the knowledge that never became data
What AI training absorbs, what it cannot — and what Angine de Poitrine signals about where creativity is heading.
Two people in oversized papier-mâché masks. Polka dots. A double-neck guitar with twice the frets of a normal instrument — modified to play notes that don’t exist in Western music. A drummer. A loop pedal controlled by feet. A language they invented, which they speak exclusively on stage and on camera.
They call themselves Angine de Poitrine — a French medical term for chest pain. They describe what they do as “mantra-rock Dada Pythago-Cubist orchestra.” They have 20 million Spotify streams and a sold-out North American tour.
Something is happening here.
A musician analysing the video made an observation I haven’t stopped thinking about. AI music systems are trained on mainstream music. Modern music is unbelievably formulaic. The AI is very good at it. What the AI cannot do is follow Angine de Poitrine into microtonal scales, Dadaist aesthetics, and time signatures that shift between four, seven, and fourteen beats in a single track.
His point was that human creativity will survive AI. Mine is slightly different.
The distinction is not between what AI can generate and what it cannot. It is between what training data can absorb and what it cannot. And what training data absorbs is everything that was ever made legible — every recording, every transcription, every documented output. What it cannot absorb is the formation that produced those outputs.
The Pythagorean ear developed through years of immersion in a specific tradition. The Dadaist sensibility acquired through reading Tzara until the logic becomes reflexive. The microtonal instinct that lives not in theory but in the hands. These things never existed as data points because the people who hold them never needed to extract them. They were how they worked.
The AI learns the surface of culture. The meaning lives in the formation underneath it.
This year, Coinbase aired a film at the Academy Awards. Director Oscar Hudson built an entire video game world in-camera — no AI, no CGI. Suits with fabric printed on them to suggest pixelation. Sets designed to look digital, built from physical materials. Extras trained to move like NPCs. Every frame captured in a studio is a deliberate act of craft at a moment when all of it could have been generated in hours. The film drove a 43% uplift in user acquisition. The audience comments weren’t about crypto. They were about what it felt like to watch something made.
A few months earlier, Apple rebranded its streaming service with a new opening mnemonic. The logo is a real object: solid glass, sculpted by hand in a London studio, lit with colour gels and macro lenses over weeks of production. Apple could have generated this in minutes. Instead, they commissioned craftspeople to build it and then made a short film showing the process — a signal to the people watching their films about what would play before them.
Neither of these is a statement about technology policy. Both are recognitions of the same thing: that in an environment where the generated and the crafted are visually indistinguishable, formation is what gives the work its quality. And that quality is felt even when it cannot be named.
Further in means two directions simultaneously.
One is upward — into the extraordinarily formed. The kind of work that requires years to produce: deep knowledge of tradition, mastery of technique, the synthesis of domains that have never been put in conversation. Formation that produces outputs the training data cannot account for, because the formation was never in the training data. Angine de Poitrine is not strange in itself. They are formed in territories most musicians have never visited.
The other direction is downward — into the rough, the imperfect, the unrepeatable. The mark of the human hand. AI training optimises for the smooth and the correct — the averaged, the unflawed. Oscar Hudson’s NPC world has the specific wrongness of real objects. The Apple logo has the light behaviour of actual glass. The frontier is also in what that optimisation removes.
Both share the same quality: the meaning lives in a formation that training data cannot contain. Not because these things are obscure, but because they were never made legible in the first place.
This is not a reassurance about human creativity in general.
Most people will be fine with the mainstream blob. The generated will become more generated. The undistinguished majority of content, music, imagery, and cultural production will become indistinguishable from machine output — because it was always operating on the same logic. The average will converge. The loop will close.
What I take from Angine de Poitrine — and from Oscar Hudson building his pixelated world by hand, and from the glass Apple logo sitting in a London studio — is a sign of what will matter.
In a world without the AI mainstream to make the distinction visible, the band might have remained a local experiment. Two people in Montreal were making strange music for a few thousand who knew where to look. The context is what gave them reach. Not because audiences were suddenly hungry for something different. Because when the generated fills the surface, the formed becomes visible for the first time.
The frontier was always there. The training data just cleared the way to see it.
The work behind the world examines the invisible systems that organise behaviour, meaning, and decision-making. The paid edition of this piece goes further — into what this means structurally for organisations whose work depends on being in the territory that formation built.
