pretty fantastic
 
I used chatGPT to select matching pairs in 30+ FETs I had measured. It took me quite a while to type all the data but after that it helped me in finding matching pairs. There was more than 4 parameters per FETs so my brain was melting. It did the matching in a few seconds. Very happy with the result. I guess a software could have done this, but it felt very convienent just to « speak » to the machine with common words and sentences. I had to explain a thing or two but it did well.
And I have tried and it’s able to « draw » some basics schematics. I’ve had it successfully found out pad values.
Obv when it comes to more complicated schematics it is not working (yet?)
But data sorting : perfect.
Couldn’t an excel spreadsheet organize this?
Mike
 
I've found it quite useful in pointing you in at least an interesting direction....When it comes to maths, I find it very strange.
It seems to be able to rationalise equations - at least simple ones - successfully, but then totally cocks up the calculations it makes based on its own conclusions!
It will then apologise when corrected - and promptly do the calculaton wrong again!.
Its basic calculator functions seem pretty poor..... Probably better at suggestions on how to approach a particular problem, rather than actually solving it accurately.
As mentoned above -- it has a way to go yet....
 
I have been bouncing a few ideas off it. Yes, some interesting directions -- not sure any of it would work, but the bot never tires of emphasising that experimentation, evaluation and reiteration with real components is in order.

For searching textbook knowledge, it seems quite comprehensive.
____
Heard that schools are faced with how to deal with it. And as students seem to use it anyhow, it could be a "tool to foster critical thinking" is a description that I like.
 
Also I assume it can't do math on purpose simply because high-school students use it for assignments ..

No, it's just that the language model has not been taught logic. It technically cannot even count. But it can tell you how counting works, even give you a tutorial in swahili pirate slang.

I have been a software engineer forever. Recently I asked it to do my job for me. With decent instruction it would write perfectly good solutions to the regular tasks we have in our current area of engineering - especially with additional guidance. Sometimes it hallucinates and "invents" things, but that's what code review is for. Its really no different to a junior engineer in that way. It has already changed software engineering landscape completely. Just that most engineers have not yet realized this.
 
No, it's just that the language model has not been taught logic.
Exactly right. It is trained on what people actually write about, so it can come up with good solutions provided enough people have written about the exact problem before. It is not always good at inferring why people write about something (e.g., the underlying mechanisms behind the words).

Strangely it does a good job on the underlying structures behind procedural programming, likely because so many people have written about it. I love it for code snippets: rather than combing through Stack Overflow, I can ask something like "In Python, given me an example on how I can loop over individual lines in a file, performing a regular expression looking for groups of 3 words separated by colons" and it will give me a snippet that does exactly that over 90% of the time without even having to modify it. It even does pretty good follow-ups like "What if I want to handle all of the common file exceptions?" and it will give you a template of code to work from tailored directly to what you are doing.

More esoteric programming it does poorly, like interrupt handlers for device drivers, etc.
 
I have a friend who oversees a design group of engineers and has numerous patents to his name. He told me that as others have said it can write perfect English composition but can come up with absolute nonsense. Also it writes small chunks of code in a very proper form better than a lot of engineers. He said the problem is when your working on a secret development for a project, it then has access to your idea depending on how much you share with it. He said he sees companies actually having there own closed systems made up of a massive amount of GPU’s in order to protect their Intellectual properties. This will create a great deal more energy use if many companies go down this path.
 
I've played around with ChatGPT quite a bit, and I must say, the depth of its answers varies based on how detailed and high-knowledge your question is. Anyway, I tried to create some scripts for UX design, and it's been a fail.

Instead, I am using the Special Character AI image program that generates some great templates based on precise and detailed prompts I give to it in a few minutes. You can see how it will work for your projects.
 
Last edited:
Back
Top