I'd say ask it for a melody and chords separately and try to improvise over what it returns. ChatGPT is, IMHO, a nice tool to get inspiration if you're stuck, but it can not, for the foreseeable future, replace coders, musicians, or ANY actual profession. But if you take everything that ChatGPT throws at you with a grain of salt then it MIGHT help you to break a creative block. Everything else is hyperbole or a gross overestimation of its abilities.
If you're in a TL:DR mode just skip to the bottom paragraph, for my suggestion. If you want to hear my reasoning, as an electrical engineer with interest in the field, read on. The AIs outputs here, to me, seem a bit like it looked up everything that fit your requirements and just kind of mashed up known songs and styles within Jazz to write this first piece. I get the same feeling with most all of the art, etc that I see from AI, as well. When you look at human artists you can usually see that they are trying to evoke or convey something, and usually something at least somewhat original, sometimes VERY original, a whole new style or genre of art. Every time I hear one of these my thought is "OK', or "Uh huh" or some other totally "not impressed" in-head response. I do that with a fair bit of human written stuff too, though, as a lot of it too, is crap to my ear. More on that later, in the penultimate paragraph. Bored yet? :-) When I hear conversations with AI, I find the AI responses to be totally bland, there is never any spark of insight or creation, it is just plain dull. I would walk away, politely, from someone giving me responses like that very quickly. I would consider them a dullard. You can quickly tell when you make some statements with an intelligent person, as the conversation evolves with interesting twists and turns, and thoughts tend to expand, non linearly, often doubling back to the original thread. With AI, you ask it a question, it gives you a dead dull answer, and then it either asks some dead dull question, often that has little real relation to the thread of the conversation, or asks some open ended crap question like "what is the meaning of life". Yeah, great. That conversation the guy from Google published, to demonstrate AI sentience, didn't even demonstrate AI stupidity, as far as I was concerned. And when you structure questions and thoughts in certain ways the AI just becomes TOTALLY stumped, and goes not only stupid, but randomly stupid. AIs tend to do really well on trivia type shows, UNTIL you design the questions so it very much does not. Then even young human contestants can demolish the thing. So far I am very impressed with things like 'AI' driving and other specific tasks, some of them, but a super brain, nope, not remotely yet. Is a super brain possible, or even probable if research and design continues. IMO, almost certainly. The conversation, or music, or art at that point that it produces will be VERY different. It will write an "Uptown Funk", or "Take 5", or Brahms Requiem, or whatever, occasionally. Right now, it NEVER does. It never even enters the planet, let alone the ballpark. I guess I would have to do a blind AB to see if that is really true. That might be a good video. Find some good not widely known music, then have the AI write some, and without saying which is which play the two, and let us guess. What do you think?
That's the response to almost everything ChatGPT produces. People are only impressed when they see output from a domain in which they don't have expertise.
this 100% reminds me of Jon Baptiste trying out terrible music gadget/toys: "man, the notes are all wrong and it has this really tinny sound that just doesn't sound right, it's small and hard to hold but I like it for what it is. 5 stars"
I spend a LOT of time testing chatGPT. It needs coaching. Sometimes it doesn't catch the context of everything you asked it for. Sometimes you need to go back and say "hey you forgot about this part that I asked about" and it'll try again. Ultimately one of the cool parts about using the tool is that you can take the thoughts that are in your brain and put them down on paper and it'll operate as a great "friend"' that you can bounce ideas off of.
@@MrFreevo nobody said it was human. They stated points where it could be improved, and how you could improve them, along with places where it could be useful b
@@harrylane4 I know that, Harrison. But some people may think it is a substitute for some professional human beings. I see that as a helping hand, not a real writer or artist.
As far as my testing goes, chat got is not capable of generating music XML files. It constantly leaves tags open or doesn't complete them correctly. And it keeps making the same mistake despite being pointed that out several times in a row.
I'm a software engineer with 14 years experience, of all the questions I asked chat gpt it got 100% of them wrong. It is good at giving sort of surface to mid level generic advice, but once it gets specific then it's just very good at being confidently incorrect
I've had it be right about plenty of things. Granted, I'm not asking it to do anything crazy, but it's surprisingly good at doing a lot of the mundane stuff. I've definitely had it be wrong, but it's not 100% wrong all the time. It's wrong like only 20% of the time for me on average. Even when it is wrong, usually only part of it is wrong, the rest is usually right or at least on the right track.
Hi! I 1000% am sure you don’t remember me, but I served you at a restaurant I used to work at one time! You let me take a picture with you, which I’m very grateful for. I just wanted to say thank you! Since then, you’ve inspired me to embrace my musical side, and I’ve now released 2 songs, working on an album!!!! You’ve taught me so much the past years. Thank you, thank you, thank you ❤
@@jasonruff1270 Nope, this AI does not copy at all. The process of abstracting an idea it's literally learning, not copying. That's why chatGPT it's wrong about a lot of things it says in the same way, a person would be wrong about something they don't know too well. It's interesting at least.
@@jasonruff1270 I take you have limited knowledge of AI huh? ChatGPT isn't even a musically trained AI, but rather a linguistic model. Also, it doesn't copy anything, it learns it. By that logic anything anyone ever does is copying.
I'm 3 minutes in and I'm more impressed with Charles' ability to take some chords and make something that sounds good than I am that ChatGPT came up with the chords.
@@droughdough While i agree with you to some extent, the advances of AI in the art department has so far pointed to the contrary. A persons art is limited to the persons experiences, impressions and personal taste while an AI is trained on a much more diverse dataset than any human alone could process during a lifetime. Whether or not this makes the art true or counterfeit is another question...
A comment written by ChatGPT: Great video Charles! It's fascinating to see how ChatGPT can understand and generate musical patterns and structure. The generated music is definitely coherent and sometimes pleasing, it's clear that GPT has a good understanding of music. However, it's also important to remember that while GPT can help with inspiration and generate starting points, it still lacks the creativity and human touch that a human musician would bring to a piece of music. I look forward to seeing further developments in this area and how it can be integrated with human musicians in the future. Keep up the great work!
@Owl Actually it doesn't 'understand' anything. it juggles data in an interesting way to find correspondences and patterns. It appears to answers questions up to a point. As long as you are not too bothered if the answers are all correct or make sense.
"You are correct, I made a mistake" is such a ChatGPT staple answer 😄 Sometimes it actually means something, other times... it actually means nothing. ChatGPT is just saying this because it's an appropriate thing to say. It's a language model after all, so its goal is to imitate people.
Once i asked it to make a list, then asked it to print this list, it forgot to add an element to the list, i added it manually. And it said, apologies, It was already on the list but i forgot to print it. My jaw dropped. It totally simulated a slightly forgetful person. I wonder how a computer program can suddenly remember something it already lost. Maybe it had a timeout, dunno.
@@TheBiggreenpig All it does is try to guess the most likely next word, over and over. It's not simulating a human mind under the hood; it just guessed some words wrong during the list, then decided that the word "forgot" was the best word to follow up a criticism. It's like trying to predict stock market trends based on past observations, but with letters instead of money.
@@flubnub266 Are you sure this is doing just that? It is obvious that it isn't equal yet, to a human mind. But if it just picking the next likely word over and over, it would always give exactly the same result. And if it was just a random guess, it would give widely different answers. Yet, this is somewhere between. It never apologised when I gave it a new list element and apologised only once, when it really did a mistake. I'm pretty sure it just timed out too early, that's why the list was unfinished.
@@TheBiggreenpig Well, each guess is based on all of the past words up to this point, so it "remembers" history. Also, generative AI usually has a randomness factor (e.g. a softmax distribution). Sorry, I thought something was missing from the middle, not the end. So yeah, it's probably just the web service self-limiting and cutting it off early. It also does that just to prevent the AI from glitching and barfing out paragraphs of garbage.
as a software engineer, ChatGPT also produces some horrifying results. More importantly it doubles down on sincerely believing what it's done is right.
@@Rocky-sn6fl Even if it was perfect you'd need someone to translate the business case into requirements it can understand, and be able to check the results. At that point it just becomes a higher level programming language.
I think what happened with the F-Dm7-G7-Cmaj7 progression it gave was a problem with its training (as most of these problems are). It is only very familiar with a ii-V-I in C, since that's the most common key to learn these things on Google (the source of its training). So it gave you a simple progression (ii-V-I in C), and *then* took your request to be in F as "being in the key of F simply means you start on F 😃" and just gave you F plus a ii-V-I in C. I really doubt it was going for a vi-ii-V-I, especially considering the preface it gave you about non-extended ii-V-I's. This kind of issue is really common, I've found. It finds one part of what you want separately from another part, so you end up not having your requests merge together as you intended.
*this* is how I think ai should be applied to the arts. Yea it spat out some chords, but the actual playing was entirely up for interpretation by the artist.
@@games528 sorry to turn your question around, but I'm curious: Do you _want_ AI to replace artists entirely? Or is it that you want AI to function as an additional avenue of artistic production, or something else? Asking because your question reads to me as though your preference would be the replacement of artists with AI, but I would hope that this isn't the case.
@@calliopeshif7581 The simplest solution that comes to mind is to just let people choose. Wanna go generate some art? Go for it. Wanna visit a human art gallery? Who am I to stop you? We care far too much about what people choose to do with their free time.
Ya but gone is the era of people loving the music because how much human emotion, creativity, and thought goes into their own writing. Knowing that a human created a great cord progression makes me love that song on deeper levels.
I love how even when it was wrong you still found what it made to sound cool and tried to make it work instead of just discrediting it as a whole. It was super refreshing!
uh no, that is just what most decent musicians do, adapt random things and find a way to make them work, that is literally what guys in bands do when bouncing ideas off each other, trying to come up with songs. But the GPT is more generic (based on existing data which the bulk of is averaged to parse out what may seem credible) but less contextual than humans, so the results not great. I'd rather deal with band members, that way it is like multiple human chatbots conversing, with more context and much more.
Great experiment, Charles, I'm also obsessing over GPT! But I think you missed GPT's confussion in your jazz question. GPT correctly identified the most common progression in jazz (II/V/I) and it provided two correct examples: one landing in Cmaj7 and one in Gmaj7. The problem appeared when it tried to generate chord progressions “in F”. Instead of transposing, GPT simply added an F chord ahead of each example, resulting in the bizarre F + Dm|G7|C and F + Abm|Db|Gb. The final answer is harmonically wrong, of course, and it doesn't look like a human mistake. But it isn’t 100% misled nor random either. It actually makes sense if you think about it in terms of statistical linguistics, which is how GPT works: in actual speech, the linguistic context of expressions like “progression in F”, “progression starting with F” and “progression in the key of F” is highly overlapping, so GPT's confusion isn't surprising… To get a better answer, you can reply: “Your examples include correct II/V/I progressions, which are indeed common in jazz, but they are not in the key of F. Try again and make sure the chord progression starts and lands in F.” In any case, GPT is evidently self-correcting quite fast - probably learning from its massive usage. I've just asked it the same question after watching your video, and the answers I received are already more concrete and relevant: Me: "Write a jazz chord progression in F major." GPT: "Fmaj7, Bb7, C7, Fmaj7" - that’s a very boring I/IV/V/I, but it is correct and it is in F. Me: Make it more interesting. GPT: "Fmaj7, Ab7, Db7, Gbmaj7, C7, Fmaj7" - this one looks pretty much like the second one you got, but now GPT introduced a dominant C7 at the end to effectively land in F!
I wonder how many levels of things which are X but not Y, having Z but not P, and in the set of A,B,C but not D,E,F, chatGPT can handle without barfing.
@@WarrenPostma I think these kinds of requests should actually be what chatGPT is best at. It is much easier for computers to understand "logic" than it is for them to understand language and abstract concepts
Yep. People do not realize how important it is for knowledgeable people to correct it for it to be able to “learn” and give better answers. Just like GIGO, it gets better with Intelligence In Intelligence Out.
@@saywhat9158 ChatGPT is a fixed model. It doesn't get modified by people continually interacting with it. If that were the case, trolls would be able to have an absolute heyday with it (like they did with the fiasco that was Microsoft's Tay bot). It's only able to utilize the results of the training that OpenAI's engineers have already put it through, plus any content from the user's currently active conversation. In the context of any given conversation, it'll behave the same as if it had just come fresh out of OpenAI's lab - even thought that individual session may "learn" new things, the model itself does not. It's stuck in the year 2021 until OpenAI train up a new one.
@@TheBeatfox what sources did you use to say that the model can not progress? According to open AI it made progress at the end of January on its capabilities to deliver facts (i still wouldn't trust it) and it's capabilities in math (didn't notice much progress when I tested it) . I saw a comment on the open AI discord saying that the update made progress across the board.
I've asked ChatGPT a couple of different things, including creating ASCII art, creating and decoding Gcode for a 3D printed or milled object, creating OpenSCAD code for3D modeling, writing general code for programming, and writing text. The fact that it knows how to make these things is astounding on its own, but for now I'd stick to text.
7:48 it probably just pasted a 2 5 1 in C after the F chord. It's likely that most input it received was in C major because that's the most common key.
Hey Charles, if you're looking to get back into looking at jazz from video games, Final Fantasy XIV has some fantastic pieces. - Akademia Anydee - Anamnesis Anyder - Smiles and Carrots All varying styles of jazz that I and I'm sure many others would love to hear your input on!
I found if you ask chat bot to do something and ask for it to give you a few suggestions for what to do next , THEN PICK A SUGGESTION IT GAVE and ask them to add which suggestion you liked. THEN amend the errors and tell it something you want from its because it inspired a new idea . and it will do 3x better or more than you were expecting once it writes up the amended attempt
I tried it out for some poetry and song lyrics, what I got was some really bad stuff that you might expect from someone in grade 4 😆 but it had some ok ideas that might be good as a starting point. So far it doesn't seem any better than many of the tools which have been out for years... it makes me wonder about the stories I've seen with teachers saying it wrote some of the best, well thought out essays they've ever read.
It fully depends on the topic. If there's been a lot written in a similar manner about the same topic, it can find a pattern and amplify it. If there are several clashing theories or ways of thinking then this AI tries to harmonize most of them which comes out as a monstrous hybrid or overgeneralised mush.
yeah the poetry is giving real “i like sandwiches with cheese / i can make them with ease / my appetite it will appease / you can have one if you please” like it seems physically incapable of doing anything non-literal or subtle even with coaching
or doing any structure that *isnt* elementary poetry class (i even asked it to be more creative with structure and that just made it add random exceptions to the meter instead of like….changing the meter so that every line isn’t the same number of syllables)
Please make this a regular series of videos! You could start off with a general sort of prompt, then continue to refine it by asking it to make it sadder or happier, more upper extensions, different structures, etc
Hey Charles, have you ever thought about making a video about Ryo Fukui? I recently stumbled upon a fascinating documentary about this legendary Jazz Pianist and the cultural significance of Jazz in Japan by STEVEM and am certainly curious what you think about him. "It could happen to you" by him is one of my favourite trio pieces. Seeing ChatGPT used for chord progressions was really interesting by the way. Thanks for the great video!
yeah me too! I stumbled upon his rendition of Autumn Leaves, its perfect, one of the best I've yet to hear! that definitely got me hooked to hear more of his
There are some fantastic Japanese artists to explore but I wouldn’t mind a look at Japanese-American trailblazer Akiyoshi who also seems to have an interesting story
Ah, Charles, ChatGPT's first attempt at a Bridge sounded very much like an old Beatles Bridge, or a Bridge from that timeframe. Honestly, with a well-written vocal, I think it would sound amazing.
An algorithm came up with a more interesting chord sequence than 99% of modern music. I hope more people use this even as a way to break out of the same boring 1, 4, 5 chord sequences.
1 4 5 is blues, not much modern music uses that unless they are doing a retro take. Most modern music is generic pop recycling, or hiphop with few if any chords
First Song feels like it's tonic in fact is Eb and it just never comes home; the chorus starts with it. The Bridge isn't C#, it's Db, which isn't up a minor third as requested, but it certainly works.
For a darker temporary feel you could change the end of the chorus to a Ddim or Ddim7. That can potentially lead to all three of the starting chords: C# (well at least C#min), Eb, and Ab. But this is all theoretically without access to my instrument.
It's important to remember that the goal of gpt was to make human-like replies. I think its ability to also approach accuracy this closely was sort of an unexpected side effect.
You should make a video asking ChatGPT about writing any jazz harmony (ofc regardless of how wrong/right it is) and make like a complete improv with it.
This video inspired me to ask ChatGPT to write some songs with chords and lyrics with suggestions like this about structure and modulation. They were incredible in the cheesiest way.
Seems to me that the reason the first progression sounds like a IV-V-vi is that the chord on the 2º degree (Bb) is major instead of minor as it "should" be. So instead of a I-ii-iii, you get a I-II-iii. The pattern of maj-maj-min suggests IV-V-vi. And I think the reason it made the II major is because your question immediately beforehand was about the Dorian mode which would, in a way, produce just that. If you were writing in A Dorian, f.ex, you would have Cmaj-Dmaj-Emin, with a major II instead of a minor ii.
Well but it actually did a great job here: It helped a musician to write a song in a couple of minutes. It gave you inspiration to do something that you might not have tried otherwise. So this is the perfect example of how an AI can be used to help us on a daily basis. Think of all the people that have blockaeds due to the preasure to deliver something in their jobs for example and are stuck in their minds. An AI being able to help you out of your blockades can save or at least change lives.
Charles, you should check out "Sonido Bestial" by Richie Ray and Bobby Cruz, or "El día que me quieras" by Eddie Palmieri. Both songs have piano solos that you might find interesting.
I have two songs I want you to check out; Dilerkefet and Asgretalos, both are by Keygen Church. Personally, I have not heard anything like them before.
It all sounded perfectly fine but the fact that it came from an AI was super interesting to me and a testament to PEOPLE knowing how to create technology that can be immensely useful IF YOU USE IT AS A SPRINGBOARD FOR FUCKING INSPIRATION AND NOT AS A DO-IT-FOR-ME TOOL
It's very easy to build chord progressions like that and you have millions of songs to build your database from and the only reason it sounded good is because we have a good piano player turning these basic chord symbols into actual music. I will be impressed when AI can create the piano parts he was playing.
To me, that first bridge isn’t in the key of C#, it’s still in Eb major but it sidesteps into Eb minor with the bVII (C#, or more correctly Db) and the bIII (F# or Gb). If you make the Eb in the bridge minor, then it’s just moving to the parallel minor. That’s how I hear it, anyway.
The second jazz chord progression it gave was basically the first where the 6-2-5 were their ‘tri-tone’ alternatives (go up/down by 6 semi tones). This is a common jazz ‘trick’ to spice chord progressions up. For example if you had a simple 1-6-2-5 chord progression in F (just bass notes omitting harmonies): F-D-G-C, a common twist would be to go through F#/Gb between C back to F because F#/Gb is the tri-tone alternative to C so: F-D-G-C Gb-F. I quite liked the pop chord progression it came up with!
Not a chatGPT request, but I was listening to Periphery's new single, "Wildfire," and immediately thought that your take on it would be really cool. It's a crazy song, and really interesting imo
This is about how good it is at writing code. Like its impressive even though it is not correct. It also needs a lot of hand holding to get things right.
I happened to ask GPT-3 to write me a simple 4 part song and give me the output as a music XML file. The AI was incapable of giving me a valid file because it would never close all the tags despite me pointing that mistake several times.
Sometimes the character limit is hit before it finishes output (and doesn’t really say as much, annoyingly). Could that have been the problem? That was an issue I had when having it write code for me.
That seems to be the deal with ChatGPT, which I've seen in lots of tests of it in a bunch of different domains: it has a high level of fluency, but low level of accuracy/truthfulness.
PLEASE cover some Ben folds! He’s one of my favorite artists and one of few that incorporate a ton of piano into their work, plus he’s got some really intricate pieces. (Gone, Ascent of Stan, Landed, Philosophy, Not the Same, etc.)
I asked it something much simpler than that yesterday. I just asked it to write a sonata in Gmaj. It said that would be best done by a professional musician; however, it thoroughly explained what a sonata is.
I ran a similar experiment with different styles of Blues, from Ragtime to Delta to Chicago. The only one it even came close to doing properly was, unsurprisingly, the simplest form - Chicago style blues. But even there it never made a bridge correctly. Always issues like this, mislabeled major/minor and/or very bad key changes.
Still close enough to be impressive between the correct root movements and Lady Bird turnaround. Make that Gb7 a dominant chord and you have a sub VI II V back to F.
Question to all the piano players here: I love how he fills all the space with all these notes, instead of just hitting chords each count like I do. I would love to develop that skill. What is it called? What should I search for on YT? On guitar it would be rhythm guitar. On piano?
Fantastic. Just a small correction I noticed. The correct nomenclature for the third minor degree of Ab is Cb, since C is the third note after A. If you say B, you are saying an augmented second.
Over the first few minutes I think I created a Sarah Mac song quietly along with you... and it feels so right... (or it could be a contemporary musical theatre song too LOL)
Even in the coding world, at first glance it looks amazing. When you start looking deeper the cracks begin to show. It will come with an answer, and most of it is pretty good.. but then you also see code that can't work; or it always falling back to the same piece of code over and over even though you told it you want something else.
I recognize the Ab -> Db -> Gb as a tritone substitution. The third of C7 == E and the seventh == Bb. Go up a tritone from C and you're at Gb7. The third of Gb7 == Bb, the seventh == Fb (aka E). So, whenever you see a V7 chord in jazz you can substitute it for its tritone twin and the third and seventh notes don't change. This particular example gives the substituted chord a little bit of a run-up with the Ab and Db chords, them being ii -> V in the key of Gb. Great content. Subscribing.
Yep, that first example was pretty much Sims2 build mode. Aaaand, new bridge is just a Mario cadence. 7:48 sums up the whole experience with AI and music.
I love not only how confident the bot is, but also how well it accepts critcism. I saw a picture of a guy asking the chat gpt bot on what was 9+4, and when it said 13, the guy said it was wrong, that the correct answer was actually 14, and from them on the bot dogmactly accepted 14 as the answer of 9+4 and nothing could make it go back! 😅
I'd love to see more of this, just asking it for more different types of music or categories and you reacting to whatever wacky (probably not actually accurate) music it has you play sounds fun.
As a musician and Music Theorist, I asked it to "learn" a new scale. It absolutely could not, because it success is restricted to short duration, short term regurgitations. It has the biggest problem with context outside of its (limited) original dataset, but it also can't "associate" things with quasi-probable outcomes (that are other than random) because it can't inject true reasoning into its output loop.
Charles, FYI it holds onto the context of your conversation. The corrections you've given it are affecting later output, possibly for the worse. For science, consider starting separate conversations when you ask it new/different things.
Do you think the algorithm understood “bridge that modulates up a minor third” from the perspective of the bridge needs to contain a modulation of a minor third and wasn’t think of it as a typical modulation that would happen at the start of the bridge? Couldn’t one argue that the movement in the middle of the bridge chord progression, from D# to F#, is the minor third modulation that fits the original prompt? Considering that D# and F# are not diatonic to each other and their roots are a minor third apart
At 3:45 the 2nd verse could start with an Fm, then Bb, Cm, and Ab - actually all verses could start with an Fm. Also, in the bridge after C# and D# I would have been expecting (had I not seen the chords) an E#m (and then we're in the key of 8#). There the F# sounds like a subtonic.
This reminds me of results I've gotten with having it solve physics problems. I'm a physics teacher and wanted to see how it would do. I've found that it gets a lot right, and is pretty good at explaining its steps, but it also often gets some things very wrong. It's rapidly getting much better, though.
Not to pick nits, but Ab to B is an augmented 2nd, not a minor 3rd. Ab to Cb is a minor 3rd. Enharmonically equivalent, but still. ;) And a better way to get from the Ab in the original chorus to the Ab in the verse might be a secondary dominant (V7 of I). The 2nd jazz progression is kinda cool, the ii-V-i of the tritone substitution of the V chord. The Gbmaj7 resolving down to F is nice, as that flat-2 tritone sub always is.
I literally tried this on Wednesday. Except I asked it to compose a piece of music using a 12 tone row, it couldn't answer the question. Now that I think about it, it wasn't a very good question.
