Much Ado About AI: An Artist, Musician, Writer, and Engineer Walk into a Debate
Kaya- The Artist, Brett- The Musician, Justin- The Writer, Vaish- The Engineer
To my true-blue, salt-of-the-earth artist friends, my work and I represent everything that’s wrong with the world. I represent technology, the TikTok algorithm, the war (all of the wars), and climate change. And they’re very good at taking creative digs that make me go home and wonder if I was being praised or made fun of.
But I do talk about AI a lot, and I kind of lean into their negative reinforcement.
However, for all the “hate” I get, I think my friends just fear AI to varying degrees. Kaya doesn’t seem to care; it simply doesn’t exist in her world. Justin says he doesn’t care for it, but over the course of the conversation, you’ll see that he’s leaning in, too, in his own way. Brett is kind of on the fence and thinks he could go both ways.
But then again, Kaya doesn’t care for it because she hasn’t used it. Justin uses it for work, so he sees the power in it, and I think it scares him to a certain degree. I showed Brett some of the use cases for ChatGPT in music, and he thought whatever I showed him was interesting. And I had hopped on the AI train when AI was called NLP and we were analyzing the hidden Markov values in Google’s “Attention Is All You Need” paper.
Fun fact: When the researchers at Anthropic looked inside Claude, they found that there are specific combinations of neurons that “light up” when they interact with relevant “ideas.” For example, they found that there’s a specific combination of neurons that lights up when they show it a photo of the Golden Gate Bridge. And not only can they identify these neurons, they can “activate” or “deactivate” them. So, if you “activated” the idea of the Golden Gate Bridge, its replies to queries would mention the Golden Gate Bridge even if it’s not directly relevant to the subject. And it does it in a way that makes contextual sense. You can ask, “Hey, how’s your day going?” And it’ll say something like, “My day feels really calm and serene, much like the weather on the Golden Gate Bridge.” And they named this LLM Golden Gate Claude.
And Golden Gate Claude is why I’m writing this. I feel a certain kinship with an idea-obsessed AI... and also because it’s really, really interesting. And I want to try and show them that side of AI: the fun side, the science side.
Bull in a china shop
I showed them an animation I made using Sora called “Bull in a China Shop.” They get the joke but hate the video.
(I didn’t want to upload the animation because I prompted and re-prompted Sora to get it as close to the style of Pixar as possible, so I’m only putting out a screenshot of the animation.)
Thoughts?
Justin: There are so many plates that haven’t been stepped on.
Huh?
Vaish: Yeah, but the models will get better with time. With advanced steer-ability, we’ll have more control.
AI art is like microwaved meals
Brett: This looks well-made since it’s drawing from what it’s learned. What if you gave it a prompt where it has nothing to draw from? Then it’s just a bastardized amuse-bouche, a collage of all these animations from the studios we’re used to. I can tell it’s sort of like this movie I’ve seen. It has this uncanny valley feeling.
Justin: The Incredibles. They have that crazy detail. Finding Nemo. They made the water molecules so realistic, they actually had to downgrade it a bit.
…
Brett: I agree with Miyazaki. The artists made these (...The Incredibles, Finding Nemo...) with so much attention to detail.
Like, if I ask it to write me a song, it feels like a microwaved meal. You don’t appreciate microwaved meals as much as you do homemade cooking.
In the future, where I can just ask it to write me a song, choreograph a dance, and make a TikTok video, where I can even ask it to handle the marketing, everyone can be a pop star instantly. And it’s going to create something viral. But once people realize you can synthesize that entire process on your own, the only thing that will matter is whether you can dance by yourself, without a camera.
Kaya: I think it’s the effort you put in and the meaning you give it. These are just pixels moving on a screen. It isn’t art. scoffs
AI should be president
Justin: I agree. It’s like how microwaved meals will never replace your mother’s home-cooked meal. It’s the same with art; you can never replace the original artist.
Think about this: they say socialism would never work because of the human element, right? So what if we made AI our president? What if we put it in charge of other things, something more important than art? If we made AI our president, it wouldn’t be swayed by greed or ego.
You could even put safeguards in place, select different scientists and operators to ensure there’s no bias programmed into it. You could set up a kind of constitution for it, like: AI cannot be used for the destruction of property or to harm another human being. Whatever.
So if we can see the progression of humanity heading in that direction, if we just ask it the right questions, like, “Here’s the homeless situation, here’s the GDP,” and feed it all the data, it would probably come up with a pretty good answer on how to solve that.
Does AI have a hard time generating lived experiences?
Which essay was AI generated?
Kaya: The first one used describing words like “yummy,” “crunchy,” etc., almost too descriptive. And it didn’t use anything personal, like my mother, we eat, we chat, we laugh. The second one isn’t personal at all either; it doesn’t talk about anyone specific. It just sounds like something written by a child: drink, fruit, treat, my friends, and my mom. Yeah, the kid one just talks about things a kid would talk about. But the second one... doesn’t. Sounds too old. And nothing personal about it. It seems to me like it has a hard time generating lived experiences.
Justin: So, does AI have a harder time putting itself in a personal environment? Because it probably hasn’t lived very much.
Brett: What child of that age group is going to say “as we sit and chat…” and “catch up on each other’s day”? No child speaks like that. I work with children all day. “As we eat and laugh and catch up…” Do you, Jimothy? Do you? No, you don’t.
Justin: “Also... I find an empty table and I sit there...” No child does that. “I’m gonna sit by myself!!”
Are non-committal answers just a function of being an artist or a requirement? Asking for some friends..
The first essay was AI generated!
Does it understand how thoughts are built?
Brett: Because it has a hard time understanding how humans synthesize thoughts. Like “it” having a thought. But from its perspective, it only has what we give it: completed ideas and thoughts. It sees nothing of the process of building those thoughts.
I do this exercise with kids in my class: “What I Like About You.” I ask the kids to write something about each member of the band. It’s never “I like that Aaron is kind and shows up on time.” No, it’s “I like that Aaron is good at drums.” To write like a little child, you have to think like a little child, less about personal, emotional thoughts that connect to other people.
A child would say, “I like Aaron’s shoes.” That’s how kids think. And you have to teach them to care about and think about other kids. So how would you imitate that disinterested way of talking that a kid would have?
And definitely not “I like that he dresses well.”
It’s actually not that different from a 4-year-old child
Vaish: I what you’re saying is really interesting, and is actually similar to how some large language models (LLMs) process information: through weights and layers! Inside an LLM, we have multiple attention layers, each responsible for progressively more abstract representations of language.
Vaish: When researchers looked inside, they found that different layers of the model learn different levels of understanding. In the early layers, the model starts by recognizing individual words and their relationships. As you go deeper into the network, it begins to grasp syntax, then sentence structure, and then semantics. Eventually, in the higher layers, it starts modeling more abstract patterns, like the tone of a paragraph, the intent behind a question, or the emotional context of a conversation.
This is very similar to how children learn, right? First they recognize words, then they understand how those words fit into sentences, then into stories, and eventually they begin to grasp the broader context and nuance.
I’d highly recommend watching 3Blue1Brown’s video on transformers. It does a great job of visually explaining how attention mechanisms work and how information flows through each layer of the model.
Kaya: Hmm... that’s actually cool.
I guess I can retire now lol.
We steer off-course here
Justin: Now... if we asked a 45-year-old man to write an essay like a 12-year-old would, how well would he write it?
So again, what is it pulling from? How many second-grade authors are there?
What if we took a paragraph from Hemingway and asked it to write in the style of Hemingway, what would happen? Would Hemingway get doubly Hemingway or Hemingway squared?
So it’s not that AI isn’t able to produce sentences that are convincingly human... it’s that it’s not able to generate sentences that are convincingly written by a second grader.
Brett: This is something they teach you in school. I learned this in jazz college. When you’re studying something like “How High the Moon,” and there have been so many artists, from Dizzy Gillespie and onward, who’ve played that tune, each of them does it through the lens of themselves as an artist. It’s not about how close I can get to this artist. It’s stealing like an artist. It’s: I’m gonna take that breath that Yo-Yo Ma takes at the beginning of Bach’s Cello Suite in G. It’s actually written in regular common time, but he gives it a long, breathy rubato sweep at the bottom of it, which is his expressive way of playing a beat that’s unique to him, and is fully himself as an artist coming through in that piece of music.
In jazz, there are standards. Everyone plays them, all the time. I’ve probably played “How High the Moon” twenty different ways. And it depends; some days it’s about my mood. How fast am I going to play it? How slow? Am I going to play it with feeling? Am I going to play it straight? Am I going to swing it? The things I’m feeling on that given day might shape how I interpret that piece of music, and it’s all through the lens of self.
And as an artist, you are beholden to that. Anything I make is going to sound like Brett. No matter the jazz tune I play, people who know me well enough will say, “That sounds like Brett.”
As an artist, everything is filtered through your own personal lens.
And that’s something I don’t think AI will ever be able to replicate. It doesn’t have its own personal voice; it only has a choir of all the other voices it’s drawing from.
Justin: Hallelujah - written and performed by Leonard Cohen, Jeff Buckley... AI can only piggyback off of what’s already been done.
Brett: Wouldn’t it be kind of useful to use AI as an additional member of my band, just one member of the band? So I’m on guitar, my friend’s on bass, I’ll do vocals, we’ve got drums and some horns. But I want that second rhythm section. I want the congas and the little this-and-thats, and I want it to be controlled entirely by AI.
So I say to AI, “Write a rhythm part for this song.” What would AI do with all these human songs as prompts? And now... how will AI write a bassline for that? That could be art, because you’re using AI as a tool. Like, “Now I’m using a world of bass players to write one part of my song.” That might be interesting.
The applications are definitely there, but I wouldn’t want it to be the beginning of the creative process.
Kaya: AI doesn’t have its own feelings and memories to draw from; it only has everybody else’s.
Justin: I think AI can be an incredible tool for creativity. If I say to AI, “I need 10 quotes from George Washington,” it can give me that. Especially the rabbit holes you can go down; that’s really fun. For research, it’s incredible.
Brett: “Give me a bunch of synonyms for this word.” Or, “Give me a bunch of synonyms that also rhyme with this word.” It can maybe help you get past that slump.
Justin: Like when the thesaurus came out, did that take over all of musical writing? No, it’s just a tool, like a dictionary. If I say, “I want to write a book on the downfall of Western society. What other authors have written about it?” If I search this, if I look it up on Google, I’ll get something like, “Uhm... Steve Jobs once said...” you know? So ChatGPT is a really great way to do research.
Vaish: Hmm.. how do you guys go from an idea to a finished project?
Justin: A lot of it is hard to describe.
I just have a feeling, an idea, and I go off of that because it interests me in some way. And then, if I decide I’m going to do it, I get into the research process.
Kaya: I feel like AI is better for logistical, cut-and-dry things. But for art, for creativity, you draw from your emotions, memories, and feelings. From your mother’s cooking, the only time you’ve seen your dad cry; these are my experiences, the things that I think of. All the things that AI can’t experience; it can only try to replicate, and not yet successfully.
So, can AI get to that point? I don’t know.
I know that AI draws from everything that’s recorded or available on the internet. But my question is, don’t we, as humans, do the same thing? Can we truly have original thoughts that don’t stem from something else?
Like, even if you think it’s original, doesn’t it still come from something you already know? You can mix it up, reshape it, and believe it’s original, but is it actually original?
Justin: Jerry Seinfeld, when talking about joke writing, asks: where do these ideas come from? And he uses the metaphor of sitting next to a huge brick wall, and every once in a while, a crumpled-up piece of paper, an idea, comes flying over the wall. It lands in front of you, and you rip it open and go, “Holy shit.” Yeah, it comes out of you, but you still don’t really know where it comes from.
Some of the ideas I come up with don’t even feel like me. I guess they’re coming out of me, but where do they really come from?
I guess ideas do just want to be expressed.
- End of Part 1 -
I guess something I'm noticing is the different lens through which we view AI. At the crux of it, it's a 4GB file with billions of weights and tunable parameters. My excitement stems from how so many different bits of math come together to make something seemingly intelligent, and how we're quite literally watching technological development unfold in real time. I guess what I'm trying to say is that I understand the complexity it took to get here, but it is still only a file.
But for my friends, it’s more than the sum of its weights. It’s this living thing that fosters discussions on the philosophical aspects of intelligence.