Can AI create a better overdrive pedal than a human?
Вставка
- Опубліковано 9 лип 2023
- In today's video, we experiment with OpenAI's ChatGPT, in an effort to design an original guitar pedal circuit. I was curious to see the kind of overdrive circuit this type of AI could help create. Our journey begins with building a guitar pedal circuit relying totally on ChatGPT.
Honestly, it didn't go as smoothly as planned. We had issues with the circuit not working at all at first and part values not being quite right which made it sound pretty bad. With these issues in mind, I asked ChatGPT for more advice and made some changes to the circuit.
After adjusting the circuit based on Chat GPT’s recommendations, it was time to test our new and improved circuit. Still, it was way off from being a great sounding overdrive circuit. A few more adjustments recommended by ChatGPT led us to a fuzzy sound, which again, not what we were looking for in a guitar pedal overdrive.
At this point, I decided to trust my own instincts and knowledge of guitar pedals. We adjusted some things, tried some new ideas, and ended up with a sound that while not perfect, was definitely better.
So what's the takeaway from all this? ChatGPT is a cool tool for bouncing around ideas and learning new things. But when it comes to making guitar pedals that really work, we're not quite there yet. As AI continues to evolve, we might see more practical applications in electronic design tools like SPICE.
To all the beginner and intermediate guitar pedal builders and engineers out there: stay curious, stay creative. AI is a fun tool, but your original ideas and knowledge make your projects unique.
If you haven't already, remember to subscribe and I'll see you in the next video where we'll keep exploring guitar pedals and guitar gear in general and perhaps even some more AI tools and techniques that may be handy to us in the guitar gear and guitar player world.
Subscribe here:
/ @wampler_pedals
Linked video: Revv chat breaker pedal • Chat GPT Takes Over Th...
Subscribe here:
ua-cam.com/channels/dVrg4Wl3vjIxonABn6RfWw.html
Linked video: Revv chat breaker pedal ua-cam.com/video/OyKzD2h-A4M/v-deo.html
Brian you should create a new pedal that's a combination of a wah and a univibe/tremolo effect. Maybe call it the Wahmpler.. Just a thought.
I liked the fuzzy one the most
I really enjoyed that fuzzy overdrive one
Me too! I would absolutely have room for that on my board.
me2 😀
1:08 Sounds Great! Very transparent.
My rudimentary understanding of chat GPT is that it’s accuracy and abilities are largely based on the training data set that was used to train it. So for instance, if someone collected all the schematics of all the overdrive pedals ever produced, the GPT would better recognize typical values and topography. It could then suggest potentially new and differs variations, but what is it solving for? “How it sounds” is subjective so the training data set would also need to include things like: this one has higher gain, this one has more top end, etc. so that it knows better what changes to the circuit match the user prompts.
Its.
I rather liked the bad fuzz sound.
I kinda liked the "Bad Fuzz" version.
Hey Brian, you should build Wampler Bad Fuzz as a $99 pedal, lot's of people seemed to like it (myself included)
I mean, I can! I just think alot of folks wouldn't like that there is so much clean signal blended in with that fuzz tone. It might make for a decent bass drive though, come to think of it.
I got that bad fuzz sound. I don't know how it was done but yrs ago I bought a modded DS-1 that said in ad "Has the Reverb mod and a Fromel mod". All I know is at lowest gain ,barely anything there. At about 9:00 great fuzz sound. Anything else above that and it's too much for me. I know a lot of others would like this but it just sits on the shelf with the other unused pedals. Only costs me $40 for a highly modded Boss pedal. It's got a large scrape in the orange paint so I painted the scrape to look like lightening bolt in green.
@@wampler_pedalscall it “bad fuzz”. But seriously I loved that sound. It almost makes a doubling effect with that clean signal sitting under a layer of mud. Maybe “mud blanket”. But really I’ll buy one.
Brian, you seem like you'd be the perfect guy for a definitive video on "noise, hum, shielding, ground loops - what is that noise and how do I get rid of it?". Some basic science and even measurements/tests would be fantastic. I'm looking for a video that I can just link to whenever someone asks.
The bad fuzz was fun! I want to buld it!!
I thought the "bad fuzz" sounded great.
I really thought the second version was a keeper.
One man’s bad fuzz is another man’s… bazz fuss?
Calling any fuzz bad is trolling
It was terrible, and that’s coming from someone who likes terrible pedals.
Informative, amusing, and musical 🤘
“Bad” Fuzz was the best sound in the video. Maybe that’s the name if the pedal.
That buffer cracked me up a bit :)
the reason why LLMs like ChatGPT struggle with tasks like this is because they are entirely based on text and language, and that's something we rarely use to describe circuitry in-depth. what it probably is pulling from is some guy on some website talking about how one component in once specific design has a certain effect, but it isn't able to recontextualize that sort of information.
I think it's definitely possible to have an "AI" system that can design circuits extremely well. but ChatGPT can't because it's not really meant to. it's essentially the wrong tool for the job.
You nailed it.
The most important idea here, in my humble opinion, is that ChatGPT, after a few iterations can help you to build an overdrive-ish pedal even if you do not have any previuous electronics knowledge, just "breadboarding".
Nice video :D
Fun video idea. I wonder if you specified you wanted an input buffer, or if it decided that on its own.
Hello, I been enjoying chatting with chatgpt for a while naw and I must say it take some time to massage the idea of what you want. I would suggest starting with general questions about overdrive pedals then move to control layout, the ask what modern features people use. And only then asked about making a circuit for it. I think it be a lot better
Wampler Bad Fuzz is a winner through my iphone speaker
Trying to ask AI how to find a good tone - is like asking what it’s favorite flavor is.
That was fun. I actually liked the "Bad Fuzz." But, there is one thing that I've been wondering about. Can AI could actually improve the accuracy snd tracking of Pitch shifting/Transposing the guitar's tuning up or down? Yes, there are pedals, plugins, and Modelers that can pitch shift, but the results are often compromised by artifacts and glitches. Perhaps AI can solve this difficult problem better than humans have done so far. 🤷
Accurate Pitch to Midi conversion without special pickups/hardware would be a related area that maybe AI could help improve.
If AI could solve those issues, it wouldn't be a matter of asking ChatGPT to design the perfect analog circuit that we've never thought of. It would be more like using AI to process our guitar signal on the fly, much like how we have current instrument-isolation AI tools available. It wouldn't really be AI solving things humans couldn't do; it would be humans harnessing AI to do something transistors or traditional programming couldn't do.
The look of pain on your face as you used the earier iterations was gold...
"Absolutely Wonky" is the next T Shirt you are wearing on a video :D
And now I need to build the "bad fuzz" 😂
very , very interesting !
You made the bad fuzz sound good. I guess it's true what they say about where the tone is at.
Me: That sounds pretty good.
Brian: It's a bad fuzz lol
Me: Makes sense.
Make sure to feed ChatGPT your findings so it learns. However, it will *never* be able to understand the subjective human appreciation of tone. Nice subject choice for a video! ;-)
very cool, what's up with that neat breadboard buddy at the top, what's that do?
It's just an easy way to add power, in and out jacks, and other voltages (+ or - 9v, +18v, 5v for digital, dig ground, buffered vref, etc)
i tried to get chatgpt's help with a circuit once and quickly realized it was essentially just agreeing with everything i suggested to it+spitting it back to me as its replies
Thank you Brian for such a great products you produce!
I’m in the “Wampler club” an is an owner of the next few OD/Dist pedals.
Euphoria
Panteon
Tumnus Delux
Ratsbane
Moxie
Pinnacle Deluxe
Looks like I have all types of drives but still trying to find which one of this pedal can be used for recreating of the Blues Diver closest sound (like a Boss BD-2 sound).
Help me with an answer on this question please and you make me happy :)
Thanks in advance
Thanks so much! I may come out with something blues driver based in the future but as of right now we don’t have anything 😞
@@wampler_pedals
Really waiting for that day!
It's the equivalent of outsourcing a task to a high-school student in my experience. It might get done, it might not, but it will be confident its good enough 😅
So dumb question, does the op amp raise the level of every frequency, then you use the capacitors for filtering EQ, and Diodes/op amps for more gain, back into a buffer and thru?
I hate this whole AI trend but I’ll still watch any Brian Wampler video
Also: everyone who is correcting ChatGPT's answers: congrats, you're working for free by feeding the complex parrot info so it can look more convincing in the future.
I thought that was a great idea for a video!
Thanks! Yeah, I was a bit surprised it wasn’t more popular
Do you think it has a learning curve and eventually will figure it out?
I've heard boutique pedals that sound worse than GPT Fuzz.
So the promise of AI is reduced labor. For the time it took to get GPT to the fuzz stage, was there any cost/time savings from using GPT vs. a novice looking up how to build one?
That's a good question! Probably not, because I had to explain what wasn't working, and why it wasn't working in order to get it to explain to me how to fix it. Keep in mind, it's explaining like this: "Connect a 1k resistor in series with a 1uf capacitor to connects to the base of an NPN transistor, with a 10k resistor from the based to ground". These type of instructions would most likely confuse a novice unless they already understood how the electronics work.
I think the problem is that a soon as you have a complex idea, it takes way longer to explain it than to execute it. I know this because I write academic text. Basically, if I am at a point where I can explain the bot what to do, I have already written the article myself right there. And by the time Mr. Wampler asked what to do next he could have probably just tried out a new mod for the circuit already.
@@wampler_pedals I kinda love the fact that the builder of my favourite pedal ever (Tumn.Dlx., own 3 already!) is the one person to demonstrate what the current problem with bots is.
@wampler_pedals you're right about confusing a novice - or at least THIS novice anyway 😅 add to that the fact that I'd probably be unable to explain what is wrong with it in a way that AI would understand and I'm pretty sure I'm better off watching your and other electronics channels until I'm confident enough to start actually building a pedal - so far I've only "modded" one that ( to my surprise ) survived the surgery AND sounds OK
Did you use Free version (3.5) or payable "Plus" v. 4.0
Paid gpt4
How much is the ref voltage in the schematic?
Vref would be 1/2 voltage, so with 9v input on this it’s 4.5v
I thought the end product sounded usable. But I am not much of a distortion "expert". I much prefer a hyper-clean tone.
I didn't think that the fuzz tone was all that bad.
That last AI version was a Morning Glory.
nah. Morning glory is a bluesbreaker pedal circuit which has an inverting op amp after the non-inverting opamp with 4 diodes symmetrically clipping in the inverting OA, with the tone control and additional low pass filter after that. This design isn't that at all.
So when are we going to see this new release Wampler 'BI' (Brian intelligence) pedal lol.
😂 no plans to release it at this time, but I appreciate the sentiment
Props for suffering through that glorified clippy to show what i already strongly suspected the result would be. Saved me the time i didn’t have …😂😂
I love your t-shirt!!!!! You sell them?
Thanks! Unfortunately I don't :( I think this came from amazon
@@wampler_pedals thanks for the reply Brian! I have to say I really love these videos you make with your breadboard and showing different sounds and all. I could watch hours and hours of this.
For what it’s worth I just made one and stuck it on my merch website, modyourownpedal.com
I appreciate the comments! 😊
no substitute for all natural intelligence
Sounds like ChatGPT can do an okay job, assuming you're an actual expert in the subject and ask it the right questions.
No, if you actually watched the video, you see that the expert has to correct it. If ChatGPT doesn't have the answer, it can't just make it appear out of nowhere. It has no ability to create, it can only repeat what info it has been fed.
Commenting for the Al Gore Rhythm.
"Bad fuzz" would probably be good for bass....
I wonder, was this the free 3.5 version of Chat GPT, and if so, would the paid version of done a better job?
no, gpt4
Saw a dude do this with a fuzz pedal.
At least your foray into AI hasn't devolved into an affair of the heart as others have experienced 🤣👍
But...give it time as this too will become as normal as...
Well as normal as having your palm read by some thin gypsy thief...?😎👍
😎✌👍❤🖖
lol! this was awesome! well played love your new triumph Pedal man! best overdrive of all time! used it in this Afro-Cuban punk project, sounds so amazing! ua-cam.com/video/AkCzJDlzCSo/v-deo.html CC is English! Cheers!
Kind of reminds me of a certain lawyer from a certain state that used ChatGPT for a certain case and their defense vaporized in court.
So it took the tube screamer & broke it.
Forget it. JUst keep making them the way you make them. My Belle, Gearbox, Plexi Mini, and my Ego are the best pedals I've owned. All on my board, and I only needed a little intelligence of my own.
Fun video, but your job is still safe.
It made a OCD on LP mode 😂
My CPU is a neural net processor.
Honestly, the fact that it could be coaxed into producing anything is a miracle. Imagine giving a "blank slate" person a traveler's phrase book and asking them to use it to do math.
I Don't Care, I still much prefer You, and other "human" makers, to design the pedals, I buy! :P :D
I’ve been breadboarding for the past 5 months and it’s obvious you need human ears to decide what sounds good.
Short answer: No. Why not? Because every overdrive pedal is predicated on an expected input signal, with expected dynamics, amplitude, and frequency response, and aims for a designated flavour of clipping based on that and the clipping elements used. By the time one has specified all of that to ChatGPT, you've basically designed the circuit yourself. And what Brian has made on the breadboard - with support from ChatGPT - is essentially a replica of something that's already out there. An instructive exercise, to be sure, but not "better than a human" design.
And it's also worth noting that many popular time-tested designs were either the product of mistakes (like the Proco Rat) or else were empirical explorations that had precious little textbook theory to guide and support the design (like the Klon Centaur). Is Chat GPT going to draw from data on what already exists, or will it go beyond? And, as we see and hear, here, the query basically delivers the strikingly familiar and banal. It's ultimately up to Brian to make it "sound decent".
Interestingly, ChatGPT suggested what has come to be known as the SWTC (Stupidly Wonderful Tone Control) as the tone control. It has the perk of not changing level in any major way as one adjusts it. The downside is that it doesn't result in powerful changes to tone. You have to really design for the ballpark tone you want, and use that tone control to compensate for speaker brightness - think of it as a fine adjustment. But with all the circuits using a Tube Screamer or Rat-style tone control, it's surprising that it suggested that.
Who even is Chet Jeepidy anyways?…
of course the guy who makes and sells guitar effects is going to say chatgpt can't do this.
😂 you literally watched me walk through it. Give it a try yourself! 😊
@@wampler_pedals I mean it's a tool for you to use, but it's probably not instantly going to do what you imagine. but you can't really say that it doesn't really work as you did what it said and i saw a bunch of comments saying they liked what it did. so i'm sure that, in some people's opinion is better than some previous overdrive they had,... iunno
also AI works differently depending on the prompt,..and you might get different results even with the same prompt so i'd say it's not impossible to get something you'd like out of it.
Did you tell the AI how good it sounds??
I would not suggest using chat GPT to design an amp. What would actually work better is training a CNN using existing schematics with the details tagged and then having the CNN generate one. ChatGPT is a text transformer. meaning it just predicts the next thing in a conversation.
As usual it creates a solution that looks good but does not work! It can not deal with context!
No. Because better is completely subjective. It's not called 'artificial' for a reason. I wouldn't worry about A.I. until quantum computing has been fully developed. Light computers, too. But for now, AI will be good for, say, business logistics. And as a concession, maybe AI can build a more efficient circuit. But better? A.I. isn't conscious. It has not concept of better.
You should be more concerned about an AI that's built and trained to design electronic circuits rather than text chat. But I don't think it'll still take your job, possibly aid you in it at some level.
This is kind of like when the Wright brothers first flew a plane and people said we would never get to space. Give it a couple of years, no human will be able to compete with AGI in anything.
How good are your prompt engineering skills? Not saying you're bad at it at all, so please dont tak offense, but it's definitely an important aspect to using GPT for any application.
You'd think we'd be better off using an AI that's designed to process audio information and simulate electronic components, rather than an AI Language Model.
In twenty, thirty years from now we all will be hiding from the machines anyway so get your ya,yas out now… 😂
I'll just lead them into a boobie trap. 😂
So - chatgpt will excel at something with somewhat vague (though potentially far-reaching) qualitative effects, when used by someone with little to no knowledge about the intended field of application. Fun times ahead.
😂😂😂
Ai can’t make anything better than humans
😂
The answer is no
"The all knowing AI" -> Big mistake!
Chat GPT is not a search engine, it is not a fact machine, it is not providing correct facts at all. If the facts are correct, that purely by chance. Providing correct facts is not the work principle of this thing.
CGPT is a linguistic machine. Based on what it has read, it guesses the most probable answer from a linguistic standpoint. It is a very eloquent trash talker. Not knowing what it talks about, but doing that in a very self-confident, eloquent, convincing way. That's the way con-men operate.
Now, CGPT has read all that rambling about schematics in forums. From that mostly sub-professional talk it derives what it may answer. Garbage in - Garbage out.
In order to give well educated answers, it would need to read through many thorough analysises of schematics and their sonic implications. Then maybe it's power would come to use.
My comment was said sarcastically.
chat gpt is just a google api.
It reminds of somebody singing songs in a language he/she actually does not speak. When they've heard the song often enough they can get very close, but there are still goof-ups that betray them immediately.
AI reminds me of “paint by numbers” that can’t count. It’s a joke by which EVERYONE claims. No substitute for creativity and imagination just those who don’t.
ChatGPT doesn't "understand" anything - it's a really clever, complex filter. The only thing it's good at is plagiarism.
No. Fuck, no.
Please remeber "Chat GPT" is a text (words) generator basing on other text created by humans. This is a piece of software, having billions of parameters and probability for each word in context of previous words. So the whole "knowledge" "AI" "smarness" comes from HUMANS. It does not "create" new knowledge or have any kind of "brain", "consciousness" .
Yep. It’s a LLM.
Chat GPT is not A.I.
ALL it did was scour for generic circuits posted around the net and recombine it into something 'original'! FAILED didn't it?
Loser AI!
Humanity 1st!
Humanity LAST!
COMPNOR FTW!
Nope Ai can’t do crap
It's not actually AI. Calling it AI is just marketing. AKA: it's an absolutely outright bald-faced lie. It's a machine learning parrot that will spew back out answers only based on what has been fed into it. If you correct it with good info, it may get better. If you feed it bad info, it'll only get worse. It cannot really do any kind of deep analysis; at best it can do light comparative analysis (ie, like to like in 1-to-1 specific comparisons). however that's also dependent on it having the info to compare with.
AI is just more bullshit for now. Probably for the best given that we don't have the infrastructure to handle how badly such a thing would fuck up the work force if it were really as competent as advertised.
Artificial intelligence yep it's as intelligent as the tv that people watch and full of just as much BS.
The first few pedals sounded like total crap. Anyone that liked them, knows nothing about pedals or guitars. They’re probably AI opinions.
You can Never take the human out of designing. “Garbage in Garbage out... “ AI would have to know it’s info for every subject that was programmed by an expert in their filed. Then they’d still lose, because AI can’t feel or judge good from bad. All the good stuff that goes into making a pedal.
At it’s very best, AI would give you generic sounding pedals. Blah!
One more thing...
Imagine the amount of ram ( memory) you’d need to store this vast amount of knowledge.
The cost would be insane.
We humans can store more and truly think, not just spit out data that may be close to a subject.
It’s just my opinion and I’m sure there are carbon units out there that think silicon units can and will out think us, but I don’t belive it.
We were programmed to have feelings ... we can hurt, have sympathy , empathy and be able to Love ❤️.
I know it’s Crazy! Even uncool these days, to belive in an omniscient power.
To a lot of us, we know that
a machine will never be able to do that.
This isn’t some religious speech.
I think that organized religions have been bastardized and corrupted for monetary goals
and other nonsense.
We don’t need them to have Faith.
You can’t be a complete human without faith. It’s not speculation.
I’ve known many that had no faith and they weren’t whole. They were empty. They knew something was missing. In the end they discovered the secret.
RIP Scott....