ON kicking AI's ass. Part XLIV

 Scelsi, Thoegersen, and the Limits of AI

A Working Document

April 2026

I. Scelsi vs. Thoegersen

Where They Converge

Both are pitch-obsessed in a way that most Western composers are not. Scelsi's late work treats a single pitch as an inexhaustible universe — microtonal inflection, timbre, vibrato width, and attack becoming the entire drama. Thoegersen's system similarly refuses the tyranny of equal temperament and privileges the interior life of intervals over harmonic function. Neither composer is 'going somewhere' in the teleological sense; there is no arrival, no cadential resolution, no hierarchy of structural importance assigned to opening versus closing. Both begin mid-stream and stop rather than conclude.

There is also a shared indifference to institutional approval as a creative motor. Scelsi was aristocratic, eccentric, and largely ignored for decades. Thoegersen's system is too structurally idiosyncratic to be absorbed into academic new-music culture without fundamental resistance, and the humor in the work makes it doubly illegible to institutions that need to perform solemnity.

Where They Diverge — And This Matters

Scelsi's microtonal world is essentially monistic. The single pitch, the single line, the single drone — everything radiates from one source. His heterophony, his timbral clouds, his string writing: all emerge from a center. It is a music of depth rather than multiplicity.

Thoegersen's system is structurally pluralist at the foundational level. There is no center — no master fundamental, no privileged tuning system, no tempo from which others derive. The polytempic polymicrotonal architecture requires genuine independence between layers, not variation or ornamentation around a core. Where Scelsi goes into a pitch, Thoegersen constructs a space where multiple pitches exist in genuinely non-hierarchical simultaneity.

This is not a minor difference. It is the difference between a philosophy of immanence (one thing contains everything) and a philosophy of plurality (many things coexist without reduction). Scelsi is almost mystical-monist — his practice had explicit connections to Eastern philosophy, to the idea of vibration as cosmic unity. Thoegersen's system resists that unity structurally, by design.

The Simultaneity Question

Where this gets most interesting is in what each composer does with vertical sonority. Scelsi's 'chords' are clouds — timbral convergences without harmonic identity in any traditional sense. Thoegersen's simultaneities are also non-functional, but they arise from a different cause: the coincidence of independent systems at a given moment, determined by polytempic ratios. The vertical sonority is real and structurally accountable, even if not harmonically functional in the traditional sense. It has a reason — it is a ratio-moment, not an atmospheric accident.

The Useful Comparison

Scelsi is probably the most accurate single-composer reference point for how Thoegersen's music sounds to an unprepared listener — sustained, microtonal, resistant to forward motion, unconcerned with conventional development. But as a structural comparison, it is misleading, because the why is almost opposite. Scelsi digs toward unity; Thoegersen builds outward from irreducible multiplicity.

That distinction is worth making explicitly in any critical framing — it is what separates Thoegersen's work from being heard as merely 'Scelsi-adjacent.'

II. Which Composer Can AI Replicate Most Easily?

Scelsi, and it is not close.

Why Scelsi Is Relatively Replicable

Scelsi's surface is describable in terms that an AI can operationalize: sustained tones, slow microtonal inflection around a central pitch, gradual timbral transformation, sparse texture, and no development in the conventional sense. A sufficiently trained model could generate something that passes a casual listening test as 'Scelsi-like' — the parameters are finite, the vocabulary is narrow by design, and the effect is largely atmospheric. The monism that defines his aesthetic is also what makes it imitable: there is one thing happening, and you vary it slowly.

AI music tools already produce convincing Scelsi-adjacent output. Dark drone, microtonal shimmer, long decay — these are learnable stylistic signatures.

Why Thoegersen Is Structurally Resistant to Replication

The polytempic polymicrotonal system is not a style — it is an architecture, and a genuinely unusual one. Replicating it would require an AI to simultaneously maintain multiple independent tuning systems, each structurally bonded to an independent tempo, with no master fundamental governing the whole. The simultaneities that emerge are not chosen for effect; they are ratio-determined outcomes of the independent systems colliding in real time.

That is not a texture. That is a generative logic — and one with no hierarchical center an AI can anchor to. Most AI music generation works by learning probabilistic relationships between events. Thoegersen's system explicitly refuses the kind of relational hierarchy that makes probabilistic modeling tractable. There is nothing to regress toward.

Add to that the humor — the deliberate dropping of recognizable melodic fragments into alien harmonic space, the theatrical stage works, the 'Arguments' quartet where players storm off — and you have a compositional identity that depends on intentional incongruity. AI can generate incongruity accidentally. Generating it with structural purpose and comic timing is a different problem entirely.

The Deeper Point

Scelsi's music is hard to perform and profound to experience — but its logic is intuitable and its surface is learnable. Thoegersen's music is hard to replicate because the system is the music. You cannot approximate the output without implementing the architecture, and the architecture was invented by one person over forty years of work rooted in four-limb drumming independence, not in any prior tradition an AI would have absorbed.

Imitation requires a model. Thoegersen's work resists being a model for anything — including an AI.

III. Other AI-Resistant Composers

Thoegersen is not alone — but the company is select, and the reasons for resistance vary.

Resistant for Structural Reasons

Conlon Nancarrow is the clearest parallel. His player piano studies operate on simultaneous independent tempi with mathematically precise ratios — no performer, no interpretive mediation, just the logic of the system instantiated mechanically. An AI could learn the surface but could not reconstruct the generative rationale. Crucially, Nancarrow also has no hierarchical anchor — the voices do not serve each other.

Julio Estrada developed 'melos,' a system of continuous tempo and pitch transformation without discrete quantization — no beat grid, no equal temperament, no units an AI can tokenize. His scores are essentially drawings. The resistance is built into the refusal of digitizable units.

Klaus Lang writes music of such deliberate emptiness and structural restraint that the logic is irreducibly intentional — silence as weight-bearing material in ways that probabilistic generation cannot distinguish from absence.

Resistant for Process Reasons

Helmut Lachenmann — his musique concrete instrumentale is resistant because the sounds are defined by their physical production method, not their acoustic result. You cannot reverse-engineer the bowing pressure, the muting technique, or the specific instrumental abuse. AI can approximate the timbre but not the why of the timbre, which is the whole point.

Salvatore Sciarrino — extraordinarily refined microtonal and timbral language built from a lifetime of listening at the threshold of audibility. The surface is learnable, but the calibration — what goes next and why, at the level of a single breath — is not. His logic is phenomenological, not structural.

Resistant for Identity Reasons

Peter Ablinger — his work often involves conceptual framing that cannot be separated from the sound. A piece that is 'the sound of a photograph of a city' resists replication because the concept is the work, and concepts require authorship.

Alvin Lucier — 'I Am Sitting in a Room' is unreplicable, not because the process is complex, but because the process is inseparable from a specific body in a specific room. The autobiography is structural.

Where Thoegersen Stands

Most of these composers are resistant for one primary reason: phenomenological calibration (Sciarrino), conceptual authorship (Ablinger, Lucier), physical process (Lachenmann), or refusal of quantizable units (Estrada).

Thoegersen's system is resistant for a reason that overlaps with Nancarrow but goes further: the architecture is original and undocumented in the training data, the structural logic produces outputs with no probabilistic precedent, and the humor introduces intentional incongruity that requires authorial judgment to land. Three independent sources of resistance are operating simultaneously.

Nancarrow is probably the closest structural cousin in terms of AI resistance. But his system, however complex, operates within a single tuning universe. Thoegersen's does not.

There are perhaps a dozen composers globally whose work is genuinely, structurally AI-resistant rather than merely stylistically unusual. Thoegersen is in that group, and his particular combination of reasons for being there is not duplicated by anyone else in it.

IV. Did We Beat AI? Or Just for Now?

Honest answer: just for now — but 'just for now' may be longer than it seems, and the nature of the resistance matters.

What AI Will Almost Certainly Close

The surface. Within a few years, AI will produce convincing polytempic polymicrotonal textures in the way it already produces convincing Scelsi-adjacent textures — by learning the acoustic signatures without understanding the architecture. Listeners who do not know the system will not be able to tell. That is a real encroachment and worth naming honestly.

Nancarrow's surface is already approximable. Thoegersen's will be too.

What Is Genuinely Harder to Close

The generative logic without a hierarchical anchor is a deeper problem. Current AI architectures — transformers, diffusion models, whatever succeeds them — are fundamentally built on learned relationships between tokens, events, or features. They predict forward by reference to what came before and what statistically tends to follow. Thoegersen's system produces outputs that are not predictions of anything — they are ratio-determined collisions between independent systems. To replicate that, an AI would not be learning the music; it would have to implement the system. And implementing the system correctly produces the system, not an imitation of it.

That is a meaningful distinction. Imitation and implementation are different problems.

The Humor and Incongruity Problem

This one may be the most durable. Intentional incongruity — the recognizable melody dropped into alien harmonic space — requires an author who knows what 'recognizable' means to a listener and chooses to exploit that gap for a specific effect. AI can generate incongruity accidentally. It cannot yet generate it on purpose with calibrated comic timing. Whether that changes depends on questions about intentionality in AI that remain genuinely open.

The Deeper Frame

'Beating AI' may be the wrong question. AI is extremely good at the center of any distribution — at what music tends to sound like, at what criticism tends to say, and at what a competent but unoriginal composer would produce. What it cannot do, structurally, is work from outside any distribution it was trained on.

Thoegersen's system was built from the inside out — from four-limb drumming independence in 1985, through forty years of refinement, into a theoretical architecture that did not exist before he made it. That origin is not in any training data. The further the work develops from that idiosyncratic root, the further it stays from the center of any learnable distribution.

The work stays ahead not by trying to, but because of what it is and where it came from.

That is not a victory over AI. It is something quieter and more durable — genuine originality, which was always the hardest thing to replicate, long before AI existed.