ChatGPT to calculate DIY speakers

[sarumbear]

It’s very interesting to read all those posts about how ChatGPT is a “joke”. How it can never do various loudspeaker calculations, that I demonstrated it can, but they don’t, or cannot, demonstrate that it fails in those calculations.

Our subject is audio, and I used ChatGPT to do a straightforward speaker calculation. Why not all those naysayers show me how ChatGPT fails in that job with examples instead of on unrelated subjects?

 

[D!sco]

We have been talking about ChatGPT this entire time and it's practical applications, if you haven't noticed. The problem is extensive and more difficult to pin down than just "it isn't good at helping us with speaker design". When you really query it for detailed assistance, as we have, it falls short.
To be incredibly specific for you, I asked it to design a box for the new Purifi 10". It has numerous problems accomplishing this and while I haven't bothered to check every individual equation, it isn't supposed to have problems with this. In fact, I'll be using the fancy new GPT-4, which hallucinates noticeably less, and has a much larger pre-training library. This is a paid product. Let's see how it responds.

Seems kinda simple but so far so good.

I'm not even going to get into how it ignored my request to calculate them as if they were wired in series and sharing a box.
As we have previously discussed, ChatGPT will mirror what you give it. Here I am insisting that this isn't enough to be accurate and impressing upon it the importance of accuracy. Nonetheless, it soldiers on. It spits out some calculations, which are fundamentally incorrect. It's reporting less than half of the value it's supposed to be calculating with those values. It is also missing the port dimensions.

Again, it's not feeling comfortable giving precise answers. It's also asking for Rg, a not-well-known T/S parameter instead of Re or Z.

Cool that it can tell the port is stupidly large, let's see if it can adjust these parameters. It's a simple volume calculation after all...

Wow, it actually suggested a passive radiator. Frankly this is decent logic, but that's not what we asked it to do. We need it to make simple calculations. So, what does it think it accomplished?

Not a bad answer, let's see what it says about the ported enclosure

Okay, so these were calculated to the resonant frequency as a standard. Let's input these values into VCAD and see what it says.

Not even close. 30hz and 11hz respectively. It's not even ballpark, frankly. There is no way to construe this as an accurate or effective way to make 'simple speaker calculations'. It is often, and fundamentally, incorrect when taken for granted and fails in unexpected ways.

 

[GaryH]

/thread

 

[sarumbear]

We have been talking about ChatGPT this entire time and it's practical applications, if you haven't noticed. The problem is extensive and more difficult to pin down than just "it isn't good at helping us with speaker design". When you really query it for detailed assistance, as we have, it falls short.
To be incredibly specific for you, I asked it to design a box for the new Purifi 10". It has numerous problems accomplishing this and while I haven't bothered to check every individual equation, it isn't supposed to have problems with this. In fact, I'll be using the fancy new GPT-4, which hallucinates noticeably less, and has a much larger pre-training library. This is a paid product. Let's see how it responds.

View attachment 292517
Seems kinda simple but so far so good.
View attachment 292518
I'm not even going to get into how it ignored my request to calculate them as if they were wired in series and sharing a box.
As we have previously discussed, ChatGPT will mirror what you give it. Here I am insisting that this isn't enough to be accurate and impressing upon it the importance of accuracy. Nonetheless, it soldiers on. It spits out some calculations, which are fundamentally incorrect. It's reporting less than half of the value it's supposed to be calculating with those values. It is also missing the port dimensions.
View attachment 292522View attachment 292523
View attachment 292519
Again, it's not feeling comfortable giving precise answers. It's also asking for Rg, a not-well-known T/S parameter instead of Re or Z.View attachment 292524View attachment 292525
Cool that it can tell the port is stupidly large, let's see if it can adjust these parameters. It's a simple volume calculation after all...

View attachment 292526
Wow, it actually suggested a passive radiator. Frankly this is decent logic, but that's not what we asked it to do. We need it to make simple calculations. So, what does it think it accomplished?

View attachment 292527

Not a bad answer, let's see what it says about the ported enclosure
View attachment 292528
Okay, so these were calculated to the resonant frequency as a standard. Let's input these values into VCAD and see what it says.
View attachment 292529
View attachment 292530
Not even close. 30hz and 11hz respectively. It's not even ballpark, frankly. There is no way to construe this as an accurate or effective way to make 'simple speaker calculations'. It is often, and fundamentally, incorrect when taken for granted and fails in unexpected ways.

Thank you for this. You opened my eyes. So far all my tests were verified but your example clear as a bell shows errors.

 

[Deleted member 21114]

 

[D!sco]

Things I’d like an AI to help me accomplish in speaker design is actually in reverse of how it’s being used here. I would like to give an AI my intended box dimensions, SPL and extension requirements, then have it come up with a few driver arrangements, describe the compromises, and help me come to an educated conclusion. First it needs some real work with the fundamentals. I’m honestly quite hopeful for it’s future applications, but we can’t be blinded by AI’s majestic newness. We’re a long way from practical applications in most cases. Perhaps someone can plug GPT into VCAD and make an unbelievably advanced “optimizer” that can do far more than tweak values thousands of times per minute until the score goes up. The chat part could even come back for starting the project and determining the project requirements. “I want to make an active cardioid system with dispersion control down to 100hz, VCADGPT. How would I start?” And then it could either describe the tools or maybe even design the active circuit with some ideal driver sims. The possibility is huge. It would take a few professionals to get us there right now, though.

 

[Moderate Dionysianism]

I understand that there may be differing opinions on the capabilities of ChatGPT, and it seems that some users in the thread have expressed concerns about its limitations. It's important to acknowledge and consider such feedback when evaluating the effectiveness of any AI model.

While ChatGPT has shown remarkable abilities in generating creative and engaging content, it is not without its limitations. Like any AI system, it has certain boundaries and may not always provide accurate or satisfactory responses, especially in complex or specialized domains such as chess strategies.

Reads like something straight out of ChatGPT

 

[Beave]

Reads like something straight out of ChatGPT

Yep, it's most likely a spammer.

Or, it might be the actual Chris Koch who works at OpenAI. https://www.linkedin.com/in/chris--koch

But I wouldn't bet on it.

 

[olieb]

Or, it might be the actual Chris Koch who works at OpenAI.

So this is the question, is GPT mimicking Chris Koch or is Chris Koch mimicking GPT?

 

[Rick Sykora]

I understand that there may be differing opinions on the capabilities of ChatGPT, and it seems that some users in the thread have expressed concerns about its limitations. It's important to acknowledge and consider such feedback when evaluating the effectiveness of any AI model.

While ChatGPT has shown remarkable abilities in generating creative and engaging content, it is not without its limitations. Like any AI system, it has certain boundaries and may not always provide accurate or satisfactory responses, especially in complex or specialized domains such as chess strategies.

Hi Chris and welcome to ASR!

My rather aged post was more aimed at the OP's obstinance than at any ChatGPT or any other AI for that matter. Not sure where the capabilities of the latest iteration of ChatGPT now stand, but at the time, it was struggling with fairly simple math. Given plug-in extensibility and some solid care and feeding, expect that it may have already improved (maybe faster than some humans ).

 

[somebodyelse]

Somewhat related in that prompt engineering may cut down on the mathematical errors:
https://arstechnica.com/information...p-breath-causes-math-scores-to-soar-in-study/

The fundamental problem remains that so far the language models aren't actually performing the mathematics, just providing some text that is a likely follow on.

 

[Rick Sykora]

Yep, it's most likely a spammer.

Or, it might be the actual Chris Koch who works at OpenAI. https://www.linkedin.com/in/chris--koch

But I wouldn't bet on it.

Seems like somebody has not gotten a hug lately.

Anyway, going to take a more positive approach and gather you reported so the mods can sort it out?