Glenn Gould is partially correct when he said that a 5 year old could compose music of the same quality as Mozart. However, only one such 5 year old ever existed and that was Mozart himself.
About Mozart, there's nothing Gould could ever have done correctly, let alone understood correctly. A cliché himself, Gould attempted to gain notoriety by slamming Mozart. Being a narcissistic showoff, Gould grabbed the wrong end of the stick. Geniuses such as Seymour Bernstein called him "a travesty", and even Leonard Bernstein had to make a disclaimer prior Gould's performances. For a vast majority, Gould is nobody. For those who are aware of his disrespectful prejudiced remarks on Mozart, he is a pathetic bigot uncapable of keeping up with playing Mozart. For an insignificant minority, he was a visionary. Mozart, however, has always been regarded one of the best three composers ever, and certainly, the most beloved by the vast majority.
😂what does that have to do with the video or op's comment? Fwiw, my main criticism is he makes great music sound like Mozart! Weird how emotional and biased your tirade is. What's the story there?? @lumenvera9452
@@lumenvera9452 I came to accept that Gould was basically trolling the classical music world, more than actually believing his most outrageous statements - or both. In any case, with most of his interpretations (that includes Bach), I never liked his sound, which was probably caused by his Yamaha piano and recording antics. I basically stopped listening to his recordings years ago. He was clearly a vastly talented musician, but he became a victim of his own intellectual eccentricity. With Gould, the line between the inspired and the grotesque is often blurry.
Really appreciate this video. I work professionally with LLM and AI as a programmer. It is essentially auto-suggestion/correction on steroids. When current AI writes music, it is using advanced text prediction, "text" being the general programming term for "tokens" (small bits, or in music, notes), or how humans cogitate, and break down complexity into smaller parts that can be modeled. It predicts what comes next based on input. We talk about "models", or the code that does the prediction. There are technical model limitations, but these limits decline as humans program better models, and as we put more compute power in. At no point do models "reason", at least not on our current reflection on how humans reason. So these models mirror back to us what humans expect. If imagination is expected, programming hasn't reached this point. (Kudos for the Schoenberg reference!)
I think its possible that AI is in fact telling us just how humans reason. Our brains are not magical; essentially they are huge webs of things that associate. So if I said River to you, you might say Dance, or Fish, or whatever. If you ever watch your thought patterns go by, its essentially just a torrent of loose associations. So in terms of thought generation, it could be doing just what AI does. The difference, of course, is that AI is trained on a massive dataset, so its like expecting a thousand human brains wired together to come up with something coherent. All the different directions its pulling conflict. Our minds have been wired on a small dataset, and it is constantly curated by the ego - what we call cognitive dissonance, where we toss out stuff that conflicts with previous input. So there's a fair amount of error "correction" going on. Frankly, I'm amazed at what AI comes up with. A lot of people on Internet forums should be so coherent.
@@themusicprofessor Watch this video: Can You Hear The Difference Between AI and Beethoven? Carl-Friedrich Welker Udio here is making beautiful music. It did not have the weird stuff you sometimes hear on Udido made stuff. It is going o get progressively better and people will be programming it detect weird-isms and leave those out. But what are we talking about here? We are talking about simulating a Mozart piece so it would be stylistically unoriginal yet original in the particular notes used. Somebody like you could work with the programmers and train it not to do certain awkward things. The programming could be even more improved if a particular program was designed to generate Mozart's music specifically, so it could be trained on more of Mozarts music than it already is. Then listen to the results and keep refining it with new rules. But how could it generate something that was more unique not just imitation of a composer's style? You train it with Mozart pieces but then add in some Chopin or Bill Evans or all three or use an occasional Indian scale. This is all subjective, you program a hybrid to your liking and in a proportion to your liking. Have it generate 10 songs. What if one was great or had a section that was great? That may be possible. What about the human feel? > that can be programmed, And the human element is you picking which song or section you like.
Does anyone rememeber when the Google doodle was an "AI" that harmonized bach chorales? It was terrible and everyone at work tried to convince me was good. But none of them actually listened to bach. Very bizarre experience
I used that google doodle to write a harmonization of "Ein Feste Burg". To do this, I ran each line a few times, and picked the best. Here is the full thing: ruclips.net/video/ByTgaLzMwgQ/видео.html
I'm a copyright law professor (from Thailand). The emergence of generative AI has challenged the copyright lawyer's perspectives on authorship. Basically, copyright law doesn't really understand (or care) about how people create a copyrightable expression. It only assumes that creation (authorship) is fundamentally a human activity. The french copyright tradition (droit d'auteur) has a doctrine called "Imprint of Personality" (l'empreinte de la personnalité) which means that a copyrightable expression must reflect a human personality behind an expression (for example, everyone draws a cat or a circle differently). The anglo-american copyright tradition relies on the "skill-judgement-labour" requirement, which I think an inferior theory to the french doctrine (and a weak Turing test for AI to pass). It's easy to be impressed with how quickly AI evolved to the point where it can generate musical compositions; and AI authority like Geoffrey Hinton claims that AI can "think" and understand like a human being. However, your demonstration in this video hints that an AI perhaps still lacks an "imprint of personality" in a sense that it does not comprehen the speech-making nature of music creation. I was looking for an interesting content about AI for my copyright law class next semester, and I think this is it. Thank you!
The French doctrine could allow for the artists, whose materials trained the generative AI, to get paid. There would have to be a way to identify which training materials reach a predetermined threshold of influence (in music, sometimes it's really obvious, but usually, I think it would need to be self-reported by the AI). This could be impossible, IDK... the copyrights for each piece of art would start to resemble mortgage derivatives, and become burdensomely complex.
I think using the work of others in the “training” dataset is in and of itself a violation of copyright. This work is created by interpolating data from the training set (ie pseudo randomly mapping an input note set to other notes in order to obfuscate the source material that’s being plagiarized). This work doesn’t sound good because the trainers don’t understand classical well enough to train out the bad results. AI doesn’t think like a human at all and the people spreading that BS are either lying or stupid. In general they are lying because they have economic interests in AI. AI doesn’t understand any of this, it just has some characteristics of the training data others are missing because the trainer lacks the musical talent/knowledge to eliminate oddities from the output.
I don't really understand the copyright issue, as no work is getting copied or distributed. And in terms of thinking like a human, you are confusing weak music generation AI with advanced LLM. Currently, it isn't a question of will AI systems match elite human levels, it's just a question of when. And what does a million super-intelligent brains combined sound like in unison - and that's less than ten years away.
Excellent content as usual. This is quickly becoming one of my favorite RUclips channels. And, as an amateur pianist who is much less of a keyboardist than Mozart was at 5, I nonetheless have had to immediately find and learn a couple of pieces I heard through this channel - namely Ravel's Prelude in A Minor & Beethoven's Bagatelles Op126 no5. Thank you!
15:00 great discussion! I especially love the line, "it can give you the symbols, but the signifier is disconnected from the reality that it's meant to symbolize"
The 5 year-old Mozart knew how to vary his material subtly. In the first four bars the melody drops down on beats two and three, and in the answering four bars he goes up. He breaks the formula - simple but effective.
I read in a book Mozart devides his tune into 5 parts . I can't remember d names .he thought 5 as a key number. His creativity has some law ,but i don't remember dt book much.
Reminds me of the old story of the "cargo cult": South Pacific islanders building wooden planes as substitutes for aircraft (now gone) that had delivered them supplies in WWII. If you don't understand the basic principles of propelled flight your simulacrum isn't going to perform. Similarly if your AI doesn't understand the workings of the human mind (or is it soul?) it is going to fall far short of a satisfactory conclusion. Interesting video. Humans 1 AI 0
Eerily reminded of Orwell describing the "soulless" music (and fiction) written by machines in "Nineteen Eighty-Four". And the "Proles" didn't even care...or rather, they didn't seem to notice.
Thank you, very interesting.I checked out the same platform shortly in the field of Jazz, which is my home base as a musician and teacher since 40 years.Basically, what AI does in the moment is pattern recognition. If you ask AI to create Jazz, from blues or bebop to free improvised Jazz, the thing you are talking about becomes even more obvious: It is just a repetition of clichés over and over, without any creative spark.A good jazz Musician is even faster, because Improvisation happens on the spot, in realtime. The best examples are all the duo improvisations of Joe Zawinul and Wayne shorter, just check out the "In a silent way" version from the live album 8:30 and remind yourself, that this is completely improvised in the moment-if you see the transcription of it, it is a perfect composition. A raw frame of the melody is the only thing it has in common with the many other versions (like the one from the last Miles Davis concert in Paris 1991).
You should do a video on some of Mendelssohn’s work. I really like his piano sonatas and songs without words. I also got to sing part of his oratorio Elijah, and I think they’d all make for an interesting video. Great content btw!!
Btw, this is the best and most detailed, and precise analysis of AI vs. Genius Composer I've seen. Thanks for the video and I'd love to see similar videos on this topic if it suits your inclinations.
Another thing to point out: on a sociological level the "unfortunate" thing about music is probably its greatest gift: it simply "sounds good." A simple chord progression sounds pretty and is pleasurable on its own and they're not hard to use so the alarms flare up when it turns out computers can make pretty music and pictures too, supposedly revealing our pretentiousness and affirming philistine doubts. But that says less about "AI catching up to Mozart" than diagnosing a culture which increasingly can't tell the difference between prettiness and art, unoriginality and familiarity, surfaces and realities (at least Gould distinguished between prettiness and art; he simply attributed the wrong one to Mozart)
In many ways I agree, but I also think that the art worlds desire to separate beauty from art, something which has been going on for a lot longer than generative AI is in many ways also a part of why this is. The point of so many more modern artists be it musicians or painters or whatever has been to treat art as a social phenomenon, which essentially makes whatever is treated as art into art. I think it's the combination of an art world which is in many ways, one might even say "ideologically" opposed to making their art beautiful and enjoyable to "the masses" being forced to recognize that people have so drastically fallen off the boat that they don't even really know what art is anymore, because it's stopped meaningfully approaching them. I think this is what leads to the lionization of certain styles which can easily be copied by AI and that style, not the art behind it becomes the art.
@@JustAGuyWhoLikesStuff. Fully agree. Though admittedly "beauty" is harder to pin down than "prettiness," I guess I could have said the difference between beauty which may include prettiness and mere prettiness. But the art world's disdain for beauty definitely contributed (though you can't really blame them) and the escalation of Modernism flung us into that rootless void we are in now where no one no longer dares to make an assertion and all discourse about art seems to end in "well, art is subjective" (imagine anyone doing that Gould program now)
Well this is a thing that the AI discussion always forgets, perhaps we are simply not very good listeners anymore. It takes good music and good listeners. In a few generations time, dumbed down by AI assistance in multiple domains, perhaps very few will even have the listening skills left to appreciate the great composers best music.
Another great presentation! The reason bar 6 sounds like bad counterpoint is the parallel 5ths between the melody and bass voices...and then Matthew points out the bar 14 'development' of that bad sound-again parallel 5ths-and the non sequiturs continue. There is absolutely no replacement for that thing that made Mozart a master musical storyteller. I think Matthew has just proved there is another level to the "Mozart Effect"...He is Impervious to AI...and so, Loki (Mozart) 2 Matthew 2 AI. 0 Gould. -1
I recently heard the 5th symphony by Beethoven's pupil Ferdinand Ries which has a 1st movement that sounded a whole lot like that other famous 5th symphony by his mentor (even motivically), and yet you could tell that there was something missing, despite it being very much in the style of middle Beethoven, but I couldn't quite put my finger on what was wrong with the music. I was wondering if you could make a video comparing the two pieces and try to explain what was Beethoven's compositional secret that his lesser contemporaries lacked (even when they consciously tried to imitate him). I think that could be very educational as I really enjoy your way of analysis and breaking down the musical elements.
Wow! I didn't know the Ries. Just listened to the first movement (ruclips.net/video/LnTf57Ix4v8/видео.html). I think it's rather brilliant actually! It's not easy to write a really good pastiche of Beethoven! Ries is very interesting though because it seems like a very clear homage to his friend and mentor (not just the 5th but also the 1st and 2nd, the 6th and 8th are detectable). So, even though it's clearly modelled on the motivic material of Beethoven 5, the Beethoven that really appeals to him is clearly the more classical even-numbered (not the visionary odd-numbered) style. This is somewhat similar to the early symphonies of Franz Schubert and I think it shows us that the Beethoven's better-behaved, more classical style was more popular with his friends and contemporaries. You'll notice that Ries's development and coda are much less ambitious, less wild and extravagant than Beethoven. It's a lovely piece. Thank you for sharing and yes perhaps I will do a comparison...
@@themusicprofessor Yes, Ries was no doubt a highly skilled composer and it wasn't my intention to put him down. It's also interesting how he has moments anticipating later romantic composers like Mendelssohn. I also noticed how Ries appears to draw most of his inspiration from early Beethoven (which is interesting since he composed most of his symphonies after Beethoven's 8th was completed), and it's even more noticeable in the 4th symphony that has an introduction clearly modeled on B's 4th, and a scherzo modeled on the scherzo of B's 1st (there's a very fine new cd on the Ondine label featuring both works which is the one I heard). And still I feel like Ries is making all the right gestures, and is capable of writing good themes and develop them like Beethoven, and yet it doesn't add up to a satisfying whole like Beethoven.
I would need to listen to more. I'm impressed by what I've heard. The style is so close to (the more classical) Beethoven that it's almost a pastiche. But it's very imaginative and beautifully crafted. I wonder what Beethoven thought about Ries's symphonies - presumably he was flattered by them! In terms of satisfying whole, Beethoven's standard is remarkably high and he pushes the art in weird and wonderful directions whilst somehow maintaining control over form. He also has a rare gift, in terms of the overall architecture, and of course the ideas themselves are so compelling, and so inventively handled and developed, and the rhythmic component, and phrase structure, is so imaginative, and all his music has a risky, visionary quality which is also unique, so generally speaking, its very hard to do better! Perhaps his only contemporary who managed it was Schubert.
@@themusicprofessor How about something on the influence of some of Haydn's symphonies on Beethoven's work, and what Beethoven does "better" than Haydn to take it further? It seems at times Haydn is unfortunate in the sense he was the giant of music whose shoulders Mozart and Beethoven stood on, and so Haydn ends up buried in the mud!
Dame Misuko Uchida analysed a Mozart sonata on TV. She played part of it and then said one note was wrong. She was right. She then showed that it made sense 30 bars later. Then it turned the piece from being mediocre to a masterpiece. She said "it is beyond the intellect, it is pure intuition, it is genius". From memory, the piece was k545, but I might be wrong.
It’s great that you engage with this on its own terms while offering your expert critique. Really interesting building on your ‘dialogue’ with GG. Good not to reject these developments out of hand, but as you say retain what the human condition uniquely contributes to music. Great you bring us back to Mozart to underline the point. 👏🏻👏🏻👏🏻
You are so right about that story-telling aspect of Mozart’s compositions! That is indeed the one thing that’s missing from AI-generated output. (Generally really, not just for music.) Thank you for that insight.
It helps that you can brutally critique the AI because you’d never go that hard on a real student but it’s illuminating to say yeah that sucks and here’s exactly why
I entirely agree with you. We oughtn't admire people merely for celebrity status. But, in so far as Mozart is concerned, Gould is quite mistaken I am afraid. I sincerely agree with you about AI. The ability to assign meaning to things and to create things with a purpose is significant only of the mind. And no mind can do so with quite so much excellence and sincerity as that of the human mind.
AI is currently on a level where it looks impressive to an amateur but experts still see the flaws. However, I think that will change soon. AI is right now a little like a person with some kind of "autistic savant syndrome". It can answer almost any question, but you have to push it a little in the right direction. E.g. you could tell the AI "Bar Y sounds a little weird because of X, how could it be improved?". I'm sure the result would be quite good. Maybe you should try that in the next video :-)
I can think of no better introduction to AI. Great show. So fascinated, I am. 3:28 This is no ending - I think it was meant to be, by AI, wondering about your question to the expert. Speaks for the encyclopedia I read last night (something RUclips's AI just cannot know) which said that any piece of music leads to its ending. This, too, speaks for your criticism of AI I yet not able to share. To me, it sounds like from the authors. I shall wathch through, though. I think I shall willingly subscribe, especially I like the clickable timeline. Maybe Ai by RUclips even understands irony that is hidden - in my case that would be great misunderstanding (to test AI, I hereby explain "because you are enabled to click forward". I do not think my comment will be deleted by RUclips, neither by you. Keep it up! Cheers! Great endeavor.
As someone who's been assembling very large databases and analyzing them using complex interlocking self-derived algorithms for decades - not to produce finished compositions, but to create successive (carefully defined) first-species counterpoint examples blending 12-tone serial conventions with basic Hindemithian harmonic theory for use as compositional source materials displaying the full range of potential 2-, 3-, and 4-voice harmonic progressions available for every possible iteration of a tone row at arbitrarily predefined frequency levels - I quite agree with the observations here offered. Careful and informed computer analysis can indeed produce theoretically correct streams of contrapuntal processes; but the infinitely various layers of immediate decision-making involved in composing any particular measure of potentially persuasive and artistic musical composition is - as of yet - far beyond the capabilities of any computer process. Computers may and will certainly continue subtly adjusting and refining their understanding of music theory and practice based on continuous, earnest feedback from human beings; but even a Beethoven, I believe, could not successfully train an AI to recognize true invention in the necessarily complex intellectual processes and emotional reactions involved in every single measure of any genuinely artistic composition.
I think it's pretty incredible. You can tear the shreds as much as you like, but I can't even begin to imagine how they came up with software to write something as superficially good is that
I tried it with the prompt, "a piano piece composed in the 1770 to 1790 time frame," and the result was utterly horrendous. The AI produced two sound clips, and one of the clips even had a piano in the background. So far, classical music is very definitely in the domain of humans.
That first cadence is hard to fix- the counterpoint mistakes aside, it feels like it can't make up its damn mind if it wants to be a half cadence or an imperfect authentic cadence. lol
Iteratively, human prompting the AI and selecting, that’s where it’s getting interesting. On my channel are some examples of what I do. I still sit several hours on one song, selecting „next few bars/notes/sounds“ out of tons of possibilities it generates.
Yeah, the AI sounds directionless. But like you, I think it's impressive it catches on to certain tropes. It's kind of like teaching a dog to press buttons that "say" a word to get a treat. For the dog, it's a routine to get something tasty, and for us, it's entertainment. For the computer, I guess it's like satisying some sort of math or command or something. (I don't know-I'm no computer scientist, but the man who taught me counterpoint over the Internet, Bernard Greenberg, was, and he actually did work on AI music!) And then for us, we hope it's art. There's some fundamental misalignment between what the computer "wants" (if we can call it that) and what we get from art. If anything, this is more scientific than artistic! I think I can accept the existence of increasingly sophisticated AI art in a vacuum, ignoring its effects on human artists, the environment, and so on. Art that I perceive to be great is special to me because it came from someone who existed somewhere. "Wow. Someone did that," my English teacher would say back in high school.
"A five-year old after a few weeks of music theory". An interesting premise, Mr Gould. Canadian kindergartens are clearly quite unlike our own. (PS, as I've said on another comment I am a real fan of Glenn Gould - at the piano or at the microphone, I love listening to him. But part of my fandom is the delight in pointing out the bits where he's totally off the rails.) Edit: Oh, you make the five-year old point, but you leave it to the end.
@@themusicprofessor As conductor George Szell said of him: "This nut's a genius!" I stand in awe of your transcription capabilities! I once prepared one of the David Cope computer-generated "Bach fugues" to do as an encore piece. Number whatever, but in Dm. 10 minutes? It took me something like four hour-long sessions stretched over four days. I can memorise, sight-read and improvise like a Banshee but I find transcription horrendously difficult. I played that fugue in a very Gouldian style (but I am by no means as wonderfully gifted) and then asked if anyone recognised it or could date it. Before saying it had been written in 1992 (or whenever), by a computer, and had taken it 3.2 seconds.
It is like the early chatbots. The larger structure isn't there, and it has used some out of context ideas. I think this is a quantity of training data problem. It's not learning what's not expected.
I think you are completely right. This is a very early stage of AI where the musical aspect ist not very strong yet, because there is very limited knowledge of music built in. The current models were built around text, not around music. So the shortcomings are explainable, I think. Creativity for me is not creating things out of the blue, but combining things in a new way. These things you have heard before and you use them as you lego bricks. So I think AI is basically creative. We as humans do not like to acknowledge that. And the author of this video is an example, sorry. I remember times when computers where bad at chess, then they were beating the world champion, nowadays they are much stronger than any human. We should acknowledge the situation. ( I am again playing chess, with fun :) .)
The point of AI is that it learns from feedback, so it would be really interesting to see if someone like you could work with it to see how far it could go to be e.g, a classical composer. As I understand it, without encouragement it could never push the envelope in terms of creativity, but there again I think there’s a lot of professional musicians who don’t either
This morning (25.10.24) I completed my Rhapsody No.10 for Piano. The work is 126 bars long. When composing the piece I had not decided how the work was going to end and had several reasonably good ideas. In the end I composed an entirely different ending and modified several of the earlier sections. I will no doubt edit the work some more! This piece is short, some seven minutes but one can see that the decisions that needed to be made are far more complex in larger forms. I have composed 97 Symphonies (and much else) to date and have gone through the same processes throughout which apart from the length of sections, one has the problems of orchestration and balance. I am afraid I find AI wanting. The human mind is endlessly inventive which is why we have those wonderful modulations that Schubert (and others) made. I have to ask a question: Do we really want AI to create music for us (or anything else which requires the creative need of humans) at all. ? Another point: Is your dog an Italian Spinone? Mine is sitting about three feet behind me! Thank you for a most interesting presentation.
Amazing video, I conpletely agree with you. People often focus on the fact that these AI's lack emotion, and that's obviously a big part of the difference between humans and computers, but I think they really lack a logical understanding of music too. I personally think they will never reach it, even tho they'll probably become good enough to fool you
I've always thought that Mozart was god-like, since I was a child because of his his sense of phrasing and melody, which at his best is just the greatest of all time. The naturalness, lucidity, and often a simultaneous feeling of both joy and melancholy. Compare his best contemporaries, such as Haydn, who are just as technically proficient and have their own thing going on, but just don't have that Mozartian feel. In the end this could be captured by an AI, I guess, but it would need to be more sophisticated than the ones we have now, and even then, it would just be derivative.
"The naturalness, lucidity, and often a simultaneous feeling of both joy and melancholy." That's the best summary of Mozart's music I've ever seen. I've tried often to put my feeling about Mozart's music into words.
Just took one of the A.I. music websites for a spin today, focusing on early-mid Romantic piano, and checking out other folks generated “pieces”. None of them managed truly effective voice leading, and few made much sense with regards to cadences/progression/macro structure. BUT: Many did have interesting harmonic lines, chord changes, runs, ornaments etc. that I probably wouldn’t have encountered otherwise, but were very inspiring. As someone trying to learn composition and piano outside the conservatory, I wouldn’t listen to any of the A.I. works on a playlist with my favorite composers, but they were pretty compelling from the standpoint of providing unique ideas to try in my improv and composition. Trying to replicate Mozart is rough though, the classical and especially baroque seem so… mathematical in their precision, not a lot of leeway to get weird. I think it seems like a useful tool to accelerate the creative process, but can’t replace the human touch
@@griffinlester9098 I actually agree- as much as I'm uncomfortable with AI and can't help but feel at least partly disgusted- I do think it can be a useful tool. For instance, I could see myself taking the first few notes of this and trying to make something out of it. It could be fantastic to break out of writer's block. But it can never replace a good, thorough knowledge of counterpoint and voice leading. I saw a video recently about the famous software that beat the top "Go" player in the world. But later, a research team studied the AI that could supposedly beat anyone, and noticed some peculiarity in its strategy. They then used that, and an amateur who'd never really played "Go" before BEAT THE AI!!! The details are less important than the bottom line- the AI could not see how this man's strategy would lead to its losing, because on a fundamental level, the AI DOES NOT UNDERSTAND THE POINT OF THE GAME. It just blindly performs operations and doesn't know what the GOAL of the game is, the way a human does. They pointed out that ANY person who understood the game would have probably immediately seen this amateur's strategy and cut him off at the pass, but the AI did not BECAUSE IT DOES NOT UNDERSTAND THE GOAL OF THE GAME. Of course... the day that it does, if it ever comes, could be very very bad for all of us....
I guess. But I'm not convinced that AI is capable of producing meaningful results. It's like the old thing about the monkeys tapping out Shakespeare over time - I'm not sure...
@@themusicprofessor@themusicprofessor There are some incredible things being produced in the pictures/video medium. As for music yeah this sounds very weird but I don't doubt for a minute that in 5 years it will be hard to tell the difference
@@themusicprofessor I take your meaning & I entertain a great amount of skepticism that AI can produce anything of meaning; However, look how far technology has progressed in the last quarter-century. I have listened to classical music for quite a while (as a kid my folks played both kinds of music: classical & jazz), but only in the last couple years have I been trying to understand it, its history, music theory, &c, but I still know that I know very little & 'meaningful' is a very ambiguous word. I value your videos highly & the information you convey, while some of it does go over my head, I am sure that in time I will understand it to a better degree. I do know that I find meaning in your discussions. Keep up the great work & I'll keep watching.
Bear in mind that I am only a stupid scientist, quite uneducated about music, but the composer who most reminds me of a storyteller is Widor in his organ symphonies.
My opinion of Leopold Mozart's son is similar to Glenn Gould's opinion, but at least on the count of "artificial intelligence" being able to write new music in Mozart's. style, Gould is wrong. The computer was able to generate something that can pass for the real thing if we don't scrutinize it too closely. Udio is presumably a large language model (LLM) like ChatGPT. It has a vast library of music written by human composers and it can identify elements that occur frequently and put them together in a plausible succession. But, like you said, it doesn't understand the why of any of those elements. The difference between Wolfgang and Udio is obvious. You hadn't even played two bars of that early composition of his and I thought "Yeah, that's actually going somewhere." But in all this admiration of Wolfgang as a magically miraculous genius, we forget that he was in a way also an LLM, for he thoroughly studied the music of his father, and his father's contemporaries, like Michael Haydn and Antonio Salieri. Unlike Udio, Wolfgang understood what those elements meant.
I suspect the AI isn't acting even as deliberately as you suspect. I'm guessing it has just looked at a big pile of music and found asked, what is the most likely thing to come next? Over and over and not paying any attention to what has gone on five bars earlier. I think the "what" that comes next is evaluated more and more abstractly as the program gets more sophisticated, but if the machine has taught itself, these abstractions are nothing a human can really put their finger on.
It'd be interesting to compare generative classical music programs that aren't driven as directly by machine learning as Loki is. Stuff that is controlled a bit more heavily by human contrived algorithms. I remember CPU Bach many years ago, and I found it both impressive and unimpressive at the time, but it was a very different experience than this. Usually those programs have kinda pre-planned structures with intended repetitions, modulations, etc. which get filled in. This... by contrast it felt like Beethoven trying to improvise while having some dramatic short term memory malfunction. It seems to be living moment to moment with no plan at all.
For me, the title in the video's thumbnail--"AI CAN'T COMPOSE"--raised three important questions. Until I realized I'd misread it: 1. _"Why has the RUclips algorithm dropped a video about gardening/recycling in my feed?"_ 2. _"Who is this Al person anyway?"_ and 3. _"Why, exactly, is he not able to compost?"_
Regardless of whether it's "better", whatever anyone might mean by that, or criteria they might use to make such a determination, what sort of person would choose or prefer to listen to music put together by an ai over a human?
Compare Mozart's piano concerto K595 or clarinet concert K622 with his early piano and violin concertos. Beethoven's piano sonata op 109, 110, 111 with the piano sonatas op 2. Or the late string quartets with the early op 18. Chopin's nocturnes op 62 and the Polonaise -Fantasie op 61 with the nocturnes op 9 and the polonaises op 26. Brahms piano pieces op 118 or the clarinet quintet with the F minor piano sonata op 5. Debussy's late piano etudes and violin, cello sonatas with the early Suite Bergamasque. It's not only that in the later works they've developed into better composers, more sophisticated, subtle, innovative in their craft, construction. The later music also has a maturity, depth, profundity, intensity, insight far greater than in the early works not only because they were geniuses but because they've lived a human life with all the personal experience and growth that entails. How is an ai ever going to do anything like that?
Hi music professor! Been listening to your lectures for a while now and absolutely loved them. This one really made me want to comment however. How you described the AI as being a bunch on non sequiturs strung together was perfect, And I felt same. It’s fascinating that it got the general idea right but was just missing that human element. I wonder if it would become more or less accurate as the styles themselves became more formulaic and less based on communication. I would be interested to see how you felt if you told the AI to compose atonal/serialist piece in the style of Schoenberg or or a minimalist piece in the style of Steve Reich
Thanks for your comment. It would be intriguing to find out what results you'd get with atonal Schoenberg. I doubt you'd get an Op. 19 piano piece! You' probably get a bunch of atonal non-sequiturs! You'd think that AI could be programmed quite successfully to write serially because it's more about finding ingenious ways to manipulate permutations, inversions, retrogrades etc. but I think the best serial composers (Berg, Webern, Boulez etc.) are far more creative at this than AI would be. Again, with Reich, I think AI could impersonate the style in a bland commodified way, but the thing that makes Reich extraordinary - e.g. the way the harmony alters at just the right moment - is beyond the reach of AI.
First, just as clarification. I am musician, I've been playing guitar for 35 years . I am also graduated in Conduction, Composition and Arrangements. ( I don't know if this is the correct term in English) So... The biggest problem about musicians (and artists in general), is EGO. They (well, we I suppose) are so full of themselves. Every single "problem" mentioned can be corrected by making a better prompt. Also, in order to achieve a great level of composition with AI, your need to re-generated, try, A LOT of times. And honestly, this isn't too different from trying to create a melody. We come with an idea...we try, change, throw away, revist, try again, until we are satisfied. The other huge bulls*hit in the whole "soul" and "connection" . This is personal, and more often than not, what the composer felt and connected when writting a song is NOT what connects to someone listening to it. The meaning is, and it always will be, personal and unique. Music is mathematics. The "feeling" is purely each ones , particular, interpretation to it (as player or as audience). AI is a fantastic tool, wise musicians will use it to enhance their works, bring inspiration and achieve things that otherwise wouldn't be possible (cost wise, for example). Instead of throwing rocks, let's understand what the new era is bringing to us. We adapt. (Sequencers, remember? some people still hate them)
Insightful video. I think that even when A.I. solves these stylistic issues enough to closely approximate the sound of human composition, it will still be deficient in the most important area of them all -- inspiration. Human art is guided by our experiences as mortal, emotional, and organic beings. The decision to write something tragic or uplifting springs from our myriad emotions, our engagement with the struggles and achievements of life, our physical pains and exultation, and perhaps most of all, from human relationships. Composers are certainly affected by the phenomenon of predictive writing and one can feel the harmonic pull at the end of phrases but we also make decisions to break these rules according to how we feel and the creative vision of the work. In short, A.I. can imitate the patterns of human art but what it constructs means nothing and comes from nowhere. A cereal commercial jingle written by a human has more meaning than a two-hour symphony constructed by A.I. As for Mozart, Gould should have stuck to playing and left composition to composers. Wolfgang's compositions are irreproachable.
Interesting and thought-provoking. A few weeks ago , I was listening to Ray Chen's clips of ai generated music and those by real composers - I could instantly tell them apart. The same here. The ai clip is bland and doesn't get anywhere. You aptly said it doesn't tell a story. Earlier I was playing k296 3rd movement. At the end, the uncanny way Mozart arrives at the opening theme is sublime - ai doesn't come close. A 5 year old can compose, but I haven't seen anyone like the young Mozart. It seems Mr Gould doesn't like Mozart - why? If Mozart had lived a normal lifespan, who knows what he might have achieved.
Mozart is now one of my favourite composers, because i came to realize, that his music is all about what you dont hear. Its like Non-Music; maybe this was Mr Goulds point. But how could an AI ever construct such a paradoxon? Beethoven in contrast jumps right into that sphere, already digging deeper, and let you expierence the impossible. Formerly Bach points towards that door. Probably AI would have less problems "composing" a Beethoven piece or something later. With Baroque im not sure, but Bach seems too difficult for AI. And i think worst thing to do with AI would be Renaissance Music, Josquin, Isaak, Palestrina.
Interesting comment. I sometimes wonder if Renaissance polyphony is actually the way in for constructing a composing AI programme, because the rules are very strict and it might be possible to painstakingly train an AI programme to gradually acquire the skill.
I'd be interested to hear you rewrite this passage into something more coherent. I feel like it's a starting point of sorts, and if it were made more elegant, more human, it'd actually be a halfway decent piece.
One of my Professors at University once said: "Even a single tone (f.e. by a singer, string or wind instrument) can be like life, with birth, growing up to full power, decline and death." How could AI and sampled instruments EVER be able to create music coming from the soul? Unless it once will have a soul. Mozart himself (!) was bringing up the issue of "composing" with just a set of his premade musical phrases consisting of measures for beginning, in between and endings plus a set of two dice. You throw the dice and Voila! there's a "Schleifer oder Walzer". It's called "Musikalisches Würfelspiel" (musical dice game). Examples can be found online.
Do they not teach it anything about voice leading or melodic shaping? The repeated notes in the melody are clumsy - as you say, it seems to get trapped, and puts together a collection of formulae. It hasn't considered the overall direction of the music.
That's not how AI works, you can't "teach" or "mentor" it to do something. It has to create the connections itself. AI is also quite limited because it only thinks about what happens next, note by note, and not actually any ideas about the entire piece. It's like musical autocomplete.
@@michaelchen2821This is a very important point, essential really. Most people apparently think that AI is "programmed", which is fundamentally incorrect.
I did my first Muzio Clementi sampling yesterday as a 3rd-year beginner and was struck by how avant garde he was as compared to Mozart. Lots of weird chromaticism?
AI is great at composing music that superficially sounds like a specific genre and doesn't hold your attention for more than 5-10 seconds. Unfortunately, that's how most people listen to music: "Hey, this sounds familiar! Anyway..."
It reminds me of the sort of work produced by keen, but essentially untutored kids who have been experimenting with Sibelius or Musescore. (Actually, I should include myself as a 9 year-old, composer, though I predated computer technology).
I’m a little tired of people commenting on the music of Mozart and the Galant period without speaking the language, which of course is partimento and schema. Gould was a phenomenal talent as a pianist, but there are many born like him, not so many like Mozart. Everyone knows Mozart; not no many Gould.
You must understand that Leopold Mozart (Amadeus’ father and illustrious court composer) would’ve overseen and guided his composition process, too, smoothing his compositional crumples. His 5-year-old-written works were very much moulded by his dad’s experiences, and Leopold was no stranger to chicanery if it contributed to his son’s exaggerated prodigious veneer. I recall hearing he lied to crowds on a number of occasions about Mozart’s age - unnecessarily, yes - to paint this picture of a true genius.
It was Beethoven's dad who lied about his age. Leopold did certainly hot-house his son but the results were astonishing (there's a very detailed account by Charles Burney who gave the little boy all sorts of ear exercises to test out his genius, at the end of which he concluded that all accounts of Mozart's talents were actually less impressive than the reality).
The fundamental issue with "AI", i.e. "artificial intelligence" is that such kind of software is indeed "artificial" but very far from being "intelligent". The software behind analyses and recognizes patterns. Based on certain algorithms and statistics, it predicts what would be the most logical next word or note. So it would not be able to have an overall plan, wouldn't know what is a musical phrase, etc. Even the most sophisticated AI program would not, to my understanding, be able to "break" the rules of harmony in a surprising and beautiful way as have been able all the great composers. Feed an AI with medieval music and all what the computer will be able to produce will sound about like that but it most probably would never be possible to reach the complexity of romantic or later music. AI may generate some music in the style of X but it may not be able to create truly interesting pieces of music with a new musical language. I certainly don't doubt that AI generated music may find its way into "fast food music" but will fail when things start getting a bit more demanding.
AI can do anything basic. It can't basically do anything (original). Original being the operative word. It may be programmed to invent a better mousetrap but it could NEVER come up with the famous statement that if a man could invent a better mousetrap the world would beat a path to his door. Or write a play called The Mousetrap - a thing to catch the conscience of the king - which is performed within a certain well-known play, which AI would first have to create. And how could AI create Metamorphosen out of bitterly recalled emotions still wonderfully alive among the ruins of a culture that visited destruction on itself? How could it first look into Chapman's Homer (or know what it is) and then come up with the most transcendent sonnet of its language? Seriously my friend, you have not thought this through.
Exactly. Good that you mention Metamorphosen - a piece that distills a lifetime of experience into an astonishing artwork. What is art? It is creative virtuosity responding to the cultural environment in such a way that the creative work reflects a deep sense of what it means to be a human being. This is infinitely beyond the scope of any intelligence except our own.
No, AI cannot do everything that a human does. Such as: understanding what feelings and emotions are, how to generate them. Also: human intelligence has an important intuitive component, which AI will never have. Experiments with that all failed. Also: human perception, interaction and communication with the environment is extremely sophisticated and complex. Until now AI cannot handle the complexity of certain traffic situations or perfectly translate any language. It will never be able to do that. Don't be fooled by the current AI hype. There have been similar waves in the past. Every now and then the IT world generates some bubble for business reasons.
Thank you. You've pulled me in from the ledge in my fear of AI. Perhaps something to keep in mind, however, is that AI has been programmed by engineers - and current-day ones at that. Aesthetic? perhaps not. AI remains a parlor trick - impressive in its expression yet soulless - at least for now.
I guess if you gave the AI more specific instructions, it would create something closer to what one specified. But you could also wonder if what AI created is reflecting AI's situation being a boxed entity that only knows what it's been told. It doesn't know what's right or wrong, it only knows what the humans have told it what's right or wrong.
The great critic Hans Keller once memorably said, 'The composer creates certain expectations, well-defined expectations, which he proceeds to meaningfully contradict.' Gould's comments on Mozart's late keyboard works, describing them as stale and lacking in a sufficient quota of meaning, seem to resonate with Keller's generalised observation. There is no meaningful contradiction of well-defined expectations. The wonder is that Gould's observations could not in any wise be valid with respect to the operas, symphonies, concertos, choral music and chamber music from this same period in Mozart's output. Now my point is that Gould, with his incredible sharpness of intellect, could well have been predicting the rise of AI compositional programs in our times, ie, cleverly designed tools that when set in motion ape certain algorithms. To be satisfied with the result is to award the prize to Beckmesser rather than Walter; to prefer a provincial Kapellmeister's well intentioned gaucheries to the St Matthew Passion.
I do a lot of experimenting with AI music. I've heard some good pop, rock, rap, etc., but I've never heard a good AI classical piece or even a good jazz piece. The AI just lacks that ability to go outside of the box and really "pop off" the way a great musician can.
Having heard Gould's horrible compositions and many recordings where he "plays" Bach's music like it's a MIDI file, hearing him say "Mozart music could be made by a 5 year old" is one of the funniest thing to me 😂
The AI music almost has a resemblance of something by Debussy in the Children’s Corner at first with starting on ii, then going on to make perfect cadences with awkward repeated notes in the melody. Very strange.
The beauty of tonal music, such as the classical style, is telling a story with a coherent structure and phrases that elicit a specific emotion. AI and the computer will never understand that level of communication and human emotions, it can only calculate and mechanically manipulate patterns. Therefore AI cannot and will never be artistically creative. It will never generate something that makes people cry or feel peaceful.
If AI doesn't know basic harmony or how to create an actual phrase, I'm sure it would fail even harder at creating something atonal that relies on more abstract things like "gestures" and "color".
Fascinating Love it Photography was going to be the death of painting Definitely fucked up portraiture Took it out of the realm of the rich Then became a separate thing in its own right AI I think will follow similar parallel tracks to be manipulated by pesky humans - eventually- to create a different art
AI can't just create music; here's some text "in the style of an Elizabethan playwright": What brave new world that has unleashed the dogs of war to take any man’s horses for the world must be peopled this Bartelmy Fair. Mockery aside, in a few years AI may go beyond regurgitating snippets of cliche and begin to utilize underlying concepts in order to create more adequate outputs. We're not there yet, and the way AI works today won't get us there (because of the training set problem) but once people realize that the GPT approach doesnt work without adding set theory, we may see more interesting results.
Maybe. But I think people seriously overestimate the power of AI. Humans are an incredibly brilliant species (for all our ridiculous faults) and computers are just brilliant human inventions but nevertheless unthinking machines. They cannot understand and they cannot invent.
If Beethoven (or any composer one could argue is great) had been able to access AI, would he have used it? Well, 1. He would have soon found it too limiting. 2. He would have designed a better program. 3. He would have abandoned the whole enterprise and set about to just compose. My point is that if an artefact can only be as good as the program that created it, why not cut to the chase and compose the artefact? It is bordering on the absurd to imagine an AI coder feverishly creating a program that produces the very result he had in mind all along. I have in mind a kind of Mel Brooks mad scientist proclaiming a new invention that saves countless hours of futile toil as it pours forth whole epoch-making, game changing masterpieces each more wondrous than the last. With reference to some of the other commenters, to criticise AI is not to heckle; but the votaries had better point to some success and soon.
" I mustn't be mean..." " Loki barks"... This is the type of educationally wholesome content about music I live for..... Godspeed to you sir, and hats off for the amazing content you produce :)
I think the AI music did some things better than I could do but ultimately would have less artistic value, even though I'm a very bad composer. It would work great to provide classical ambiance that one was not intended to listen to at all but just hear passively. I don't think it would be worth paying enough attention to that one would actually remember any of it, because nothing is really memorable since it does lack meaning as you've said.
@@themusicprofessorindeed, and on reflection, the way Beethoven exploits that ambiguity is also impossible to imagine AI doing at present. That whole sonata movement is 'about' the opening, in a way which no AI generated music can be said to be 'about' anything. I think that's why none of it has any grip, so to speak, no narrative.
As someone who loves classical music but has absolutely zip musical talent, I can say that what I've heard and read about mediocre composers and even the juvenilia of great composers is, "Why do you limit yourself?" This is extremely evident n the AI music that is formulaic and lacks creativity. It could be a long time (if ever) that AI can actually compose something as great as Mozart's works.
16:00 : In my opinion, it’s really not a matter of time but more a matter of training. AI could absolutely imitate any kind of narrative as long as you feed it with enough material that contains this narrative. Of course no one has done that with classical music because it ins’t a field that present enough potential for people to invest the millions of dollars it takes to train a solid AI. The problem of these kind of turing test of AI generated classical music is that they dont feature programs trained on classical material (since it doesn’t exist) but rather pop (hence the title Whispers of Elegance which is clearly wrong for a classical composition, the beginning on subtonic or the weird rythms). However, if you ask the same program to generate a pop song it will give you a way more convincing result because it was mostly trained to recreate pop codes. With this in mind, I think that Gould was right : AI could replicate Mozart style (maybe not up to the level of sophistication of most of his pieces, but enough to fool most of us), but so could it do for any composer as long as it’s shown enough of the style in which the composer wrote. Anyway thank you for this video and more broadly for your content on RUclips which is always very qualitative. Have a good day
Of course, an AI can replicate existing patterns of classical composers. But it will not do what the greatest of the composers did, it will not develop new patterns that work and evolve, revolutionize music. Additionally, it will not, by all means, produce music that has a true meaning for the listener (like a lot of Chopin's wildly emotional, deep music, or the programme music of most Romantic composers.) Music having a meaning, making you feel one or many emotions, hear musical stories. No AI in the world will produce an - equally unique - piece as emotionally powerful as Ballade No.1 of Chopin because a computer, for all its clever processing power, can't comprehend what was going on in the composer's mind when he put those specific harmonies together to inject specific feelings into the listener. And comparing Pop song patterns with classical music.... Please. Literally any mediocre amateur musician (like me) can write a convincing Pop song. 90% of modern Pop contains around 1% of the complexity of a Mozartian or Beethovian symphony or even an art song from Schubert.
This video won't age well! haha I'd say AI creating Bach-sounding compositions would be almost child's play since there's kind of a very small set of musical rules and theory to learn. Train the AI by feeding it enough Bach pieces and it probably will start to generalize and understand all of the ways it could make notes that make sense in a timeline. oh, I just did a little search. Bach Bot. Already there is one that can create these kinds of streams of notes and harmonize them.
Obviously I disagree: this video is simply me critiquing AI composition in 2024. It's an accurate representation of the situation now. But I also think that AI has a very long way to go to produce anything of genuine compositional interest.
“That doesn’t really work, does it?”
You beat me to it :)
Too many notes!
Ohhh, I see what you did there!
@@manco828😂😂😂😂
"That doesn't really work, does it? Did you try...shouldn't it be a bit more, or is this this Better? What do you think?"
Glenn Gould is partially correct when he said that a 5 year old could compose music of the same quality as Mozart. However, only one such 5 year old ever existed and that was Mozart himself.
About Mozart, there's nothing Gould could ever have done correctly, let alone understood correctly. A cliché himself, Gould attempted to gain notoriety by slamming Mozart. Being a narcissistic showoff, Gould grabbed the wrong end of the stick. Geniuses such as Seymour Bernstein called him "a travesty", and even Leonard Bernstein had to make a disclaimer prior Gould's performances. For a vast majority, Gould is nobody. For those who are aware of his disrespectful prejudiced remarks on Mozart, he is a pathetic bigot uncapable of keeping up with playing Mozart. For an insignificant minority, he was a visionary. Mozart, however, has always been regarded one of the best three composers ever, and certainly, the most beloved by the vast majority.
*pArtIally
Talented 5 year olds are great, but I’m more interested in what they’re doing 20, 30 or 40 years later.
😂what does that have to do with the video or op's comment? Fwiw, my main criticism is he makes great music sound like Mozart!
Weird how emotional and biased your tirade is. What's the story there?? @lumenvera9452
@@lumenvera9452 I came to accept that Gould was basically trolling the classical music world, more than actually believing his most outrageous statements - or both. In any case, with most of his interpretations (that includes Bach), I never liked his sound, which was probably caused by his Yamaha piano and recording antics. I basically stopped listening to his recordings years ago. He was clearly a vastly talented musician, but he became a victim of his own intellectual eccentricity. With Gould, the line between the inspired and the grotesque is often blurry.
Really appreciate this video.
I work professionally with LLM and AI as a programmer. It is essentially auto-suggestion/correction on steroids. When current AI writes music, it is using advanced text prediction, "text" being the general programming term for "tokens" (small bits, or in music, notes), or how humans cogitate, and break down complexity into smaller parts that can be modeled. It predicts what comes next based on input. We talk about "models", or the code that does the prediction.
There are technical model limitations, but these limits decline as humans program better models, and as we put more compute power in.
At no point do models "reason", at least not on our current reflection on how humans reason.
So these models mirror back to us what humans expect. If imagination is expected, programming hasn't reached this point.
(Kudos for the Schoenberg reference!)
Thank you for your really interesting insights.
I think its possible that AI is in fact telling us just how humans reason. Our brains are not magical; essentially they are huge webs of things that associate. So if I said River to you, you might say Dance, or Fish, or whatever. If you ever watch your thought patterns go by, its essentially just a torrent of loose associations. So in terms of thought generation, it could be doing just what AI does. The difference, of course, is that AI is trained on a massive dataset, so its like expecting a thousand human brains wired together to come up with something coherent. All the different directions its pulling conflict. Our minds have been wired on a small dataset, and it is constantly curated by the ego - what we call cognitive dissonance, where we toss out stuff that conflicts with previous input. So there's a fair amount of error "correction" going on. Frankly, I'm amazed at what AI comes up with. A lot of people on Internet forums should be so coherent.
@@riverstun I would say Phoenix
@@themusicprofessor Watch this video:
Can You Hear The Difference Between AI and Beethoven?
Carl-Friedrich Welker
Udio here is making beautiful music. It did not have the weird stuff you sometimes hear on Udido made stuff.
It is going o get progressively better and people will be programming it detect weird-isms and leave those out.
But what are we talking about here? We are talking about simulating a Mozart piece so it would be stylistically unoriginal
yet original in the particular notes used. Somebody like you could work with the programmers
and train it not to do certain awkward things.
The programming could be even more improved if a particular program was designed to generate Mozart's music specifically, so it could be trained
on more of Mozarts music than it already is. Then listen to the results and keep refining it with new rules.
But how could it generate something that was more unique not just imitation of a composer's style?
You train it with Mozart pieces but then add in some Chopin or Bill Evans or all three or use an occasional Indian scale.
This is all subjective, you program a hybrid to your liking and in a proportion to your liking.
Have it generate 10 songs. What if one was great or had a section that was great? That may be possible.
What about the human feel?
> that can be programmed,
And the human element is you picking which song or section you like.
Does anyone rememeber when the Google doodle was an "AI" that harmonized bach chorales? It was terrible and everyone at work tried to convince me was good. But none of them actually listened to bach. Very bizarre experience
Yes! Glad someone else shares the same opinion on that.
I spent way too long attempting to tweak it to produce a single beautiful thing
I used that google doodle to write a harmonization of "Ein Feste Burg". To do this, I ran each line a few times, and picked the best. Here is the full thing: ruclips.net/video/ByTgaLzMwgQ/видео.html
It's not like anything has happened in the field of AI since 2019, AMIRITE?
@@eti313 Or, frankly, since the Analytical Engine in 1837.
I'm a copyright law professor (from Thailand). The emergence of generative AI has challenged the copyright lawyer's perspectives on authorship. Basically, copyright law doesn't really understand (or care) about how people create a copyrightable expression. It only assumes that creation (authorship) is fundamentally a human activity. The french copyright tradition (droit d'auteur) has a doctrine called "Imprint of Personality" (l'empreinte de la personnalité) which means that a copyrightable expression must reflect a human personality behind an expression (for example, everyone draws a cat or a circle differently). The anglo-american copyright tradition relies on the "skill-judgement-labour" requirement, which I think an inferior theory to the french doctrine (and a weak Turing test for AI to pass). It's easy to be impressed with how quickly AI evolved to the point where it can generate musical compositions; and AI authority like Geoffrey Hinton claims that AI can "think" and understand like a human being. However, your demonstration in this video hints that an AI perhaps still lacks an "imprint of personality" in a sense that it does not comprehen the speech-making nature of music creation.
I was looking for an interesting content about AI for my copyright law class next semester, and I think this is it. Thank you!
Superb comment. Thank you for your insights.
The French doctrine could allow for the artists, whose materials trained the generative AI, to get paid. There would have to be a way to identify which training materials reach a predetermined threshold of influence (in music, sometimes it's really obvious, but usually, I think it would need to be self-reported by the AI). This could be impossible, IDK... the copyrights for each piece of art would start to resemble mortgage derivatives, and become burdensomely complex.
I study copyright law in graduate school. Fascinating with the development of AI. Great video.
I think using the work of others in the “training” dataset is in and of itself a violation of copyright. This work is created by interpolating data from the training set (ie pseudo randomly mapping an input note set to other notes in order to obfuscate the source material that’s being plagiarized). This work doesn’t sound good because the trainers don’t understand classical well enough to train out the bad results. AI doesn’t think like a human at all and the people spreading that BS are either lying or stupid. In general they are lying because they have economic interests in AI. AI doesn’t understand any of this, it just has some characteristics of the training data others are missing because the trainer lacks the musical talent/knowledge to eliminate oddities from the output.
I don't really understand the copyright issue, as no work is getting copied or distributed. And in terms of thinking like a human, you are confusing weak music generation AI with advanced LLM. Currently, it isn't a question of will AI systems match elite human levels, it's just a question of when. And what does a million super-intelligent brains combined sound like in unison - and that's less than ten years away.
Excellent content as usual. This is quickly becoming one of my favorite RUclips channels. And, as an amateur pianist who is much less of a keyboardist than Mozart was at 5, I nonetheless have had to immediately find and learn a couple of pieces I heard through this channel - namely Ravel's Prelude in A Minor & Beethoven's Bagatelles Op126 no5. Thank you!
15:00 great discussion! I especially love the line, "it can give you the symbols, but the signifier is disconnected from the reality that it's meant to symbolize"
Cheers! Yes - AI doesn't understand meaning!
@@themusicprofessorLacan comes to mind in the mentioned sentence, although it also comes from linguistics/de Saussure I believe.
@@christophedevos3760John Searle
@@bs92491 right, seems very interesting, I need to read this man.Also in regard to his critique on reductionism. Thanks for pointing this out.
It's funny to me, but as a composition teacher I have this overwhelming urge to fix it. "Funny little tune, but it yielded some good things..."
Yes, something like that. Amadeus of course.
The 5 year-old Mozart knew how to vary his material subtly. In the first four bars the melody drops down on beats two and three, and in the answering four bars he goes up. He breaks the formula - simple but effective.
He's amazing - even at 6.
@@themusicprofessorhe was better composer in six than almost anyone might even hope to ever be.
@@unwrought9757 That one might be a bit of an overstatement.
@@thepotatoportal69 it might indeed. But I might be right as well:-)
I read in a book Mozart devides his tune into 5 parts . I can't remember d names .he thought 5 as a key number. His creativity has some law ,but i don't remember dt book much.
Reminds me of the old story of the "cargo cult": South Pacific islanders building wooden planes as substitutes for aircraft (now gone) that had delivered them supplies in WWII. If you don't understand the basic principles of propelled flight your simulacrum isn't going to perform. Similarly if your AI doesn't understand the workings of the human mind (or is it soul?) it is going to fall far short of a satisfactory conclusion. Interesting video. Humans 1 AI 0
Eerily reminded of Orwell describing the "soulless" music (and fiction) written by machines in "Nineteen Eighty-Four". And the "Proles" didn't even care...or rather, they didn't seem to notice.
Yet another example of Orwell's genius for perceiving potential horrors.
Машины там не создавали ничего, кроме "жанров" для пролов , такое действительно ИИ может писать не хуже. Но ничего большего, чем ничтожество.
Thank you, very interesting.I checked out the same platform shortly in the field of Jazz, which is my home base as a musician and teacher since 40 years.Basically, what AI does in the moment is pattern recognition. If you ask AI to create Jazz, from blues or bebop to free improvised Jazz, the thing you are talking about becomes even more obvious: It is just a repetition of clichés over and over, without any creative spark.A good jazz Musician is even faster, because Improvisation happens on the spot, in realtime. The best examples are all the duo improvisations of Joe Zawinul
and Wayne shorter, just check out the "In a silent way" version from the live album 8:30 and remind yourself, that this is completely improvised in the moment-if you see the transcription of it, it is a perfect composition. A raw frame of the melody is the only thing it has in common with
the many other versions (like the one from the last Miles Davis concert in Paris 1991).
You should do a video on some of Mendelssohn’s work. I really like his piano sonatas and songs without words. I also got to sing part of his oratorio Elijah, and I think they’d all make for an interesting video. Great content btw!!
Btw, this is the best and most detailed, and precise analysis of AI vs. Genius Composer I've seen. Thanks for the video and I'd love to see similar videos on this topic if it suits your inclinations.
Another thing to point out: on a sociological level the "unfortunate" thing about music is probably its greatest gift: it simply "sounds good." A simple chord progression sounds pretty and is pleasurable on its own and they're not hard to use so the alarms flare up when it turns out computers can make pretty music and pictures too, supposedly revealing our pretentiousness and affirming philistine doubts. But that says less about "AI catching up to Mozart" than diagnosing a culture which increasingly can't tell the difference between prettiness and art, unoriginality and familiarity, surfaces and realities (at least Gould distinguished between prettiness and art; he simply attributed the wrong one to Mozart)
Perceptive comment. Yes, our era is in peril of becoming deadened to quality in a sea of corporate blandness.
Great comment. Could not agree more.
In many ways I agree, but I also think that the art worlds desire to separate beauty from art, something which has been going on for a lot longer than generative AI is in many ways also a part of why this is. The point of so many more modern artists be it musicians or painters or whatever has been to treat art as a social phenomenon, which essentially makes whatever is treated as art into art. I think it's the combination of an art world which is in many ways, one might even say "ideologically" opposed to making their art beautiful and enjoyable to "the masses" being forced to recognize that people have so drastically fallen off the boat that they don't even really know what art is anymore, because it's stopped meaningfully approaching them. I think this is what leads to the lionization of certain styles which can easily be copied by AI and that style, not the art behind it becomes the art.
@@JustAGuyWhoLikesStuff. Fully agree. Though admittedly "beauty" is harder to pin down than "prettiness," I guess I could have said the difference between beauty which may include prettiness and mere prettiness. But the art world's disdain for beauty definitely contributed (though you can't really blame them) and the escalation of Modernism flung us into that rootless void we are in now where no one no longer dares to make an assertion and all discourse about art seems to end in "well, art is subjective" (imagine anyone doing that Gould program now)
Well this is a thing that the AI discussion always forgets, perhaps we are simply not very good listeners anymore. It takes good music and good listeners. In a few generations time, dumbed down by AI assistance in multiple domains, perhaps very few will even have the listening skills left to appreciate the great composers best music.
Another great presentation! The reason bar 6 sounds like bad counterpoint is the parallel 5ths between the melody and bass voices...and then Matthew points out the bar 14 'development' of that bad sound-again parallel 5ths-and the non sequiturs continue. There is absolutely no replacement for that thing that made Mozart a master musical storyteller. I think Matthew has just proved there is another level to the "Mozart Effect"...He is Impervious to AI...and so,
Loki (Mozart) 2
Matthew 2
AI. 0
Gould. -1
I recently heard the 5th symphony by Beethoven's pupil Ferdinand Ries which has a 1st movement that sounded a whole lot like that other famous 5th symphony by his mentor (even motivically), and yet you could tell that there was something missing, despite it being very much in the style of middle Beethoven, but I couldn't quite put my finger on what was wrong with the music. I was wondering if you could make a video comparing the two pieces and try to explain what was Beethoven's compositional secret that his lesser contemporaries lacked (even when they consciously tried to imitate him). I think that could be very educational as I really enjoy your way of analysis and breaking down the musical elements.
Wow! I didn't know the Ries. Just listened to the first movement (ruclips.net/video/LnTf57Ix4v8/видео.html). I think it's rather brilliant actually! It's not easy to write a really good pastiche of Beethoven! Ries is very interesting though because it seems like a very clear homage to his friend and mentor (not just the 5th but also the 1st and 2nd, the 6th and 8th are detectable). So, even though it's clearly modelled on the motivic material of Beethoven 5, the Beethoven that really appeals to him is clearly the more classical even-numbered (not the visionary odd-numbered) style. This is somewhat similar to the early symphonies of Franz Schubert and I think it shows us that the Beethoven's better-behaved, more classical style was more popular with his friends and contemporaries. You'll notice that Ries's development and coda are much less ambitious, less wild and extravagant than Beethoven. It's a lovely piece. Thank you for sharing and yes perhaps I will do a comparison...
@@themusicprofessor Yes, Ries was no doubt a highly skilled composer and it wasn't my intention to put him down. It's also interesting how he has moments anticipating later romantic composers like Mendelssohn. I also noticed how Ries appears to draw most of his inspiration from early Beethoven (which is interesting since he composed most of his symphonies after Beethoven's 8th was completed), and it's even more noticeable in the 4th symphony that has an introduction clearly modeled on B's 4th, and a scherzo modeled on the scherzo of B's 1st (there's a very fine new cd on the Ondine label featuring both works which is the one I heard). And still I feel like Ries is making all the right gestures, and is capable of writing good themes and develop them like Beethoven, and yet it doesn't add up to a satisfying whole like Beethoven.
I would need to listen to more. I'm impressed by what I've heard. The style is so close to (the more classical) Beethoven that it's almost a pastiche. But it's very imaginative and beautifully crafted. I wonder what Beethoven thought about Ries's symphonies - presumably he was flattered by them! In terms of satisfying whole, Beethoven's standard is remarkably high and he pushes the art in weird and wonderful directions whilst somehow maintaining control over form. He also has a rare gift, in terms of the overall architecture, and of course the ideas themselves are so compelling, and so inventively handled and developed, and the rhythmic component, and phrase structure, is so imaginative, and all his music has a risky, visionary quality which is also unique, so generally speaking, its very hard to do better! Perhaps his only contemporary who managed it was Schubert.
@@themusicprofessor How about something on the influence of some of Haydn's symphonies on Beethoven's work, and what Beethoven does "better" than Haydn to take it further? It seems at times Haydn is unfortunate in the sense he was the giant of music whose shoulders Mozart and Beethoven stood on, and so Haydn ends up buried in the mud!
Another great analysis Matthew, thank you. It sounded to me like someone doing a good impersonation of language they can't actually speak.
Or they know the words and throw them together badly or in the wrong place, so they say good afternoon when they need to say goodbye.
Dame Misuko Uchida analysed a Mozart sonata on TV. She played part of it and then said one note was wrong. She was right. She then showed that it made sense 30 bars later. Then it turned the piece from being mediocre to a masterpiece. She said "it is beyond the intellect, it is pure intuition, it is genius". From memory, the piece was k545, but I might be wrong.
It’s great that you engage with this on its own terms while offering your expert critique. Really interesting building on your ‘dialogue’ with GG. Good not to reject these developments out of hand, but as you say retain what the human condition uniquely contributes to music. Great you bring us back to Mozart to underline the point. 👏🏻👏🏻👏🏻
You are so right about that story-telling aspect of Mozart’s compositions! That is indeed the one thing that’s missing from AI-generated output. (Generally really, not just for music.) Thank you for that insight.
It helps that you can brutally critique the AI because you’d never go that hard on a real student but it’s illuminating to say yeah that sucks and here’s exactly why
I bet he would haha
@@JazzGuitarScrapbookagree and he'd be doing that student a service by providing that critique.
I'm much kinder to students - and they produce much more interesting work than this!
But the computer deserves it, ever so much! 😈
@@themusicprofessor i really doubt that
I entirely agree with you. We oughtn't admire people merely for celebrity status. But, in so far as Mozart is concerned, Gould is quite mistaken I am afraid. I sincerely agree with you about AI. The ability to assign meaning to things and to create things with a purpose is significant only of the mind. And no mind can do so with quite so much excellence and sincerity as that of the human mind.
I remember playing the Minuet in F 18 years ago✌️
The joke's on you! The computer wrote its own "Ein musikalischer Spass."
Except it's not funny!
@@themusicprofessor Neither is Ein musikalischer Spaß
It seems like a nice way to get ideas for melodies and themes.
Humans can produce better melodies and themes than these!
@themusicprofessor that is very true. But these aren't bad. Humans can make worse too
Nice analysis! Thank you!
Brilliant, thx for your work.
AI is currently on a level where it looks impressive to an amateur but experts still see the flaws. However, I think that will change soon. AI is right now a little like a person with some kind of "autistic savant syndrome". It can answer almost any question, but you have to push it a little in the right direction. E.g. you could tell the AI "Bar Y sounds a little weird because of X, how could it be improved?". I'm sure the result would be quite good. Maybe you should try that in the next video :-)
I can think of no better introduction to AI. Great show. So fascinated, I am.
3:28 This is no ending - I think it was meant to be, by AI, wondering about your question to the expert.
Speaks for the encyclopedia I read last night (something RUclips's AI just cannot know) which said that any piece of music leads to its ending. This, too, speaks for your criticism of AI I yet not able to share. To me, it sounds like from the authors. I shall wathch through, though. I think I shall willingly subscribe, especially I like the clickable timeline. Maybe Ai by RUclips even understands irony that is hidden - in my case that would be great misunderstanding (to test AI, I hereby explain "because you are enabled to click forward". I do not think my comment will be deleted by RUclips, neither by you.
Keep it up! Cheers! Great endeavor.
As someone who's been assembling very large databases and analyzing them using complex interlocking self-derived algorithms for decades - not to produce finished compositions, but to create successive (carefully defined) first-species counterpoint examples blending 12-tone serial conventions with basic Hindemithian harmonic theory for use as compositional source materials displaying the full range of potential 2-, 3-, and 4-voice harmonic progressions available for every possible iteration of a tone row at arbitrarily predefined frequency levels - I quite agree with the observations here offered. Careful and informed computer analysis can indeed produce theoretically correct streams of contrapuntal processes; but the infinitely various layers of immediate decision-making involved in composing any particular measure of potentially persuasive and artistic musical composition is - as of yet - far beyond the capabilities of any computer process. Computers may and will certainly continue subtly adjusting and refining their understanding of music theory and practice based on continuous, earnest feedback from human beings; but even a Beethoven, I believe, could not successfully train an AI to recognize true invention in the necessarily complex intellectual processes and emotional reactions involved in every single measure of any genuinely artistic composition.
I think it's pretty incredible. You can tear the shreds as much as you like, but I can't even begin to imagine how they came up with software to write something as superficially good is that
I tried it with the prompt, "a piano piece composed in the 1770 to 1790 time frame," and the result was utterly horrendous. The AI produced two sound clips, and one of the clips even had a piano in the background. So far, classical music is very definitely in the domain of humans.
robots can only produce robotic music - so let the robots listen to it
That first cadence is hard to fix- the counterpoint mistakes aside, it feels like it can't make up its damn mind if it wants to be a half cadence or an imperfect authentic cadence. lol
Great content, I enjoy it immensely! And Gould is wrong!
Iteratively, human prompting the AI and selecting, that’s where it’s getting interesting. On my channel are some examples of what I do. I still sit several hours on one song, selecting „next few bars/notes/sounds“ out of tons of possibilities it generates.
Yeah, the AI sounds directionless. But like you, I think it's impressive it catches on to certain tropes. It's kind of like teaching a dog to press buttons that "say" a word to get a treat. For the dog, it's a routine to get something tasty, and for us, it's entertainment. For the computer, I guess it's like satisying some sort of math or command or something. (I don't know-I'm no computer scientist, but the man who taught me counterpoint over the Internet, Bernard Greenberg, was, and he actually did work on AI music!) And then for us, we hope it's art.
There's some fundamental misalignment between what the computer "wants" (if we can call it that) and what we get from art. If anything, this is more scientific than artistic! I think I can accept the existence of increasingly sophisticated AI art in a vacuum, ignoring its effects on human artists, the environment, and so on. Art that I perceive to be great is special to me because it came from someone who existed somewhere. "Wow. Someone did that," my English teacher would say back in high school.
I am new to the channel, but instantly subscribed. I was wondering if there was any way to access your transcriptions?
"A five-year old after a few weeks of music theory". An interesting premise, Mr Gould. Canadian kindergartens are clearly quite unlike our own.
(PS, as I've said on another comment I am a real fan of Glenn Gould - at the piano or at the microphone, I love listening to him. But part of my fandom is the delight in pointing out the bits where he's totally off the rails.)
Edit: Oh, you make the five-year old point, but you leave it to the end.
Yes, and a true Gould fan knows that - he wants us to disagree!
@@themusicprofessor As conductor George Szell said of him: "This nut's a genius!"
I stand in awe of your transcription capabilities! I once prepared one of the David Cope computer-generated "Bach fugues" to do as an encore piece. Number whatever, but in Dm. 10 minutes? It took me something like four hour-long sessions stretched over four days. I can memorise, sight-read and improvise like a Banshee but I find transcription horrendously difficult.
I played that fugue in a very Gouldian style (but I am by no means as wonderfully gifted) and then asked if anyone recognised it or could date it. Before saying it had been written in 1992 (or whenever), by a computer, and had taken it 3.2 seconds.
It is like the early chatbots. The larger structure isn't there, and it has used some out of context ideas. I think this is a quantity of training data problem. It's not learning what's not expected.
I think you are completely right. This is a very early stage of AI where the musical aspect ist not very strong yet, because there is very limited knowledge of music built in. The current models were built around text, not around music. So the shortcomings are explainable, I think.
Creativity for me is not creating things out of the blue, but combining things in a new way. These things you have heard before and you use them as you lego bricks. So I think AI is basically creative.
We as humans do not like to acknowledge that. And the author of this video is an example, sorry.
I remember times when computers where bad at chess, then they were beating the world champion, nowadays they are much stronger than any human.
We should acknowledge the situation. ( I am again playing chess, with fun :) .)
The point of AI is that it learns from feedback, so it would be really interesting to see if someone like you could work with it to see how far it could go to be e.g, a classical composer. As I understand it, without encouragement it could never push the envelope in terms of creativity, but there again I think there’s a lot of professional musicians who don’t either
Interesting thought.
This morning (25.10.24) I completed my Rhapsody No.10 for Piano. The work is 126 bars long. When composing the piece I had not decided how the work was going to end and had several reasonably good ideas. In the end I composed an entirely different ending and modified several of the earlier sections. I will no doubt edit the work some more! This piece is short, some seven minutes but one can see that the decisions that needed to be made are far more complex in larger forms. I have composed 97 Symphonies (and much else) to date and have gone through the same processes throughout which apart from the length of sections, one has the problems of orchestration and balance. I am afraid I find AI wanting.
The human mind is endlessly inventive which is why we have those wonderful modulations that Schubert (and others) made. I have to ask a question: Do we really want AI to create music for us (or anything else which requires the creative need of humans) at all. ?
Another point: Is your dog an Italian Spinone? Mine is sitting about three feet behind me!
Thank you for a most interesting presentation.
Amazing video, I conpletely agree with you. People often focus on the fact that these AI's lack emotion, and that's obviously a big part of the difference between humans and computers, but I think they really lack a logical understanding of music too. I personally think they will never reach it, even tho they'll probably become good enough to fool you
I've always thought that Mozart was god-like, since I was a child because of his his sense of phrasing and melody, which at his best is just the greatest of all time. The naturalness, lucidity, and often a simultaneous feeling of both joy and melancholy. Compare his best contemporaries, such as Haydn, who are just as technically proficient and have their own thing going on, but just don't have that Mozartian feel. In the end this could be captured by an AI, I guess, but it would need to be more sophisticated than the ones we have now, and even then, it would just be derivative.
Yes, it would not be original, it would not evolve and it would not make you FEEL the same.
"The naturalness, lucidity, and often a simultaneous feeling of both joy and melancholy." That's the best summary of Mozart's music I've ever seen. I've tried often to put my feeling about Mozart's music into words.
10:07 Quintet string in C major KV 515, 3. Menuetto? It begins with a D, although it's in C major.
It does - but I'd say the harmony there is essentially V - I.
@@themusicprofessor Sure; as a translator, I think it's a relief that AI can't compete with Mozart 🙂Its 'compositions' are nonsensical
No one and nothing is or even might be better than Mozart.
I love Mozart ❤ Beautiful pieces that I play on my clarinet. Just 4 months in learning piano so I'm nowhere near to playing that kind of music yet 😢
If this was an assignment submitted to me or any of MY harmony, counterpoint, or voice leading teachers... it would have a LOT of red ink all over it.
Just took one of the A.I. music websites for a spin today, focusing on early-mid Romantic piano, and checking out other folks generated “pieces”. None of them managed truly effective voice leading, and few made much sense with regards to cadences/progression/macro structure. BUT: Many did have interesting harmonic lines, chord changes, runs, ornaments etc. that I probably wouldn’t have encountered otherwise, but were very inspiring. As someone trying to learn composition and piano outside the conservatory, I wouldn’t listen to any of the A.I. works on a playlist with my favorite composers, but they were pretty compelling from the standpoint of providing unique ideas to try in my improv and composition. Trying to replicate Mozart is rough though, the classical and especially baroque seem so… mathematical in their precision, not a lot of leeway to get weird. I think it seems like a useful tool to accelerate the creative process, but can’t replace the human touch
@@griffinlester9098 I actually agree- as much as I'm uncomfortable with AI and can't help but feel at least partly disgusted- I do think it can be a useful tool. For instance, I could see myself taking the first few notes of this and trying to make something out of it. It could be fantastic to break out of writer's block. But it can never replace a good, thorough knowledge of counterpoint and voice leading. I saw a video recently about the famous software that beat the top "Go" player in the world. But later, a research team studied the AI that could supposedly beat anyone, and noticed some peculiarity in its strategy. They then used that, and an amateur who'd never really played "Go" before BEAT THE AI!!! The details are less important than the bottom line- the AI could not see how this man's strategy would lead to its losing, because on a fundamental level, the AI DOES NOT UNDERSTAND THE POINT OF THE GAME. It just blindly performs operations and doesn't know what the GOAL of the game is, the way a human does. They pointed out that ANY person who understood the game would have probably immediately seen this amateur's strategy and cut him off at the pass, but the AI did not BECAUSE IT DOES NOT UNDERSTAND THE GOAL OF THE GAME.
Of course... the day that it does, if it ever comes, could be very very bad for all of us....
Great video yet again, just wondering, do you happen to do feedback on compositions?
I do, although I am busy. We are planning to do something like this on the channel soon.
Well, as you Brits are fond of saying; 'it's early days, yet'.
I guess. But I'm not convinced that AI is capable of producing meaningful results. It's like the old thing about the monkeys tapping out Shakespeare over time - I'm not sure...
@@themusicprofessor@themusicprofessor There are some incredible things being produced in the pictures/video medium.
As for music yeah this sounds very weird but I don't doubt for a minute that in 5 years it will be hard to tell the difference
@@themusicprofessor I take your meaning & I entertain a great amount of skepticism that AI can produce anything of meaning; However, look how far technology has progressed in the last quarter-century. I have listened to classical music for quite a while (as a kid my folks played both kinds of music: classical & jazz), but only in the last couple years have I been trying to understand it, its history, music theory, &c, but I still know that I know very little & 'meaningful' is a very ambiguous word. I value your videos highly & the information you convey, while some of it does go over my head, I am sure that in time I will understand it to a better degree. I do know that I find meaning in your discussions. Keep up the great work & I'll keep watching.
Bear in mind that I am only a stupid scientist, quite uneducated about music, but the composer who most reminds me of a storyteller is Widor in his organ symphonies.
My opinion of Leopold Mozart's son is similar to Glenn Gould's opinion, but at least on the count of "artificial intelligence" being able to write new music in Mozart's. style, Gould is wrong. The computer was able to generate something that can pass for the real thing if we don't scrutinize it too closely.
Udio is presumably a large language model (LLM) like ChatGPT. It has a vast library of music written by human composers and it can identify elements that occur frequently and put them together in a plausible succession. But, like you said, it doesn't understand the why of any of those elements.
The difference between Wolfgang and Udio is obvious. You hadn't even played two bars of that early composition of his and I thought "Yeah, that's actually going somewhere."
But in all this admiration of Wolfgang as a magically miraculous genius, we forget that he was in a way also an LLM, for he thoroughly studied the music of his father, and his father's contemporaries, like Michael Haydn and Antonio Salieri. Unlike Udio, Wolfgang understood what those elements meant.
I suspect the AI isn't acting even as deliberately as you suspect. I'm guessing it has just looked at a big pile of music and found asked, what is the most likely thing to come next? Over and over and not paying any attention to what has gone on five bars earlier. I think the "what" that comes next is evaluated more and more abstractly as the program gets more sophisticated, but if the machine has taught itself, these abstractions are nothing a human can really put their finger on.
Glenn was so cringe for his takes on Mozart.
It's just Shimano vs Campag.
It'd be interesting to compare generative classical music programs that aren't driven as directly by machine learning as Loki is. Stuff that is controlled a bit more heavily by human contrived algorithms. I remember CPU Bach many years ago, and I found it both impressive and unimpressive at the time, but it was a very different experience than this. Usually those programs have kinda pre-planned structures with intended repetitions, modulations, etc. which get filled in. This... by contrast it felt like Beethoven trying to improvise while having some dramatic short term memory malfunction. It seems to be living moment to moment with no plan at all.
Loki is the dog!
@@themusicprofessor LOL I totally got my wires crossed on the name of the AI program. Well he certainly doesn't deserve that criticism.
@@themusicprofessorLoki could be a very suitable name for an Ai program as well, the trickster god of imitation and make believe.😊
For me, the title in the video's thumbnail--"AI CAN'T COMPOSE"--raised three important questions.
Until I realized I'd misread it:
1. _"Why has the RUclips algorithm dropped a video about gardening/recycling in my feed?"_
2. _"Who is this Al person anyway?"_ and
3. _"Why, exactly, is he not able to compost?"_
Well I doubt AI can compost either.
Very nice insight into this controversial subject..thank you professor.
Regardless of whether it's "better", whatever anyone might mean by that, or criteria they might use to make such a determination, what sort of person would choose or prefer to listen to music put together by an ai over a human?
Let AIs listen to AIs
Compare Mozart's piano concerto K595 or clarinet concert K622 with his early piano and violin concertos. Beethoven's piano sonata op 109, 110, 111 with the piano sonatas op 2. Or the late string quartets with the early op 18. Chopin's nocturnes op 62 and the Polonaise -Fantasie op 61 with the nocturnes op 9 and the polonaises op 26. Brahms piano pieces op 118 or the clarinet quintet with the F minor piano sonata op 5. Debussy's late piano etudes and violin, cello sonatas with the early Suite Bergamasque. It's not only that in the later works they've developed into better composers, more sophisticated, subtle, innovative in their craft, construction. The later music also has a maturity, depth, profundity, intensity, insight far greater than in the early works not only because they were geniuses but because they've lived a human life with all the personal experience and growth that entails. How is an ai ever going to do anything like that?
Absolutely right! Superbly insightful comment.
Hi music professor! Been listening to your lectures for a while now and absolutely loved them. This one really made me want to comment however. How you described the AI as being a bunch on non sequiturs strung together was perfect, And I felt same. It’s fascinating that it got the general idea right but was just missing that human element. I wonder if it would become more or less accurate as the styles themselves became more formulaic and less based on communication. I would be interested to see how you felt if you told the AI to compose atonal/serialist piece in the style of Schoenberg or or a minimalist piece in the style of Steve Reich
Thanks for your comment. It would be intriguing to find out what results you'd get with atonal Schoenberg. I doubt you'd get an Op. 19 piano piece! You' probably get a bunch of atonal non-sequiturs! You'd think that AI could be programmed quite successfully to write serially because it's more about finding ingenious ways to manipulate permutations, inversions, retrogrades etc. but I think the best serial composers (Berg, Webern, Boulez etc.) are far more creative at this than AI would be. Again, with Reich, I think AI could impersonate the style in a bland commodified way, but the thing that makes Reich extraordinary - e.g. the way the harmony alters at just the right moment - is beyond the reach of AI.
First, just as clarification. I am musician, I've been playing guitar for 35 years . I am also graduated in Conduction, Composition and Arrangements. ( I don't know if this is the correct term in English)
So...
The biggest problem about musicians (and artists in general), is EGO. They (well, we I suppose) are so full of themselves. Every single "problem" mentioned can be corrected by making a better prompt.
Also, in order to achieve a great level of composition with AI, your need to re-generated, try, A LOT of times. And honestly, this isn't too different from trying to create a melody. We come with an idea...we try, change, throw away, revist, try again, until we are satisfied.
The other huge bulls*hit in the whole "soul" and "connection" . This is personal, and more often than not, what the composer felt and connected when writting a song is NOT what connects to someone listening to it. The meaning is, and it always will be, personal and unique.
Music is mathematics. The "feeling" is purely each ones , particular, interpretation to it (as player or as audience).
AI is a fantastic tool, wise musicians will use it to enhance their works, bring inspiration and achieve things that otherwise wouldn't be possible (cost wise, for example).
Instead of throwing rocks, let's understand what the new era is bringing to us. We adapt. (Sequencers, remember? some people still hate them)
Insightful video. I think that even when A.I. solves these stylistic issues enough to closely approximate the sound of human composition, it will still be deficient in the most important area of them all -- inspiration. Human art is guided by our experiences as mortal, emotional, and organic beings. The decision to write something tragic or uplifting springs from our myriad emotions, our engagement with the struggles and achievements of life, our physical pains and exultation, and perhaps most of all, from human relationships. Composers are certainly affected by the phenomenon of predictive writing and one can feel the harmonic pull at the end of phrases but we also make decisions to break these rules according to how we feel and the creative vision of the work. In short, A.I. can imitate the patterns of human art but what it constructs means nothing and comes from nowhere. A cereal commercial jingle written by a human has more meaning than a two-hour symphony constructed by A.I. As for Mozart, Gould should have stuck to playing and left composition to composers. Wolfgang's compositions are irreproachable.
Correct. Inspiration is essential. AI can never be inspired.
“Ai can’t compose!”
“Ok so we set the bar very low, maybe at Mozart level, and see what happens ok?”
Interesting and thought-provoking. A few weeks ago , I was listening to Ray Chen's clips of ai generated music and those by real composers - I could instantly tell them apart. The same here. The ai clip is bland and doesn't get anywhere. You aptly said it doesn't tell a story.
Earlier I was playing k296 3rd movement. At the end, the uncanny way Mozart arrives at the opening theme is sublime - ai doesn't come close.
A 5 year old can compose, but I haven't seen anyone like the young Mozart. It seems Mr Gould doesn't like Mozart - why?
If Mozart had lived a normal lifespan, who knows what he might have achieved.
Mozart is now one of my favourite composers, because i came to realize, that his music is all about what you dont hear. Its like Non-Music; maybe this was Mr Goulds point. But how could an AI ever construct such a paradoxon? Beethoven in contrast jumps right into that sphere, already digging deeper, and let you expierence the impossible. Formerly Bach points towards that door. Probably AI would have less problems "composing" a Beethoven piece or something later. With Baroque im not sure, but Bach seems too difficult for AI. And i think worst thing to do with AI would be Renaissance Music, Josquin, Isaak, Palestrina.
Interesting comment. I sometimes wonder if Renaissance polyphony is actually the way in for constructing a composing AI programme, because the rules are very strict and it might be possible to painstakingly train an AI programme to gradually acquire the skill.
I'd be interested to hear you rewrite this passage into something more coherent. I feel like it's a starting point of sorts, and if it were made more elegant, more human, it'd actually be a halfway decent piece.
Interesting idea.
AI: "I'll have you know, I'm playing all the RIGHT NOTES, just not necessarily in the right order."
One of my Professors at University once said: "Even a single tone (f.e. by a singer, string or wind instrument) can be like life, with birth, growing up to full power, decline and death." How could AI and sampled instruments EVER be able to create music coming from the soul? Unless it once will have a soul.
Mozart himself (!) was bringing up the issue of "composing" with just a set of his premade musical phrases consisting of measures for beginning, in between and endings plus a set of two dice. You throw the dice and Voila! there's a "Schleifer oder Walzer". It's called "Musikalisches Würfelspiel" (musical dice game). Examples can be found online.
Yes - I'm planning to look at the Dice game some time soon. It's like Mozart knew all about AI back in the 1780s!
Do they not teach it anything about voice leading or melodic shaping? The repeated notes in the melody are clumsy - as you say, it seems to get trapped, and puts together a collection of formulae. It hasn't considered the overall direction of the music.
That's not how AI works, you can't "teach" or "mentor" it to do something. It has to create the connections itself. AI is also quite limited because it only thinks about what happens next, note by note, and not actually any ideas about the entire piece. It's like musical autocomplete.
@@michaelchen2821This is a very important point, essential really. Most people apparently think that AI is "programmed", which is fundamentally incorrect.
Including the video author at 8:39
I did my first Muzio Clementi sampling yesterday as a 3rd-year beginner and was struck by how avant garde he was as compared to Mozart. Lots of weird chromaticism?
He's good. Lots of fine piano sonatas (Beethoven claimed he preferred them to Mozart's).
Beethoven Sonata Op. 31 #3 starts on the supertonic, but it's a 7th chord
AI is great at composing music that superficially sounds like a specific genre and doesn't hold your attention for more than 5-10 seconds. Unfortunately, that's how most people listen to music: "Hey, this sounds familiar! Anyway..."
It reminds me of the sort of work produced by keen, but essentially untutored kids who have been experimenting with Sibelius or Musescore. (Actually, I should include myself as a 9 year-old, composer, though I predated computer technology).
I’m a little tired of people commenting on the music of Mozart and the Galant period without speaking the language, which of course is partimento and schema. Gould was a phenomenal talent as a pianist, but there are many born like him, not so many like Mozart. Everyone knows Mozart; not no many Gould.
You must understand that Leopold Mozart (Amadeus’ father and illustrious court composer) would’ve overseen and guided his composition process, too, smoothing his compositional crumples. His 5-year-old-written works were very much moulded by his dad’s experiences, and Leopold was no stranger to chicanery if it contributed to his son’s exaggerated prodigious veneer. I recall hearing he lied to crowds on a number of occasions about Mozart’s age - unnecessarily, yes - to paint this picture of a true genius.
It was Beethoven's dad who lied about his age. Leopold did certainly hot-house his son but the results were astonishing (there's a very detailed account by Charles Burney who gave the little boy all sorts of ear exercises to test out his genius, at the end of which he concluded that all accounts of Mozart's talents were actually less impressive than the reality).
@@themusicprofessor Ah, thanks for the clarification. Great video, by the way! I’m astounded I haven’t come across this channel sooner!
The fundamental issue with "AI", i.e. "artificial intelligence" is that such kind of software is indeed "artificial" but very far from being "intelligent". The software behind analyses and recognizes patterns. Based on certain algorithms and statistics, it predicts what would be the most logical next word or note. So it would not be able to have an overall plan, wouldn't know what is a musical phrase, etc. Even the most sophisticated AI program would not, to my understanding, be able to "break" the rules of harmony in a surprising and beautiful way as have been able all the great composers. Feed an AI with medieval music and all what the computer will be able to produce will sound about like that but it most probably would never be possible to reach the complexity of romantic or later music. AI may generate some music in the style of X but it may not be able to create truly interesting pieces of music with a new musical language. I certainly don't doubt that AI generated music may find its way into "fast food music" but will fail when things start getting a bit more demanding.
IMHO, AI can basically do anything....the equivalent of a Turing test for music is not a problem.
It really can't.
AI can do anything basic. It can't basically do anything (original). Original being the operative word. It may be programmed to invent a better mousetrap but it could NEVER come up with the famous statement that if a man could invent a better mousetrap the world would beat a path to his door. Or write a play called The Mousetrap - a thing to catch the conscience of the king - which is performed within a certain well-known play, which AI would first have to create.
And how could AI create Metamorphosen out of bitterly recalled emotions still wonderfully alive among the ruins of a culture that visited destruction on itself? How could it first look into Chapman's Homer (or know what it is) and then come up with the most transcendent sonnet of its language? Seriously my friend, you have not thought this through.
Exactly. Good that you mention Metamorphosen - a piece that distills a lifetime of experience into an astonishing artwork. What is art? It is creative virtuosity responding to the cultural environment in such a way that the creative work reflects a deep sense of what it means to be a human being. This is infinitely beyond the scope of any intelligence except our own.
No, AI cannot do everything that a human does. Such as: understanding what feelings and emotions are, how to generate them. Also: human intelligence has an important intuitive component, which AI will never have. Experiments with that all failed. Also: human perception, interaction and communication with the environment is extremely sophisticated and complex. Until now AI cannot handle the complexity of certain traffic situations or perfectly translate any language. It will never be able to do that. Don't be fooled by the current AI hype. There have been similar waves in the past. Every now and then the IT world generates some bubble for business reasons.
Thank you. You've pulled me in from the ledge in my fear of AI. Perhaps something to keep in mind, however, is that AI has been programmed by engineers - and current-day ones at that. Aesthetic? perhaps not. AI remains a parlor trick - impressive in its expression yet soulless - at least for now.
I guess if you gave the AI more specific instructions, it would create something closer to what one specified.
But you could also wonder if what AI created is reflecting AI's situation being a boxed entity that only knows what it's been told. It doesn't know what's right or wrong, it only knows what the humans have told it what's right or wrong.
The great critic Hans Keller once memorably said, 'The composer creates certain expectations, well-defined expectations, which he proceeds to meaningfully contradict.'
Gould's comments on Mozart's late keyboard works, describing them as stale and lacking in a sufficient quota of meaning, seem to resonate with Keller's generalised observation. There is no meaningful contradiction of well-defined expectations. The wonder is that Gould's observations could not in any wise be valid with respect to the operas, symphonies, concertos, choral music and chamber music from this same period in Mozart's output.
Now my point is that Gould, with his incredible sharpness of intellect, could well have been predicting the rise of AI compositional programs in our times, ie, cleverly designed tools that when set in motion ape certain algorithms. To be satisfied with the result is to award the prize to Beckmesser rather than Walter; to prefer a provincial Kapellmeister's well intentioned gaucheries to the St Matthew Passion.
I do a lot of experimenting with AI music. I've heard some good pop, rock, rap, etc., but I've never heard a good AI classical piece or even a good jazz piece. The AI just lacks that ability to go outside of the box and really "pop off" the way a great musician can.
Having heard Gould's horrible compositions and many recordings where he "plays" Bach's music like it's a MIDI file, hearing him say "Mozart music could be made by a 5 year old" is one of the funniest thing to me 😂
Big Tech: "Well, you can go off and do something else then". Like eat and breath perhaps? Brahms: J.S. Bach is my daily bread.
13:37 It clearly shows Loki didn't like the computer version
The AI music almost has a resemblance of something by Debussy in the Children’s Corner at first with starting on ii, then going on to make perfect cadences with awkward repeated notes in the melody. Very strange.
The beauty of tonal music, such as the classical style, is telling a story with a coherent structure and phrases that elicit a specific emotion. AI and the computer will never understand that level of communication and human emotions, it can only calculate and mechanically manipulate patterns. Therefore AI cannot and will never be artistically creative. It will never generate something that makes people cry or feel peaceful.
If AI doesn't know basic harmony or how to create an actual phrase, I'm sure it would fail even harder at creating something atonal that relies on more abstract things like "gestures" and "color".
Fascinating
Love it
Photography was going to be the death of painting
Definitely fucked up portraiture
Took it out of the realm of the rich
Then became a separate thing in its own right
AI I think will follow similar parallel tracks to be manipulated by pesky humans -
eventually- to create a different art
Totally agree.
AI can't just create music; here's some text "in the style of an Elizabethan playwright": What brave new world that has unleashed the dogs of war to take any man’s horses for the world must be peopled this Bartelmy Fair. Mockery aside, in a few years AI may go beyond regurgitating snippets of cliche and begin to utilize underlying concepts in order to create more adequate outputs. We're not there yet, and the way AI works today won't get us there (because of the training set problem) but once people realize that the GPT approach doesnt work without adding set theory, we may see more interesting results.
Maybe. But I think people seriously overestimate the power of AI. Humans are an incredibly brilliant species (for all our ridiculous faults) and computers are just brilliant human inventions but nevertheless unthinking machines. They cannot understand and they cannot invent.
"UdioAI - please compose a piece of classical music in the style of SunoAI composing a piece in the style of Mozart."
If Beethoven (or any composer one could argue is great) had been able to access AI, would he have used it? Well, 1. He would have soon found it too limiting. 2. He would have designed a better program. 3. He would have abandoned the whole enterprise and set about to just compose. My point is that if an artefact can only be as good as the program that created it, why not cut to the chase and compose the artefact? It is bordering on the absurd to imagine an AI coder feverishly creating a program that produces the very result he had in mind all along.
I have in mind a kind of Mel Brooks mad scientist proclaiming a new invention that saves countless hours of futile toil as it pours forth whole epoch-making, game changing masterpieces each more wondrous than the last.
With reference to some of the other commenters, to criticise AI is not to heckle; but the votaries had better point to some success and soon.
" I mustn't be mean..."
" Loki barks"...
This is the type of educationally wholesome content about music I live for..... Godspeed to you sir, and hats off for the amazing content you produce :)
Thank you so much!
I think the language word which the AI is missing is syntax or meaning.
I think the AI music did some things better than I could do but ultimately would have less artistic value, even though I'm a very bad composer. It would work great to provide classical ambiance that one was not intended to listen to at all but just hear passively. I don't think it would be worth paying enough attention to that one would actually remember any of it, because nothing is really memorable since it does lack meaning as you've said.
The only chord ii beginning i can think of is Beethoven Op 31, 3. But that's (in)famous precisely because it's so unusual.
Yes. It's wonderfully ambiguous.
@@themusicprofessorindeed, and on reflection, the way Beethoven exploits that ambiguity is also impossible to imagine AI doing at present. That whole sonata movement is 'about' the opening, in a way which no AI generated music can be said to be 'about' anything. I think that's why none of it has any grip, so to speak, no narrative.
As someone who loves classical music but has absolutely zip musical talent, I can say that what I've heard and read about mediocre composers and even the juvenilia of great composers is, "Why do you limit yourself?" This is extremely evident n the AI music that is formulaic and lacks creativity. It could be a long time (if ever) that AI can actually compose something as great as Mozart's works.
16:00 : In my opinion, it’s really not a matter of time but more a matter of training. AI could absolutely imitate any kind of narrative as long as you feed it with enough material that contains this narrative. Of course no one has done that with classical music because it ins’t a field that present enough potential for people to invest the millions of dollars it takes to train a solid AI.
The problem of these kind of turing test of AI generated classical music is that they dont feature programs trained on classical material (since it doesn’t exist) but rather pop (hence the title Whispers of Elegance which is clearly wrong for a classical composition, the beginning on subtonic or the weird rythms). However, if you ask the same program to generate a pop song it will give you a way more convincing result because it was mostly trained to recreate pop codes.
With this in mind, I think that Gould was right : AI could replicate Mozart style (maybe not up to the level of sophistication of most of his pieces, but enough to fool most of us), but so could it do for any composer as long as it’s shown enough of the style in which the composer wrote.
Anyway thank you for this video and more broadly for your content on RUclips which is always very qualitative.
Have a good day
Of course, an AI can replicate existing patterns of classical composers. But it will not do what the greatest of the composers did, it will not develop new patterns that work and evolve, revolutionize music.
Additionally, it will not, by all means, produce music that has a true meaning for the listener (like a lot of Chopin's wildly emotional, deep music, or the programme music of most Romantic composers.) Music having a meaning, making you feel one or many emotions, hear musical stories. No AI in the world will produce an - equally unique - piece as emotionally powerful as Ballade No.1 of Chopin because a computer, for all its clever processing power, can't comprehend what was going on in the composer's mind when he put those specific harmonies together to inject specific feelings into the listener.
And comparing Pop song patterns with classical music.... Please. Literally any mediocre amateur musician (like me) can write a convincing Pop song. 90% of modern Pop contains around 1% of the complexity of a Mozartian or Beethovian symphony or even an art song from Schubert.
4:20 uncanny - AI always makes believe that it makes sense but it ends up making none. It's much easier with English language though.
Замечательно сказано, что суть великого искусства заключается в человечности!
8:20 Clementi op 36 no 5 part 1 has grace note triplet rhythm throughout
Ah - but his rhythm has the grace note preceding a triplet, whereas this is a triplet followed by a grace note which just sounds weird.
This video won't age well! haha I'd say AI creating Bach-sounding compositions would be almost child's play since there's kind of a very small set of musical rules and theory to learn. Train the AI by feeding it enough Bach pieces and it probably will start to generalize and understand all of the ways it could make notes that make sense in a timeline. oh, I just did a little search. Bach Bot. Already there is one that can create these kinds of streams of notes and harmonize them.
Obviously I disagree: this video is simply me critiquing AI composition in 2024. It's an accurate representation of the situation now. But I also think that AI has a very long way to go to produce anything of genuine compositional interest.