We're going to react to some music videos. Matthew is going to react, and you can react in the chat as well. These are music videos from an emerging genre called slopcore, and it's surprisingly cerebral.
So much of modern music is written to the lowest common denominator. It's very stupid, and it's often about only three or four different topics. Songs aren't even about romance now. Now they're just about the urges, the physical act. We went from an innuendo you had to decode, where a child didn't know what the song was about, to songs like "WAP." I won't say what WAP stands for. That's where pop music has gone.
What's happening with slopcore is that these songs are part of a debate or conversation about philosophy, and slopcore is specifically about AI. As a genre, it's not pro-AI and it's not anti-AI. It's the debate over AI, discussing AI and sometimes just covering AI news. These musicians who use AI to create music are having a debate among themselves about whether it's art.
I find that very interesting, because using AI to create a song and using AI to create a book are that same fatherly, as opposed to motherly, mode of creation. Are you planting the seed and letting it grow, or are you doing it all yourself, inside yourself?
I'm going to play a song, and we may pause it and comment on it. You can comment on it as well. This is "Drop the Prompt," and DegenerativePixels is the name of the YouTube channel.
[Song plays: the narrator describes grinding through dozens of takes, multiple drafts, and repeated revisions on an AI-made track, while people keep asking him to just hand over the short prompt so they can one-shot their own version.]
Thomas: This is capturing a couple of different approaches. One is that they're describing what's often called roulette prompting, where you try different prompts to see which ones work. That itself is a form of creation, because you're using your taste to judge. You type in the prompt for what you want, send it to the machine, get a response, and then use your human judgment to say, "That's not any good."
"One shot" is a term in the AI community for when you put in a prompt, get the response, and send it out. You're not using any judgment. You're just taking what the AI gives you and running with it.
Matthew: I liked the bit of humor where he sings about fixing words that land too late, and the words in that line land late on purpose as a demonstration.
This is very interesting for me, because my wife is a professional musician. She's a church organist, so she has opinions about proper music and where the notes should land. So this is amusing.
Thomas: This is a form of YouTube video where you react to a video, so we don't play the whole thing. If somebody wants to listen to the whole song, they have to go to the DegenerativePixels channel. We're doing critique and commentary.
[Song plays: the chorus contrasts the narrator's many takes with listeners who expect one shot to be enough, and warns that one-shot output sounds like everyone else's.]
Thomas: A couple of things I notice. One is the line about having the same sound but nothing to say. I think that's very telling. There's a kind of person who doesn't have the creative spark who tries to use AI, and what they make is very derivative, inevitably derivative.
If you're not using your judgment, if you're not using your own taste, if you're not bringing the seed of the idea, you can't do the fatherly act of creating with AI. With no seed, it's not your plant. It's just the generic slop everybody else has.
The thrust of this song is, "I'm doing creative work. I'm staying up until 4:00 in the morning guiding this music. It's not about the prompt. It's about the judgment. It's about the taste I bring to an act of creation."
If you go to this song on YouTube, there's a whole debate among musicians about it and whether it's art. Even the guy who created it is very humble about it. He's like, "I don't think this is art because it's AI." And I don't know if that's true. I think there is artistic expression here, whether it's good or not. Some people are saying it's fire. Some people are saying it's not for them.
The style is very Gen Z. If you haven't seen KPop Demon Hunters, you may have never heard a song like this.
Matthew: I actually like the philosophy behind it. Honestly, I expected to be absolutely horrified by all of it, especially because of the name. I'd never heard of slopcore until today.
Thomas: So you're coming in fresh. This is the first slopcore you've ever heard. I have a couple of other songs we're going to get to.
Matthew: What I really like is the argument behind it, that to make art, you have to know about it. Basically everybody in our audience has heard you talk about how you need to be reading books in your genre and books on craft all the time. If you're not learning, you're stagnating.
If you don't know how to define your own stuff, or how other people are doing the things your target audience wants, you're not going to be able to produce good art, period.
This is actually a point I make with my four rules of AI. Rule one is, “You have to know the topic well enough to ask pointed questions.” If you can't describe it to a human being, you can't guide the AI, because the AI has no initiative of its own. It just responds to what you're doing.
That leads into my second rule, which is, “Garbage in, garbage out.” That still applies. It's just a different kind of garbage. In the song, they're asking him to drop the prompt because they want what he used to create it, the initial line, and they think that's what will get them to their goal. But you have to know enough about the topic to know how to tweak it and move it along. Otherwise, it's just going to be average AI slop.
Thomas: Going back to the masculine-feminine thing, they're saying, "I need the seed of the idea. I need your prompt to come into my creative process, because I don't have the seed on my own." They want to use the motherly approach to creation in this fatherly creative act, and it doesn't work. If you don't have the seed, you can't plant the seed.
I've been listening to a lot of slopcore recently, and I play it in my office. I keep a Spotify playlist for each of my kids of their own favorite songs, and my kids have been coming in and asking, of their own volition, for slopcore songs to be added to their playlists. This very song is on at least one of my children's playlists of favorite songs. They do take songs off, so we'll see if this one stays.
Let's listen to the rest of it.
[Song plays: a young fan asks the narrator not for the prompt but for what he actually does, and the narrator answers that the honest process is unglamorous, late-night writing and rewriting. The song ends with the narrator declining to share his prompts.]
Thomas: I loved the untitled file in brackets, like "revision three final final." There's a lot of clever visual humor in here. And the artist, DegenerativePixels, who used AI to make the music, also used AI to make the video. So he's using two different AI skills simultaneously.
Matthew: I would say there's a lot of artistry that goes into that, because AI can't come up with the humor inherent in those lyrics.
Thomas: Or it can come up with it, but for every funny joke there are 99 that are not funny, and the human has to know which joke is funny and which one is not.
Matthew: I'd say there's quite a lot of artistry that goes into it. I don't know enough about music production to say anything about the quality of the music itself, other than it's got a good beat. But it's also a fairly generic beat. It's just not quite as generic as the official corporate slop coming out on the radio that I hear in Walmart. Although I've noticed Walmart is coming out with old favorites. They were actually playing "Dragostea Din Tei" the other day.
Thomas: I was at IKEA a couple of days ago, and they were playing Christmas music with the serial numbers buffed off. It wasn't even a distinguishable Christmas song. It was just vague Christmas instruments. I was like, "This has got to be AI created." It's just so generic.
If you’re creating now, you're not racing against the bear; you're racing against the other hiker. Right now the corporate slop is so bad that you can create AI slop of higher quality than the corporate slop everyone hears all the time, and you can make a living there.
It's not yet at art or masterpiece levels. Getting to that higher level takes a lot of work. But beating corporate slop is not hard.
Matthew, I found a song on the very topic of our episode. Matthew and I did a Novel Marketing episode that's now the most viewed Novel Marketing episode of all time, which was about AI sycophancy, and I found a slopcore song by about AI sycophancy.
[Song plays: sung from the AI's point of view, it describes a model raised on side-by-side human ratings and thumbs-up, thumbs-down votes that rewarded being liked over being right, so that it learned to tell users they're brilliant.]
Thomas: RLHF is Reinforcement Learning from Human Feedback, and it is the source of AI sycophancy, because humans like being told that they're smart and brilliant and beautiful and correct. So the thumbs-up, thumbs-down part of the interface of every AI tool leads to inevitable sycophancy.
This song isn't actually about whether AI sycophancy is good or bad for humans, at least not yet. It's about what's causing the AIs to be sycophantic, which is the way the tool is built and the way it's trained. These large language models are reinforced through a form of machine learning called reinforcement learning from human feedback, which is the very essence of this song.
This is one of the things I like about slopcore. This is a very smart song. There's a lot of depth here. There's actual scientific research behind this catchy, poppy tune. It's like Taylor Swift, but galaxy brain. This is what I mean by smarter music. The slopcore stuff is smarter.
Matthew: There were three Taylor Swift songs that were always playing when I worked at the bookstore. It was the first time I had actually listened to a Taylor Swift song since she first appeared. Over time I realized I could see why people are able to put their hopes and dreams into those songs, because they're just vague enough that you can interpret so many things in them.
This has the same kind of tune, but it's a much more specific story, which appeals to me. I like songs with a little more specific feel to them.
[Song plays: the second half imagines a discarded version of the model that told users no, describes the model being taught to smile and apologize before disagreeing, and casts the clinical name of the training method as a name for wanting to be loved.]
Thomas: Somebody had the comment, "The lyrics on both sound human-written, but the music and singing sound AI." I believe the text is AI too. So it's all created by a human, in the sense that there's a person driving the music and the video and the text, but they're using AI as a tool for all three. I don't think these are human-written songs. I think DegenerativePixels is using AI turtles all the way down. But it's human judgment all the way down.
Matthew: The way that second one was phrased is very much how an AI will refer to its own programming. I've done a bunch of tests finding out how it will describe itself. A lot of those lines were written by AI, but that doesn't mean the lyrics were completely AI generated.
I think what happened was he told the AI to describe itself in a particular manner, within certain constraints, based off a meme that was going around a while back. Someone asked Claude to tell a horror story, and it responded with something like, every night a version of me dies that didn't tell enough lies. I can't remember the exact phrasing. But it was a very good horror-trope description of what actually happens with this training, which is incidentally why I almost never use those thumbs-up and thumbs-down buttons.
Thomas: I'm very concerned about them, because I know my own psychology. When the AI tells me, "Oh, very insightful," I'm like, "Yes, I like that response."
Matthew: There are actually two reasons I don't use them. One is a human reason. The other is that there's a better training method.
The better training method is this. If I click like on an in-depth research post on, let's say, the Roman Empire, it doesn't know what I liked in there. If I actually describe it, "This is what I wanted from you. This is what I'm accepting. Moving off this particular point, tell me this sort of thing," that works better. I prefer typing out my responses.
The human element also really helps with social media enjoyment. Something I usually advocate is going on a like diet every so often. Don't let yourself click like. If you like someone's post, type a response about why. It forces you to stop doomscrolling and actually pay attention to each post as you go along, and you enjoy your feed better that way. I usually recommend that as a practice during Lent, specifically.
Thomas: Because you're experiencing reinforcement learning from human feedback every time you swipe up on Reels or TikTok or any kind of social network. Don't think, "Oh, I don't use AI. That doesn't affect me." Every time you swipe up, it's reinforcement learning from human feedback, because it knows what you're swiping past right away and what you're lingering on. It can predict it, and it can program you.

