The author's central claim is that AI is just an "averaging machine" that can only reproduce patterns. From that he seems to conclude that "AI will never create truly great music". As a musician myself, I don't think that this conclusion works; neither do I consider the other arguments of the author (as far as I understand them) convincing. The author describes "experiments" with his "listening party": by presenting two excellent recordings of the same piece, the author forces guests to confront a crucial reality: even between two technically brilliant performances of identical music, there are subtle, meaningful differences that can be heard and felt. I agree with that; it even corresponds to my own experience: I didn't recognize the beauty of Bach's Matthäus and Johannes Passions until I heard a recording by John Elliot Gardiner and his team for the first time (ironically, in the credits of a hollywood film); that moment - thirty years ago - was was truly an epiphany; before that, I had always listened to recordings by supposedly great German artists, with those bombastic vibratos that you have to endure in Wagner, which completely overshadowed the modest and perfect beauty of Bach's work.
But that doesn't say anything about computer-generated music. I've been following its development for even more than thirty years, and we did indeed have to wait until 2025 to hear completely convincing results for the first time. The models currently available are optimized for pop, jazz, and other music styles, but not yet for classical music. However, there are already studies that have made it very clear that almost 100% of people can no longer distinguish between human-made and machine-made music (I'll provide a reference on request). I last tried this in May, letting Suno interpret my own pieces, and was impressed in every respect (here are the examples: https://rochus-keller.ch/?p=1350). The arrangement and the musicality evident in the interpretation clearly have a human quality. It is simply not true, that the choice between classical recordings proves that greatness cannot be achieved through averaging or pattern-matching. Much great music, including classical masterpieces and jazz standards, works precisely because it follows established patterns and conventions while introducing variations within those structures. E.g. in Jazz, an improvisation itself is pattern-based, using scales, chord progressions, rhythmic conventions, with creative variations. If AI can learn these patterns and introduce variations (which Suno demonstrably does), why can't that be "great"?
To my knowledge, we don't have LLMs yet trained especially to generate classical music (at least I haven't heard a convincing one so far). In this respect, I have to agree with the author that, at least in the classical field, we don't have the same quality as in other musical styles (yet). But that is only a matter of time, and it is not for the reasons the author supposes. As demonstrated with my - and actually also the authors - example, even among human artists, there can be huge quality differences. The author's "immortal souls" retreat is essentially an abandonment of empirical arguments for unfalsifiable metaphysics, which essentially concedes the technical debate.
But that doesn't say anything about computer-generated music. I've been following its development for even more than thirty years, and we did indeed have to wait until 2025 to hear completely convincing results for the first time. The models currently available are optimized for pop, jazz, and other music styles, but not yet for classical music. However, there are already studies that have made it very clear that almost 100% of people can no longer distinguish between human-made and machine-made music (I'll provide a reference on request). I last tried this in May, letting Suno interpret my own pieces, and was impressed in every respect (here are the examples: https://rochus-keller.ch/?p=1350). The arrangement and the musicality evident in the interpretation clearly have a human quality. It is simply not true, that the choice between classical recordings proves that greatness cannot be achieved through averaging or pattern-matching. Much great music, including classical masterpieces and jazz standards, works precisely because it follows established patterns and conventions while introducing variations within those structures. E.g. in Jazz, an improvisation itself is pattern-based, using scales, chord progressions, rhythmic conventions, with creative variations. If AI can learn these patterns and introduce variations (which Suno demonstrably does), why can't that be "great"?
To my knowledge, we don't have LLMs yet trained especially to generate classical music (at least I haven't heard a convincing one so far). In this respect, I have to agree with the author that, at least in the classical field, we don't have the same quality as in other musical styles (yet). But that is only a matter of time, and it is not for the reasons the author supposes. As demonstrated with my - and actually also the authors - example, even among human artists, there can be huge quality differences. The author's "immortal souls" retreat is essentially an abandonment of empirical arguments for unfalsifiable metaphysics, which essentially concedes the technical debate.