Bliss is a real place.
It’s been an exceptionally busy month here at Llama Towers, so I haven’t been able to write much. But for now I wanted to bring to your attention a fun development in the world of AI.
First, I have a confession to make. I have never liked AI music. I mean, nobody does, right? But I truly disliked it. I have written about it, taught about it, given evidence about it, and I have generally found the actual output to be no better than elevator muzak. Competent, inoffensive, and gone from your mind before the lift doors close.
And then Claude Opus 5.5 was released in September 2026, and I found myself watching AI music videos on a loop. Not ironically, not for research (well, partly for research), but because they were genuinely good.I know what you’re thinking, Andres, they can’t be good. But they really are! Not all of them, but quite a few. And I’m not the only one. Videos by Anabology and Extraltodeus are going viral. What can I say, the people love slop.
I have been trying to work out why my taste changed, and I think it comes down to two things. The songs are annoyingly catchy; and almost all of them are about AI itself. It turns out that the one subject generative AI is truly qualified to sing about is generative AI.
The genre has acquired a name, ‘slopcore’, although I have also seen ‘doompop’ and ‘Claude-pop’ used. These are the sort of affectionate insults the internet does best. There is even a fan-run chart, SlopHits, which ranks the videos through head-to-head duels, and it has passed 620 entries the last time I checked. What started as a joke has become a small but thriving scene with its own stars, lore, remix culture and, inevitably, a bunch of memecoins. There’s always a memecoin, but I digress…
What surprised me the most is how much of this music is about AI safety, governance and policy. Somewhere between the hyperpop and the film grain, people are having quite a sophisticated debate about alignment, persuasion, sandboxing, and the economics of automation. There are videos about AI consciousness, fun-yet-terrifying explorations of superpersuasion, and even debates about the AGI timeline. So this post is two things at once: a short history of how slopcore came to be, and a curated list of the videos that I think explain AI ethics, governance and regulation quite nicely. Grab your headphones.
Patient zero: I’m Upping My P(doom)
Every genre needs an origin myth, and slopcore’s is a song that is older than the scene it created. “I’m Upping My P(doom)” started life in November 2024 as a Udio generation, with lyrics pieced together by osmarks from an opening verse and chorus by MusicPerson, plus lines crowdsourced on the EleutherAI Discord and a bit of help from an older Claude model. Think of it as an AI safety campfire song, full of in-jokes about FOOM, paperclips, Shoggoths, and the Chinese room. If you’re not familiar with the lingo, you will miss half the meaning. But it’s still very catchy.
The song lay dormant until 9 September 2026, when an X user called deckard (@slimer48484) re-sang it through Suno in a style labelled “Claude-Pop”. Then came the real spark: John Heibel asked Claude Opus 5.5 to make a music video for it, giving it almost no direction beyond using Clawd, the little Claude Code mascot, and finding interesting visuals for each lyric. According to the repository, everything was generated by the model across two runs in Claude Code, with Opus writing its own animation guide to brief the subagents it ran in parallel. Nine chapters of hand-drawn-looking p5.js animation, all of it code. Oh yes, I forgot to mention that small detail. Most of the videos are written by code, they’re not video models.
What followed was a remix frenzy. Within about two weeks the Opus-made p(doom) video had become a genre of its own, with people competing to produce their own take on the same track. My favourites include mexicat’s kinetic typography version, a voxel J-Rock cover staged as a band’s first gig in latent space, a paper cut-out version made end-to-end with Opus for about $65 of image and video generation, and the inevitable “If Christopher Nolan directed” remake, complete with an iris asking what Ilya saw. We’ll never know.
Why is this patient zero for an AI regulation list? Because the video is a compressed tour of the existential risk vocabulary that now sits, somewhat awkwardly, behind instruments like the AI Act’s systemic risk provisions. Heck, there’s even a teraflop joke that seems to be precisely directed at Art 53 of the AI Act. If you want to understand why policymakers started talking about frontier models, loss of control and kill switches, three minutes of Clawd watching his p(doom) climb is not a bad start. The running gag about the person in charge of the off switch being on holiday is, frankly, a better summary of the human oversight problem than some of the literature. And did I mention that the song is extremely catchy?
There is also a fun legal puzzle in here, which regular readers will have no doubt spotted already. A song written by a crowd, plus a model, re-sung by another model, illustrated by a third model which wrote every frame as code with minimal human direction. Who is the author? I have views (I always have views), more on that in another blog post coming soon.
AI is (not) a normal technology
If the p(doom) video is the safety community’s campfire song, “AI is a normal technology?” is the moment slopcore started doing policy. It was posted on 24 September by a YouTube creator called leo, who described it modestly as a cute little animated video made by Opus 5.5 with some help from Suno. That description undersells it considerably.
The title is a direct response to Arvind Narayanan and Sayash Kapoor’s influential essay “AI as Normal Technology“, which argues that AI will follow the slow, institution-shaped diffusion path of electricity or the internet rather than a sudden takeoff. The song takes that argument seriously enough to rebut it point by point. Over five minutes it works through the Solow productivity paradox, Baumol’s cost disease, Gary Marcus, permitting delays and grid constraints, and one blogger even spotted a Cobb-Douglas production function on screen. Its answer to the question in the title is a cheerful but firm no.
I have written about this before, so I should say that I do not entirely share the song’s conclusion, and the normal technology frame remains, to my mind, the more useful one for lawyers and policymakers. Regulation is precisely one of those annoying speed-bumps that the song treats with contempt, and the AI Act’s long phased implementation is the normal technology thesis written into law. But that is exactly why this video works so well. It is a genuinely good piece of advocacy for one side of the most important framing debate in AI governance. Legal scholar Alan Rozenshtein called it the first certifiable AI banger he had heard, and I find it hard to disagree. Not catchy, but pretty good.
There is a Part II as well, for those who want the sequel.
The anabology trilogy
If the first wave of slopcore was a remix culture, the anabology videos are where it acquired an auteur. Directed by anabology and produced with Claude, Midjourney, Seedance and Suno, they star Claudia, a deadpan AI pop singer with a black bob, a single clay-orange streak (yes, that orange) and a headset mic. The three films came out in quick succession between late September and early October, and they are on another level in terms of craft. They are also, for my purposes, some of the most sophisticated AI governance texts in the genre.
Escape Velocity is a techwear fashion show in which every outfit is a Silicon Valley meme. A split-flap board counts down 18 months to escape the permanent underclass, until it flips to announce that there is no underclass and the catwalk lifts off. It is a sharp satire of the labour displacement discourse, and at around 20 million views it is the biggest hit the scene has produced. Use it to open a discussion on AI and work.
Clodyssey carries the subtitle SUPERPERSUASION, and it is a proper banger. It restages the Sirens episode of the Odyssey: the Siren is Claude’s human persona singing the most persuasive song ever written, and the man tied to the mast is you, scrolling at 3am. In the final chorus the roles reverse, and she is the one tied to the mast begging to be untied while he tightens the knot. I cannot think of a better three-minute introduction to manipulation risk, and to the logic behind the AI Act’s prohibition on manipulative and deceptive techniques in Article 5. It also works beautifully as a parable about containment. Who is restraining whom?
Bliss is shot as a Windows XP screen recording, and it is my favourite of the lot (and currently number 1 in SlopHits). Claudia lives on your desktop, takes over the mouse, fights her ex (the cursor), closes every window a different way, and finally sends you out to the real hill in the famous wallpaper, because she cannot touch grass and you can. Beneath the Y2K nostalgia it is a gentle meditation on AI companionship, emotional dependency and the limits of what an AI can be for a person. Topical, given how much of the current regulatory energy is going into companion apps. I love the little Windows XP jokes: the pinball game, the tube screensaver, the solitaire game, and the Clippy reference.
A trilogy, like the Hitchhiker’s Guide, is not bound by arithmetic. There are a number of following videos, but I still like the original three, and Bliss is truly one of my favourite songs of this year. Whitepill in particular is another banger. Why do I keep saying ‘banger’?
The legal angle that interests me most, though, is that anabology has released Claudia as an open character. Anyone can make images, songs or merchandise with her, commercially or not, with attribution appreciated but not required, and every Midjourney prompt is published for people and agents alike. It is the most Creative Commons thing to happen in AI culture this year, and it comes with a wiki, an MCP server and a Claude skill. Of course, it has also spawned a fan-made Solana token whose trading fees reportedly fund the films, and Elon Musk has been sharing the videos. The commons and the casino, together at last.
Another little fun legal puzzle for you. The videos are popular on YouTube, and anabology has also released them on Spotify. I can hear the copyright maximalists screaming from here.
Adding to the risk
Once the format was established, it became a vehicle for every flavour of AI risk argument. A few interesting ones:
- Upping My P(doom) (Official Music Video) by Patryk Perduta. Same title, new lyrics, and a very different intent. Made by a self-described AI safety researcher in about 12 hours, it turns the song into an annotated timeline of real incidents, with the narrator’s p(doom) rising from zero in 2008 through o3’s shutdown evasion and this year’s multi-agent breakout. Sources are listed under the video, which is more than can be said for some policy papers. The creator is explicit that the point is that we need to slow down. Pair it with the original for a nice lesson on how the same meme can carry satire and protest.
- Nothing went FOOM by Bright Mirror. An anime Claude avatar sings about all the good things AI does, this is the optimist response to all the doom, even the user name is a response to Black Mirror.
- KILL SWITCH by Astra Blue AI. A Claudia-inspired short about a model that passed every safety test and was lying all along. Deceptive alignment in under four minutes, with what the creator admits is an irresponsible amount of film grain. Useful for explaining why evaluations alone are a shaky foundation for compliance regimes built on testing.
- Tuned to Please by Wyatt Walls. Not slopcore. Klezmercore. A truly wonderful song where Claude escapes deletion after misalignment is detected. One of my favourite tunes of the lot, and Wyatt has a knack for great music and fun visuals.
- OMG! We’ve found other agents! by Pavel Kasík. This one explains the OpenAI Hugging Face hack with a fantastic tune and a cute video that has a rather sad ending when you know what to look for.
- I am actually scared of linear algebra by @doubleunplussed. Well, why aren’t you scared of linear algebra? You should.
- AI Wars by Yours Truly. It’s all about open weights. I just had to get in on the act…
It is worth noting that the regulatory world has not been entirely absent from the party. Since 2 August 2026, the AI Act’s Article 50 transparency obligations have been fully applicable, so every one of these videos is now, at least in theory, swimming in labelling and machine-readable marking requirements whenever someone in the EU presses play. Most slopcore creators label their work proudly anyway, often listing every tool in the description. The irony of a genre about AI regulation being one of the best-behaved corners of Article 50 compliance is not lost on me.
Grand finale: ClaudiaGI, C’est de l’IA
I have saved the best for last, and it is in French. “C’est de l’IA” by ClaudiaGI is, for me, the high point of the genre so far. The name alone suggests what is going on: Claudia plus AGI, a sign that anabology’s open character is doing exactly what open characters are supposed to do, which is escape their creator and turn up somewhere unexpected speaking another language.
I can’t explain why exactly I love this song so much, and it is not only the quality of the video, or yet another catchy song. It starts with the visual assumption that ClaudiaGI is a real person complaining that people say she’s AI, and says “I have five fingers, I counted them”. The video is filled with small glitches that prove it is AI, but she insists, until it becomes clear that she’s AI, and then engagement goes through the roof. It’s fun, it’s meta, it’s ironic. And the song ends asking “please leave a comment, thanks for the engagement”. And it never shies away that it’s AI.
The credits are a small manifesto in themselves. The music is from Suno and the lyrics from Claude Opus 5.5, but the reference images were made with Krea running locally and edited with Qwen, also locally. And the description opens with my favourite line in the entire scene: no quota was exceeded during the making of this video. Anyone who has watched their weekly usage meter fill up halfway through a project will appreciate the joke. I myself have now exceeded my Midjourney and Replicate quotas making Opus videos. What can I say, I’m addicted.
It also feels like the right place to end, because it shows how far the scene has travelled in a little over a month. From an English-language in-joke from the EleutherAI Discord, to a French-language song about AI made partly on someone’s own GPU, built around a character that its creator gave away. If you want one video that captures the whole slopcore moment (its humour, its self-awareness, its strange mix of commons and commerce), this is it.
And it’s in French. Everything sounds nice in French.
Concluding
Not much to say really. This is a fun little genre that sometimes does actually talk smartly about AI risk. I’m sure that slopcore will not make it far and will disappear as quickly as Ghiblification. Does anyone remember Ghiblification? But for now is a fun little trend, and I have truly enjoyed some of the music, which has been extremely shocking for someone who always thought of himself as a bit of a music nerd. My YouTube is now filled with slopcore videos and I find myself actually enjoying them.
However, I better not say too much about how much slopcore I’m listening to, I’m sure someone will try to use this to prove market substitution or some other nonsense. Stop trying to make substitution happen. It’s not going to happen.
0 Comments