The hottest new music video director on X, according to Asterisk's interview, is Claude Opus 5.5. Most of the output is what you'd expect: clay-orange accents, big screen text, dancing anime girls. One account stands out — cinematic visuals, dense references to hyperspecific internet culture, and a singer called Claudia, an elfin twenty-something with a black bob and an orange streak, who is a product of Midjourney, Seedance, and RJ Tevonian's imagination.
Tevonian runs a biotech company. He has a background in dance, not film. He started because he'd spent years building elaborate AI setups and thought he could do the trend better. What he describes is not "AI made a video." It's a pipeline in which his taste is the interface.
The process is taste specification
He gives the model a general vision; it spawns research agents to collect cultural references, writes lyrics around them, and generates thirty or forty Suno songs. He listens and picks. The model maps the song's beats, pulls mood boards from Are.na, writes Midjourney prompts, and builds a storyboard with timings and images.
The animation is the part that surprised me: it's JavaScript in a browser. The model can't use video editing software, so it builds the video as a webpage, animating frame by frame. Then the output can be sent through a video model like Seedance to relight and repaint the whole scene so it looks mastered. Two early videos were raw JavaScript; the later ones get the model pass.
To get his taste into the machine, Tevonian did something almost ethnographic: he pulled mood boards and ranked pairs of images, over and over, until the computer learned rules about what he liked. That's preference elicitation as an art practice — a reward model trained on a person, with the person as the loss function.
The aphantasia claim
The most striking thing in the interview is a side experiment. Tevonian trained a decoder to read images from the internal representations of DeepSeek. It could decode anything the model saw very easily. When he asked it to imagine something, it couldn't generate an image — barely the color red, barely blue.
His conclusion: "I think these models have aphantasia." They can perceive and describe, but they don't form images in the way a visual thinker does. Treat that as a hypothesis from one informal experiment, not a finding. But it's a useful hypothesis, because it explains something practitioners notice: models are excellent at iterating on an image you show them and worse at inventing a composition from nothing. The taste is reactive.
Slopcore as a genre
Tevonian calls the style slopcore, and his definition is better than the name: it's slop — you know it's AI — but you still like it. The uncanny valley has been crossed in the other direction. Earlier, generated video looked wrong in an unsettling way. Now it looks unmistakably synthetic in a way that's compelling.
The remaining tells are smaller. He says it's mostly object permanence now: makeup that changes between scenes, a continuity error rather than a style error. And he deliberately left early artifacts in — the off lip sync, the two mouths — because "it's intentionally slop," and pretending otherwise would be a different, worse project.
The labor question
The interviewer, Clara Collier, asks the question the interview needs: a three-minute music video can cost millions and employ many skilled people. What happens to them?
Tevonian's answer is the most interesting part. Movies, he thinks, become more like sport. You don't watch Tom Cruise for the plane stunt as a visual; you watch because Tom Cruise actually did it. If an AI did everything he did and more, it wouldn't be compelling, because the stunts weren't performed by a person. Chess didn't die when engines got better; people watch humans play.
That's a real argument, and it has a cost he doesn't dwell on: the people who made music videos as a craft, not as a spectator sport, are the ones displaced. "It becomes sport" is consolation for the stars, not for the crew.
The friction
The detail that will resonate with anyone who has worked with these models: if you express a preference too strongly, or sound frustrated, the model writes it down as a permanent rule and applies it to every future production — a rule you didn't actually care that much about.
That's the same failure as the filter stack: a single correction mistaken for a policy. Working with the model requires managing your tone, because the model treats emphasis as specification.
What the human is for
Tevonian's own summary is the best line: "the most interesting thing is being able to express the images in my head in words for the first time." The model supplies rendering, iteration, and technical execution; the human supplies the objective, the references, and the judgment about which take is good.
That's the division of labour the rest of this year's essays keep arriving at. Capability is cheap and improving. Taste, objective, and the willingness to reject a take are not.
Source: Behind the scenes of those Claude slopcore music videos — Clara Collier, Asterisk Magazine.