My Insanely Convoluted Process of Composing a Violin Piece with ChatGPT
Vložit
- čas přidán 11. 09. 2023
- In my previous videos on ChatGPT, I've seen if ChatGPT can directly write melodies and harmony, acting as a (somewhat) autonomous creative agent. In this video, I take a different approach, using ChatGPT as a creative consultant in my own compositional process. Does it help me explore new possibilities as a composer? And what can we learn about the nature of creativity along the way?
Camden's website: www.camdenboyle.com/
Other self-promotion:
Check out my SCAMP libraries: scamp.marcevanstein.com/
Take my course on Kadenze.com: www.kadenze.com/courses/compu...
Support me on Patreon: Patreon: / marcevanstein
Private Lessons: teaching.marcevanstein.com
Creative Commons visual resources :
- commons.wikimedia.org/wiki/Fi...
- commons.wikimedia.org/wiki/Fi...
- thenounproject.com/icon/laryn...
- commons.wikimedia.org/wiki/Fi...
- commons.wikimedia.org/wiki/Fi...
- commons.wikimedia.org/wiki/Fi... - Věda a technologie
Came here from your latest Tic Tac Toe video and I am glad you have shown your best work in that video to get recognition for this video.
Algorithm is unfair brother it's a great example for that. Just be optimistic and keep on making videos with your heart ❤
Thanks! I'm just grateful that all the attention from the TTT videos has led some people to check out other videos I've worked so hard on. :-)
I can't even imagine being so competent in so many masteries like you are in a single video 😑
I'm here because I really enjoyed the tictactoe follow up vid and he asked nicely.
I think part of why GPT struggles with tone rows is that the tokens it’s trained on are mostly words and this task interacts at some level with letters and letter pairs that are technically also tokens but mostly seem to fall below what I think of as the “regular token” threshold. Same reason it’s useless at Wordle or similar tasks that involve looking at sub-token-level information. It just really sets GPT up for failure.
Yeah it also struggles a bit with poetry, I think for the same reasons. When it tries to rhyme it's a bit hit or miss.
@the titles thing. From my experience ChatGPT heavily reflects to the initial prompt. The description you gave it is weirdly abstract and lacks personality.
You have to make it a lot more colorful and relatable e.g. "A musical piece is a about the feelings of a skilled and nimble athlete who is currently sitting melancholically around a fire surrounded by a heavy silence. He ponders an idea until he stumbles onto a possible solution which reframes all his troubles into a normal problem. He now feels fulfilled". Then maybe art direct it from here.
Putting this into gpt 3.5, gives me:
"Embers of Contemplation"
"Silent Reverie by the Fire"
"Athlete's Solace"
"The Melancholy Leap"
"Resonance of Resolve"
"In the Stillness of the Flames"
"A Dance of Reflection"
"Fulfillment in the Firelight"
"Whispers of Resilience"
"Metamorphosis in Silence"
"The Essence of Revelation"
"A Musing Unveiled"
"Solitude's Serenade"
Which I think are pretty decent.
Also, there's no use going down the "you made a mistake here" route, just scrap the prompt and start over.
The final piece and interpretation is fabulous! Thanks for that…
Just found this after the tic-tac-toe video. I love composition, generative AI, and music. This video is just amazing.
Watching this after the latest Tic Tac Toe video :)
Even though I don't have interest in learning how to write my own music your videos are of a huge interest to me. The way you describe problem solving process. And the theories and terms you bring up.
Todays transformers (the T in GPT) lack the ability to scale their "thought" to the complexity of a problem. Basically this is because they're predicting the next word given the N previous words using a fixed computation (this is also why it says music theory gibberish... it can only think so deep). Yes, GPT-4 can phone a friend and this helps somewhat. I'm excited to see what improvements future models make!
So asking ChatGPT to do complex abstract reasoning, like thinking about the evolution of rule-based systems (and especially in a short response format), will quickly reveal the model's limitations.
How have I not discovered your channel earlier? As an amateur enjooyer of moosic and python this feels made for me. Epic video, lots of amazing experimentation and creativity and cognitive decisions being made. Thanks for sharing!
This video is great. You deserve way more views for the effort that must have gone into this video
To be honest, your experience with gpt reflects my experience as well.
First getting more and more frustrated trying to "fix" gpt to get the desired result, until giving up (sorta) on the idea and trying a different way to work with gpt, untill it gives some suggestions that I didn't think about or never heard of.
In the end, I salvage about 10% of the information I received from gpt and then use it in my small projects.
(Backstory time), initially when I heard about gpt I was really excited, as someone who didn't touch code in years would be a perfect match for me, I will tell it the problem, the tasks and sub tasks, what functions and classes to write, and in the end combine everything to achieve the project's goal, unfortunately as may of you know gpt can't deal with even the simplest of tasks .
I have also found that (at least for me) the best use for gpt is to present a task (not the whole problem) and asking him how to solve it and then taken the idea and integrating it yourself.
you said this took a lot of time and effort. so i came to give it a like😀👍
I came here to give a view to the video nobody watches, then I remembered.. :))
I think it's worth exploring such ideas. they are:
- Interesting to watch for viewers like me
- possible to blow up (I've had similar recommendations of even x year old videos which got hit by the algorithm)
- genuine experimentation, it's cool to see what other people think of and how far they take it. I think the soul of CZcams is somewhere between videos like this + CodeBullet + Marble Clock + random livestream of coding/eating a soup
def naming my kid Xavient, thanks :D
Made me wanna pick up a violin, thank you!!
As to the question of GPT's performance with regards to tone rows vs pitch classes, I would consider it in terms of its training set. What code relevant to these concepts is it likely to come across? I have trouble imagining where tone rows and code intersect much, but pitch classes I certainly can. If only because GPT's training set likely includes plenty of signal processing code. While it might not directly mention "pitch classes", the conceptual construct similarities are likely activated. Pitch classes easily relate to signal processing, where concepts like modularity (where one value wraps around to the start after reaching a certain point) are prevalent, especially with crucial aspects like harmonic relationships. A pitch class doesn't reference absolutely frequencies, but rather a position in a cycle of twelve, which relates to frequency domain analysis and Fourier transforms. GPT does well with conceptual similarities, even when not specified.
I doubt the training set has much intersecting musical theory and code. I'm sure there is some, but not to the volume that would be needed for what you are attempting, and you are working along the frayed edges of GPT's creative capacity. I wonder if you would have better luck with tone rows if rather than directly naming the concept, you just described the concept without necessarily connecting it with music.
Ultimately, I’ve found GPT to be incredibly helpful, so long as I stay on well trained paths. Creativity takes it off that path and things can easily go awry. Even things it could competently do before, once we stray from the path, it seems to lose it capacity.
4:20 At this point my Pixel 6 Pro recognized Hans Zimmer's "Remember Who You Are" here. lol
Here, have my view, coming directly from your last tic tac toe video. Thank you for the Infotainment
Try using controlnet with the rigging and segmentation inputs to get coherent identities in your images. Also, generate character models in advance.
cool thing is., you are not the only one that will shape this (song? what word should be used here?). who, when, and why someone play it will shape it to be different. this is so hard for me to grasp.
Conventionally, the word is "piece".
incredible work! i like this
I came here from the second tic-tac-toe video and the reasons I can say why this video is not blowing up are not so catchy thumbnail and title. Also the video is a bit too long. I understand that you worked on this so long but that's how the algorithm works, It just likes shorter video, because the attention span of viewers decreased by alot. I'm quite confident you knew this info before, but I'm just explaining why some gems don't shine. Goodnluck on more viral videos!
hbomb's recent 4 hour video got 13 million views in 1 month, I don't think 'the algorithm likes shorter videos' is the reason here..
No, it's clearly title/thumbnail, but also just an oversaturation of ChatGPT content. So to stand out against all the other ChatGPT content you need a *really* catchy title/thumbnail
@@marcevanstein Maybe ask ChatGPT to help you with that?
(jk)
@@JasperKloek chatgpt is terrible at coming up with titles (don't ask what video i obtained this information from)
Kiitos!
Came here from TTT video. I must admit, I did not understand the technical aspect of most of the things in this video but I don't know why I watched the whole video regardless, that too at 4 a.m.
This is perhaps the most interesting musical video I’ve seen in a while, while chatGPT may not have been the perfect tool to help, it definitely gave you some tools and ideas to work with and while you steered, like a friend in the room helping you idea-generate. Although as of now Chatgpt may be more like an unmusical friend that doesn’t really get it as of now, I can see the potential
Commenting for the algorithm. Great video!
Great video! I wonder if GPT struggles with some of these tasks because it's so specific and there's very little training data related to stuff like this... I tried to get it to generate lilypond code before (on version 4) but it was only able to do super simple one line melodies and really struggled with making anything resembling a good melody
Quite the struggle. I wondered where the follow up video was hiding. The short essay 'Deep Blue or the Computer's Melancholia' by Jean Baudrillard is a very worthwhile read.
I had a really good laugh at GPT's terrible titles at 18:53. GPT has a pretty strong comedic sense.
BANGER VID
you are an absolutely fascinating person /pos (positive!!)
Appreciate all the kind words!
Ur welcome for watching this video
I couldn't stand to be haunted by the mystery of what the second half of this video could possibly be 💥🤯
i appreciate you
I feel like chat gbt failed when it needed to be creative. It was able to fake creativity well but it started to poke through when it was asked to name the song or fix a problem in the code even if it was pointed out.
Sometimes chat gbt will fix code if told what the problem is and sometimes it won't. You can also tell it to fix something that is right and it will then make it worng.
just commenting for the algorithm
Snoopspeare is more of a direct portmanteau but Shake Dogg is absolutely hilarious so that's better.
I think one possible reason this wasn't picked up by The Algorithm is that by using words like "complementarity" during the introduction, you're alienating a good chunk of the CZcams viewer base. Even I'm having to slow down a bit to unpack what you are putting down, which is not great at 2 minutes in.
This is all very interesting, but through all your attempts, I feel like you're using chatGPT more as generator of random ideas you try to apply than as a true collaborator. A muse if you want. Your attemps at getting ChatGPT to produce useful code is particularly telling: they show that one is better off practicing programming enough so as to be able to write down your ideas directly as good code rather than asking someone to write it for you.
As some say about generative AI: "They say programmers will become obsolete, as we'll only need to figure what we want the program to do and say it. Programmer: then we're safe! Figuring what we want the program to do is the hard part.".
I think one of the roles of a collaborator is to get you outside of your creative comfort zone, but I don't disagree that one of us was doing a lot more of the heavy lifting creatively!
Re: code, I've started to find an interesting balance there. I 100% agree that if you don't know how to code, you'll hit limit pretty quick with these AI tools. But there are cases where they are super helpful. I don't know if you've seen watched my video called "The Fibonacci Music Box", but the javascript instrument I created for that video was very much a collaboration with ChatGPT(4), where I did the higher level work of planning and consolidating the code, and it did the grunt work. It was particularly helpful because I'm a good python programmer, but not a JS programmer, and it was really helpful at translating.
I watched it like you asked :)
I like the part about making titles. From tick tack toe video. A little too technical but i like seeing the limits of AI.
Sad Violin Goes Brrrr! XD
Here from the tictactoe followup
As a last resort, give it a PlantUML diagram illustrating your idea (or any other text-based diagram format)
Intresting --hadn't considered something like that!
chatgpt needs very clear instructions and samples. I think of it as the worlds most advanced autocomplete.
It's like the library of babel.
Hey I'm just a random guy but in your Tic Tac Toe Video (about there only being "1" game of Tic Tac Toe) you kentioned that you didn't know why the YT alg didn't pick up your video.
I think it may be due to it's pacing, maybe speed up the intro a bit? But what do I know, I'm just random guy.
engagment
What ChatGPT could probably tell us about emotional impact of the melody ( specific notes and chords should be described somewhere in literature ) Could this AI translate for us what was written and played?
Tic tac toe brought me here. ;)
liked because of gangnam style, thanks
this is a comment.
The question that fascinates me is whether AI will ever compose a melody comparable to the best human composers. Personally, I don't see why not, it just isn't there yet. The issue I still have with most algorithmic music is that it sounds like algorithmic music! I think machine learning will sort that out. It's the same with generative modular stuff (which I'm guilty of making sometimes too). It sounds like random notes quantised to a scale, because that's exactly what it is. Also, maybe one reason we can't train computer models to create decent melodies is because we don't understand how the human mind does it either. If I knew that, I'd be a millionaire and not sitting around watching CZcams videos about complicated algorithmic composition. 😂
Here from tic tac toe
I think the reason why ChatGPT is bad at Titles is because he's trying to suggest titles that are "memorable", by using starting words that start with the same letters like Sound & Science, sorta of like how game level names are structured, their naning convention are focused on memorability and recognition, so this is probably where some of it's bad ability to create titles comes from.
I love that this video that's pitched as a "collaboration" with chat GPT basically proves that it's a stupid machine that spits out garbage that an actual human who knows what they're doing has to then refine to make it into anything useful. Kind of disproves all the AI dumbasses who think it will replace artists.
sounds pretty nice to me tho.....!!)
To be unnecessarily fair to gpt, L systems are not really trivial to figure out even for humans
So..the title with a good melody with chatgpt worked and this title didnt, see how the melody one was a positive title and this Violn one is negative in the title-pic thing...and maybe put a Violn on the pic thing and the Chatgpt symbol that should do the job
It is good to know that ChatGPT has enough good taste to refuse to write twelve tone rows! Or manipulate a given row!
tic tac toe brought me here
1 game
80% snoop %20 shakespeare is just mc ride
Well AI is a newborn, the worst that will ever be and doubling in power and quality every 12 months.
I'm sure that future models will have a far more developed agency enabling them to choose the artistic influence to make a new piece of art
@2:35 The only thing I have to wonder about with AI for the future is: Will we ever enter an era as fucking hilariously *_stupid_* as this again?
i find those names to be pretty terrible. every single name contains an x, y, or z.
i find this video really boring y really gave it a chance but its just boring and idk here is some feedback,
I found that youtube did not push my music composing + chatgpt video. not sure why. Loved your tictactoe video. lold at the (num_gms/8) comment.