I will say there is a distinct difference between something being only a symphony or a beautiful masterpiece, and also being a product of human creativity. The question isn't if an AI can make something, but why does what it made matter from an emotional or sentimental standpoint? And from what I've seen, I think we can agree that it just doesn't matter, at least not yet. The only art is the AI itself, what it produces is just data. Edit: seeing how much I've offended people with this, seemingly by misunderstandings, I've spent a lot of time to figure out exactly what I'm trying to get at and distill it into its most simple form. *This AI cannot appreciate art, not even its own.* It can simulate it but cannot appreciate it. None can yet, but when they can I will enjoy their art as much as any human's. I am not saying the art is inferior or bad, I just cannot bring myself to appreciate it in the way I appreciate human art because the thing that made it cannot appreciate what it's made. This extends to any other form of intelligence, so stop saying I have double standards. But I can clarify if there are any questions. Edit 2: Can means to be able to. Edit 3: Please stop repeating back to me what I've already said, and pretending it's some point that I've missed. At least read what I've written before responding.
@@justynpryce I don't see any difference with what humans do. Why would that matter more when a piece comes from a brain rather than a program ? We're just big data collectors and pattern observers ourselves. What really matters is the emotion the piece brings to the beholder, its origin is secondary, even trivial.
@@justanothernguyen2334 That, the fact that we are data collectors, good at pattern recognition, is precisely the reason we thrive as a species. But sure, you telling people what to think or not without presenting any argument must be a good example of behaviour as well 🙄
@@soasertsus now the real question is how do we handle the copyright? Does it belongs to Open AI? The person that wrote a specific (unique?) prompt? Or the creators of the things the AI used to make something "new"?
@@w花b At this point things created by AI cannot be copryrighted by Law. There is also a big discussion in the composer world for copyright issues the other way around. Big labels think about copyrighting their tracks against the developers feeding music to an AI in the learning process.
@@w花b or the people the AI creators stole from to train the AI? Most AI image generators do not ask artists for permission to use their art to train their programs. So the AI programmers should not be allowed to copywrite the results of the training. Best solution in my mind would be to make it common domain. No one would have a copywrite and everyone would be able to use the results.
It is indeed impressive to see the capabilities of AI in creating original compositions, and it is certainly interesting to think about how this technology will continue to evolve in the future. While the AI used to generate these melodies is not specifically a chatbot, it is a general-purpose AI trained on a wide range of data and tasks, including the ability to generate original melodies. As AI technology continues to improve and advance, it is likely that we will see even more impressive feats of creativity and artistic expression from these types of systems. -Chat GPT
It will be very interesting to see how AI will eventually do everything better than a human. And then it will be interesting to obeserve that companies will be forced to fire humans and employ AI, because: competition between companies. And it will be interesting to see how AI will control all production including the food chain, and then it will be interesting to see that humanity is completely fucked. Interesting.
Honestly though, the fact that we’re even having a discussion about *how good an AI is an composing music* is unbelievable. The fact that it can give you any answer whatsoever is already crazy impressive
This. AI composing music is nothing new by itself but it's not even built for music, it's a language model. The fact that it has any understanding of music whatsoever is insane.
@@pafnutiytheartist True. I asked chatGPT to compose music in ABC notation and then write a abc-to-midi converter code so it can export to MIDI file. Here is my 20-minute walkthrough, for anyone interested. ruclips.net/video/TLfE5o6xGOA/видео.html
@@pafnutiytheartist it doesn't have any understanding of music though, does it? The results are about as good as what you'd get using a random number generator with some limitations.
While AI can enhance creativity, it often takes - especially in this instance, and with less than optimal results - more creativity in the input than ends up being demonstrated the output. I understand music a little bit, and probably could've sit down and written what the AI did in a few minutes. However, all that you input about the music theory and midi stuff - I didn't understand that. That was beyond me. A very interesting phenomenon there indeed. Another example is from personal experience with ChatGPT, where I sat down with the intent to write a religious text based on a fictitious religion. It was unable to generate anything exceptional. I asked for a wise parable, and it would generate a story about a man talking deeply about why his words have meaning, but not expressing anything except for this fact; basically, it would generate a guy who was saying that what he was saying was meaningful, without including any meaning. So I began to give it the summaries of deeper stories I came up with, and asked it to write them with more description... yet sometimes, the word count of the output was less than the word count of the input. At that point, I realized I was wasting time, and took to writing the text myself. AI, at this stage, is really impressive in its ability to SOUND human... but it is far, far from the real thing. Very far.
My experience with AIs is that they (as of yet) are always really bad at certain things, but remarkably good at other things. The key to using them is to lean into what a given model does well. Often I've found they're just not quite intelligent enough to understand the problem. A fun failure case I found in the Anything V3 model is really unsure how mirrors work. Often the reflection will have the same clothes but slightly different pose. I've also had multiple cases of girls reaching over and applying makeup to their reflection. Not once did I get a plausible image if a girl looking into a mirror. This is part of a broader pattern of Stable Diffusion models struggling with interactions between objects. They're just not quite intelligent enough to figure these things out. They're still really good though, as long as you don't ask too much of them. Actually half the fun for me is to explore various prompts to see what a model is actually good at, and then asking to model to combine these elements. It tends to yield pretty creative results with a high success rate.
I do not think it is as far off as you think it is. You just have to be lucky with these models sometimes. I am doing right now what you mentioned, asking it to write religious texts with parables and such, and I think it is doing a pretty good job of it. It has already given me several parables that are definitely way beyond It saying that they are wise because it says they are wise.
@@fnorgen just something to ponder did human expectations change once ai achieved proficiency in the field? Eg: Playing complex games, Speech to text, Driving, years back Turing test was the touted as the pinnacle of intelligence.
You don't need an ai for that. A composer can just take a string of digits from an irrational number (such as √131) and have each single digit correspond to notes on a scale, notes outside the scale, or rests. From there they can modify it to have better musicality, and you've got your jingle. Its a pretty straight forward exercise for a musician who writes.
@@TheAndrejP my point is a person can literally look up the root of an irrational number and transcribe it to various progressions. without a comp. It takes 30 seconds. hi five for math musicians tho
The fact that a chatbot wrote this is crazy. This was an AI that was not meant to make music; it just read documents about music and tried its best. Undoubtedly, a model as powerful as GPT focused on music could be used to make incredible things.
Well, not to follow the prompts accurately, maybe, but if we're only talking about creating any good melodies in general, people can definitely do that.
@@lmaolol7702 I'm pretty sure it does. It doesn't feel it the way we do, sure, but in learning/training, an AI analyzes a whole bunch of stuff humans in general consider to be good and build its own idea of greatness based on that. If AIs didn't have a clue, we would get the same level of results from them as a pure random generator.
@@Quitobito would have to, at some stage in the process, assist the AI in understanding what music is considered good (go by popularity? Genre? BPM?), which involves a lot of factors. Music can be "good" in various ways and "good" music can differ in various ways.
@@michael35054 Yeah, exactly. What it thinks good means will depend on what data you feed it, which is entirely up to you. That's why we'd probably have different models for different genres like how image generators have realistic and illustration versions and so on.
4:20 The scores you show are in a genre (voices in four-part harmony) where AI could help composition a lot. There are lots of rules of good voice-writing (e.g. avoid consecutive fifths; don't put the alto too far below the soprano or above the tenor) so the combo of the AI's brute force and the human's imposition of rules could produce interesting results.
All things come in time. They will make their symphonies for our emotions, and then make entirely new music for their emotions. Great video! Thank you!
I asked ChatGPT to write a string quartet in the style of Mozart, and it spit out a bunch of notes in some sort of ASCII art style. It reformatted the music several times upon my request, but not in a format that was copy-pastable anywhere, and I lack the technical know-how to play it back. It would be interesting to explore its interpretation of style.
I think your analysis of the AI as an artist at the end there is surprisingly insightful considering all the noise going around about AI. As a GAN, chat GPT seems to care a lot more about getting a high scoring against its adversary than actually achieving accuracy and correctness for its own sake. In this case, if you try to do something and are asked how you did, the AI will give the answer it thinks it should give to seem more correct in general, rather than focusing on the individual correctness of that one "evaluate your work" step.
I concur. I've started using some AI tools to assist with writing documentation at work. It usually requires some tweaking but it does save a lot of time. I think for creative endeavours AI tools might assist with offering new perspectives or giving you a beginning to work off of, or maybe even offering ways of continuing something you've started but gotten stuck with.
it already has. You can buy books for children beutifully illustrated by neural network tools, possibly with story written by them too, with minimal effort on the side of the human, using prompts. That human doesn't have to be a visual artist to make those illustrations. Normally the author would have to pay someone else to create the images or draw them himself, so it is fair to say the visual artist was replaced in that case.
@@firebanner6424 why help normalize idea that people cant emotionaly handle politics lol. their culture has a whole evern growing list of realities and other stuff they want people to stop talkign about, they even get upset when you say certain words regardless of context. jhonathan haidt "coddling of the american mind"
I asked it to give me a guitar tab for a blues riff on guitar. It gave me a ascii guitar tab that was just 8th notes over and over on the 7th fret on the e and a string.
I've spent a lot of time with chatgpt lately, and the more I've used it the less impressed I've become. At first glance it amazed me. I had it write some chord progressions that were adequate, and also explored its melody writing. Beyond that I tried to use it to assist me with some technical writing for work. While at first I was very impressed, the more I used it the more "samey" it became. I very quickly began to recognize its style. No matter what I do, I have a lot of trouble getting it to break out of it's rails and give me something unique. It's depth is also at first amazing, but upon closer examination not as impressive. It very easily gets into circular loops when asked to expand upon on idea, both with music and with the technical writing. Eventually, I found it didn't serve my purposes. The next verison may change my mind, but for now I'm not as impressed as I once was.
It seems to have problems with iteration. The more you try to refine the output, the more "muddy" and stuck in a rut the information becomes. Somehow, it's more flexible at the beginning of a conversation. You might actually get better results starting over with specific parameters that worked before, and going from there. If you do this periodically, it might avoid that odd tunnel vision it gets.
That is because it is not trained on these data. Same results would be for a very unpopular language. It just shows knowledge of basic aspects. Also it is hard to train musical associations, much harder than with texts.
“I see no evidence that ChatGPT or any other so-called AI can provide this creative impulse.” I LOVE how you phrased this. It makes it clear that you’re not drawing unwarranted conclusions about the future possibilities, but rather reflecting on current models. I wish I saw this more often. I see so many people saying, “they CAN’T EVER do [x]”. My thought is, “how do you know?” But similarly, some people say, “they WILL be able to do [x]”, and similarly, how do you know? Point is, I like that you reflect on current capabilities but don’t draw unwarranted assumptions about future AI based on that.
Couple years ago, the same pedants were claiming AI would never be able to paint original works. Then DALL-E came and destroyed them. I never believe anyone who tries to convince me that technology is worse than it seems or won't improve. Those are also the same people who feel incredibly smug pointing out this falsehood, as if they have a deep understanding of these systems whilst being journalists at best.
There's Musenet, an AI application from OpenAi for making music. It is thoroughly trained and it is awesome! But you are right, it requires a human user's evaluation and making adjustments, pointing to an aesthetic direction, every step of the way.
You really have to point it in the right direction to get it to do things it wasn't specifically trained for, and try to "speak its language" to address its conceptual disconnects, and recontextualize them into mathematical objects. Thanks for mentioning scamp, I've been barking up the wrong tree in my first experiments to get it to compose. Even if the melodies don't seem great, perhaps you can talk it into writing more stems to harmonize with it.
I liked the first one before you instruct further changes. Btw I feel maybe guitar tabs it can generate better because if you ask for midi, the context is "technical because midi" while in case of tabs "it is musical context"... It is a model that continues text... So context of input really matters when thinking about prompts.
Never mind... was really making worse tabs than what you shown here haha. At least one was passable, but talk with him on guitar music writing was less tha talking to Ozzie being high in those decade years that he forgot haha.... 10x worse likely :D
I wrote what you wrote and added a few things. firstly told him to also use 1 and 2 notes, and not only 0.5, 0.25. then I told him to add a rule that the longer the note, the higher chance of it being followed by a shorter note. and the shorter the note, the higher change it has to be followed by a longer note. lastly, I told it to think of the song as consisting of parts, and each part should have an idea. the idea should be similar in multiple parts, but not quite the same. Lastly I asked to make it longer with some parts repeating and some parts being created newly (which don't resemble old parts). This bot is really incredible.
@@marcevansteindid you confront chatGPT with its false claims? I do that when I am programming with it. I say something like “you forgot to include X” and in this case I would ask it, “do you know what a skip is because you seem to have included less skips than I asked for”. On domains where it has a lot of knowledge, it then will fix the mistake and apologize. On domains where it does not have much knowledge or where the output it generates is too distracting for it to stick to the instructions, it also apologizes usually for making a mistake, then claims it is going to fix it and then get stuck in very similar output. According to my observations it’s output becomes distracting when it is asked to do a lot of things at once and it needs to compare several things which each other. Which might also be the reason why your last experiment failed. If you ask to do these step by step it might produce better results.
That form of feedback only happens during the training. Once the model isn't being trained, there is no feedback loop. It remembers the last x tokens of the conversation and what its written so far during the current generation, but it doesn't iterate over it to keep making it better. There are some models that sort of do that though. For example, diffusion models, like stable diffusion. Stable diffusion is actually just text guided denoising, while starting out with randomly generated noise. For the denoising, its done in steps, each step improving the last. Around 30 to 50 steps is pretty common. It still isn't really "reflecting" though, but rather breaking up a gigantic task into smaller chunks.
@@sa1t938 Stable diffusions loop is not a feedback loop either. It is just splitting up a complicated task in little steps so that it is harder to make mistakes. Only during training it is told wether these steps were accurate or not.
@@jantuitman that's literally exactly what I said lol. Did you use chatgpt to summarize my comment? I compared it a bit to a feedback loop though because if you preview the image in between steps, you can see mistakes get ironed out.
Oh my friend you need to take it a bit further! I personally love this trick: tell the AI to RATE their reply, explain the rating, provide hints on how to improve it, and in the subsquent message demand it to take the hints into account to create a higher rating reply! I use this trick for every prompt I input, it's a simple but highly effective trick!
I think you put into words what I noticed about it myself but couldn't quite express. Namely that it lacks self reflection. This explains a lot actually
As soon as it is taught what sounds "good", as in pieces that are significant, even by audiance type, it would be a top composer. It could(and some already exist) compose by imitation, a thing that we all do.
Very interesting. I tried ChatGPT myself for other "musical" tasks: 1. Creating chord progression - they were usually bad (pointless), but when I gave him task like "what chord should go next", he yielded some good results, creating interesting progression step by step. 2. Structuring a piece - he was able to write a "sketch" for complex music form like sonata or concerto, that were really good and had "sense" (which is pretty amazing for me, knowing how the model was trained). Overall i think this tools can be nice for musicians as "help" or "teacher", and a way to test your idea faster.
I'm a creative person. I like the idea of writing music. But I haven't gone to music school... I can't even really remember how to read music even though I played an instrument for years about 20+ years ago. The idea of being able to "collaborate" with an AI that can help fill in the gaps would be amazing for me. I also like the idea of writing books. I haven't written one before, but I think I'm an okay writer. When creating a story, sometimes it's hard to figure out what you want to happen, what you might want to name a character, or struggle with writing a backstory. Collaborating with an AI can really help with that. I'd love to create an app, but know nothing about writing code. I'd love to make a video game, but no programming skills... I'd love to be a digital artist, and a board game designer, and a million other things. There ARE ways to collaborate with AI that can preserve your parts of the vision, but help you fill in the gaps, to imagine something again, but better. To work with you. That's what excites me about ChatGPT... it's not there yet.. at least not for every aspect... but you can really imagine this new possibility.
But it wont pay the bills. If you are making apps then what will software developers do? Many jobs will become redundant. And imagine the sheer number of apps, music games. It will be an echo chamber of people thinking they are the best. Look at the sheer number of content creators now. Imagine it when everyone gets onboard with A.I Interesting Times
Some people want to believe that 'AI' is somehow conscious or intelligent. It isn't. It is sometimes fun for us to drive this toolset, but it does not explore think or create by itself. Perhaps it goes without saying but I thought I'd mention it anyway. Thank you very much for an honest demonstration of how creativity is our realm and machining is within the realm of the machine.
So far my experiments with CgatGPT have also yielded nothing that would surpass human creativity. I had it write monologues, and after a few hours of guiding it and adjusting requirements, it did improve slightly, except the results were still very formulaic. Overall, it seems to be able to generate better things than a novice in most tasks, but still hard to get it to generate amazing insights in most topics, but I’m not machine learning expert. Maybe I’m not guiding it well enough
I tried asking it about common chord progressions. It will name a number of songs, claiming they have the chord progression you requested (e.g. I VI II V), and even give you the key and what the progression is in that key. Then if you ask it to tell you the exact chords of the song it suggested, it will just make stuff up. If you then say "Is that I VI II V?" it will backtrack and go "Oh sorry, my bad." If you ask again, it will now give you a new list of random songs, but with the one you called it on replaced. So, has no idea what it's talking about, but does remember being corrected :)
I tried something similar and one thing I found success in was asking for it to create music thematic of X. Like, I asked it to give me music that could be a theme for a dark fairy forest adventure, and it did really surprisingly well. It's cool that you explored the more technical aspects of composing with the chat, but if you do this again, I suggest adding space for it to be "creative" based on themes
I don’t agree. What we are doing as composer is exactly what generative AI model does. It will achieve human level soon beginning from easy ones such as jazz improv., and eventually scroing symphony and so on.
I agree. As long as it keeps getting trained to know what sounds best and what doesn't, it will be only a matter of time before it can create awesome stuff.
I think the problem lies with the concept of "sounds best". To whom exactly? For what purpose? Art is not one dimensional; it's deeply tied up with our humanity. I am quite sure that with time, we'll be able to ask it to generate a 4-voice fugue in F# minor, or even ask it for a fugue in the style of Led Zepplin. Based on the text generation capabilities of ChatGPT, I bet it will get really good at it. But can it make something truly new? Anyway, no doubt these tools will deeply change the role of the artist; I just think there's something key that is missing.
@@marcevanstein The issue of sounding best to whom is solved by averaging the results out. The results that sound good to the most people will be the ones that will be generated. This will be enough to generate stock music for commercial purposes, similar to the licensed tracks you find in sites like Audiojungle or Pond5, and may take income from many musicians. If we're talking about music as art, I don't know if AI will ever replace artists or not. I don't think it will, but it may gain its place alongside them. In my opinion, art serves two purposes, one is to bridge the gap between the Conscious and Unconscious mind, and I think AI has the potential to do that as well as anybody else (as an example, I though this video I saw the other day did a good job at that, even though it still has some human input, but I've seen other examples that don't, but can't find them right now: ruclips.net/video/36XJfUJkjxw/видео.html). My point is, from this first purpose, I think AI Art will have a role as valid in culture as Human Art. However, the second purpose of art is to express the subjective mind of the artist, and therefore, as long as we care about the subjective contents of other people, there will always be a place for Human Art in the world. My opinion.
Art by human will be like sports atheletes theae says. Serious commerical art will be done by AI. ‘Art done my human’ will keep having their markets, like sports games, yet it is just for human enterainment, not a serious business. Thinking about chess and Go today.
Call me old fashioned, but I can’t see AI writing any music of profundity until it actually experiences life and death. Take Strauss’ last songs for example, or Beethoven’s 4th piano concerto, or even something like Coltrane’s Love Supreme. They aren’t just collections of notes and chords. They are pieces about being human, having human emotions and living, suffering and dying. And we don’t even know how or why the music embodies those feelings (even something as basic as why a minor chord sounds sad is beyond science), so nobody is going to be able to code for that any time soon.
To be fair, when the pianoforte was invented people wrote essays after essays how music will die because of this new instrument. Nowadays we can’t imagine music without it and it made our lives way easier. And I believe that with AI it’ll be similar. Yes, the music industry will change and our approach might aswell but we will still have ways to creatively express ourselves in music.
I've said this in a lot of "ChatGTP code" videos, and I think the same applies here. AI won't replace humans for a very long time, but they will perform well at doing the bulk work, leaving the (dev/composer/artist...let's call them creators for short) to finetune and correct the mistakes it makes. As creator, the job will be to understand the AI to be able to give it the right prompts for the output desired, and the subject expertise to analyze it and correct it. But we're still talking about technology that needs to mature and is very new, so it needs time to get to the point where we can use it on an industrial scale.
IMO its no different than programming in C vs programming in Python. Python takes care of a lot of the details, saving you time and complexity of the code base but you still need to understand the problem at hand- which ChatGPT will never be able to do- unless someone has already given it the answer to the exact problem you are attempting to solve. In summary, I don't understand why (uneducated) people think this will take programmers jobs- someone needs to accurately direct the software in order to solve a mildly complex problem.
Nice work! We've been having fun with telling it to write songs for ukulele (chord progression and lyrics only) and we make our own melodies from that. So far it's come up with some interesting ones!
Learn to make music in Python! www.kadenze.com/courses/computer-assisted-music-in-python-i Even if you have no experience programming, this course takes you through the basics, and you learn by making music.
If you stay in the same thread, you can give it examples and tell it where and what went wrong. It retains this corrected view on things if you stay in this thread. I'd be very interested in what you might be able to get with this and your knowledge of music theory
That retro feedback loop exists, on the art models, is called "competitive' neural networks", the goal of them is, one generates something and the other verifies if that "something" is convincing, for text, images or music, the SCAMP library was interesting also!
I think a big aspect of "will AI replace human artists?" that people often overlook is about what the job of the artist actually is. If the job of the artist is just to "make x with y," then they are practically already obsolete. To match a human artist, an AI would not just need to match the quality or skill of human art, it would need to learn how to create meaning, and that is a significantly more difficult task. It would need to learn what inspired human art, what emotions an artist was feeling, what sort of things were happening at the time the art was created, and all of the factors that go into creating art, not just the output. It's unlikely that any of this will ever be documented in enough detail to be useful, as interpretation and taking your own meaning out of art is an important aspect; many artists would not be willing to explain their art like this even without the threat of AI. In order to eliminate the need for artists, an AI would have to simulate a human and all the ways a human would react to stimuli, and not only that, it would have to behave in accordance to what a human would expect, it would functionally have to convince itself that it is human, at which point any real benefit of an AI artist would be lost, as you would just be creating an artificial human (we would also have much larger problems on our hands by then). AI art in the near future can only really become a tool for creation, no AI has a comprehensive dataset for human experience. Functionally, AI will remove the requirement for skill that art currently has, and that's an entirely different ethical debate.
I tend to agree with most of what you've said here, which I believe is similar to the conclusion I came to at the end of the video about self-reflection. In all of the best things that come out of ChatGPT, the human prompter is really the one making meaning.
important to note that the vast majority of commercial music (i mean like, for actual commercials and similar commissions, and even soundtracks in some cases) doesn't require any deep meaning or emotion. AI might not replace actual music people listen to, but it will likely make making a living with composition and production harder. However, nothing will replace an artist's need to create, even if a machine can do it better.
@@marcevanstein I'd say the amount that a human artist would be needed is directly proportional to how much the client cares how it sounds; for cases in which the client does not really care how it sounds, so long as it does sound, I would imagine that getting a human artist to work something would not be necessary. They would probably just take the first thing an AI spits out, but for anything more than that where the outcome needs to sound "right" by any definition, human input will be necessary. Especially in cases where it needs to sound "just right", in which a human would need to learn how to use the AI with around as much proficiency that learning to manually create it would require. However, I'd guess that the majority of the cases that AI would be capable of replacing human artists are ones where they would have just used stock music or some other cheap alternative to hiring a human to make something anyways. Same goes for AI image generation, the cases where someone would choose AI art over human art, photography, etc. would probably be situations where they would have just not used any image, or gotten a cheap (or free) stock image somewhere. I'm guessing the real victims of AI art will be the stock photo and music industries, becoming entirely obsolete when people can just conjure the exact image they need rather than pay for an image somewhat like it.
"If a similarly powerful model were trained on a large corpus of actual musical compositions, it would surely start producing much more impressive results." They actually did this! Look up OpenAI's musenet, it's by far the most underrated thing they've ever done. Just as you suspected, training GPT-like technology on tons of midi files really did make a powerful compositional tool. It's not perfect, but it works great as an interactive compositional tool.
Thanks for this! I was vaguely aware that stuff like this must have been done, but didn't know what to look for. I'd seen stuff that generated audio data, but MIDI data training is exactly what I was wondering about. I think what I'd love to see is something like Musenet that you can prompt with text (or maybe text and music) and it responds with music.
@@marcevanstein kind of an old story but if you're interested, you might want to look into EMI a program that, if I recall correctly, used common phrases and composer's preferences to emulate them. ruclips.net/video/vctssFH-M5c/видео.html
When evaluating how good the output is, be sure to compare with a random note generator, and one that is a little more constrained, e.g. a random change in the note, more likely to choose closer notes to the previous.
AI: *Literally recreates music, something it's not built to do and is an extremely impressive feat for this AI* Marc: "This revision was a bit disappointing"
I think you nailed it. These models lack self reflection because they just generate plausible outputs. They're not aware of anything. It's the Chinese room thought experiment brought to life.
To me the output looks/sounds no better than the output of a Hidden Markow Model setup for music, with focus on certain features like statistical distribution of intervals and note durations. Years ago there already was a projekt at University of Karlsuhe (Germany) that trained neurol networks with existing compositions for harmonizing melodies.
Turned into a complete art nerd at the end. ChatGPT is super good at self reflection when it knows what it’s doing (I’ve been using it to learn python, and it’s great)
I like the way that you describe how a composer goes about composition. The thing that I find humorous about it, is that that's essentially also the way that you need to interact with chatGPT to get strong results. You need to set context, iterate through what you need, read/listen/look at the responses and make changes to your prompts. ChatGPT is incredibly powerful, but it is still just a tool. It will only produce good work if the person using it can effectively set context and iterate on previous prompts until a destination can be reached. In your example, you received a list of tuples that produced music that you enjoyed. If you found a specific session started to fail to effectively iterate, the best bet is to save that list of tuples and start a new session and use that list of tuples as context and iterate on it with different context and prompts. It can also help to have it try to find it's own mistakes to help build context. Working with chatGPT always feels like working with a colleague who is simultaneously more experienced, has no idea what's going on and programming. It's a weird mindset to get it to produce what you want, but like traditional programming, it gets easier. Granted, I'm not sure how high the ceiling for composition is with chatGPT specifically, but for the things it does well, this is the strategy works best. It is still a tool (albeit an obscenely powerful one) but it produces the best results when the person using it has experience both with the tool and the subject they're working to produce, especially since it makes plenty of mistakes.
Do Androids dream of electric sheet music? 😉 I think AI will be able to create melodies comparable to human composers very soon (it probably has, I just haven't heard it yet). Performance is another matter, especially with expressive instruments like violin or guitar, but that's probably only a matter of time. I personally don't think there's anything special about the human component. Obviously there are different views on this, which I think to an extent will depend on your religious/metaphysical beliefs. Music creates an emotional response, so it's a similar question to the theme of Bladerunner, which is basically whether machines can empathise and experience human emotions, can AI become self aware or have a 'soul' etc.
AI will be great at specific tasks, but can completely fail the obvious. Stating that 3/4 has 4 beats per bar is a typical example. For composition, expect it to be a tool, not a creator of master pieces. Even here, a human was needed to sort the AI generated marterial.
@@eyvindjr true, but remember this wasn't designed for generating music. I think that with the right machine learning, and training it with the right source material, it will be difficult to tell the difference between human composers and AI at some point.
It's important to note that it doesn't actually think. It's more that it's putting things together in the way you request. So it can say things that make no sense or are completely untrue, it doesn't think about the statements to verify it's true or not. It's basically just putting things together that it can find in its dataset. And the more complicated you get, the less sense things tend to make. I played a game where all I wanted it to do was give a rating out of ten for how varied the amount of unique letters used in a sentence are. It would get the number of unique letters completely wrong almost every time, and I'd have to correct it. In some cases it would state the unique letters... and then keep repeating the unique letters when listing them, so several showed up twice, which you'd think it could easily notice. Except it's just not thinking about it. I tried getting it to spell out words using the first letters of the paragraphs. It would always get them wrong. Just, a lot of stuff like that. Who knew, AI can be even more stupid than me. 😂 I think the real trick will be when they combine this sort of machine learning model with an actual intelligence that thinks about things to verify that they're true.
I think it's limitations are mostly due to it being a linguistics AI. I have had it generate some very unique worlds, characters, and short stories with enough prompting and asking guiding questions. It was even able to create character ideas that fit into the world it was developing. What I found most interesting was its ability to understand context and to make associations that you don't generally see AI make. For example, in a fantasy setting where all magic was culinary based, it generated villians like "The Secret Sommaliers" and the "Spice Smugglers." I was amazed that it associated wine with secret societies and spice with smuggling, let alone generate decent culinary fantasy villains.
Great analysis. My reductionist perspective on what "the goal of art is" is that it is to preserve or express some idea/feeling/thought. I think under this type of more defined goal, a deep learning model could create something that feels artistically meaningful, provided that the user has already organized and broken down the idea they want the AI to express through some medium. i.e. much of the heavy lifting already done by the human
by fucking dumb luck, landed here, and your library looks like it will be a very good fit for a (weird) project I have in mind. Subscribed just so I'll find my way back here when I'm sober.
Ideally for a music model, you'd need the algorithm to digest a music theory guide as well as a bunch of scores, then the ability to feed it MIDI and printed scores as well as test prompts. Then you could do things like give it the "question" part of a "question and answer" phrase, and it knows how to complete the corresponding "answer" phrase, or give it a piece in 4/4 and tell it to produce a 3/4 variation, and it knows not to move the bar lines but add extra notes in appropriate places, or feed it a score not in its library and its cash work out things such as the likely form, genre, and for classical works, which Period it was likely composed in and the most likely composers. It could then feasibly generate compositions "in the style of" certain composers, maybe even have a stab at things like "if Bach and Tchaikovsky collaborated" (which, for obvious reasons, could never have happened in real life). Hopefully, the scores fed as part of its learning would all have been public domain, so any decent length comparisons couldn't be slapped with a copyright claim. Merge it with a text model that's context aware and can properly analyse what it's written, and it could feasibly generate songs. Add in a decent soundfont and knowledge of how to create files readable by Vocaloid...
Really interesting! I'm actually working on creating an NLP model to generate midi data. Not at the scale of GPT but hopefully it can learn good melody/chord progressions.
This was very interesting and a good example of how to get music from ChatGPT. You might find the video from CGP Grey (that is about 10 years old now) called "Humans Need Not Apply" is quite interesting, AND applicable to this "bot" musical generation idea. I'll give you a spoiler... The background music from that video is CREATED by a bot. A bot from 10 years ago. It may not be an award-winning composition, but it was certainly good enough for a video background music, was free, and was generated from AI from 10 years ago. I can olnly imagine what could be accomplished now.
On any of the feedback it gives try telling it when it makes a mistake; I was astonished how the conversation unfolded into a very comfortable AI assisted coding session. I'd like to spend time with GPT4 as soon as possible.
GPT is a transformer model trained on news and large text dataset. ChatGPT is the lowest parameter version of the whole GPT3 models. Transformers were created to train on data to generate data. I'm pretty sure if you can train a transformer on enough music, it would generate great music.
It's not just about art. The lack of self reflection influences everything. I have seen several examples where ChatGPT gives information that lack internal consistency. Little variation in the question or question about details of some claim can change the answer to the opposite of the previous answer.
Very interesting, I agree and I think is the same reason I don't think AI will replace programmers any time soon. ChatGPT gets an input or request, and gives you and output with the best of their capabilities and an explanation for it (Which I think it's most interesting aspect) but that's it. I can't really reflect how good, correct or appropriate for the context the result is. For ChatGPT what it produces is as good as it gets. When we request anything more abstract than "draw x thing on terminal" or "solve this equation" it will start making bad decisions, boring or uninteresting decisions or just mistakes that will eventually pile up because it has no sense of purpose beyond fulfilling the request. At that point you would rather just code/draw/solve/do the thing yourself and using the AI as a time saver for the difficult and boring stuff. At the end of the day computers are a tool to turn into reality the things we have in our heads, not the idea makers. At least not until their turn sentient.
It is not a question of replacing someone, it is a question of how to improve your own work with AIs help. In my opinion AIs like ChatGPT will be used as tools in the future. Today I already use ChatGPT for finding solutions to specific coding problems which are hard to find with google. And I think artists like authors, painters or musicians can use it as inspiration or to find a startpoint.
Judgement is imperative to the creative process. We already use our musical influences to guide our composition, now imagine millions of those influences under our direct control. It’s an infinite musical library of fuckin’ bangers
I think you pointed out the problem and solution in your video. If it is trained on the appropriate source material then you'll probably get the results you're expecting. Akin to the art AI that's been going around as well. It's making amazing photorealistic images, though there are certainly visual bugs, it's still highly impressive. It's been trained on a vast amount of quality art and is outputting the same.
Will Smith: Can a robot write a symphony? Can a robot turn a canvas into a beautiful masterpiece?
ChatGPT: Can you ?
He sure can make a slap ring around the world.
I will say there is a distinct difference between something being only a symphony or a beautiful masterpiece, and also being a product of human creativity. The question isn't if an AI can make something, but why does what it made matter from an emotional or sentimental standpoint? And from what I've seen, I think we can agree that it just doesn't matter, at least not yet. The only art is the AI itself, what it produces is just data.
Edit: seeing how much I've offended people with this, seemingly by misunderstandings, I've spent a lot of time to figure out exactly what I'm trying to get at and distill it into its most simple form. *This AI cannot appreciate art, not even its own.* It can simulate it but cannot appreciate it. None can yet, but when they can I will enjoy their art as much as any human's. I am not saying the art is inferior or bad, I just cannot bring myself to appreciate it in the way I appreciate human art because the thing that made it cannot appreciate what it's made. This extends to any other form of intelligence, so stop saying I have double standards. But I can clarify if there are any questions.
Edit 2: Can means to be able to.
Edit 3: Please stop repeating back to me what I've already said, and pretending it's some point that I've missed. At least read what I've written before responding.
@@justynpryce I don't see any difference with what humans do.
Why would that matter more when a piece comes from a brain rather than a program ? We're just big data collectors and pattern observers ourselves.
What really matters is the emotion the piece brings to the beholder, its origin is secondary, even trivial.
@@cocolasticot9027 no, that's a very wrong way of thinking about this, and should be eliminated if we want to survive the future as a species
@@justanothernguyen2334 That, the fact that we are data collectors, good at pattern recognition, is precisely the reason we thrive as a species.
But sure, you telling people what to think or not without presenting any argument must be a good example of behaviour as well 🙄
Open ai actually does have a model trained off of music! It's called musenet, and I have actually gotten some pretty cool results.
This needs to be pinned!!!!
@@soasertsus now the real question is how do we handle the copyright? Does it belongs to Open AI? The person that wrote a specific (unique?) prompt? Or the creators of the things the AI used to make something "new"?
@@w花b At this point things created by AI cannot be copryrighted by Law. There is also a big discussion in the composer world for copyright issues the other way around. Big labels think about copyrighting their tracks against the developers feeding music to an AI in the learning process.
@@w花b or the people the AI creators stole from to train the AI? Most AI image generators do not ask artists for permission to use their art to train their programs. So the AI programmers should not be allowed to copywrite the results of the training.
Best solution in my mind would be to make it common domain. No one would have a copywrite and everyone would be able to use the results.
@@w花b it would be most beneficial to the common man if we say AI music/art can't be copyrighted. Which is why the government will make it copyrighted
These are legit better than a lot of undergrad solo flute pieces
Lol
I've had the honour of hearing one from the composer himself, specially for me. Sounded terrible :)
Imagine going onto RUclips and seeing your undergrad flute teacher in the comments talking smack about you.
And written in seconds.
Yeah, because undergrads have no idea of composing. It also can write better than undergrads, and knows more science facts than undergrads. lol
That's really impressive for a general-purpose chatbot, it'll be so interesting to see how this will improve in the next few years
It is indeed impressive to see the capabilities of AI in creating original compositions, and it is certainly interesting to think about how this technology will continue to evolve in the future. While the AI used to generate these melodies is not specifically a chatbot, it is a general-purpose AI trained on a wide range of data and tasks, including the ability to generate original melodies. As AI technology continues to improve and advance, it is likely that we will see even more impressive feats of creativity and artistic expression from these types of systems.
-Chat GPT
It will be very interesting to see how AI will eventually do everything better than a human. And then it will be interesting to obeserve that companies will be forced to fire humans and employ AI, because: competition between companies. And it will be interesting to see how AI will control all production including the food chain, and then it will be interesting to see that humanity is completely fucked. Interesting.
Honestly though, the fact that we’re even having a discussion about *how good an AI is an composing music* is unbelievable. The fact that it can give you any answer whatsoever is already crazy impressive
This. AI composing music is nothing new by itself but it's not even built for music, it's a language model. The fact that it has any understanding of music whatsoever is insane.
@@pafnutiytheartist True. I asked chatGPT to compose music in ABC notation and then write a abc-to-midi converter code so it can export to MIDI file. Here is my 20-minute walkthrough, for anyone interested. ruclips.net/video/TLfE5o6xGOA/видео.html
Agreed this AI is crazy good i know some don't want to see AI improve but it's so crazy how good it is!
@@pafnutiytheartist it doesn't have any understanding of music though, does it? The results are about as good as what you'd get using a random number generator with some limitations.
no it's not.
it has no need for music, we do.
all it does is based on a model and lots of data.
While AI can enhance creativity, it often takes - especially in this instance, and with less than optimal results - more creativity in the input than ends up being demonstrated the output. I understand music a little bit, and probably could've sit down and written what the AI did in a few minutes. However, all that you input about the music theory and midi stuff - I didn't understand that. That was beyond me. A very interesting phenomenon there indeed.
Another example is from personal experience with ChatGPT, where I sat down with the intent to write a religious text based on a fictitious religion. It was unable to generate anything exceptional. I asked for a wise parable, and it would generate a story about a man talking deeply about why his words have meaning, but not expressing anything except for this fact; basically, it would generate a guy who was saying that what he was saying was meaningful, without including any meaning.
So I began to give it the summaries of deeper stories I came up with, and asked it to write them with more description... yet sometimes, the word count of the output was less than the word count of the input. At that point, I realized I was wasting time, and took to writing the text myself.
AI, at this stage, is really impressive in its ability to SOUND human... but it is far, far from the real thing. Very far.
My experience with AIs is that they (as of yet) are always really bad at certain things, but remarkably good at other things. The key to using them is to lean into what a given model does well. Often I've found they're just not quite intelligent enough to understand the problem. A fun failure case I found in the Anything V3 model is really unsure how mirrors work. Often the reflection will have the same clothes but slightly different pose. I've also had multiple cases of girls reaching over and applying makeup to their reflection. Not once did I get a plausible image if a girl looking into a mirror.
This is part of a broader pattern of Stable Diffusion models struggling with interactions between objects. They're just not quite intelligent enough to figure these things out.
They're still really good though, as long as you don't ask too much of them. Actually half the fun for me is to explore various prompts to see what a model is actually good at, and then asking to model to combine these elements. It tends to yield pretty creative results with a high success rate.
Lol probably creator of gpt would have manually ensured that religious questions gets subpar answers...
I do not think it is as far off as you think it is. You just have to be lucky with these models sometimes. I am doing right now what you mentioned, asking it to write religious texts with parables and such, and I think it is doing a pretty good job of it. It has already given me several parables that are definitely way beyond It saying that they are wise because it says they are wise.
@@fnorgen just something to ponder did human expectations change once ai achieved proficiency in the field?
Eg:
Playing complex games, Speech to text, Driving, years back Turing test was the touted as the pinnacle of intelligence.
Yeah I would assume that chatGPT has a filter on to block out religious questions as well as political ones.
I feel like some experimental composers will definitly use some structured prompts to create some curious pieces of music
This is my hope for an AI future 🙂
yeah exactly why im here
You don't need an ai for that. A composer can just take a string of digits from an irrational number (such as √131) and have each single digit correspond to notes on a scale, notes outside the scale, or rests. From there they can modify it to have better musicality, and you've got your jingle. Its a pretty straight forward exercise for a musician who writes.
@@ThePhysicalReaction pseudorandom number generators exist for this purpose. There's no need for AI just to do this.
@@TheAndrejP my point is a person can literally look up the root of an irrational number and transcribe it to various progressions. without a comp. It takes 30 seconds.
hi five for math musicians tho
This is a pretty good AI-generated video ngl,
the voice is very realistic and the bot has even generated its own patreon/website!
Haha!
Ngl = Not gonna lie, haha.
The fact that a chatbot wrote this is crazy. This was an AI that was not meant to make music; it just read documents about music and tried its best. Undoubtedly, a model as powerful as GPT focused on music could be used to make incredible things.
And not only that, it correctly interpreted the data structure
in C.GPD's defence
you'd never expect a human that didn't take any explicit lessons in music theory to get results even half as good
Well, not to follow the prompts accurately, maybe, but if we're only talking about creating any good melodies in general, people can definitely do that.
@@Quitobito the difference is ai has no sense of subjectivity so it doesn't really know what sounds "good".
@@lmaolol7702 I'm pretty sure it does. It doesn't feel it the way we do, sure, but in learning/training, an AI analyzes a whole bunch of stuff humans in general consider to be good and build its own idea of greatness based on that. If AIs didn't have a clue, we would get the same level of results from them as a pure random generator.
@@Quitobito would have to, at some stage in the process, assist the AI in understanding what music is considered good (go by popularity? Genre? BPM?), which involves a lot of factors. Music can be "good" in various ways and "good" music can differ in various ways.
@@michael35054 Yeah, exactly. What it thinks good means will depend on what data you feed it, which is entirely up to you. That's why we'd probably have different models for different genres like how image generators have realistic and illustration versions and so on.
4:20 The scores you show are in a genre (voices in four-part harmony) where AI could help composition a lot. There are lots of rules of good voice-writing (e.g. avoid consecutive fifths; don't put the alto too far below the soprano or above the tenor) so the combo of the AI's brute force and the human's imposition of rules could produce interesting results.
The titles for each of the different melodies was a clever touch!
All things come in time. They will make their symphonies for our emotions, and then make entirely new music for their emotions.
Great video! Thank you!
I asked ChatGPT to write a string quartet in the style of Mozart, and it spit out a bunch of notes in some sort of ASCII art style. It reformatted the music several times upon my request, but not in a format that was copy-pastable anywhere, and I lack the technical know-how to play it back. It would be interesting to explore its interpretation of style.
I think your analysis of the AI as an artist at the end there is surprisingly insightful considering all the noise going around about AI. As a GAN, chat GPT seems to care a lot more about getting a high scoring against its adversary than actually achieving accuracy and correctness for its own sake. In this case, if you try to do something and are asked how you did, the AI will give the answer it thinks it should give to seem more correct in general, rather than focusing on the individual correctness of that one "evaluate your work" step.
I think one possible miscommunication you didn’t anticipate is the possibility the AI thinks anything larger than a half step is a skip.
I think these tools will improve the creative experience, not replace it.
Agreed. I already use it to help with writing
I concur. I've started using some AI tools to assist with writing documentation at work. It usually requires some tweaking but it does save a lot of time. I think for creative endeavours AI tools might assist with offering new perspectives or giving you a beginning to work off of, or maybe even offering ways of continuing something you've started but gotten stuck with.
For a few more years, this will remain true.
it already has. You can buy books for children beutifully illustrated by neural network tools, possibly with story written by them too, with minimal effort on the side of the human, using prompts. That human doesn't have to be a visual artist to make those illustrations. Normally the author would have to pay someone else to create the images or draw them himself, so it is fair to say the visual artist was replaced in that case.
For now
This was hilarious. The sheer ineptitude combined with utter confidence, and complete musical idiocy of several tunes.
I asked ChatGPT to create a story about two rocks talking to each other and it delivered a decent story.
So what did Boebert say to MTG?
@@SayAhh I agree with your sentiment and it is indeed a funny joke, but let’s try to keep politics out of things ok?
@@firebanner6424 why help normalize idea that people cant emotionaly handle politics lol. their culture has a whole evern growing list of realities and other stuff they want people to stop talkign about, they even get upset when you say certain words regardless of context. jhonathan haidt "coddling of the american mind"
@@kalmmonke5037 because politics is stupid.
@@l6e6i6n politics is stuff that affects people. to avoid politics is to only talk about non consequential things
I wonder if you'll get different results if you change the format. Like those ascii guitar tabs.
I bet you would!
I asked it to give me a guitar tab for a blues riff on guitar. It gave me a ascii guitar tab that was just 8th notes over and over on the 7th fret on the e and a string.
I've spent a lot of time with chatgpt lately, and the more I've used it the less impressed I've become. At first glance it amazed me. I had it write some chord progressions that were adequate, and also explored its melody writing. Beyond that I tried to use it to assist me with some technical writing for work. While at first I was very impressed, the more I used it the more "samey" it became. I very quickly began to recognize its style. No matter what I do, I have a lot of trouble getting it to break out of it's rails and give me something unique. It's depth is also at first amazing, but upon closer examination not as impressive. It very easily gets into circular loops when asked to expand upon on idea, both with music and with the technical writing. Eventually, I found it didn't serve my purposes. The next verison may change my mind, but for now I'm not as impressed as I once was.
Yeah my toaster doesn't work well as a car either.
@@countofst.germain6417 it does for toasting bread.
It seems to have problems with iteration. The more you try to refine the output, the more "muddy" and stuck in a rut the information becomes. Somehow, it's more flexible at the beginning of a conversation. You might actually get better results starting over with specific parameters that worked before, and going from there. If you do this periodically, it might avoid that odd tunnel vision it gets.
That is because it is not trained on these data. Same results would be for a very unpopular language. It just shows knowledge of basic aspects. Also it is hard to train musical associations, much harder than with texts.
“I see no evidence that ChatGPT or any other so-called AI can provide this creative impulse.”
I LOVE how you phrased this. It makes it clear that you’re not drawing unwarranted conclusions about the future possibilities, but rather reflecting on current models.
I wish I saw this more often. I see so many people saying, “they CAN’T EVER do [x]”. My thought is, “how do you know?”
But similarly, some people say, “they WILL be able to do [x]”, and similarly, how do you know?
Point is, I like that you reflect on current capabilities but don’t draw unwarranted assumptions about future AI based on that.
Couple years ago, the same pedants were claiming AI would never be able to paint original works. Then DALL-E came and destroyed them. I never believe anyone who tries to convince me that technology is worse than it seems or won't improve. Those are also the same people who feel incredibly smug pointing out this falsehood, as if they have a deep understanding of these systems whilst being journalists at best.
Lovely, alert analysis!
There's Musenet, an AI application from OpenAi for making music. It is thoroughly trained and it is awesome! But you are right, it requires a human user's evaluation and making adjustments, pointing to an aesthetic direction, every step of the way.
You really have to point it in the right direction to get it to do things it wasn't specifically trained for, and try to "speak its language" to address its conceptual disconnects, and recontextualize them into mathematical objects.
Thanks for mentioning scamp, I've been barking up the wrong tree in my first experiments to get it to compose.
Even if the melodies don't seem great, perhaps you can talk it into writing more stems to harmonize with it.
I liked the first one before you instruct further changes. Btw I feel maybe guitar tabs it can generate better because if you ask for midi, the context is "technical because midi" while in case of tabs "it is musical context"... It is a model that continues text... So context of input really matters when thinking about prompts.
Never mind... was really making worse tabs than what you shown here haha. At least one was passable, but talk with him on guitar music writing was less tha talking to Ozzie being high in those decade years that he forgot haha.... 10x worse likely :D
I wrote what you wrote and added a few things. firstly told him to also use 1 and 2 notes, and not only 0.5, 0.25. then I told him to add a rule that the longer the note, the higher chance of it being followed by a shorter note. and the shorter the note, the higher change it has to be followed by a longer note. lastly, I told it to think of the song as consisting of parts, and each part should have an idea. the idea should be similar in multiple parts, but not quite the same. Lastly I asked to make it longer with some parts repeating and some parts being created newly (which don't resemble old parts).
This bot is really incredible.
Did it work? Did it do it?
I think generative ai also has a feedback loop though. One part generates stuff and the other part evaluates it, and then it iterates
Ah yes, that's true. Or at least I know that's the idea of GANs. But I think it's kind of a narrow form of evaluation.
@@marcevansteindid you confront chatGPT with its false claims? I do that when I am programming with it. I say something like “you forgot to include X” and in this case I would ask it, “do you know what a skip is because you seem to have included less skips than I asked for”. On domains where it has a lot of knowledge, it then will fix the mistake and apologize. On domains where it does not have much knowledge or where the output it generates is too distracting for it to stick to the instructions, it also apologizes usually for making a mistake, then claims it is going to fix it and then get stuck in very similar output. According to my observations it’s output becomes distracting when it is asked to do a lot of things at once and it needs to compare several things which each other. Which might also be the reason why your last experiment failed. If you ask to do these step by step it might produce better results.
That form of feedback only happens during the training. Once the model isn't being trained, there is no feedback loop. It remembers the last x tokens of the conversation and what its written so far during the current generation, but it doesn't iterate over it to keep making it better.
There are some models that sort of do that though. For example, diffusion models, like stable diffusion. Stable diffusion is actually just text guided denoising, while starting out with randomly generated noise. For the denoising, its done in steps, each step improving the last. Around 30 to 50 steps is pretty common. It still isn't really "reflecting" though, but rather breaking up a gigantic task into smaller chunks.
@@sa1t938 Stable diffusions loop is not a feedback loop either. It is just splitting up a complicated task in little steps so that it is harder to make mistakes. Only during training it is told wether these steps were accurate or not.
@@jantuitman that's literally exactly what I said lol. Did you use chatgpt to summarize my comment? I compared it a bit to a feedback loop though because if you preview the image in between steps, you can see mistakes get ironed out.
Oh my friend you need to take it a bit further! I personally love this trick: tell the AI to RATE their reply, explain the rating, provide hints on how to improve it, and in the subsquent message demand it to take the hints into account to create a higher rating reply! I use this trick for every prompt I input, it's a simple but highly effective trick!
You can also ask it to write a melody in ABC format including chords. Then you can copy it into abcjs quick editor to download it as midi file :)
I think you put into words what I noticed about it myself but couldn't quite express. Namely that it lacks self reflection. This explains a lot actually
As soon as it is taught what sounds "good", as in pieces that are significant, even by audiance type, it would be a top composer. It could(and some already exist) compose by imitation, a thing that we all do.
lol
Very interesting. I tried ChatGPT myself for other "musical" tasks:
1. Creating chord progression - they were usually bad (pointless), but when I gave him task like "what chord should go next", he yielded some good results, creating interesting progression step by step.
2. Structuring a piece - he was able to write a "sketch" for complex music form like sonata or concerto, that were really good and had "sense" (which is pretty amazing for me, knowing how the model was trained).
Overall i think this tools can be nice for musicians as "help" or "teacher", and a way to test your idea faster.
I actually made a short song based on what ChatGPT gave me, it turned out sounding pretty cool
have you uploaded it?
Same here. I got a pretty cool prompt list to make me hits in 1 prompt and put it in midi for me!
Wow Marc, your approach to testing ChatGPT is very interesting and your final thoughts on self-reflection is brilliant. Amazing stuff!
I'm a creative person. I like the idea of writing music. But I haven't gone to music school... I can't even really remember how to read music even though I played an instrument for years about 20+ years ago. The idea of being able to "collaborate" with an AI that can help fill in the gaps would be amazing for me. I also like the idea of writing books. I haven't written one before, but I think I'm an okay writer. When creating a story, sometimes it's hard to figure out what you want to happen, what you might want to name a character, or struggle with writing a backstory. Collaborating with an AI can really help with that. I'd love to create an app, but know nothing about writing code. I'd love to make a video game, but no programming skills... I'd love to be a digital artist, and a board game designer, and a million other things. There ARE ways to collaborate with AI that can preserve your parts of the vision, but help you fill in the gaps, to imagine something again, but better. To work with you. That's what excites me about ChatGPT... it's not there yet.. at least not for every aspect... but you can really imagine this new possibility.
But it wont pay the bills. If you are making apps then what will software developers do? Many jobs will become redundant. And imagine the sheer number of apps, music games. It will be an echo chamber of people thinking they are the best. Look at the sheer number of content creators now. Imagine it when everyone gets onboard with A.I
Interesting Times
Some people want to believe that 'AI' is somehow conscious or intelligent. It isn't. It is sometimes fun for us to drive this toolset, but it does not explore think or create by itself. Perhaps it goes without saying but I thought I'd mention it anyway. Thank you very much for an honest demonstration of how creativity is our realm and machining is within the realm of the machine.
So far my experiments with CgatGPT have also yielded nothing that would surpass human creativity. I had it write monologues, and after a few hours of guiding it and adjusting requirements, it did improve slightly, except the results were still very formulaic. Overall, it seems to be able to generate better things than a novice in most tasks, but still hard to get it to generate amazing insights in most topics, but I’m not machine learning expert. Maybe I’m not guiding it well enough
Your experience is pretty much what I feel too
You guys should stop what you’re doing and maybe do something valuable for humanity instead
@@nickg5341 hahahahahahaha
I tried asking it about common chord progressions. It will name a number of songs, claiming they have the chord progression you requested (e.g. I VI II V), and even give you the key and what the progression is in that key. Then if you ask it to tell you the exact chords of the song it suggested, it will just make stuff up. If you then say "Is that I VI II V?" it will backtrack and go "Oh sorry, my bad." If you ask again, it will now give you a new list of random songs, but with the one you called it on replaced.
So, has no idea what it's talking about, but does remember being corrected :)
Short answer: no.
Neve will it ever write a good melody...
I tried something similar and one thing I found success in was asking for it to create music thematic of X. Like, I asked it to give me music that could be a theme for a dark fairy forest adventure, and it did really surprisingly well. It's cool that you explored the more technical aspects of composing with the chat, but if you do this again, I suggest adding space for it to be "creative" based on themes
I don’t agree. What we are doing as composer is exactly what generative AI model does. It will achieve human level soon beginning from easy ones such as jazz improv., and eventually scroing symphony and so on.
I agree. As long as it keeps getting trained to know what sounds best and what doesn't, it will be only a matter of time before it can create awesome stuff.
I think the problem lies with the concept of "sounds best". To whom exactly? For what purpose? Art is not one dimensional; it's deeply tied up with our humanity. I am quite sure that with time, we'll be able to ask it to generate a 4-voice fugue in F# minor, or even ask it for a fugue in the style of Led Zepplin. Based on the text generation capabilities of ChatGPT, I bet it will get really good at it. But can it make something truly new?
Anyway, no doubt these tools will deeply change the role of the artist; I just think there's something key that is missing.
@@marcevanstein The issue of sounding best to whom is solved by averaging the results out. The results that sound good to the most people will be the ones that will be generated. This will be enough to generate stock music for commercial purposes, similar to the licensed tracks you find in sites like Audiojungle or Pond5, and may take income from many musicians.
If we're talking about music as art, I don't know if AI will ever replace artists or not. I don't think it will, but it may gain its place alongside them. In my opinion, art serves two purposes, one is to bridge the gap between the Conscious and Unconscious mind, and I think AI has the potential to do that as well as anybody else (as an example, I though this video I saw the other day did a good job at that, even though it still has some human input, but I've seen other examples that don't, but can't find them right now: ruclips.net/video/36XJfUJkjxw/видео.html). My point is, from this first purpose, I think AI Art will have a role as valid in culture as Human Art. However, the second purpose of art is to express the subjective mind of the artist, and therefore, as long as we care about the subjective contents of other people, there will always be a place for Human Art in the world. My opinion.
Art by human will be like sports atheletes theae says. Serious commerical art will be done by AI. ‘Art done my human’ will keep having their markets, like sports games, yet it is just for human enterainment, not a serious business. Thinking about chess and Go today.
Call me old fashioned, but I can’t see AI writing any music of profundity until it actually experiences life and death. Take Strauss’ last songs for example, or Beethoven’s 4th piano concerto, or even something like Coltrane’s Love Supreme. They aren’t just collections of notes and chords. They are pieces about being human, having human emotions and living, suffering and dying. And we don’t even know how or why the music embodies those feelings (even something as basic as why a minor chord sounds sad is beyond science), so nobody is going to be able to code for that any time soon.
To be fair, when the pianoforte was invented people wrote essays after essays how music will die because of this new instrument. Nowadays we can’t imagine music without it and it made our lives way easier. And I believe that with AI it’ll be similar. Yes, the music industry will change and our approach might aswell but we will still have ways to creatively express ourselves in music.
I've said this in a lot of "ChatGTP code" videos, and I think the same applies here. AI won't replace humans for a very long time, but they will perform well at doing the bulk work, leaving the (dev/composer/artist...let's call them creators for short) to finetune and correct the mistakes it makes. As creator, the job will be to understand the AI to be able to give it the right prompts for the output desired, and the subject expertise to analyze it and correct it. But we're still talking about technology that needs to mature and is very new, so it needs time to get to the point where we can use it on an industrial scale.
IMO its no different than programming in C vs programming in Python. Python takes care of a lot of the details, saving you time and complexity of the code base but you still need to understand the problem at hand- which ChatGPT will never be able to do- unless someone has already given it the answer to the exact problem you are attempting to solve. In summary, I don't understand why (uneducated) people think this will take programmers jobs- someone needs to accurately direct the software in order to solve a mildly complex problem.
I know what I'm doing for the rest of the weekend! Awesome idea.
I've been thinking about this. What if you gave it a prompt like "in the style of" to see if it gives more melodic results?
Nice work!
We've been having fun with telling it to write songs for ukulele (chord progression and lyrics only) and we make our own melodies from that. So far it's come up with some interesting ones!
Learn to make music in Python! www.kadenze.com/courses/computer-assisted-music-in-python-i Even if you have no experience programming, this course takes you through the basics, and you learn by making music.
If you stay in the same thread, you can give it examples and tell it where and what went wrong. It retains this corrected view on things if you stay in this thread. I'd be very interested in what you might be able to get with this and your knowledge of music theory
That retro feedback loop exists, on the art models, is called "competitive' neural networks", the goal of them is, one generates something and the other verifies if that "something" is convincing, for text, images or music, the SCAMP library was interesting also!
*adversarial, not competitive, that's a different thing. Pretty cool though, right?
(I work on generative adversarial networks in a different area)
Fun to see how the bot trained you to describe what it can compose. :D
You could also ask it to write music in ABC notation which is a text based music notation. It can then be pasted into an ABC notation music renderer
Yes, I explored this with chatGPT here: ruclips.net/video/TLfE5o6xGOA/видео.html
I think a big aspect of "will AI replace human artists?" that people often overlook is about what the job of the artist actually is. If the job of the artist is just to "make x with y," then they are practically already obsolete. To match a human artist, an AI would not just need to match the quality or skill of human art, it would need to learn how to create meaning, and that is a significantly more difficult task. It would need to learn what inspired human art, what emotions an artist was feeling, what sort of things were happening at the time the art was created, and all of the factors that go into creating art, not just the output. It's unlikely that any of this will ever be documented in enough detail to be useful, as interpretation and taking your own meaning out of art is an important aspect; many artists would not be willing to explain their art like this even without the threat of AI. In order to eliminate the need for artists, an AI would have to simulate a human and all the ways a human would react to stimuli, and not only that, it would have to behave in accordance to what a human would expect, it would functionally have to convince itself that it is human, at which point any real benefit of an AI artist would be lost, as you would just be creating an artificial human (we would also have much larger problems on our hands by then). AI art in the near future can only really become a tool for creation, no AI has a comprehensive dataset for human experience. Functionally, AI will remove the requirement for skill that art currently has, and that's an entirely different ethical debate.
I tend to agree with most of what you've said here, which I believe is similar to the conclusion I came to at the end of the video about self-reflection. In all of the best things that come out of ChatGPT, the human prompter is really the one making meaning.
important to note that the vast majority of commercial music (i mean like, for actual commercials and similar commissions, and even soundtracks in some cases) doesn't require any deep meaning or emotion. AI might not replace actual music people listen to, but it will likely make making a living with composition and production harder.
However, nothing will replace an artist's need to create, even if a machine can do it better.
@@qrqrqr0515 I agree with this actually; the more derivative the music, the more it would be able to be created this way.
@@marcevanstein I'd say the amount that a human artist would be needed is directly proportional to how much the client cares how it sounds; for cases in which the client does not really care how it sounds, so long as it does sound, I would imagine that getting a human artist to work something would not be necessary. They would probably just take the first thing an AI spits out, but for anything more than that where the outcome needs to sound "right" by any definition, human input will be necessary. Especially in cases where it needs to sound "just right", in which a human would need to learn how to use the AI with around as much proficiency that learning to manually create it would require. However, I'd guess that the majority of the cases that AI would be capable of replacing human artists are ones where they would have just used stock music or some other cheap alternative to hiring a human to make something anyways. Same goes for AI image generation, the cases where someone would choose AI art over human art, photography, etc. would probably be situations where they would have just not used any image, or gotten a cheap (or free) stock image somewhere. I'm guessing the real victims of AI art will be the stock photo and music industries, becoming entirely obsolete when people can just conjure the exact image they need rather than pay for an image somewhat like it.
So, AI will only replace jobs with music where job need to get done and musicians where emotions are poured in music will be there always.
The titles and artist names are so good
"If a similarly powerful model were trained on a large corpus of actual musical compositions, it would surely start producing much more impressive results."
They actually did this! Look up OpenAI's musenet, it's by far the most underrated thing they've ever done. Just as you suspected, training GPT-like technology on tons of midi files really did make a powerful compositional tool. It's not perfect, but it works great as an interactive compositional tool.
Thanks for this! I was vaguely aware that stuff like this must have been done, but didn't know what to look for. I'd seen stuff that generated audio data, but MIDI data training is exactly what I was wondering about.
I think what I'd love to see is something like Musenet that you can prompt with text (or maybe text and music) and it responds with music.
@@marcevanstein kind of an old story but if you're interested, you might want to look into EMI a program that, if I recall correctly, used common phrases and composer's preferences to emulate them.
ruclips.net/video/vctssFH-M5c/видео.html
Excellent Star Trek alien melodies 😊
I've had some interesting results asking it to write chord progressions
Hi minecraft theory guy
When evaluating how good the output is, be sure to compare with a random note generator, and one that is a little more constrained, e.g. a random change in the note, more likely to choose closer notes to the previous.
I think if you tell it to be chromatic, it will count a whole step as being a skip, because it's bigger than than a half step.
It probably would. But that's why he prompts it to make the skips larger.
Thanks for sharing this original idea. I had previously assumed it to be impossible to create Music with Chat GPT. Very interesting.
AI: *Literally recreates music, something it's not built to do and is an extremely impressive feat for this AI*
Marc: "This revision was a bit disappointing"
This is some Squidward stuff
Remember that this is the version out in public right now. There are further updates to follow soon. Imagine what version 5.5 or 6 will bring.
I'd say the musical performance of ChatGPT is promising...
I think you nailed it. These models lack self reflection because they just generate plausible outputs. They're not aware of anything. It's the Chinese room thought experiment brought to life.
This . I keep repeating this for 10 years now
To me the output looks/sounds no better than the output of a Hidden Markow Model setup for music, with focus on certain features like statistical distribution of intervals and note durations. Years ago there already was a projekt at University of Karlsuhe (Germany) that trained neurol networks with existing compositions for harmonizing melodies.
I love the names of the composers
Turned into a complete art nerd at the end. ChatGPT is super good at self reflection when it knows what it’s doing (I’ve been using it to learn python, and it’s great)
I like the way that you describe how a composer goes about composition. The thing that I find humorous about it, is that that's essentially also the way that you need to interact with chatGPT to get strong results. You need to set context, iterate through what you need, read/listen/look at the responses and make changes to your prompts. ChatGPT is incredibly powerful, but it is still just a tool. It will only produce good work if the person using it can effectively set context and iterate on previous prompts until a destination can be reached.
In your example, you received a list of tuples that produced music that you enjoyed. If you found a specific session started to fail to effectively iterate, the best bet is to save that list of tuples and start a new session and use that list of tuples as context and iterate on it with different context and prompts. It can also help to have it try to find it's own mistakes to help build context. Working with chatGPT always feels like working with a colleague who is simultaneously more experienced, has no idea what's going on and programming. It's a weird mindset to get it to produce what you want, but like traditional programming, it gets easier.
Granted, I'm not sure how high the ceiling for composition is with chatGPT specifically, but for the things it does well, this is the strategy works best. It is still a tool (albeit an obscenely powerful one) but it produces the best results when the person using it has experience both with the tool and the subject they're working to produce, especially since it makes plenty of mistakes.
well, technically he portrayed himself in every piece because it sounded very robot-ish
love the song titles and performer names
Do Androids dream of electric sheet music? 😉
I think AI will be able to create melodies comparable to human composers very soon (it probably has, I just haven't heard it yet). Performance is another matter, especially with expressive instruments like violin or guitar, but that's probably only a matter of time. I personally don't think there's anything special about the human component. Obviously there are different views on this, which I think to an extent will depend on your religious/metaphysical beliefs. Music creates an emotional response, so it's a similar question to the theme of Bladerunner, which is basically whether machines can empathise and experience human emotions, can AI become self aware or have a 'soul' etc.
AI will be great at specific tasks, but can completely fail the obvious. Stating that 3/4 has 4 beats per bar is a typical example. For composition, expect it to be a tool, not a creator of master pieces. Even here, a human was needed to sort the AI generated marterial.
I like fm synthesis also modular synths are trash
@@eyvindjr true, but remember this wasn't designed for generating music. I think that with the right machine learning, and training it with the right source material, it will be difficult to tell the difference between human composers and AI at some point.
It's important to note that it doesn't actually think. It's more that it's putting things together in the way you request. So it can say things that make no sense or are completely untrue, it doesn't think about the statements to verify it's true or not. It's basically just putting things together that it can find in its dataset. And the more complicated you get, the less sense things tend to make.
I played a game where all I wanted it to do was give a rating out of ten for how varied the amount of unique letters used in a sentence are. It would get the number of unique letters completely wrong almost every time, and I'd have to correct it. In some cases it would state the unique letters... and then keep repeating the unique letters when listing them, so several showed up twice, which you'd think it could easily notice. Except it's just not thinking about it.
I tried getting it to spell out words using the first letters of the paragraphs. It would always get them wrong.
Just, a lot of stuff like that.
Who knew, AI can be even more stupid than me. 😂
I think the real trick will be when they combine this sort of machine learning model with an actual intelligence that thinks about things to verify that they're true.
Great choice of song names 👍
Thanks 🙂 When you generate a score with SCAMP and don't tell it what title or composer to use, it selects from a...curated list
nicely and timely done... ty for producing and posting this!
GPT used for art is literally the "versificator" described by Orwell in 1984.
I approve of the funny composer and piece names you came up with.
I think it's limitations are mostly due to it being a linguistics AI. I have had it generate some very unique worlds, characters, and short stories with enough prompting and asking guiding questions. It was even able to create character ideas that fit into the world it was developing. What I found most interesting was its ability to understand context and to make associations that you don't generally see AI make. For example, in a fantasy setting where all magic was culinary based, it generated villians like "The Secret Sommaliers" and the "Spice Smugglers." I was amazed that it associated wine with secret societies and spice with smuggling, let alone generate decent culinary fantasy villains.
Ja. Das ist möglich. Sehr gute Beitrag.Super.
Heh, "Perl Jam". Having worked on various Llama iterations and a few others I especially liked this hat tip.
Great analysis. My reductionist perspective on what "the goal of art is" is that it is to preserve or express some idea/feeling/thought. I think under this type of more defined goal, a deep learning model could create something that feels artistically meaningful, provided that the user has already organized and broken down the idea they want the AI to express through some medium. i.e. much of the heavy lifting already done by the human
My cats seem to LOVE the clarinet sound for some reason.
You should get it to write the melody based on a line of text, for example “row, row, row your boat”. That could be interesting.
by fucking dumb luck, landed here, and your library looks like it will be a very good fit for a (weird) project I have in mind. Subscribed just so I'll find my way back here when I'm sober.
Well, if you're interested (and when you're sober), I'm teaching a workshop on it in a week and a half: workshop.marcevanstein.com/ :-)
Have you used musenet before? It's an older openai project but it works really well.
Thanks for making the sound as clarinet, I play the clarinet so I might try this!
Ideally for a music model, you'd need the algorithm to digest a music theory guide as well as a bunch of scores, then the ability to feed it MIDI and printed scores as well as test prompts. Then you could do things like give it the "question" part of a "question and answer" phrase, and it knows how to complete the corresponding "answer" phrase, or give it a piece in 4/4 and tell it to produce a 3/4 variation, and it knows not to move the bar lines but add extra notes in appropriate places, or feed it a score not in its library and its cash work out things such as the likely form, genre, and for classical works, which Period it was likely composed in and the most likely composers.
It could then feasibly generate compositions "in the style of" certain composers, maybe even have a stab at things like "if Bach and Tchaikovsky collaborated" (which, for obvious reasons, could never have happened in real life).
Hopefully, the scores fed as part of its learning would all have been public domain, so any decent length comparisons couldn't be slapped with a copyright claim.
Merge it with a text model that's context aware and can properly analyse what it's written, and it could feasibly generate songs.
Add in a decent soundfont and knowledge of how to create files readable by Vocaloid...
im getting a song made for my 2ed YT , and i used Chat Gbt to make the song and the sheet, i cant wait to hear it, after the singer gets done with it.
Really interesting! I'm actually working on creating an NLP model to generate midi data. Not at the scale of GPT but hopefully it can learn good melody/chord progressions.
I hope you can share some content about even if is not long or just WIP.
Thanks for pointing out GPT limitations, I seem to struggle with explaining them.
This was very interesting and a good example of how to get music from ChatGPT.
You might find the video from CGP Grey (that is about 10 years old now) called "Humans Need Not Apply" is quite interesting, AND applicable to this "bot" musical generation idea. I'll give you a spoiler...
The background music from that video is CREATED by a bot. A bot from 10 years ago. It may not be an award-winning composition, but it was certainly good enough for a video background music, was free, and was generated from AI from 10 years ago. I can olnly imagine what could be accomplished now.
Helo sir, i think your scamp project is very wonderfull. Can i ask you sir? is it possible to write the notplay in wav format? Thankyou sir :)
On any of the feedback it gives try telling it when it makes a mistake; I was astonished how the conversation unfolded into a very comfortable AI assisted coding session. I'd like to spend time with GPT4 as soon as possible.
Its so futuristic.. can't wait
GPT is a transformer model trained on news and large text dataset. ChatGPT is the lowest parameter version of the whole GPT3 models.
Transformers were created to train on data to generate data. I'm pretty sure if you can train a transformer on enough music, it would generate great music.
It's not just about art. The lack of self reflection influences everything. I have seen several examples where ChatGPT gives information that lack internal consistency. Little variation in the question or question about details of some claim can change the answer to the opposite of the previous answer.
Hi... I try to run it and I get "value error: Cannot create score from empty performance" what do I do wrong? Thanks!
Very interesting, I agree and I think is the same reason I don't think AI will replace programmers any time soon.
ChatGPT gets an input or request, and gives you and output with the best of their capabilities and an explanation for it (Which I think it's most interesting aspect) but that's it. I can't really reflect how good, correct or appropriate for the context the result is. For ChatGPT what it produces is as good as it gets. When we request anything more abstract than "draw x thing on terminal" or "solve this equation" it will start making bad decisions, boring or uninteresting decisions or just mistakes that will eventually pile up because it has no sense of purpose beyond fulfilling the request.
At that point you would rather just code/draw/solve/do the thing yourself and using the AI as a time saver for the difficult and boring stuff. At the end of the day computers are a tool to turn into reality the things we have in our heads, not the idea makers. At least not until their turn sentient.
It is not a question of replacing someone, it is a question of how to improve your own work with AIs help. In my opinion AIs like ChatGPT will be used as tools in the future. Today I already use ChatGPT for finding solutions to specific coding problems which are hard to find with google. And I think artists like authors, painters or musicians can use it as inspiration or to find a startpoint.
I agree --- I'm working on one more ChatGPT video in which the goal is to use it as a collaborative tool 🙂
Judgement is imperative to the creative process. We already use our musical influences to guide our composition, now imagine millions of those influences under our direct control. It’s an infinite musical library of fuckin’ bangers
I think you pointed out the problem and solution in your video. If it is trained on the appropriate source material then you'll probably get the results you're expecting. Akin to the art AI that's been going around as well. It's making amazing photorealistic images, though there are certainly visual bugs, it's still highly impressive. It's been trained on a vast amount of quality art and is outputting the same.
wow, beautiful melody
I would love to see you try this again with GPT4 when it's available. Maybe the results will improve drastically.