I would like to offer a few comforts to anyone inclined to be terrified by programs like EMI.
First is the simple and obvious point that EMI could not have composed like Chopin if Chopin had not existed. Chopin, on the other hand, wasn't copying anyone.
Second, Hofstadter's experiment contains a crucial flaw. He had to use a very obscure Chopin piece, lest the professional musicologists in the audience recognize it. But surely the piece deserved its obscurity! So the comparison was between EMI's mashup of Chopin's best work — probably not even including lesser pieces in its database — and a throwaway that Chopin probably tossed off one day when he had a hangover. Not fair at all.
And thirdly, reading your description of how EMI worked, it's clear that this is not AI composing music from scratch on its own; this is David Cope writing a program incorporating a considerable amount of his own knowledge about composition, along with stylistic elements from great composers, to generate music. In short, the agency and the genius, if it was genius, lay within Cope, not the machine. If computers didn't exist, Cope could in principle have done the same thing with pencil and paper and dice; would that have meant the dice were composing music?
Perhaps these observations will help people ... uh ... cope :-)
Nice post! To elaborate briefly on your third point, Cope's approach was vastly different from those used in LLMs and attempts at so`-called AGI in numerous respects. It was single purposed, had meticulously curated data sets and was undoubtedly highly tuned. Hardly intelligent in any sense.
I imagine generative AI, suitably trained on sheet music, can make something that sounds much more like bach, most of the time (with a few bars of pure absurdity now and then). But it sounds like copes code is a lot more useful if we want to actually understand Bach. I miss the days when it seemed like these goals of reproducing and understanding seemed inextricably linked. Spectacular thougn they are, interesting though they are, sometimes LLMs feel like part of the plot of "idiocracy".
I might be missing the mark here, but I feel that this highlights exactly the difference between AI research of the latter 20th/early 21st century and modern generative AI. Although I'm not familiar with Cope's work specifically, it sounds as though he created domain-specific rules that a computer could use to "think" narrowly and deeply (about Baroque composition in this case), whereas generative AI takes the whole squishy bulk of its training data, usually with some stochastic finishing touches, and then tries to do something "narrow" out of this extreme generality. I have no opinion on which is better, but generative AI has definitely attracted more hype recently.
Really interesting. But is asking professional musicians to evaluate two pieces, as if doing a beer tasting, really a test of creativity? This kind of test is notoriously tough even with human compositions. You can go into a souvenir shop and find paintings that are technically better executed than the work of the masters. But those paintings *depend* on the masters. Likewise, the EMI mazurka couldn’t have been created before Chopin’s. So we need to evaluate creativity in its proper historical context. No doubt, a machine will one day create original work that opens up a whole new artistic direction and makes us see the world in a different way.
I'm a big fan of classical music, but it's sad to me how few people attempt to compose music that sounds like Bach nowadays. It's like nobody is even trying. You can't really make money doing it, so modern classical composers are mostly making movie scores.
It would be great if AI could finally achieve this. More music like Bach, or ideally even better than Bach. You can question whether AI-written music is "really creative", but you could a/b test and if people who enjoy Bach prefer the LLM piece, then that is "Turing equivalent" to a success, and the philosophers can continue to debate forever.
LLMs seem like they turn this impossible dream into something that's actually achievable. But it still seems pretty expensive. And is anyone really working on it? Hopefully someone does eventually.
At the 1998 AAAI conference I remember walking through a garage, or elevator or such and he was walking right in front of me. Then pretty soon I heard his presentation, which was fascinating. I think most of his work was symbolic AI, which seems a bit more innocent than hoover-it-up deep learning, because we can see the code. Thanks.
An enlightening and moving reflection. I've been thinking lately of this in the context of 'The Death of the Author.' There seems to be so much written and expressed about art and music, in fact all of human output, centered around the glory of humanity. It feels like a significant bit of "scaffoldry" to appreciate art for the artist rather than the art. I know it doesn't have to be binary (pun intended) but I still wonder is the goal of art empathy with the artist?
Western music really is a cultural meme and our understanding of the emotions stems from a longstanding cultural relevance that carries forward, at least that is my contention. I am not convinced that these sounds ex culture would have the same effect, and the stickiness of the themes lend toward their accretion.
I often wonder about this fear that people express. It feels a bit like the death of the ego, that which is dying is the thing that is afraid. It is not the end of humanity, or the creativity of humanity. But it does raise the specter of the post truth world, where truth is a bygone perspective infused with a cultural dna that claims primacy.
And maybe what is lost is the long accrued interest on the investment and time that went to generate that deep culture. The care and feeding of it that became a part of our identity. To lose that feels a bit like death.
But maybe we've been fooling ourselves this whole time and in fact it isn't turtles all the way down.
If anyone feels the generated "Bach" fugue sounds like the real thing at times, it may be because it's quoting an actual Bach work (the D major fugue from the Well Tempered Clavier, book 2) almost verbatim, several times, starting at around 1:50 in the clip. Compare to the the original, https://www.youtube.com/watch?v=injtWmIcYfk, from around 1:00.
As a classically trained musician with some experience in composing baroque-style fugues, the generated music to my ears sounds like a technically competent but musically tasteless improvisation. It's worth knowing in this context that improvisation, not written composition, was the bedrock of western musical training until the mid 19th century - that is, up to and including the time of Chopin. Improvising a fugue or dance movement wasn't an astonishing feat; anyone with decent keyboard training could do it, but the result would be rather formulaic and directionless, except in the hands of the great masters.
Thus the obscure works of the great composers might well be obscure because they were the sort of thing they could just sit down and play, or jot down on paper as quickly as the ink would dry, on a bad day. That an AI system can do that isn't so distressing to me. With enough investment, I'd expect AI composers to be able to produce works of similar quality as, say, G.P. Telemann, a prolific contemporary of J.S. Bach. But there's a reason why Telemann's work today gets only a fraction of the attention of Bach's, despite the enormous quantity of preserved, technically very well written music of the former. And I can't help but ascribe the difference to the divine inspiration of the latter.
While taking a different approach than today's AI, it's an ominous precursor to the general idea of non-human entities emulating human creativity. I've been thinking a lot about the concept of "appreciation", and wondering if over time AI will be able to emulate this human characteristic. (BTW, just finished listening to the Complexity Podcast: Nature of Intelligence, fabulous, highly recommended!)
AI can still only produce works and not evaluate them, so they are not creative on their own, but only in the hands of human beings. For example, you could put a computer to work to write novels. Hundreds, thousands, millions of them in a year perhaps. But if there's no-one there to evaluate the works, they're essentially meaningless. So all work, even AI work, only gain meaning in a human context. The same holds true for new generated music in the style of Bach, Chopin, The Beatles, what have we. It's not the AI that is creative, but creativity is created in the human context of evaluation and reflection.
I love this piece and the reflections on the purpose of art. I am reminded of a piece of research (perhaps sociological) that I cannot find on how the invention of the photograph was perceived by artists of the time. Since the renaissance you could say that artists were striving for ever more realism and invented a great deal of techniques for capturing a scene as close to reality as possible. The invention of the camera, one could say, made painting as an art form obsolete. Certainly this is true if you view art’s purpose as the production and consumption of artistic objects. There was a demand for a type of object (realistic images) that could now be satisfied by a new means of production (photograph). Of course, we know that human creativity is an expression of something deeper, a profound yearning for connection between the artist and audience. As a result, two things happened. First, photography became a medium of expression in its own right. Second, artists began painting in ways that were not realistic finding a whole new way of connecting with their audience. In fact much of what we love about modern art is that it’s freed from the constraints of photo-realism.
Thanks for posting this. EMI was a huge inspiration for me when I wanted to go in to computer music generation. I was introduced to it as an undergrad by the same professor (Paxton) who had taught a seminar on Hofstadter's GEB. I immediately bought all of Cope's books and read them with great excitement. 15 or so years ago, I was visiting a friend in San Francisco, who agreed to give me a ride to Santa Cruz so that I could visit with Dave Cope. I spent a pleasant afternoon in his office discussing topics ranging from SPEAC analysis in EMI, his ideas about novel board game rule systems, to the birds nesting outside of his office's window.
After that he shared a draft manuscript for one of his later books for me to provide comments. I printed it out to read when I had time. I was dragged to a boring party/event in Bloomington, IN one time and I luckily had brought a big portion of the manuscript with me, so I snuck off to a chair in a corner somewhere and read most of it and wrote up many of my comment there. A fond memory!
Cope's work was groundbreaking, and I think a great example of a hybrid AI system: not hybrid in the sense of symbolic + deep learning, but instead hybrid in terms of domain-specific rules for how to extract useful information from a large corpus and reassemble it into novel musical material. His work on the SPEAC system, a hierarchical analysis that seems to me like a more grounded and practical sort of Schenkerian analysis, is interesting on its own, and then his work on composer's characteristic "signatures" showing up in cadences was also quite useful in characterizing musical style. I still find it inspiring when I think of new ways to make computer music systems to compose music, and I expect it to continue to inspire future research.
That sounds like a pretty simple little Bach-type fugue to me. Very impressive, especially for its day, but lacking the complexity of much of Bach. Of course, writing like Bach is a very high bar, and one that I suspect that an AI system will one day (soon?) reach.
To tell you the truth, I'll be more impressed when we hear excellent AI music not done 'in the style' of Bach, Mozart, or Chopin, but in the style of the AI system itself, and something we've never heard before but can recognize its creativity and greatness.
RIP David Cope, UCSC star professors back in the mid-1970s then was all about Gregory Bateson, John Grinder, Norman O. Brown. David quietly was building a national following .
I would like to offer a few comforts to anyone inclined to be terrified by programs like EMI.
First is the simple and obvious point that EMI could not have composed like Chopin if Chopin had not existed. Chopin, on the other hand, wasn't copying anyone.
Second, Hofstadter's experiment contains a crucial flaw. He had to use a very obscure Chopin piece, lest the professional musicologists in the audience recognize it. But surely the piece deserved its obscurity! So the comparison was between EMI's mashup of Chopin's best work — probably not even including lesser pieces in its database — and a throwaway that Chopin probably tossed off one day when he had a hangover. Not fair at all.
And thirdly, reading your description of how EMI worked, it's clear that this is not AI composing music from scratch on its own; this is David Cope writing a program incorporating a considerable amount of his own knowledge about composition, along with stylistic elements from great composers, to generate music. In short, the agency and the genius, if it was genius, lay within Cope, not the machine. If computers didn't exist, Cope could in principle have done the same thing with pencil and paper and dice; would that have meant the dice were composing music?
Perhaps these observations will help people ... uh ... cope :-)
Nice post! To elaborate briefly on your third point, Cope's approach was vastly different from those used in LLMs and attempts at so`-called AGI in numerous respects. It was single purposed, had meticulously curated data sets and was undoubtedly highly tuned. Hardly intelligent in any sense.
I imagine generative AI, suitably trained on sheet music, can make something that sounds much more like bach, most of the time (with a few bars of pure absurdity now and then). But it sounds like copes code is a lot more useful if we want to actually understand Bach. I miss the days when it seemed like these goals of reproducing and understanding seemed inextricably linked. Spectacular thougn they are, interesting though they are, sometimes LLMs feel like part of the plot of "idiocracy".
I might be missing the mark here, but I feel that this highlights exactly the difference between AI research of the latter 20th/early 21st century and modern generative AI. Although I'm not familiar with Cope's work specifically, it sounds as though he created domain-specific rules that a computer could use to "think" narrowly and deeply (about Baroque composition in this case), whereas generative AI takes the whole squishy bulk of its training data, usually with some stochastic finishing touches, and then tries to do something "narrow" out of this extreme generality. I have no opinion on which is better, but generative AI has definitely attracted more hype recently.
Really interesting. But is asking professional musicians to evaluate two pieces, as if doing a beer tasting, really a test of creativity? This kind of test is notoriously tough even with human compositions. You can go into a souvenir shop and find paintings that are technically better executed than the work of the masters. But those paintings *depend* on the masters. Likewise, the EMI mazurka couldn’t have been created before Chopin’s. So we need to evaluate creativity in its proper historical context. No doubt, a machine will one day create original work that opens up a whole new artistic direction and makes us see the world in a different way.
I'm a big fan of classical music, but it's sad to me how few people attempt to compose music that sounds like Bach nowadays. It's like nobody is even trying. You can't really make money doing it, so modern classical composers are mostly making movie scores.
It would be great if AI could finally achieve this. More music like Bach, or ideally even better than Bach. You can question whether AI-written music is "really creative", but you could a/b test and if people who enjoy Bach prefer the LLM piece, then that is "Turing equivalent" to a success, and the philosophers can continue to debate forever.
LLMs seem like they turn this impossible dream into something that's actually achievable. But it still seems pretty expensive. And is anyone really working on it? Hopefully someone does eventually.
At the 1998 AAAI conference I remember walking through a garage, or elevator or such and he was walking right in front of me. Then pretty soon I heard his presentation, which was fascinating. I think most of his work was symbolic AI, which seems a bit more innocent than hoover-it-up deep learning, because we can see the code. Thanks.
An enlightening and moving reflection. I've been thinking lately of this in the context of 'The Death of the Author.' There seems to be so much written and expressed about art and music, in fact all of human output, centered around the glory of humanity. It feels like a significant bit of "scaffoldry" to appreciate art for the artist rather than the art. I know it doesn't have to be binary (pun intended) but I still wonder is the goal of art empathy with the artist?
Western music really is a cultural meme and our understanding of the emotions stems from a longstanding cultural relevance that carries forward, at least that is my contention. I am not convinced that these sounds ex culture would have the same effect, and the stickiness of the themes lend toward their accretion.
I often wonder about this fear that people express. It feels a bit like the death of the ego, that which is dying is the thing that is afraid. It is not the end of humanity, or the creativity of humanity. But it does raise the specter of the post truth world, where truth is a bygone perspective infused with a cultural dna that claims primacy.
And maybe what is lost is the long accrued interest on the investment and time that went to generate that deep culture. The care and feeding of it that became a part of our identity. To lose that feels a bit like death.
But maybe we've been fooling ourselves this whole time and in fact it isn't turtles all the way down.
If anyone feels the generated "Bach" fugue sounds like the real thing at times, it may be because it's quoting an actual Bach work (the D major fugue from the Well Tempered Clavier, book 2) almost verbatim, several times, starting at around 1:50 in the clip. Compare to the the original, https://www.youtube.com/watch?v=injtWmIcYfk, from around 1:00.
As a classically trained musician with some experience in composing baroque-style fugues, the generated music to my ears sounds like a technically competent but musically tasteless improvisation. It's worth knowing in this context that improvisation, not written composition, was the bedrock of western musical training until the mid 19th century - that is, up to and including the time of Chopin. Improvising a fugue or dance movement wasn't an astonishing feat; anyone with decent keyboard training could do it, but the result would be rather formulaic and directionless, except in the hands of the great masters.
Thus the obscure works of the great composers might well be obscure because they were the sort of thing they could just sit down and play, or jot down on paper as quickly as the ink would dry, on a bad day. That an AI system can do that isn't so distressing to me. With enough investment, I'd expect AI composers to be able to produce works of similar quality as, say, G.P. Telemann, a prolific contemporary of J.S. Bach. But there's a reason why Telemann's work today gets only a fraction of the attention of Bach's, despite the enormous quantity of preserved, technically very well written music of the former. And I can't help but ascribe the difference to the divine inspiration of the latter.
While taking a different approach than today's AI, it's an ominous precursor to the general idea of non-human entities emulating human creativity. I've been thinking a lot about the concept of "appreciation", and wondering if over time AI will be able to emulate this human characteristic. (BTW, just finished listening to the Complexity Podcast: Nature of Intelligence, fabulous, highly recommended!)
AI can still only produce works and not evaluate them, so they are not creative on their own, but only in the hands of human beings. For example, you could put a computer to work to write novels. Hundreds, thousands, millions of them in a year perhaps. But if there's no-one there to evaluate the works, they're essentially meaningless. So all work, even AI work, only gain meaning in a human context. The same holds true for new generated music in the style of Bach, Chopin, The Beatles, what have we. It's not the AI that is creative, but creativity is created in the human context of evaluation and reflection.
I love this piece and the reflections on the purpose of art. I am reminded of a piece of research (perhaps sociological) that I cannot find on how the invention of the photograph was perceived by artists of the time. Since the renaissance you could say that artists were striving for ever more realism and invented a great deal of techniques for capturing a scene as close to reality as possible. The invention of the camera, one could say, made painting as an art form obsolete. Certainly this is true if you view art’s purpose as the production and consumption of artistic objects. There was a demand for a type of object (realistic images) that could now be satisfied by a new means of production (photograph). Of course, we know that human creativity is an expression of something deeper, a profound yearning for connection between the artist and audience. As a result, two things happened. First, photography became a medium of expression in its own right. Second, artists began painting in ways that were not realistic finding a whole new way of connecting with their audience. In fact much of what we love about modern art is that it’s freed from the constraints of photo-realism.
Thanks for posting this. EMI was a huge inspiration for me when I wanted to go in to computer music generation. I was introduced to it as an undergrad by the same professor (Paxton) who had taught a seminar on Hofstadter's GEB. I immediately bought all of Cope's books and read them with great excitement. 15 or so years ago, I was visiting a friend in San Francisco, who agreed to give me a ride to Santa Cruz so that I could visit with Dave Cope. I spent a pleasant afternoon in his office discussing topics ranging from SPEAC analysis in EMI, his ideas about novel board game rule systems, to the birds nesting outside of his office's window.
After that he shared a draft manuscript for one of his later books for me to provide comments. I printed it out to read when I had time. I was dragged to a boring party/event in Bloomington, IN one time and I luckily had brought a big portion of the manuscript with me, so I snuck off to a chair in a corner somewhere and read most of it and wrote up many of my comment there. A fond memory!
Cope's work was groundbreaking, and I think a great example of a hybrid AI system: not hybrid in the sense of symbolic + deep learning, but instead hybrid in terms of domain-specific rules for how to extract useful information from a large corpus and reassemble it into novel musical material. His work on the SPEAC system, a hierarchical analysis that seems to me like a more grounded and practical sort of Schenkerian analysis, is interesting on its own, and then his work on composer's characteristic "signatures" showing up in cadences was also quite useful in characterizing musical style. I still find it inspiring when I think of new ways to make computer music systems to compose music, and I expect it to continue to inspire future research.
He'll be missed!
Thanks, Eric. I never met him but still felt he played a big role in my thinking on AI and intelligence / creativity in general.
Thanks for sharing this.
That sounds like a pretty simple little Bach-type fugue to me. Very impressive, especially for its day, but lacking the complexity of much of Bach. Of course, writing like Bach is a very high bar, and one that I suspect that an AI system will one day (soon?) reach.
To tell you the truth, I'll be more impressed when we hear excellent AI music not done 'in the style' of Bach, Mozart, or Chopin, but in the style of the AI system itself, and something we've never heard before but can recognize its creativity and greatness.
This matters. Great work
RIP David Cope, UCSC star professors back in the mid-1970s then was all about Gregory Bateson, John Grinder, Norman O. Brown. David quietly was building a national following .
Hi Melanie, thanks for the nice obit. Wow, a true pioneer.