Saul, David: Musikalisches Würfelspiel
Mark From: lilypond-user-bounces+carsonmark=ca.rr....@gnu.org <lilypond-user-bounces+carsonmark=ca.rr....@gnu.org> On Behalf Of Saul Tobin Sent: Monday, October 21, 2024 6:40 PM To: David Olson <dadadha...@dslextreme.com> Cc: lilypond-user <lilypond-user@gnu.org> Subject: Re: Das Glasperlenspiel coding question First of all, I just want to say I love almost everything about this, and especially I love the Hermann Hesse reference! I am quite curious what you mean by saying you use Lilypond to write poetry. Do you mean that you use Guile Scheme to generate the text and use Lilypond to typeset it as top level markups, with no musical score? I can see why Lilypond fits the spirit of your literary idea, as a text-based format for encoding music, and especially one based on LISP. It would definitely be cool to see Lilypond popularized a bit more in that way. That said, I can't help but wonder whether specifying the use of Lilypond is necessary for your literary purpose. If your reason for specifying the use of Lilypond is technical, I outline below some thoughts on its suitability to be the language of musical LLMs. I am a bit skeptical that an LLM is actually what you want for your story. LLMs as they currently exist need to overcome some major technical hurdles before they will be able to compose sheet music competently, which I'll outline below. Maybe it can be assumed that for fictional purposes the practical challenges are solved, but even then, I wonder if there might be other types of algorithms that are a better fit both technically (more thoughts on that below) and in spirit. Admittedly my sense of the Glass Bead Game is quite clouded by memory, but I would have thought that to get at Hesse's idea, you would want a mechanism that is "conscious" of the relationships of ideas, which is pretty much the opposite of what an LLM does. For your literary purposes, do you even need to specify the type of AI algorithm? Serialist systems are the sort of generating scheme for musical material that LLMs are the very worst at, but serialism also lends itself potentially to other sorts of AI techniques, such as genetic algorithms. I wonder also if the strength of the serialist constraints you place on the output of each successive token would in some way render the LLM equivalent to other, non-LLM based algorithmic methods; by crude analogy, if I asked an LLM to calculate for me the output of some very complicated numerical function, and then performed validation filtering out all the incorrect answers, the resulting output would be the same as if I had entered that function into a calculator. I wouldn't claim to have a deep understanding of LLMs, but I have played around with them quite a bit in trying to get them to generate music in Lilypond format. My strong impression is that even a very powerful current-generation LLM trained on a large dataset of musical repertoire in a parsable format would not be well suited to generating new musical scores, except perhaps if they were limited to a single melodic voice. The reason is that while LLMs are capable at suggesting appropriate continuations to an existing text, which makes them a good tool for some types of creative writing, non-local relationships between semantic tokens are a critical aspect of musical scores to a degree that they are not in text. LLMs are by design not well equipped, for example, to validate that the rhythms in each part add up to the correct total durations, or to check harmonic and contrapuntal relationships. Lilypond actually makes this particularly hard because Lilypond input files are typically written "horizontally," whereas some other notation programs store data "vertically" across all instruments measure by measure or beat by beat. In order to suggest a musical continuation of a polyphonic score, the LLM would first need to parse the Lilypond code sufficiently to understand what the semantic tokens are from a musical perspective across all the simultaneous parts, which is quite different than parsing the tokens within the code itself or even from parsing the notes and rhythms, since musical ideas can consist of complex interrelationships. That said, the same mechanisms that are already being developed to allow AIs to, for example, perform logical validation of software code or mathematical derivations, will also provide the needed baseline capability for composing music. However, music is likely to test AI logical reasoning capabilities to a greater extent than even most mathematical and engineering use cases, due to the fact that in a musical score, the relationship of every note to every single other note in a given passage must be considered. Even if the logical reasoning challenge were solved, that would only bring LLMs to the starting line of being able to coherently add a new polyphonic musical idea onto an existing sequence of polyphonic musical ideas. But this would be the musical equivalent of the type of trite, contentless drivel that GPT-3 was famous for. To write actually competent music requires a whole additional set of layers of non-local relationships, namely those between ideas. Strong LLMs are capable of imitating such relationships in text because they are trained on datasets that contain them; if you ask an LLM to write a short story about characters named Adam and Eve, it will know that in sentences where the name of a person is appropriate, it should prefer one of those names to some arbitrary name. But that is not the same thing as knowing that the story is about those characters, and the difference becomes apparent in situations where a reasonable choice of words based purely on probability leads to nonsense or inconsistency. This, it seems to me, is the very sort of understanding that is critical to composing music of any quality. Now, it might be that with a sufficiently large training dataset, the model would be able to make good enough guesses that its lack of non-local understanding wouldn't matter in any practical sense. Clearly that is the case at least for certain types of prose writing. But it's difficult for me to imagine that being the case with music, simply because even if the entire canon of published sheet music were available as training data, that dataset would still be orders of magnitude smaller than what is available for prose. Maybe an LLM could be expected to compose new Christmas carols, possibly even baroque fugues or bel canto arias, but for instance if one wanted it to write a completion of the Mozart requiem, I am doubtful that Mozart's catalogue would provide sufficient training data for the model to succeed in picking up the thread of musical development where Mozart leaves off, let alone to continue the thread of development in a convincingly Mozartian manner. On the subject of training data, one might naively think that text based formats like Lilypond or MusicXML would be well suited to training LLMs on sheet music, since interacting with code is an existing core capability of LLMs. Personally, I think this is unlikely to be the path toward a model that can actually compose, because the repertoire available in these type of text formats is a tiny fraction of published sheet music. In order to create a training dataset of sufficient size, either an implausibly large human data entry effort would be required, or advanced OCR for sheet music would be required. But OCR would probably itself use some type of AI model to do the translation into text. And let's remember that Lilypond is ultimately a language that is compiled into music notation. So if our AI model is able to use OCR to parse sheet music into Lilypond and then write Lilypond that compiles into sheet music, which can then be parsed back into Lilypond by the AI...why not just let the AI read and write music notation directly? The main benefit of Lilypond for algorithmic composition is that the Scheme representation of music is useful for programmatic manipulation, but this isn't the type of internal representation used by LLMs. All this to say, IMO, if we ever get an LLM that can competently compose sheet music, I think it is almost certainly going to be trained directly on music notation without the need for a text representation of the music. I'm including the list in my reply, since I suspect the overall question of AI composing music is of interest. Please pardon my ramblings. Saul On Mon, Oct 21, 2024 at 4:45 PM David Olson <dadadha...@dslextreme.com <mailto:dadadha...@dslextreme.com> > wrote: Dear Lilypond Composers, Questions are regularly posed on this list that suggest that many Lilyponders are composers who use Lilypond to generate music based on creative conceptions, rather than the traditional way. I'm writing a faux-philosophical novel in which characters occasionally speak speculatively about Hermann Hesse's "glass bead game". It is kind of an updated Search for the Holy Grail (actual Glasperlenspiel). How to decode Hesse's many hints. What are the rules? What if Lilypond is used to play the game? One of the characters says: The Glass Bead Game, while being set far in the future, is nothing other than sets of Leibniz's characteristica universalis identified by competing future large language model artificial intelligence, wherein the coding teams who have developed each LLM compete by generating serial music using dadaist algorithms for their own LLM AI, in which, not the notes, but the large language model universal characteristics (post-Platonic forms) can only appear once. (the character's nickname is "Nothing Other Than"; he's an academic philosopher) 1. Can this general idea (scenario) be tweaked to be more specific? (more interesting to the coding community) 2. Is there way to mention Lilypond specifically in this paragraph? (e.g. replace everything that follows "wherein") 3. Is there any objection to mentioning Lilypond specifically in the novel? 4. Are there any other novels that mention Lilypond? 5. How might Lilypond be mentioned in some earlier chapter, to better prepare the reader? Honestly, No. 5 is entirely my responsibility, but having grown up in a Daoist country, I like Sets Of Five. On the other hand, if someone has had an humorous personal experience that could be imported into a narrative, I'm all ears. Maybe that would be a personal response. Hesse and dada. Hesse and dharma. I also welcome personal correspondence about thoughts Lilyponders might have about Das Glasperlenspiel (The Glass Bead Game). Because the protagonist is a member of a Freemason-like fraternity that has a karaoke room in every lodge, often used to sing a cappella rather than karaoke, and the protagonist travels frequently, it's easy to introduce a variety of characters with different views about music. Historically, Freemasons liked to sing together and had their own songbooks. Until 1991, The Sacred Harp contained a 4-page anthem titled "Masonic Ode." I'm not a mason, but I'm fascinated with Masonic music. For the record, I don't use Lilypond to compose music. I use it to write poetry. I feel that my first draft of Nothing Other Than's description is unsatisfying to people who actually write code. Suggestions welcome. Pond-fraternally yours, David Olson Los Angeles
