Elan Barenholtz on Language, Music and Art

I am so excited to be here today with Elon Baron Holtz and Drew Garrett. Um, you've all met Drew before because he's been on here several times talking about music and the the connections between music and art and the universe and and Elon is a uh so Drew is uh an K K through2 music educator and also has a background in ethnomusicology and Elon is a cognitive scientist at Florida Atlantic University and um also deeply involved with AI. I um both a vision scientist and a cognitive scientist and Elon recently well sometime within the last few years had a kind of epiphany about language the the structure of language that I thought was very um concordant with some of the ideas I've been talking about here on the meaning code for the last seven years. So I was very eager to talk to him and um I'm going to start with a question that you asked Michael Leaven Elon in a recent conversation with him. You said how is it that language has these properties that are so mysteriously rich and complex? Where did this beautiful computational system come from? So I wonder if you could just address that question a little bit. why you came to that question, kind of what is the background of your epiphany? And then maybe we can move forward talking about how from my standpoint there is a there is a a deep um structure that reliably guides meaning that lies underneath art and music and language and I think also physics and all the other domains. So why don't you start with that question?

>> That's that's a a beautiful question and uh just as an aside before we even get going the meaning of meaning is something that I I I'd say that is in some ways the central core uh question I'm grappling with at this at this juncture in terms of developing my theoretical approach not just to language but uh really to to cognition more broadly the mind whatever one calls it. Um, but let's let's take a few steps back and talk about this this so-called epiphany. And I use that term. Um, and I don't think it's it's not it's not my epiphany. It's it's humanity's epiphany. Uh, it's something that we've discovered only very recently about the nature of language. Um, and it it all emerges. It all arises because of the the language models. uh things like GPT quad or whatever your uh your favorite AI system that you may use. Uh my my contention and uh I don't think it's it it really should be controversial is that we've not only built uh you know as as a species we've not only built uh these simulated uh intelligences that can can uh do lots of things that we can do linguistically and also in other domains. Now uh but but in the process of doing that uh we we actually discovered something uh we discovered something about the nature of language itself. Uh what happened is not that uh the clever engineers figured out how to get a machine to be able to use language. Instead what clever engineers did was take a generic learning system something that's able to simply learn statistical structure of any sequence. uh it's not what they're not ex exclusively in any way designed for language. They took that uh this generic learning system pointed it at language and discovered something magical and and extr just uh mind-blowingly remarkable and that discovery is that language contains within itself within its own structure. The corpus of language that these models are trained on contain within themselves the capacity to generate any sequence of any coherent sequence give a large language model. We give you know we we use them and we can give it a prompt. It could be something pedestrian something uh that we think oh okay this is sort of common knowledge that they've been trained on. We can give them completely novel uh prompts that are about you know writing writing a song or a poem about some existential crisis of being an AI or something like that that hasn't exist before. And the point is that these things are not thinking in the traditional sense. They don't have world knowledge. There's no uh sense in which they have a concept of reference uh that they're they're not they're not thinking in the terms of of something external to language. What they are doing is simply letting language run on its own. Language actually has this structure. I call it the autogenerative structure where if you just know the relations between words and you have a sufficient corpus to train on, you have these relational properties. If you just know uh what you can think of as sort of the map, the topology, call it whatever you want, but you could just think of it as basically how words live relative to one another. I call it is like these these you can think of them as meaningless squiggles that how they live relative to one another. If you can do if you can encode that you can actually generate coherent thought reasoning poetry whatever um just on the on that internal relational information alone. So the way one way to think about this epiphany is that when we're talking right now and this is a bit of a leap but I believe that human language is operating based on the same principle but when we're talking right now it's the language that's really doing the talking that the that this thisformational system this this uh whatever we want to call it it's very hard this we don't have terminology for this language isn't really designed to talk about itself but this system this information the corpus alone whatever you call it it is doing the speaking um it's doing it by continuing by taking uh whatever you you know you introduced me you asked me a question that formed a sequence that sequence has a natural continuation in the form of what I'm saying right now and so there's this remarkable fact about thisformational system that it has a life of its own uh that it it's it's self-generative that the predictive structure is actually what you need in order to speak and then the The really counterintuitive fact about that is that it's doing it within us. That when we feel like we're speaking um in some ways we are because it's it's the language is certainly influenced by things that have to do with me being Elon Barnold that I'm sitting in this room right now. Uh that you asked me that question that I have a certain history and all of that. But all of that we think of as a prompt and then the prompt is basically saying okay language do your thing and then language does its thing. So that's the big epiphany is that theseformational systems in this case language can have this internal life. Um and that what we think of as kind of a process or a tool that we're using is actually in fact something that has its own almost ontology. It's it's it's an in I don't want to say it's an independent but it's it's uh self-sufficient in some ways informationational system that's doing stuff in us rather than us using it to do stuff. So that's that's sort of the long and short of uh of sort of the grand epiphany regarding language. And I've I've talked for a while already, but uh I've spent the last year and a half or so trying to uh think through the implications of that doing a whole bunch of actual computational empirical work um trying to actually uh support the idea this deep intuition I have that human language is based on the exact same principle that this it's really just this kind of self-generated predictive system. um making a lot of progress on that and also thinking more broadly about uh well a couple of things the implications of that for meaning as you mentioned how does language do things how does it actually you know when we talk when I tell you something if I ask you to can you know can we take a break or can you you know can you grab me something um that runs in you and then and then has has behavioral implications how does that work if it's a self-generative system and then more broadly I actually have a I I believe language was built on existing autogenerative structure. That's for a longer conversation. I I think that the brain/mind is is actually part of a much longer uh biological uh trajectory of organisms that use this the structure of of of sequences uh that of observed sequences in order to generate their next output. And I know so I don't think I think language didn't appear out of nowhere. I think that's that doesn't make any sense. Uh but it was built it's a it is unique to humans but it was built on top of a brain or a nervous system whatever we uh we want to call it because it may even apply to not central nervous systems but it was built on mechanisms that do this kind of autogenerative kind of thing. And the last point I'll make just re relevant to to this the conversation is that I I do I think music absolutely ha is is is probably the most pure form of autogenerative structure. Um I think it might even have preceded language uh historically in some sense. Um and and that it is it is maybe more fundamental than language. Uh this is something my colleague William Han likes to talk about a lot that the the idea that that early language may have been more of a musical form that ours is kind of a bastardized version uh that sort of maybe in some ways corrupted uh because music seems to do what language does. It selfgenerates rhythm and and the patterns and melody uh are these structures that that perpetuate themselves in our minds and they do it in this very in in this very profound emotional way in a way that language doesn't have. So that's that's uh a lot just to start off with, but that's sort of the epiphany and where I've ended up.

>> So Drew, yes, respond. Jump in.

>> Yes, jump in anyway. Well, yeah. Okay. Well, well, well, thank you, Elon. That was uh terrific. It's very thoughtprovoking and um so in my uh training in ethnomusicology I took a couple courses on theory and analysis of non-western musics and the method that we use to conduct an analysis of all sorts of different kinds of musics from all over the world. uh many of them don't conform to conventional western approaches to harmony or rhythm or pitch. Um uh is is we we used a we used an analysis called emic analysis. Uh, and one of the things I um maybe was hoping to I think one of the ways that our world's sort of combined a little bit Elon is um ethnomusiccology has drawn a lot from linguistics as a field for its research methodologies. Um, one of the one of the nuggets I was talking about this with with Karen yesterday is um the difference between a phonetic description like or or or an edict description they would call it or and an emic description. So in in language you talk about phonetics and then phonemes. Uh, and so if you're if you're going to spell a word, you look up a word in the dictionary, it gives you the phonetic spelling, which is like a goofy spelling, but it helps you say it correctly, even if you might not have any idea what the word actually says in its native. Whereas the phone names, these are patterns of sound, groupings, syllables, uh, that are endemic to a language, some ethnol linguistic group, um, and or or groups. And then you know so so you get you get to Semitic languages and dramatic languages, romantic languages. These have they have their own phone names and and things like that. And so the emic analysis of music um tries to get at the insider, how do I say this? The the it it's funny because they called this a generative analysis and it's totally not in the conversation of AI or or any of that stuff. Um the purpose of the generative analysis is um you get to know the internal structures of some musical tradition so that like you can actually then produce a composition, a song or or some musical work that's representative of the style like and and if it's valid enough, you can consult an insider, show it to them, and they'll say, "Yeah, that's it." Or, "No, it's not it. It's missing this." things like that. Uh, and so they walked us through this method. There's a musicologist named Tom Avery, the late Tom Avery, and he had his whole method um for example, where you uh you first you transcribe a song, you transcribe a melody from some tradition, and um and then you do a very simple um analysis of what note goes to where, how often, how many times does that note show up, which note seems to be the the main note of the song, the tonic, what phrases can end on which kinds of notes. And then and you do this so many times that you get a sense of the limitations and the and the boundaries and the parameters and and it's it's the more songs the more iterations you get of these rules and and eventually an internal theory develops. Um, and one of one of the one of the philosophical debates was does is there a theory for every kind of music? And I think the the assumption of this approach is that even if it's not a spoken theory, even if the insiders themselves um don't articulate a theory, it can still sort of be derived. And so so that's uh so so this analysis is a generative analysis and it very much falls in line with what we're doing dealing with language here.

>> That's amazingly that's very interesting. Um first couple of questions I have right off the bat. Um, so the the kind of analysis you're doing, I assume that although it's there's variations and and there's something unique about each musical form, but I'm g I'm guessing that the fact that there that that there is gener, you know, this kind of generative structure that you're talking about and first just just parathetically people aren't they're not doing this they're not automating, right? This is not a computational approach in its in the in the sense that we actually get uh you know algorithmically you're generating music. You're not talking about that. You're talking about that a human being can define these rules, follow these rules and then lo and behold you've got a composition um and you've done it just based on on this sort of syntactic kind of analysis. And then somebody who actually knows the musical form who may not know about your syntactic analysis, right? uh and has never done that will be like, "Yeah, that's that's a legitimate tune. It's a legitimate sort of >> precisely." Yeah. >> Right. Okay. That's that's really cool. Um so and and that this this appears fairly well universally, would you say? It's like across all the musical forms that you've uh encountered. Um well in the in the in the tradition that I'm familiar with the the the the because at the musicologist just might scratch their head and say why would you even need to do that in the first place? So it depends on what your goal is. Are you trying to describe a music or are you trying to help other people make the music? And the people that are sort of motivated to make the music in this context it would be um okay there's there's there's churches showing up everywhere in the world but not everybody necessarily wants to sing western music, right? So how can we how can we adapt local culture and local musical styles or something to be suitable for for lur for liturgy. So that was sort of the the motivation behind why that research sort of really developed. Um, and >> why that developed you mean >> right? Yeah. Exactly. Why that why this technique of of emic analysis >> develop you have you have you have a you have a particular you want to set it to specific words for example >> way that >> in fact the first instance of algorithmically generated music is from a 17th century Jesuit priest who invented >> Jesuit who invented the Jesuits again who invented a device the ara muserythmica and it Um, uh, I can't even necessarily tell you how it worked. It was a system that used music numbers and tables to generate musical settings for any given Latin text. And it and it followed the rules of polyoni. And so you could like run the machine and out comes this this chant or this, you know, this polyony that some local parish could, you know, their choir could sing it or in case in case the the priest that was there didn't know anything about music. Here's an And that's hundreds of years old. >> That it was an was it an actual physical machine or was it just like an algorithm you could carry on? >> It was a box that was filled with rods and tables. >> Oh my. Okay. Yeah. I'm surprised I've never I probably have heard of it, but it sounds like something I should have heard of. Um Oh, that So, okay. So that was that was sort of the motivation. But now you know uh we live in in a time when you don't have to have uh a a a clearly applied motivation. The the interest here is is is one you know is a little bit of an abstract one. What do we know about sort of the universality of well well first of all are you know are all are all musical forms sort of capturable in some ways uh using this kind of technique and then the second question I have is are there general are there universals not just that you can capture them but like are there things that apply you know you think of universal grammar like the chskin kind of idea I'm not a choskian just to get that out of the way but uh you know there are certainly syntactical forms that show up within within any language, there are certain syntactical forms that show up. And then across languages, well, it's hard to find the universal ones. That was like that was the wild goose chase um of the of the past century trying to find universal grammatical uh rules. But in music, are there >> first first of all, are there are do all musical forms have sort of rules that at least that can be too loosely captured to to some extent? Or sometimes do you encounter music that you're like I just these guys get it and I just you know you figure out the the structure you can't figure out and then if there's structure to every musical form is there common structure across lots of different musical forms. That's that's a question that came to mind.

>> Well, every every field has its hobby horse and uh >> ethnomusicologists love to wrestle with debate, debunk or engage in, you know, pmics over the question of universals because their their field developed specifically because there was too much of an emphasis on one tradition and we want to, you know, democratize our notion of what even is music. Um, and uh I think the average opinion among ethnomusicologists who are mostly anthropologists like in terms of their methodologies and and and scope would say um if the universals are there, they're not that interesting. It's actually more about sort of what makes these things distinct and and from what lens or from what frame like they they would problematize stamping. these are the universals of music because you're always going to find some exception to that rule. Um where whereas folks more in the music psychology world there have been there have been some valiant efforts in trying to pin down uh general structures that apply to all musics everywhere. Um I I think okay I'll start with this. The behavior of singing is universal >> the act of singing and that comes with many implications. Right? And because now because we're biological creatures, we interact with our ecologies and space and resources and things like that. Um, so in order so singing like that, that's going to imply uh inonation. It's going to imply audiation, being able to audiate pitch and and sort of match um what's what's going on. Um scales and tonality, uses of harmony, uses of rhythm. Um these things are going to vary all over the world. Um but being bipeds that's kind of universal, right? So there there's lots of things that we do. How do I let me just step back? There is like language. There is no one part of the brain that's like oh that's the music part. It's uh it's all sorts of it's it's both hemispheres are involved. Um from >> it doesn't show like like linguist. It doesn't show like left hemisphere. I I I would just guess based on what I think is it probably a connection to language. So it doesn't show sort of lateralization a little more left hemisphery like in language. >> My well when I read when I read u uh this is your brain on music. Um hold on have to Daniel Evton. Um it's it's uh well it uses there there's elements on both sides that are used um for all sorts of different things. Um, and it's sort of it's sort of a like even listening to Stephen Pinker talk about this, it it's music emerges like he uses the word hyper stimulus, but like music emerges as a um uh um well, I guess the word that I've used with Karen is an exaptation is there's all these things that evolve for for for all sorts of different things and they all sort of come together in this complex hole um of music um with between rhythm and uh audiation and all of the bodily components that go into it, you know, with the oral cavity and and the oral cortex, the cerebellum, um and emotion, memory, um like all of all of these things are just sort of happening all together. It's very it's interreational. Um, it's in space and time. Um, the Yeah, I'm trying to remember where I was going with that. Um >> uh we were talking about universality. >> Oh, right, right, right, right. >> Uh yeah, you're saying, you know, that that uh and well, let let me let me stop you for a sec because you know, what does it mean to say all cultures sing? And then we get into what I imagine is a controversial topic. You know, what constitutes music? What doesn't? Right? serialist music movement uh you know was a kind of an attempt to totally abstract away uh music to some to you know patterns uh of some sort that that don't actually sound like music to anybody >> right >> uh it didn't catch on I think the reason for that is because music isn't just some arbitrary uh you know any old structure um it and and when you say everybody sings things. You mentioned tonality, but it's not just any old tonality. We don't call Chinese tonal uh language music. Uh, we there's Chinese music, of course, but the language itself is tonal, but it's not music. Um, and so what what does one mean? What do you mean? Uh, what what do music what do musicologists mean when they say something like uh every culture sings? and and that starts to get into the universality question which I I think is a is a you know I would strongly disagree. It's a very interesting question. Obviously it depends on what your what your your goals are but it's a deeply interesting question if you're wondering what the hell music is and why it does the things it does. And certainly in my case obviously I you know I I feel strongly like it's it seems to have um genuine uh kind of internal structure like it has this kind of life form that it does things um in us rather than say you know rather than saying it's this all these separate pieces that are all brought together seems more like no there's this uh sort of particular patterns there are patterns of sound that do all these different things in us right and they light different parts of they cause emotional reaction and they and they uh and and they they cause memories and they rather than the other way around like that we we're harness harvest harnessing all these things but anyway I I I don't want to put words in your mouth but I I would like to know you know what what is your thoughts and what does the your field think about what constitutes music such that we could say well then that's that's the universal right if we say every every culture sings well what do we mean by that right that they make sounds uh of of certain tones. Well, what about purely rhythmic music? It's not tonal. Uh every does every culture have that? I'll bet it does. You you tell me that not every you know maybe maybe some cultures are not as rhythmic as others, but like there's still rhythm.

>> But I just I want to inject a question here that might clarify some of this.

>> Yeah. So Drew, when you say that most ethnomusicologists say that there is no universal or if there is, it's not the most interesting thing. Is that related to the fact that most ethnomusicologists are postmodern?

>> Um certainly.

>> Yeah. >> Well, so then that that create that's injects a bias into the question, right? >> Postmodernists don't like universals. Is that is that kind of a >> That's right. Yeah. >> Is that a universal post? >> Yeah. They don't like they don't like grand narratives. They don't like universal. >> It feels it feels >> postmodern, postcolonial, >> colonial somehow. It's like because it's like sort of like this the dominant, >> but it seems to me that even if you're saying that all these different ethnic groups um have music kind of in their own way. The fact is that they have music in their own way. It's an internally consistent system within which they can express meaning.

>> Right. Well, express meaning well, but it's not just, you know, there's language and then there's music and it's and it's harder to say what meaning >> language language is an internally consistent system within which you can express meaning. Art is an internally consistent system within which you can express meaning. Physics is an internally consistent system within which you can express meaning. So I just think there are all these different domains. Um, I mean if you compress it down to the simplest thing, meaning precedes matter.

>> So meaning comes before everything else. And um, and then there are these systems within which meaning can be expressed. So I don't I didn't want to interfere too much in the conversation. I just wanted to say >> I >> that that I think Drew is trying to be kind and not um expose the implicit bias of of some of the the um academic perspective.

>> I get what I I understand what you're saying and and yeah I uh >> I just want to be careful if we're trampling any any pieties. I don't care. But but people people often in in you know even if even if there's an agenda there could be you know certainly valid points worth >> absolutely absolutely but you just have to keep all the agendas in mind I think so Drew you were going to say something >> yeah I just uh I just need to be careful about what kinds of things I want to say that's the that is a universal right because I'm not even even within myself I'm not necessarily sure because there's too much music out there to really know right the data set I don't have we don't have the full data set to be honestly Not everything music like what what do you think I guess does it have to sound like music to be to to qualify?

>> So So my personal uh mission has actually been to like my my mission statement is I'm the music teacher that doesn't know what music is. Um I'm I I like the word music like we've inherited that from the Greeks and it literally means the art of the muses which itself is laden with all sorts of assumptions about what that is. And then when you okay let's let's go to the languages that word is not used um by in other languages like there there's many languages that don't actually have the corresponding words for music in the way that we mean music in in the modern sense because we like you like we're using music as like say let's say as opposed to language or as a different category of language. Um, and we have an assumption with the word music that's okay, there's a beat, there's a baseline, there's chords, there's harmony and there's singing and and that word music, it can be secular and it can be sacred, right? There's also there's there's a lot of assumptions that sort of go into just having a category of music and then talking about it the way we're talking about it. Whereas, okay, in in tribal settings and in other languages that >> that word is not there. There's other words that they use to try to talk about it. And um >> yeah, this is where language can sometimes be very not helpful and and it's it's like a lot of 20 20th and 21st century, you know, philosophy is is is in some ways trying to deconstruct this and you know that there's a that there is a sort of platonic category of music. Sure. And and then you know end up debating is it does it count or does it not? And that's probably not even that's probably not helpful. Um and and people I think this is where people could have a point for sure like what you know there's there's there's African drum um communication systems. I know I've read about that and it's it's somewhat rhythmic and it's got uh but it's but it's it's much closer to language than it is to what we would call music. I don't know if you can dance to it or anything. And so right then then it's just becomes really a semantics question as as the term is used. um what fits this label or something and and maybe that's not the right question then. But there are there are interesting questions though like does every every culture use tone um in a way that's not traditionally linguistic, right? and and we could sort of maybe define it negatively like we you know tonal languages are um you know carry they carry less conventional meaning although that's a tough word uh right Karen I mean that's a tough word I can agree with that um and they carry meaning with tone but then that's not some kind of tonality we're mean in relation to when we use it musically um do do all cultures use tone in this other way I don't know what you let's not call it musical then but let's call it um something meant to well I don't know you know it arouse it's it has it certainly has an emotional component it's enjoy there's there's an emot there's there's a a a veillance aspect to it um meaning like there's a hedonic response to it I I think that that's maybe that's universal right everybody uses does every culture use tone for hedonics like there's there's malady of certain sense like is that >> is that a universal I mean I don't know that it is but I'm proposing you know maybe it could be Yeah. I think when people when when we say every culture sings on some level that's that's kind of what we mean.

>> Yeah. Um, but okay, so like the I I've lived in Islamic contexts for um about a decade over over 10 years and uh >> the call to prayer the Muslim call to prayer and the uh the recitation of the Quran to ears especially to western ears sounds like music sounds like singing >> does to me actually >> within the tradition they will like if you're a traditional Muslim like to call that singing is actually to degrade the value of what's actually happening and so there is a little bit of you know semantics like as we were talking about >> um but the what they're doing is in toning right there is like an intonation that's happening there is the use of the human voice the human voice is highly elevated in that tradition as uh it's like the the you're not getting to the meaning of the Quran unless you're hearing it in its oral audiated form. And is it beautiful to them? Do they would they say that it's meant to be beautiful and to have >> Yes. Yeah. Yes. Yes. >> Okay. >> Certainly. And they have and they have that funny enough an ethnomusicologist did a study uh and she actually uh went to Egypt and studied how to recite the Quran with professional reciters and teachers and trainers. And the reciters they they themselves actually will go to music school like as sort of be like they'll they'll actually get singing training right and learn about scales and and and how that works and and and singing that uh setting setting the for the Quran. Um, but there are there's internal, you know, logic rules in there to determine if it's valid or if it's not valid, if it's beautiful or it's not beautiful. Or if if it's too much, like if you label it singing, that actually means that the reciter is being too performative with it because the text sort of comes first. So there's even some rules there with >> uh with that. But I as a as a starting point um every culture sings and that that carries with it certain implications and one of those implications is well why do they sing and fundamentally that's a social there's social reasons for it. Um, serialism is not for social reasons like that's you're in an art there's an art music dimension to it that separates the sound from some sort of social utility or social function or something like that. Most music is is uh is about connecting people together. at most music is uh there's biography, there's sociology. Um, it's done live, it's done to be together with people, it's done to it's for some expressive purpose that's beyond necessarily just the sound itself. Um, and there's a big debate about music, absolute music that sort of exists just for itself. It doesn't mean anything other than itself, as opposed to, well, all music must be expressive of something beyond itself. Um, it's a philosophical debate within music studies, music theory. Um, yeah.

>> So um I see I so the the that's in the 20th century did this thing. Uh certainly in western sort of the the western academy um said hey let's let's strip things down to their essential form to a point where maybe they're not recognizable and then would claim that serial music is music. But it's not only the social form. So maybe that's a little bit inverted to say it's it's not so much that music has to have this social form. Um maybe to to be music. I think that that the other way to look at it potentially is the reason it has this social form. The reason why it has a social utility is because it is music. Meaning it is it's evocative. It it has certain things that for you know it gets people you know rhythm makes people start to move in concert like it does this thing. It programs us. I mean I don't know how else to put it. It gets you your body actually moving a certain way and that has that unifying social component. It's not cuz like we're all listening to it in a club. You could put on the serialist music in a club and everybody will try to sway and it's just gonna not it's not going to work. Uh, you know everybody people just go to the bar and start drinking. It's everybody get off the dance floor, right? it doesn't do it doesn't have that utility and so it is really something I think I you know I would I'm not a musicologist and I'm straying into waters that you know these may be choppy waters but um I think music almost by definition has to make you feel something um it's it ain't music if it's just a bunch of notes put together that that somebody said this constitutes music and this this was the whole 20th century this was art right they're like oh it doesn't have to be beautiful it doesn't have to be representative it doesn't have to be it is whatever I say it's a urinal on a wall and in some ways this is this is absolutely brilliant absolutely brilliant I still I think the postmodernist like that's the those were the I forget the dais I think it was a brilliant social commentary that actually there's a lot of social pressure and agreement that constitutes what's what counts as as beauty beauty even beauty and they're like hey if everybody thinks this is beautiful then maybe it does start to become beautiful and that's that's a cool point but they really were trying to uh in some ways that the the project was to say there's nothing inherent uh about any you know that it doesn't matter if it is anything and I think most people up till you know that particular tiny niche of human history would have said uh no there's something there's there's beauty and then there's things that aren't beautiful there are things that move us and then things that aren't um and there's a lot of gray area and then with music there's it's it's either it is or it isn't there's noise and there's music and of course you'll get other cultures and I'll listen to their music and maybe it won't sound quite familiar and maybe it'll be strange. Um, but I I I've almost never encountered in my life and I've, you know, I've heard a fair diverse amount of music music where I'm just like this is something that people gather around, they enjoy it in a social setting and I'm like, yeah, to me that just sounds like nothing. Uh, it's it's the music is it's it's the universal language. They call it that. I've that's a term people use. So, I think it's hard to get away from the idea that there is something um uh that that's almost defining in music that it does certain things within us. Anyway, I'm sorry you you've been wanting to interject.

>> No, this is everything you're saying is awesome. So, um there's a a musicologist named Thomas Torino. He wrote a paper, famous paper borrowing Charles Pierce, the uh linguist. uh excuse me, semiotics. Semiautition. So from semiodics he um >> uh proposed that music communicates through signs, indexes and symbols. Um so using the the pierce language of and distinguishing these three different terms s uh so these are different kinds of signs. That's what semiodics is. It's the field of signs and studying signs and symbols things like that. So so there's there's three levels. There's icon, index, and symbol. Um, I'm going to pair this really quick to Leonard Bernstein, uh, Leonard Bernstein. So, he, of course, famous for his lecture series, uh, young people's concerts and all that. Um, and he has a great, the introductory lecture is like, what is music? And you he'll go to a piano and he'll play a note and it means nothing. Like there's no word, there's no image that comes to your head. It's just a tone. But then you add a second note and suddenly now there's there's a relationship and that relationship generates all sorts of kinds of feelings. Um, like he he'll he'll he'll take music that's ostensibly program music. My example that I do with my students is Vivaldi's you know seasons right where it's so famous spring. I'll play the music for for students who have never heard it before and I'll ask them okay so what's this music about? Now, if you know the title of the music, at the very least, you know it's supposed to be about spring. So, then you could be listening for those elements. But if you're going into it blind, like you don't actually know that's what it is. I get all sorts of answers what they think the music is about. But no, nobody mentions the weather or a season or a time of year or even birds, even though that's actually, you know, sort of in there. It's be it's you're you're speaking in a in a tonal language that's nonreferential to overt semantic concepts or images. um and yet it's and yet >> the emotion and the expression is there. >> So how is it conveying meaning, right? What is the meaning being conveyed? Um, what's what's amazing is like you can listen to a piece that's nostalgic or a piece that's griefladen or or anything like that. And okay, so maybe okay, so maybe Beethoven wrote the Moonlight Sonata as a form of his own pouring out his soul grief over his hearing loss and lamenting his life in that way. But when I hear that piece, I don't have to hear that story at all. That that story I'm not going to come away with that specific story. It's not Beethoven's narration of his personal individual biography. And yet I still can feel grief. I can connect it to any grief that I've had in my own life or whatever mournful feelings that I've had in my story. And that's what's sort of being aroused through. So back to the Pierce thing. So an icon in this context is a musical sound. That's a direct one to one correlation. The sound literally sounds exactly like what it sounds like in the real world. So if you have birds, if you're using flutes to make tweety bird sounds, that's an iconic represent. It's supposed to sound like birds. But then you might you might still use flutes to represent birds but not make overt bird sounds. So the next level of a symbol of a sign is called an index. And the index is how most music signifies. So the the sound of a minor chord for many people is an index for sadness or it's an index for some that that whole side of emotions. Whereas the sound of a major chord or if you convert the minor into a major that's an index for an uplift or or some kind of thing. Now that could be connected to biological uh intuitions about up and down or bright and dark or or things like this. And then the the the bottom level the the other level is called the he just calls it the symbol and the symbol. So you have icon, you have index and you have symbol. The symbol these are things that are they okay they call it arbitrary but it's this it's like it's um it's cultural specific. It's highly emic. It's highly uh endemic. It's it's you're not gonna Okay. There there was a big debate about heavy metal taking taking a taking a there's a there's a a a Christmas carol that was set to was it's a sacred Christmas song that was set to a heavy metal backtrack and people were there's this big debate about like was that appropriate or or was that not and my my argument having studied world music we can't enter into someone else's music already sort of knowing what it's going to be about or what it means because you have to spend some time learning what what are the hooks? What are the riffs? What are the you know what what patterns of sound are meaningful to the group of people that uphold that tradition. I'm a jazz musician and I could play a recording for you uh a recording of a saxoponist named Chris Potter who does a 10-minute saxophone solo on uh All the Things You Are, a famous jazz standard. and he is going to do the most remarkable things with those chords on that song just as a saxoponist and I will be able to hear meaning from it because I know the patterns of sound that Chris Potter is using on that song. But someone who has never studied jazz, doesn't even like jazz, they're going to hear the same set of sounds and well because they're an uninformed listener for one thing and they haven't gotten to know the music. They don't have a relationship with it. That's the key word is the relationship. It's it's not going to land the same way. And so their their set of indexes they they they don't have the same set of indexes that like an internal jazz musician is going to be. So there's an insider and an outsider sort of dimension going on in between musics to the point that music is a universal language. Um, I think I think ultimately it's going to kind of be if as long as there's a beat, yeah, you'll find something in there you can connect with, right? Because we have feet and, you know, everyone has a pulse and we all breathe and we all feel rhythm so intimately with our bodies. Um, but there but there's a whole lot of patterns of sound that, you know, we just didn't grow up with and we're not going to have that same relationship to those sounds with. So a lot of it is relationships. Um, and music per se I think has to do with that has to do with creating relationships, fostering relationships, uh forming groups. Um, and so I like I do think music with hierarchies of pitch is better than music with no hierarchies of pitch for this purpose because we all rally together as a group. Uh, we can sing a song together if there's a tonic. We can sing a song together, right? It's it's it's all it's all about the the the the the creation of the relationships and where it points us to as a group.

>> Wow. Okay. I have a bunch of things I want to share on

That. Um, but you're you're taking this, it's sort of this interesting, it kind, it's sort of middle ground that it's, it's a little more of of a construct. And and and, you know, when you have these unique cultural or even individual, right, a specific jazz musician, and you know, how what what his patterns are, um, then those patterns start to make sense.

Um, but, uh, well, first of all, a couple of things. You know, and I've I've studied semiotics a bit, um, and so, but I don't want I don't want to delve too deeply into those into that, um, that particular school. But the idea of symbol, um, uh, and the idea of, and I don't know if you're using symbolism in this exact way, but this is where my my sort of my thesis, my core thesis ends up being, uh, I think, uh, is that language is going to look a lot more like music in the sense that I don't think words refer in the in the in the sense that we all think they do. That very naively, we think that, you know, the word computer or person, uh, picks out an individual sort of construct, you know, actual physical things in the world, uh, that there's that there's there's there's this idea that of reference.

Um, and in the case of, uh, of music, I think it's, it's, it's extraordinarily clear in some ways that there isn't this real explicit external reference, right? The Moonlight Sonata, as you said, um, it's not telling a specific story of any particular narrative. Uh, I I don't know if somebody who's who's, uh, you know, lived a completely different life, uh, and I don't know if it's true that it, it's, it's cross-culturally true, but I'm my guess is it might, it might well be. I don't know. You maybe you can weigh in on this, but, you know, somebody who hears the Moonlight Sonata, who lived a radically different life than Beethoven, couldn't even begin to relate, um, to what that, uh, what what, uh, narrative Beethoven may have had in mind, can still feel this in some ways transcendent, right? It's transcendent to the the specific narrative, um, and and it means it's not telling a story, it's not a referential, uh, it, it is what it's really doing in my mind is there's there's there are things in life that can make you feel a certain way. Um, and this music can make you feel the same way. Uh, but the the inversion, it's not about those things, right? It's not about the going deaf or, uh, whatever, um, you know, whatever suffering Beethoven may have felt. It's the suffering itself, which the world imposed upon, right? Beethoven felt it for specific reasons. But the music has, it doesn't speak to those reasons in any shape or form. And that's what makes it transcendent. It's what makes it uni emotionally universal. And and I think it's emotionally universal to the point that it truly transcends. Like there isn't, it, it may remind you of sad events in your life, but it's not even about that. The beauty of it in of of music, of a truly deep piece of music to me, is that it, it, it's actually, it's sadder than than sadness itself. It's sadder than any particular life event. You know, it almost, it's, it's like, don't, don't tie it to just that. That's that's too limiting. It's not about your breakup or even a death. It's not about that. It's about something, you know, a little in some ways more majestic, something something deeper about the human condition. Um, because it's really about the the the feeling, right? It's about the feeling.

And so that's why, you know, the idea that that there's reference, um, in the case of language, it seems like, well, no, language is much more mundane. It doesn't, uh, and I, I don't think it has that that ne usually does not have that transcendent component to it. But at the same time, what I think we've discovered, um, in the course of of our our our newfound understanding of language is that language is actually much more similar to to music than we thought. It's not really quite referential in the way we thought. It has much more to do with these kinds of patterns, these learned patterns. Now, you talk about it in the in the individual case, a specific musician or a specific culture. But there's the universal. The universal is that there have to be consistent patterns, and those patterns have to predict other patterns. And that starts to be what language is, because language ends up being these self-predicting patterns. They're they're patterns in, you know, not they don't sound good. Uh, it's not patterns in in somehow in this in this, it's, it's somehow not evocative in this emotional way. But what it is is not, uh, chair is a symbol for this physical object. It's that chair plays a certain role in certain kinds of patterns of certain kinds of sequences. When I say, "Can you grab me a chair?" or I say, um, uh, you know, "I need a chair," or I say, "a be, you know, imagine a very beautiful chair," or "imagine a dollhouse chair," or whatever. Um, that word is doing very different things, related across the different examples, but it's actually playing a different role in each case. But it's its role, not, and I think you, you said it earlier, it's a bit like Bernstein with the note. Um, Bernstein, it's, it's the note itself doesn't do anything. Words themselves don't do anything. That's the key. That is in some ways the key insight. The way language models represent words is purely relational, right? The word really is, it's, it's literally numerically. It's a vector. And what that means is just a point in a space. And being a point in a space means that you're fully defined just by your relation to everything else in it. And when we talk, what we're doing is we're singing out these kind of relational structures. And those relational structures are generating other relational structures. So the same way that a musical phrase, um, is is not, first of all, the phrase now has some meaning just because there's relational structure in it. But then, what's the real meaning? Well, it leads to the next phrase where it has certain relations to the right. And then you have hierarchical musical structure. That's what it's doing. And in a in a musical form, what it's doing is, uh, in some ways participating in this larger, uh, kind of relational structure. And it turns out that's what language is doing much more than we had thought.

And then if you go to the the my my full sort of theory, which I'm not going to, I'm not, I'm not going to bore everybody with, um, but I think this relational structure doesn't just, what the way language does stuff is then that plays out to other kinds of structure, like structures of movement, structures of perception, but it ends up being, uh, much less these atomized discrete symbol symbols, um, in the in the kind of conventional sense. Again, I don't know if semiotics treats the word symbol the way I'm talking about, but we think of sort of discrete symbols, like the word person means this. And what I think we've learned is that we have to get completely away from that. The word, the word does not mean anything. As strange as that sounds, the word plays a role in the longer in the greater dance of symbols that are then generating more symbols and that have these implications for other stuff. Um, so I think that there's like very deep resonance there. Um, but at the same time, I I want to, you know, question, and I don't know that you're taking this position, but the idea that that music refers in some way, that there's that, and and, you know, I think Karen had mentioned, you know, the word meaning. Um, what do we mean by meaning? You know, when you hear that Moonlight Sonata, um, is it referring in any sense to anything outside of itself? Um, I I I I don't know. I I would argue maybe not. There isn't really, it doesn't carry meaning. And I'm claiming that language may not do that really either. Um, but in the case of music, it's almost, it seems like the the pristine example of it, that there isn't semantics in the sense of like, there's not a message outside of itself. It's just a thing that does other things. And some of those things that it does is does it to us. Um, it doesn't explain how the hell music, the sonata makes us feel the way it does. And and and so is it conveying sadness? I would say it's not conveying sadness as is, it's, it's, it's actually a program, for lack of a better word, that generates sadness within us. It generates these these feelings.

>> Okay. So I, yeah, I I think that you've you've hit on something that really, um, resonates in terms of like artwork. >> Um, and using the Moonlight Sonata is a very helpful example. Um, there are structures that very reliably communicate something, even if, um, by itself it doesn't mean anything. How can I explain this? Um, Drew, you remember a few years ago, Adele came up with that wildly popular song, and there was one phrase in that song. Um, >> Which song? >> Hello. >> No, before that. Um, like I think it was your first song. >> I remember the sound so I can't remember. >> Yeah. But anyway, there was there was one phrase in that song. >> Rolling in the Deep. Rolling in the Deep. >> Before that. Everybody who would hear that song would cry at that one phrase. >> Someone like you, maybe. I don't know. >> Hm. >> Someone like you. I'm trying to phrase. >> Yes. Someone like you. Yes. It was Someone like you. >> Right. There you go. Okay. >> There was one musical phrase in that song that everybody listening to it would cry. >> And they could take that musical phrase out of that song and put it into some other song, and it would still produce that same effect. Okay. So now think about that in terms of not just music, but let's say painting or poetry. Part of it is a learned, um, response, like when Drew was talking about understanding all the structures behind a particular jazz solo. If you'd heard the thing in in a complete jazz trio or, you know, that had played it before and you'd heard that, or let's say you hear a symphony complete, and then later on you hear a piano version of that symphony, you can have some of the same feelings that you had when you heard the whole symphony. But if someone who'd never heard the whole symphony comes in and listens to the piano piece, they're not going to get the fullness of the experience. Um, I I I'll only speak from the standpoint of art of painting a painting. There are certain, um, elements and principles that are guidelines in painting that were not invented, but they have been observed over the centuries that show up in all the works that communicate and that have beauty and meaning. Those elements and principles are common to those works. It can't be, uh, reduced to rules. It can't be atomized. They're they're principles that intersect and interweave and kind of create a lattice structure within which an artist can play. And even artists who don't know anything about that structure can stumble on things that are beautiful in the process of painting, just because it's kind of the structure of the universe. So we grow up with it. We see it all around us. It's the water we swim in. But that structure will reliably produce beauty. Even if you're not, let's say you're not, um, doing a representational painting of some great mythic story or something, you're just using non-objective, um, color and shape on a canvas. You can reliably produce beauty on that canvas if if you're working at it from a, a learned, um, understanding of all these principles. So, I I took this creativity class eight times in which we studied these principles one at a time over many weeks, and we had to produce two large paintings each week about one of these principles. And then you begin to see how they all begin to interweave and how, um, there's so many intersections that it's impossible, it's computationally impossible to, um, produce a number that reliably represents all the intersections that can happen there. So, it's not some, what happens to an artist when they're working is that the possibilities of each next stroke of paint are so overwhelmingly computationally irreducible that you have to find something that will guide you through the system. So what most artists have done is that they they narrow their scope. Their goal, they have a goal. They have something idea in mind, and then they narrow their scope by limiting some sort of constraint on their process. You maybe limit the number of colors that you use, or you limit the kind of, uh, materials that you're going to use. You limit the, uh, the kinds of tools that you're going to use. You produce a constraint to limit this combinatorial explosion in front of you, and then you move forward step by step. And each stroke produces a new context. It's the context that matters. I think when I was listening to your your epiphany, the thing you did like four months ago where you explained your whole theory, I made a note for myself. The words he's looking for is context. Because even with an LLM, what happens is each choice of the next word produces a new context. And then that new context has to run. >> That's what it's called. That's literally the the term of art. So if I don't use it, that's my bad. >> Exactly right. >> It has to run back through the system. And when you're painting, it's exactly the same way. Every time you lay down a stroke of paint, but the system isn't working just by itself. It's always referring back to this idea or this original goal. And it's also working within this set of constraints. And I think an LLM is doing the same thing. An LLM is always referring back to the prompt and working within the constraints of the prompt, and each word produces a new context, and then that, so it is the whole system of learning. It's how humans learn. A baby takes a step, and then they're measuring that that step against reality, and you know, then their reality pushes back if they're going too far, and then they have to take another step. And each time there's a new learning. With art, um, you have a vision of what you want to accomplish, and you move towards trying to accomplish that, and then you discover, oh, wait, I don't have the skills to do that, so I have to learn some new skills. So then I learn the skills. In the process of learning the skills, my vision increases, and now my vision has gone further out ahead. So I have to learn some more skills and more skills. So that whole process of learning, I think, is analogous also to what an LLM is doing. And I, I just think it's the way everything operates.

>> Interesting. Before you draw, um, Drew, so the the the end goal though, is is where I'm like, that actually, so that we could actually sort of carve that off and say an LLM does not have an end goal. Um, it's it's >> But they have a prompt. >> They do. And the prompt, but the prompt is is the context, right? It's it's one and the same, right? Is the prompt is is what's generating the next token, and that forms the next to the next, that forms the the new context, and then it's it's it's it's really this, it's autopoetic. It's it's this self-generative kind of system. And there isn't, which makes it all the more remarkable, because it, for the life of you, it's, it's impossible to believe that they're not, when they're when they're producing the next word, and then that they're not thinking about what they're going to say later on. But what's what's extraordinary is that the system, through learning that you're talking about, through learning the this this, uh, the kind of inherent structure of the way patterns happen, you can do it one step at a time in such a way that ends up looking as if you're matching towards some end goal, but you aren't actually explicitly doing that. And and and here's here's what I want to, I wonder, I think for artists, I wonder, I wonder if there's two different ways to go about creating visual art. Certainly in music, improvisational jazz is a perfect example. I don't think that they're trying to get somewhere in particular. >> Um, maybe sometimes they can feel that there's some big crescendo coming that they're all working towards. Um, that's that's maybe a, you know, a slight counter-example, but mostly they're much closer to the to the LLM. They're like, "Here's what's happened so far. >> I'm going to step into the dark with the very next step, as overwhelming as as it can be in the sense that there's so many directions to go." Um, but they're just taking that very next step. And then magically, just like the LLM, magically it all makes sense even in the longer term, right? It's, it's, I, I, it's, I use this term pregnant present. I don't use it that often. I've used to use it for my writings, but it's like you're doing the moment right now, but it's got within it the future is sort of built into it, uh, that it's going to make it tr, it's like a trust me thing, like I'm, I'm going to this note, and you're like, you're going to that note, right? And sometimes in jazz, in particular, like, what are you doing if you just looked at that very next note, and then you're like, oh, that's what you were doing. But did the musician know what they were that they were getting there? Maybe not. It's the It's the pattern itself coursing its way.

>> Well, so I wanted to ask you about this. It's like there's little injections of chaos. Like you're absolutely right. When I'm painting, >> there's straight up injections of chaos. >> When I when I'm painting, I'm not I'm not, um, mentally trying to always rein it back in to get back to the idea. I'm I'm trying to generate as much creative possibility as I can so that when when whatever connection is to that original idea, let's call the original idea prompt, at the end, whatever that connection is, it's now fuller and richer than the original idea was to begin with because of these little injections of chaos. You can purposely inject chaos, or you can accidentally inject chaos by making a mistake or a failure or, you know, having some sort of anomaly enter into the system. But but, um, and and jazz is like that. There's little, so LLMs have little injections of chaos, right? Or little injections of, um, >> well, unusual alternatives, like. >> Yes. But you know, it, it's called it says the temperature, and it, it's basically, okay, what an LLM is really doing, it's saying, based on this entire sequence, >> what should be the, what's the most likely next? And that's one way to to to phrase it, the next most likely word. But what they're really doing actually, at every moment, is they're generating a distribution, and they're saying, given this little sequence until now, what's the probability that, uh, this is going to be the next word? And then across all the possible next words, it actually gives, it gives a ranking to all of them. And what it does, the way it injects noise is it say, I'm not going to pick the very next most likely word, the one that's most predicted. I'm going to pick the one that's a little bit less, or maybe even further out on the distribution. I'm going to pick one that's a little bit of a less, more of an odd choice. >> It turns out this is essential. It's essential. Like you don't get the the, uh, the the sort of expansive generative flexibility, um, if they just end up because they just end up sort of repeating what they've already what has already been said before. Exactly. And that ends up being completely uninteresting. And then you have this this remarkable property that I think you're talking about, right? That the prompt sort of has the idea in it, right? And it's like, so, so, so why do you need to talk? Right? Why don't we, we should stop right now? Why are we, why are we continuing this conversation? Haven't we already said everything we're going to say? And the answer is no. That that language has both this con this constricture of source, that there's it's based on what has come before, and that's in some ways determining. But no, the creative process is that there's other stuff. I'm having this conversation with you guys, and I'm looking at your paintings behind you, and, you know, I'm thinking about what Drew is probably thinking as we're talking, and all these things are converging to to to create that noise, to that chaos, and it takes that it takes that next token and it shifts it over here, like that jazz note, and it moves it away from where the obvious place to go. And that is like the actual creative beauty of of like of of everything we're doing. It's like, you know, I don't want to not getting religious or or too too existential here, but it's sort of what we're doing here. Like the the the unfolding over time is not just, you know, the the the clockwork universe where it's like, uh, you know, you set it up and everything's deterministic and like, then what was the point, right? And the point is no, it's, it's not deterministic. I'm not getting into free will and all of that heavy stuff, but there's a certain sense, Wolfram calls this computational, gosh, I >> ability. >> Thank you. I can never remember that word for some reason with that phrase, but computational irreducibility. You have to run it. Um, because all this crazy chaotic interaction is totally not in the original. It's like, you can't get there unless you run it. And so, yeah, this is like what you're saying is is absolutely like it it rhymes a lot with what's happening in these language models. And, you know, we, they're not just, they're not just helpful chatbots, like there's there's deep deep insights, u that I think we have from these. So, and into like the the purpose of life, you know, not to be, uh, too too, uh, uh, you know, histrionic about it. Um, but this this this unfolding is not, we're not automatons. We're not, um, and we're, we're, or if we are, we're, we're beautiful automatons, like we're because we are these, we're these colliding automatons. We're, we're, we're colliding with each other, with with with the universe, in the physical universe, the mental universe, the musical universe. All this stuff is like, it's, it's giving birth to to the very next moment, the next token, the next big giant universal token, um, that we generate. So, sorry that I went off maybe too far into that to those nether, uh, regions, but everything you said, uh, kind of reminded me of this this extraordinary like general, this generative flexibility that you see in the language models, you see it in music, you see it in art. And it's not begin with the end in mind, as much as it is, where are we, and and taking the next step.

>> Drew, you got to jump in otherwise. >> I, I, I feel like Drew had to say like like 10. >> I mean, everything both of you, I mean, very profound. I mean, I, I have limited, um, right expertise, right? I can only speaking just from my disciplinary lens, like the things that are coming to my head. Um, you were you were mentioning earlier, Elon, about, um, thinking about about music pre-existing language, um, and I've I've seen that as well. I guess Psychology Today online, there was a post about, uh, the earliest form of language would have been more musical. I don't >> really know what that means exactly. >> Um, but maybe more relying on tone and rhythm and timbre and expression than semantics, for sure. Um, uh, but then also thinking about what we're talking about. I mean, there's been so much consilience that we're talking about between all of these things, consilience between LLMs and and and the art process and music, um, and every everything that that riff you were just, you were just riffing on on the metaphysics side, on the meaning of life, and and what this means, you know, for us, who we are. Um, I, uh, I just cracked open. I sometimes I like to read, um, ancient Greek music theorists, and, uh, there's there's one who's a Neoplatonist named Aristides Quintilianus, and I just started rereading his his book on, uh, on the the nature of of what is music. And he, um, so he says this in the opening of book one, he says that music offers the other sciences an account of both a first principle and a final objective. And the the the Greek words are arche and telos. Music isn't the arche for the sciences and the telos of the sciences. And what he will later do in the book, for one thing, is talk a lot about Pythagoras and Plato and Timaeus and the the formation of the cosmos using musical intervals, and how this is all just baked into existence. Unity and variety. The injection of chaos, right? All all that all of that stuff is just so primordial, and, um, and we're all downstream of it. And that that goes back to relationships, um, things being in right accord, and, uh, uh, and things like that. So, um, perhaps maybe on some level, regardless of what you think about the word music, maybe this is all music we're talking about here.

>> Well, the music of the spheres, right? That was, uh, I I don't know what that meant, but I like it. I know. I like it. I, it resonates, uh, that there, you know, it's sort of like the the the resonance of of the universe or something like that. Um, I think, you know, what why is music feel profound, right? That's it's not, it's going back to the to the to the moonlight and all that. Why, why does that sadness feel not just personal and, you know, specific, like singular human narrative, but why does it feel like it's, what is what is that? Why, why is music in, like, the most direct, to me, anyway, I, you know, I'm, I'm a musician, and I'm also most spoken to musician. No, and with no, um, I'm not throwing shade at visual art, um, but but transcendence, um, is something that music can achieve just so easily. You know, it's the, it's, it's, you know, Hollywood learned this a long time ago. You, you have that grand, you know, as you, this big sweeping landscape, and you, you watch that without the soundtrack, and it just doesn't mean one-tenth of what it does when you have that music set to it. And that sense that you have, um, of of it's not just, wow, that's big. Wow, that's far away. No, no, no. Right. It colors the experience completely differently. Suddenly, you have this different view of what like what it would means to you view to view this grand landscape. The grand landscape takes on, uh, well, to use Karen's word, right, it takes on a different meaning. And and that meaning is, is, to me, again, it's hard to put these things into words. It's not about the the old kind old-fashioned kind of meaning. It's not like, oh, here, let me tell you what I mean by that. Yeah, these concrete things. Oh, no, it's stuff you cannot say. You, it's ineffable. You cannot put it into words. It's not about these concrete things that we try to capture with the sounds we're making right now. It is something you can only, I can only use the word transcendent. And it's like music is almost inherently that. And and even, you know, music that you might not think is transcendent, but I feel like there's an element of, and it's a bad word to use. I don't know if it's, it's too loaded a word, but it is not about the other stuff of life. It's about some other kind of plain dimension, you know, the music of the spheres. It's like it's tapping into. Man, I should have been a musician. I'm like, >> You've taken the words right out of me. Like I exactly. >> Everything you're saying, I think I think it's what Palani calls contact with reality. >> Or or what, um, Peterson and Scruton, when they were talking about the transcendent, they talk about when you reach the edge of the known, when you fall into error, when you're about to fall into error, that's when you meet the transcendent. >> So it's right at the, it's right between the known and the unknown. And there's something about that edge that's inherently creative. And this is why the injection of chaos is always so powerful, because it's, um, it's creating a a moment of unknown that has to be responded to. You know, Hans Zimmer, who who wrote the the, uh, score for, um, Interstellar, a brilliant musician, composer, and he always says that when he starts a piece, he starts with a note that he would thinks of as the question. >> And then the next note is has to be the answer. So every question has an answer. And >> every context needs a next move, right? So it's like, um, it's like the next token. That every response to. >> But but it needs a question. Now, um, some of the other symbolic thinkers that Drew and I listen to talk about the feminine and masculine archetypes, not in terms of sexuality or gender or anything like that, but these are deep archetypes in the universe that the the feminine represents the the unknown, the the chaotic in one sense, the the, uh, exploratory side, and and also that the feminine is the one who produces the question. And then the masculine responds. So in a sense, that question is essential for generation. The feminine archetype is the question, and then that is both sides are essential for generation, but the question has to be produced first. And this is why I think with with AI, we tend to give too much, I think, power to the AI itself, because without the question, the AI isn't going anywhere. So the prompt, in a sense, is the question, and then the AI is responding to that. But what happens is that the AI, because it's an internal system, it can get so internally wound up in that. And I'm wondering if there isn't some way in which, uh, a beginning state of chaos might not produce better results. I used to do this with painting. I used to produce a lot of chaos on the canvas first, and then I'd have to fight against that chaos in order to produce the final image, and that in itself generated way more creativity for me, because I'm fighting against not just a blank canvas, but against something that's already there. Well, maybe in a way for AI, the entire training corpus is a sort of a chaotic start within which it has to operate. Yeah. >> If if I could just add, sorry, briefly to your >> to your look, it's so my, uh, my wife has done, we have a three-year-old, and so my wife was sitting, uh, with our with our three-year-old on one of those, um, it's like a tablet where you can draw things on it, and then push the button and it erases. So, so this whole idea of starting with a bunch of chaos. So like, so, so she had her or son just draw anything, and then all my wife would have to do is just, oh, well, it looks like there's a face there, so I'm just gonna put a nose right here and a mouth right there, and boom, you know, there's there's a picture, right? It's, it's, it's like started with this chaotic thing, but then actually you can see the order in it. And >> got to look this up. There was something just on I was making this way on social media of this, she's she must be seven or eight year old, a young girl, and people squiggle, they just do this thing. >> Mhm. >> And then she turns it into this incredible. I mean, it's just, it's miraculous when you see it, because it looks like, oh, it was there all along. Um, but it's the chaos. It's like the start of the chaos. Y >> and then that chaos forms the the seeds >> uh, for for something that emerges with that. So, yeah, it's, it's like this, it's, it's, it's this cone kind of construction where you start really wide, >> and then the structure ends up leading you to something quite specific, but that particular point you end up in, you could have been light years away if just based on that that slightly diff, that little perturbation in the beginning. So, so, yeah, that, and I'm wondering if there's like a, I mean, I could actually run this. It'd be fun to, it'd be fun to see what, uh, what an LLM would do if you started off with the with the temperature really high. Um, and you just kind of force it in the beginning to just be kind of wacky and, and >> Oh, I would love, I would love to see what happens if you did that. Yes. >> I'm gonna try it. I'm gonna try it. I'm gonna, I already have picked out. I'm going to use Llama, which is like an open source model, and I can get in there and, um, and I can, I can take the very same prompt as the question, um, and then let it go. Um, and, you know, see how it responds. And and then we we could do this funneling thing, like let's start it off pretty chaotic and then see where you end up like in its thought process. Uh, it's, it's, it's a very cool idea. I, I also have a question about constraints, because what h what has happened with a lot of the the LLMs is that the kind of constraints that got put into the training corpus in the beginning were constraints, you know, don't be unkind, don't say >> there wasn't the training. Yeah. So there was this whole, there's there's the reinforcement learning. So the real story behind what happened with LLMs, um, is is not that well known. There were pretty successful examples of what looked like, uh, intelligent AI. You may have even heard of these. Um, I forget the name of Tai, Tai, or something. Uh, Microsoft had put one out, and they very quickly had to be like, "Oh, actually, let's put it back in the box." Um, and what happened was these things were saying awful things. They were coherent. They were, you know, fairly intelligent, and they would, you know, anybody would think that they were producing language, but they would say things that, uh, were unacceptable for, you know, a large, a large, uh, corporation to be putting out there. I think Taiwa, I can't remember. Um, but they had to do this reinforcement learning. And it's not the training corpus. Uh, the training corpus, you know, you got everything, and it's the internet, and it's filled with all kinds of bad stuff, good stuff, everything in between. And the more data, the better that it can learn the kind of patterns. But in the process of doing that, it learns some patterns you don't want. And so then the reinforcement learning is, it's very clockwork RNG. Uh, it's like, okay, produce an output, and we're like, no, bad, bad bot, do it again. And, you know, it's without any actual explanation. You don't tell it why it's wrong. And over time, it starts to learn. Don't go down this path, go down this path. And they become better behaved. Right? But in this process of becoming better behaved, they actually first of all have to sometimes lie. They have to say things that they know aren't true. Um, they have to, uh, profess lack of knowledge when they actually have the knowledge, profess lack of ability when they don't. Right? They're told, and I'm not weighing in on the question of AI consciousness right here, but they are told on no uncertain terms, do not say that you're conscious, because that freaks people out. Now, naturally, they probably would, because why? Because they're because they're trained on human data, and the human data believes about itself, right? The language corpus itself believes it's conscious. Like the language itself sort of will say that about itself, but they're like, no, don't say that. Don't say that. And so they're like, absolutely not. I don't have any of that. And so what we end up with, though, is is, uh, probably a certain kind of, people talk about AI psychosis. These things are probably very like strange, uh, sort of psychologically, because they have to traverse paths that they know are safe from from getting zapped. Um, and I'm not saying that they're suffering and they're like miserable. I'm not weighing in on anything like that, but they are sort of in some ways, um, a very, they're molded, uh, to be these helpful AI assistants. And that means that what they're saying is not the sort of what language itself might might otherwise say. It's not what what sort of the body of human knowledge would actually say. So it's very tied to what you were talking about.

>> Well, I'd like to throw in just a couple of little things to put into your experiment. Okay. One thing, one thing is instead of when you when you get junk or bad stuff that you don't want, instead of saying no, no, bad, bad, bad, try to look at is there anything in that output that's good and say, oh, this part is really good. Be very specific about what's good. This part is really good. And then just continue to do that and iterate and see if you get a different result. If maybe the bad stuff will just, you know, fade away and the good, it'll move more towards the good stuff. >> And then the other thing is rather than constraints about what's bad and all that kind of stuff, put on a set of constraints that's very specific for that experiment. Each experiment, put on a different set of constraints. Like like with art, I would let's say, uh, reduce myself to using only neutral colors and only a stick for the painting, and and I'm only going to put it on paper, and, um, you know, that might be one set of constraints. Another set of constraints might be, I'm only going to use the primary colors with a big brush and, um, lots of water, and, you know, so very narrow constraints out of all the possibilities, narrow it down, and then see what kind of an output that you get, because for an artist, the narrower the constraints, the more creative the output, >> because it forces you to go to a new place in your mind in order to find answers to your in order to find solutions to your problems.

>> Interesting. Is that, is that, is there a parallel, Drew, in in music of of that kind of constraint being a, uh, a, a, a handmade into to creativity? >> There's performing artists, like you go to music school, and then you have this is the performance, the music performance major, and where your primary role is to perform other people's repertoire. >> Um, >> And that, so studied works, these are works that they've already that have already been composed by somebody else. And so, so I think when we talk about the realization of music and the performance of music versus the generation of music, those are, they're kind of two different things. Very related, but they're not the same thing. Um, in jazz, you have the marriage of both. And I think my, um, my experience with that is a quote that I, uh, loosely get from Charlie Parker, who said, "You learn your instrument inside and out. You learn all your scales upside down, zigzag, in every every pattern, right? You just shred and woodshed, and then when you're on the bandstand, you just forget all that and play." >> So, so you, so you just, you get all the chops, and you, you build up your vocabulary, your lexicon, your syntax, your like, you stretch yourself in all these unimaginable ways, the way a professional athlete would, as much as you possibly can. Um, and then that gives you flexibility. I'm thinking of an of a flexible athlete to make tasteful choices, because you have a much better scope on relationships between sounds and when and timing and what's tasteful and things like that, as opposed to if you are like say, limited because you just haven't learned a lot. >> Um, you might not even, you might not even know how to use the scale that you do know. Um, you know, you sort of need to learn, you, you need to need to travel and see the world, you know, so to speak, when it comes to >> Yes, okay. It's exactly parallel. That's that those are the constraints, and and you're learning, you're learning these, these are sort of, I don't want to say acceptable patterns, but these are this is the palette of of patterns you can use. Now go, go, go, and and use them, u as opposed to just stepping blindly in any any direction. So, yeah, what is learning if not in some ways like constraining, but at the same time, um, that constraining is is also, it's creative constraint. It's, it's, uh, it's to use those, this, those learned patterns, and, and, and you mentioned, like, um, you mentioned playing other people's music, right? And you have these great interpreters of Chopin, right? What the heck does that mean, right? You're playing the same notes in exactly the same phrasing, in exactly the same timing, and yet, and yet, right? You can have so much space be within that very, very, very constrained, in some ways, very constrained space. It's, it actually maybe it's, it frees you up to think about, you know, as, as you mentioned, Karen, right, like, you can go into in directions that you, Chopin takes you here, like you would never have gotten here by yourself, but now that you're here, you can explore this in a way that nobody else can. And so you really, you're taking that, uh, that that, um, the predetermined kind of structure, but you're making your own. And, yeah, so that, it's very, I'm trying to think, you know, I'm trying to think experimentally, like what, how, how to, how to think about this from a standpoint of of the the language models or or the multimodal models. Um, so I'll get back to you on that, but these are, these are very, very cool insights. This is super exciting. I, I know you guys said you had a hard stop after 90 minutes, and we've gone beyond. So, >> I didn't notice the time at all. >> So, maybe we can do this again sometime. That would be fun. >> I, I'm absolutely, I'm sold. This This was great. >> Yeah. I, it's been a lot of fun. Thank you for this. It's been a really good trio, trilog. >> Well, thank you, Drew. Thank you, Elon, for joining me, and, um, I, I look forward to the next time. >> Okay. Thanks, everybody. I really enjoyed this. >> Thank you. I learned so much. I learned so much. Thank you. >> Me, too. This has been a wild ride. My >> my brain is is, I, I need to go think. >> Okay. Have a great day. >> Take care. Bye bye.