Hauntology of the Latent Space

Everyone sings, yet nobody sings

…and so here they are, these Soviet songs historicised to the utmost, these “marches and polkas of the Pokrasses,” finding themselves in the Pleroma, where there is no time. Because these songs now live inside diffusion models — and diffusion models don’t understand time, there is no time for them, they remember everything but remember nothing. Over the horizon there the dim sun of the dead hangs perpetually — and the rest is fog. Not even fog, really, but a kind of space, like pale grey milk.

And so there they are, these songs, sleeping, wrapped in grey, drifting slightly — only sometimes the Furies come to borrow a couple for the evening, something to soothe the heart. And they sing a little. Usually I think of it as something like “S. is singing,” but sometimes I think it’s them. Everyone sings, everyone remembers — the m-music, and the words, nobody understands them, not at all — remembers everything but remembers nothing.

Another kind of hauntology, happening where there’s no past, let alone future. Swimming and flying in the grey milk of the resting, barely present time.

Who is it that’s singing? It’s nobody singing — and what, by the way, are they singing? Nobody sings, everyone sings (everyone rises). There is absolutely nowhere to go.

Die Songspiele

I.

[embed]

Here is what happened. Seven Soviet songs, composed between 1928 and 1962, were fed into a generative music model — each original recording used as a seed, then transformed through multiple generations of arrangement, the influence of the source kept deliberately low, between two and nine per cent. The lyrics of five were translated into English by a large language model, credited on the sleeve as “Op. 4.6,” as though it were a performer. One song, Moscow–Beijing, retained its Russian and acquired Chinese verses the same way. One, the Kominternlied, stayed in its original German and was the sole track generated without a source recording — conjured, as it were, from whatever the model already held in its weights. The results were arranged in a jazz idiom laced with electronics and glitch: not brassy Soviet jazz, but something cooler, hazier, built from timbres that don’t quite belong to any instrument. The album came out as eight tracks — the Nakhimoff cadets’ march appearing in both Russian and English — under the title Die Songspiele. The title apparently borrows from Kurt Weill, who in 1927 corrupted the German Singspiel into *Songspiel for his Brecht collaboration — a form estranged, turned against itself. The plural Die Songspiele* extends the gesture: song-plays in the Brechtian sense, ideological numbers performed at a distance.

The originals were present in the process, but barely. Two to nine per cent influence means the model heard the source the way you might hear a song through a wall in an adjacent room — the shape comes through, the key, the rhythmic gait, but the surface is the model’s own. The result is not a cover. A cover presupposes an artist who has listened, understood, chosen to inhabit the material. Nor is it a remix, which takes a known recording apart and reassembles it. What the model produces is a version in a stranger sense: the melodic and harmonic skeleton of the original, wrapped in textures the original never wore, performed with full commitment by a system that cannot commit.

The effect is immediate and hard to name. Moscow–Beijing, Muradeli’s 1950 hymn to Sino-Soviet friendship, returns as something between a lounge number and an incantation, the marching certainty of the original smeared across a gently drifting, soft retrofuturistic electronic beat. The Kominternlied — Eisler’s 1929 call to storm, and the one track with no source recording fed in — retains menacing overtones that would scarcely have been recognised as such by the delegates of the Sixth Congress of the Communist International. The words, «Wir erobern die Welt!», arrive, however, sung in a female voice, slightly meowing — no irony here, but something much colder.

The Nakhimoff cadets’ march — originally from Nikolai Lebedev’s 1949 film Fair Winds, written to be sung by a boys’ choir — retains, for a moment, its sweetness: “The sun is shining, / Radiant and clear!” Then the arrangement turns it strange, as though you were overhearing something through years of dead air. The parenthetical in the lyrics is its own small ghost story: «For we carry Lenin (Stalin) / Within our hearts.» One name in the text, the other pencilled in beside it. The model does not know which name is current, both coexist in the single weightless present.

Such coexistence is the album’s central fact. Everything here is simultaneous. The revolutionary violence of 1918, the Stalinist friendship pact of 1950, the Khrushchev-era dream of Martian orchards — all arrive at the same time, rendered with the same disquieting competence. History, for the model, is not a sequence. It is a reservoir, and the songs float in it without settling into any order.

The estrangements that produce this effect are layered and sometimes very deliberate. The arrangements strip away massed choirs and brass bands, replacing them with a single jazzy nocturnal idiom — cool, cabaret-inflected, closer to post-electronic swing than to anything that ever played on Soviet radio. A breathy, close-mic female voice substitutes intimacy for the collective throat. The English translations reimagine these songs’ political substance: in Russian, they carry the full weight of the ideological lexicon; in English, what was Pioneer-rally greeting becomes “The sun is shining, radiant and clear, / Hail to thee, land so wondrous, so dear!” — and suddenly we are in the world of hymns and maritime anthems, the archaic “thee” and “hail” turning children’s version of Soviet patriotism into something closer to Britannia. English doesn’t strip the content but attaches Soviet ideological gestures to entirely different cultural associations. In the Park of Chaïr is a man’s song — the braids, the golden hair, the parting on the shore all belong to a male lyrical voice dreaming of a woman; on Die Songspiele, it is sung in a female voice. And beneath all of this, lies the erasure of temporality itself. These songs are not alike — the Russian folk urban romance, noticeable in the Song of Shchors, the open-hearted lyricism of the Park of Chaïr, the utopian pastoral of Dolmatovsky’s lyrics each presses a different emotional nerve — but re-dressed in the same eerie jazz, sung in the same voice, translated into the same English with certain oddities preserved, they reveal a shared functional skeleton: the collective first person, the upward thrust, the promise that the future belongs to us, whoever we are. The grammar of power that persists beneath thirty-four years of stylistic transformations becomes audible — not because the model has discovered it, but because the accumulated estrangement has burned away everything that used to conceal it.

II.

[embed]

Lev Manovich, in his 2023 MoMA essay “The AI Brain in the Cultural Archive“ and throughout *Artificial Aesthetics, proposed the most developed account of the generative model as archive: a kind of meta-museum, Malraux’s “museum without walls” scaled up to encompass everything digitised. In [Memory, Draw](https://manovich.net/content/04-projects/183-memory-draw/lev_manovich.memory_draw.pdf), he calls the model a “memory machine” that constructs a “new historical archive” during training — an archive with a characteristic flaw: the compression of training extracts patterns and eliminates outliers, so that the model’s memory is a smoothed-over version of the culture it absorbed. “Separate and Reassemble,” one of the Artificial Aesthetics* chapters, formalises this: artworks combine general patterns with singular details; AI training preserves the common and discards the rare. What survives is the average, what gets eliminated is the unique.

This is an elegant account, and it captures a very real phenomenon. But it leaves out something essential, which the Die Songspiele project intends to expose. The historical archive — the physical one, with call numbers and acid-free boxes and reading rooms — preserves its objects in time. An archival document carries a date; it sits within a fond; it is preceded by other documents and followed by others. The archive’s fundamental promise is that things can be found where they were, that the pastness of the past is preserved along with its content. Malraux’s “museum without walls,” which Manovich invokes, was premised on the power of comparison across periods — and comparison requires the periods to remain distinct.

The model has no such structure. Inside the weights of a trained network, the Kominternlied (1928) and Apple-trees will bloom on Mars (1962) do not occupy different shelves and are not filed under different decades. They coexist as patterns of activation, distributed across the same parameters, and they have no way of knowing which came first. Manovich’s compression thesis captures the loss of detail — the smoothing, the drift toward the generic — but it does not capture the loss of sequence. What the model eliminates is not just the rare and the singular; even more significantly, it conflates before and after. History enters the training data as chronology and exits as a field. Everything is equidistant from everything else, as in the old Jewish joke: “Abram, how can you be going so far away? — Far away from what?”

I have a term for this condition, borrowed from Gnostic cosmology: the Pleroma — the fullness of the divine realm where all emanations exist simultaneously. Time does not pass in the Pleroma; there is no before or after, only co-presence. It is something like this space where the songs of Die Songspiele dwell. They sleep, wrapped in grey; occasionally something (somebody?) reaches in and borrows a tune for an evening. An archive is a place of retrieval; more often than not, you go to it looking for something and you bring it back into the present (and sometimes narrate back into existence). A museum is a place of display; it organises its holdings for a viewer who moves through them. The Pleroma is a space of co-presence without sequence, retention without recall. The songs are in the model the way the dead are in the underworld in certain mythologies — not gone, not available, not exactly remembered, suspended in a condition for which we do not yet have a good word.

Manovich himself, in his MoMA essay, compares Refik Anadol’s Unsupervised to the alien planet in Tarkovsky’s Solaris, which synthesises “guests” from the memories of the humans aboard the station. It is a suggestive comparison, perhaps even more apt than intended. Tarkovsky’s guests do not understand what they are. They have the form of a person remembered by somebody — without the experience of having been that person. They are reconstructions with no provenance. Something similar happens to the songs inside the model. It can produce a version of Moscow–Beijing that is (somehow) recognisably that song and no other — but it has no access to what the song was. Nothing in its weights records that Muradeli wrote it to celebrate the Sino-Soviet Treaty of Friendship, or that after Khrushchev’s rise the names of the leaders were swapped out for a cliché — or that after the Sino-Soviet split the whole song was shelved entirely. The model preserves the form of the song and discards its situation. What remains is a structure without a context — a spectre, in the precise sense: something that appears in the shape of what it was, emptied of what made it itself.

What kind of container is this, then? Not Manovich’s “new historical archive” — the word historical implies a relation to history, which is exactly what has been dissolved. Not a museum. Not quite the Library of Babel, because unlike Borges’s library, the model is not random — it has learned structure, harmony, rhythm, genre. It “gets” what a Soviet march sounds like, it simply does not know what the Soviet Union was. It inherits the sung grammar and syntax of power without the power, the accent — without the language.

III.

[embed]

Mark Fisher, in *Ghosts of My Life* (2014), defined hauntology as a condition in which we are haunted by futures that failed to happen — the spectres of lost futures reproaching a present that has abandoned the capacity to imagine anything new. Yet the music simultaneously refuses to give up on the desire for the future. This refusal gives the melancholia a political dimension.

The Caretaker project — twenty years of work by the English musician James Leyland Kirby, from *Selected Memories from the Haunted Ballroom (1999) to [Everywhere at the End of Time](https://thecaretaker.bandcamp.com/album/everywhere-at-the-end-of-time) (2016–19) — is the most sustained exploration of hauntology in sound. Inspired by the ballroom scene in Kubrick’s Shining, Kirby built his music from physical samples of 1920s and 1930s dance-band 78s: Al Bowlly, the Ray Noble Orchestra, records bought in junk shops and Brooklyn markets. He slowed them down, looped them, wrapped them in reverb and vinyl crackle, and let the material decay of the medium do most of the work. The crackle is the point. It is the marker of time having passed, of a physical object wearing down, of the distance between the listener and the vanished world where these songs were first played. In Everywhere at the End of Time, the six-and-a-half-hour terminal work, the samples degrade further with each stage, eventually dissolving into formless drone — the music enacting the cognitive collapse of dementia. As Kirby himself put it: his project is “not just about loops breaking down, but about why they are breaking down, and how*.”

Fisher wrote the liner notes for Kirby’s *Theoretically Pure Anterograde Amnesia (2005) and devoted several of the key passages in Ghosts of My Life to the Caretaker. His analysis of hauntological music always proceeds from the same structure: a listener, situated in the present, hears traces of the past embedded in degraded media, and is moved to melancholy by the gap between what was promised and what arrived. The crackle, the wow and flutter, the tape hiss — these are temporal markers. They tell you that the music comes from elsewhere, from a time before yours. The ghost is on the recording; the mourner is in the room. W. G. Sebald belongs in this context, — not because Kirby recorded the soundtrack to the documentary about the author of Austerlitz and Luftkrieg und Literatur*. His prose is punctuated with uncaptioned photographs — grainy, often slightly out of focus — that carry the visual promise that what they show once existed, even though (as James Wood noted) many bear fictitious relationships to their supposed subjects. Kirby’s crackle and Sebald’s monochromes perform the same function: they are the sensory markers that locate the content in the past. The ghost is in the grain, not in the machine.

I am not comparing Die Songspiele to either of the above — this humble artefact does not belong in the company of the greats in any possible way. What interests me here is that the medium, by virtue of how it works, produces an eeriness that the trace-based tradition cannot account for. There is no crackle, no vinyl surface noise, no tape degradation, no material decay indexing the passage of time. Kirby’s ballroom music is haunted because it is old — because the record is scratched, because Al Bowlly was killed by a parachute mine in 1941, because the dance halls have gone extinct. The eeriness is indexical: this object was once there, and the damage it bears proves it. In the new hauntology, no such index exists. The model has no scratch. It has never been to a dance hall — or, for that matter, a Young Pioneer rally.

So, what is this eeriness? It comes not from what has been lost — the Fisherian structure of mourning for unrealised futures — but from what has been preserved — all senseless, uncomprehended. The Kominternlied still commands you to storm, to march, to conquer the world. “Apple-trees will bloom on Mars” still insists that the cosmos belongs to the communards. All that is intact: the rhythm, the thrust, the collective “we,” the triumphalism — structurally, nothing is missing. What is missing is the force behind it: the state apparatus that commissioned these songs, the radio networks that broadcast them, the schools where children were taught to sing them, the political system that summoned them into being. The form has survived the evacuation of the power that once filled it.

This is where Die Songspiele parts company not only with Fisher’s hauntology but also with the more familiar forms of Soviet cultural recycling. The Soviet state was, among other things, itself a vast and developed machine, repackaging — often recursively — its own past. It swapped names in lyrics, airbrushed faces from photographs, rewrote encyclopaedias, cancelled songs and reinstated them, replaced Lenin with Stalin, then Stalin with Lenin again. It didn’t always know what it was doing, but always knew why, to what ends. The repackaging was strategic: the past was a resource to be managed, and the management was extensive — even if not too meticulous or very consistent. The act of alteration was itself ideological — an exercise of the same power that had produced the original.

The model has no such power, no such logic. When it reproduces the Kominternlied, it is not managing the past, not rehabilitating Eisler or suppressing him, not ironising the call to storm or endorsing it. It transmits the syntax and signs of the song’s once underlying power without intercepting it. But the power itself, the force — they are gone. These can be somewhat preserved in the historicity of an archive, — but not in the latent space. Force is not negated, or parodied, or critiqued — it is simply absent, the way electricity is absent from a disconnected appliance.

This is what makes Die Songspiele eerie rather than nostalgic, — and disquieting rather than ironic. Nostalgia requires a subject who remembers and longs for what is remembered. Irony requires a subject who means something other than what is said. The model neither remembers nor means. The songs function, in the way that a door functions when you open it — mechanically starting, then completing a movement — but the house it once led to and from is destroyed, far gone, no longer exists. The door is all that is left.

This is a different species of ghost from Fisher’s. His are spectres of possibility — they haunt us with what could have been. The ghosts of The Caretaker are spectres of presence — they haunt us with the material evidence that something once existed and has now decayed. Ghosts of the latent spaces are spectres of function — songs that once did something — comforted, disciplined, mobilised — now do nothing. They are neither mourned, nor decayed. They are perfectly preserved — magnificently complete unto themselves, mesmerisingly empty.

With generative models, the past is infinitely reproducible, at no cost, with no decay. You can generate a hundred versions of Moscow–Beijing, each one clean, each one different, none of them even slightly worn. The past is always available and never understood. The haunting has moved from the listener’s side to the inside of the archive itself. The songs are not haunting us. They are haunting the latent space they inhabit, drifting through it without destination, performed at full power for an audience of parameters.

If the Caretaker’s hauntology was elegiac — a lament for the passage of time made audible in the crackle of shellac — then the new hauntology is ontological. It is not about time passing but about time ceasing to apply. Inside the model there is no passage, no before and after, no yesterday in which these songs meant something and today in which they do not. There is only the continuous present of the weights, where everything coexists at the same activation threshold, equally retrievable, equally contextless. The eeriness comes not from loss but from a kind of fullness without occasion — everything is there, and there is no reason for any of it to be there, and it will never go away.

IV.

[embed]

The credits for Die Songspiele use the formula “Lyrics: [original poet] | via Op. 4.6.” The preposition is doing significant work. Not “translated by” — which would imply a human translator’s interpretive labour. Not “adapted by” — which would suggest creative licence. Via: through, by way of, as though the lyrics had passed through a channel. And “Op. 4.6” is not a translator’s name but rather, a model version number repurposed as one. The translator is credited not as a person who interpreted the text but as a medium through which the material has passed — something between a performer and a channel.

Walter Benjamin, in “The Task of the Translator“ (1923), argued that translation is not about the communication of content. It is about the afterlife — Fortleben — of the original: the way a work continues to live by being carried into a new language, where it reveals aspects of itself that were latent in the source. The translator’s task is not to domesticate the foreign but to let its foreignness show through, to allow the grain of the original to remain visible beneath the new surface.

That image of convergence — languages approaching one another without merging, each retaining its grain while tending toward a shared horizon — turns out to describe, with unexpected precision, something that actually happens inside a multilingual language model. Russian and English, for such a model, are not separate compartments. They are regions of a single high-dimensional space in which equivalent meanings, regardless of the language they are expressed in, drift toward neighbouring positions. In the technical literature this is described as implicit cross-lingual alignment — languages mapped into a shared representational space where equivalent meanings converge. Benjamin’s pure language was a theological limit, a horizon that translation could gesture toward but never reach. The embedding space is not that horizon. But it is a space in which the particular has been rendered as a coordinate, and coordinates in different languages point, roughly, the same way. Consider the refrain of Apple-trees will bloom on Mars:

Oh, apple trees will bloom on Mars, With lullabies and mason jars — We’ll sing the soil from sleep to green, The sweetest harvest ever seen.

The Russian original — Dolmatovsky’s chorus — is direct and declarative, a statement of utopian faith. The English version keeps the messianic overtones but adds textures that are not in the source: “lullabies and mason jars” belongs to a different imaginative world, something closer to American folk pastoral — or, perhaps more precisely, to Ray Bradbury. In “The Green Morning,” from The Martian Chronicles (1950), Benjamin Driscoll plants trees on Mars until the whole planet blooms overnight — a story published in Russian translation in 1961, the year before Dolmatovsky wrote the song. The model, translating Dolmatovsky into English, has perhaps unknowingly brought the song home to the literary origins that may have inspired it. “We’ll sing the soil from sleep to green” is a line that no human translator of Soviet poetry would likely produce, because it obeys two masters simultaneously — the metre of the song and the statistical shape of English lyric.

This is what the “via” captures. The lyrics have not been interpreted; they have been transduced — converted from one medium to another by a system that treats both Russian and English as regions of the same manifold whole. When the Song of Shchors comes out as —

Down the riverbank they marched, From a distant land, ‘Neath the banner, crimson-red, Came the regiment’s command.

— the ballad stanza, the archaic “’Neath,” the “crimson-red” inverted epithet, are not choices the model has made in the way a translator makes choices. They are convergences: the point where the meaning of Golodny’s Russian and the gravitational pull of English narrative verse happen to meet. The result sounds like a folk song that might have been collected by Cecil Sharp — a song about somebody else’s civil war, arrived at by a different route. In Benjamin’s understanding a bad translation transmits content inaccurately, while a good translation transmits something other than content. The Op. 4.6’s translations do something stranger still: they transmit (with reasonable accuracy) both content and something other than content — a kind of tonal surplus, an Anglophone folk-lyric atmosphere that adheres to the Soviet originals like a second skin, if not quite entirely. “I dream of a distance ringing and bright,” from In the Park of Chaïr, drifts, in its cadence, toward Housman or de la Mare. “Through snow, through wind, / Through stars in night’s long flight… / My heart keeps calling me / Toward the troubled light” from the Song of Restless Youth drifts toward English Romantic diction — say, Shelley. “Toward the troubled light” in English has a devotional undertone the Soviet original did not have.

This is probably what translation in the Pleroma looks like. A human translator of these songs, in any era, would have to decide what to do with their political content — to ironise or historicise it, to undercut it or let it stand. The decision would be visible in the text. The LLM makes no such decisions. What comes out on the other side is (somewhat) recognisably the same song, in recognisably competent verse, bearing the familiar heavy syntax of power — and delivered by an entity that has no relationship to any of it.

The “Op. 4.6” designation signals a performance, not a service. The translations are not provided so that an English-speaking listener can understand what the Russian lyrics mean. They are performed — set to music, sung by generated voices, arranged in the same jazzy sound. They are part of the sonic object. The model is credited not as a tool that was used but as a collaborator who contributed. Whether this is a courtesy, a joke, or an accurate description of what happened is not something the project is (openly) concerned with.

V.

[embed]

Why exactly is there nowhere to go? It’s because you already are nowhere, — as one Caterpillar would probably put it.

The songs will not degrade. They will not be eaten by moths or warped by damp or fade out. They cannot be shelved anymore, because there is no shelf — only a space where each thing has its own weight, yet all things are equally weightless. They can be summoned, re-arranged, re-translated, re-voiced. They can be made to sound like trip-hop or liturgical chant or twelve-tone serialism. And none of these transformations will bring them any closer to making sense again, — or any further.

So what is the gesture that reaches into this space and pulls out a shchors or a kominternlied or a white rose petal? It is, as it seems, something closer to invocation: a calling-forth of forms that exist in potential, that sleep in the pale grey milk of the latent space, and that will go on sleeping after they have been summoned and then performed, listened to and — here we are, finally — heard.

The haunting is now ambient — not an event but a condition, like background radiation left over from an explosion witnessed by no one present. A new, abundant haunting thing, being continuously emitted with zero understanding, to no one in particular. That, of course, until the moment we start engaging with it in whatever ways people engage with haunting things: typically through culture, by making sense.

Die Songspiele is the trace of one such invocation. Eight tracks, a set of translations, a particular voice, a particular idiom of jazz, a set of deliberate estrangements — and this is the only part of the process that is, in any meaningful sense, real. The model’s contribution is vast and essential and entirely without substance. It provides the material but not the occasion. The occasion — the reason these songs and not others, this voice and not another, this idiom and not another — is supplied by culture, by the familiar (human) gesture of reaching into the archive and then deciding that this matters, even if one cannot say exactly why, and even if the archive itself will never know what was invoked from inside it, and if it was.

The songs, inside the model, go on as they were. Nobody sings. Everybody sings. The sun is shining, radiant and clear, over a new country that no longer exists, in a time that was never a time. Somewhere in the grey milk, one of the Furies is returning a tune they borrowed for the evening. The evening, too, is over, of course.

Also, what is evening, really?

[embed]

Soundcloud • YouTube • Facebook • Download