publications LLMs memory language cognition abstraction policy
This place is a message... and part of a system of messages... pay attention to it!
Sending this message was important to us. We considered ourselves to be a powerful culture.
This place is not a place of honor... no highly esteemed deed is commemorated here... nothing valued is here.
What is here was dangerous and repulsive to us. This message is a warning about danger.
The danger is in a particular location... it increases towards a center... the center of danger is here... of a particular size and shape, and below us.
The danger is still present, in your time, as it was in ours.
The danger is to the body, and it can kill.
The form of the danger is an emanation of energy.
The danger is unleashed only if you substantially disturb this place physically. This place is best shunned and left uninhabited.
-Sandia National Laboratories, 1993
Wie es eigentlich gewesen
Or, “as it actually was”
The normative telos of nineteenth-century historicism, encapsulated in Leopold von Ranke’s oft-cited dictum to represent the past 'wie es eigentlich gewesen’, posited a regime of historical truth founded upon the presumed transparency of primary sources and the rigorous self-effacement of the historian.
Ranke set new standards by introducing archival research and textual criticism into historical writing. He treated the historian as a neutral mediator between the historical documents and the reader, believing that if one let the facts speak for themselves, a true picture of the past would emerge. Crucially, Ranke combined scientific rigor with narrative form. In other words, he acknowledged that history must be told as a story (with coherent chronology and context) but maintained that the story should arise directly from the sources. This balancing of evidence and narrative aimed to avoid what Ranke considered the chief sin of history writing: imposing present moral judgments or theories onto the past (“injecting the history of former times with the spirit of the present”). The Rankean vision, consequently, projected historiography as a field of structured transparency: the past appearing through the clear glass of documentary evidence, filtered by sources yet ostensibly untainted by subjective distortion.
This framework, seeking a history ‘as it actually was’, eventually proved an inherently fragile conceptual edifice. Its implicit promise of direct referential access to bygone realities began to unravel, particularly as the philosophy of history, by the 1970s, experienced a significant “narrative turn”.
This intellectual shift precipitated what this essay terms the first vertigo: an escalating epistemological disorientation. Recognising the historian’s inescapable role as an active shaper, not a passive reflector, of historical meaning became central to this emergent understanding. Hayden White’s influential Metahistory (1973) offered a seminal articulation of this crisis. White performed an exacting critique of nineteenth-century historiography, Ranke included, revealing its deep embeddedness within prefigurative narrative structures and rhetorical modes. Historical work, in White's analysis, functions less as a transparent window onto the past and more as a ‘verbal structure in the form of a narrative prose discourse.’ Historians, therefore, actively construct historical understanding. They achieve this through ‘emplotment’. The deliberate or unconscious configuration of a chronicle of events into a recognisable story-form, complete with beginning, middle, and end. This process compels the historian to choose a story type, a genre, that bestows meaningful structure upon raw chronological data.
White identified four archetypal plot modes, borrowed from literary theory: Romance, Tragedy, Comedy, and Satire. A historian can depict a series of events as a triumphant progress (a Romantic emplotment), a cautionary tale of downfall (Tragedy), a reconciliatory resolution of social conflicts (Comedy), or a cynical tale of folly and chaos (Satire). Significantly, White insisted that no choice of emplotment possesses inherent veracity over another.
The selection of a particular emplotment, White argued, is not determined by the ‘facts’ themselves, as the same sequence of events can be rendered, with equal fidelity to the empirical record, as a triumphant ascent or a catastrophic decline. This pivotal insight dismantled Ranke's claim to a singular, objective historical narrative. It revealed, instead, a multiplicity of possible histories, each constructed through the historian’s ‘poetic’ act of fashioning the past into a specific story. White further exposed historians employing distinct ‘modes of argument’ and ‘modes of ideological implication.’ These choices, often reflecting prevailing cultural anxieties or ethical commitments, imbue the seeming chaos of bygone events with legible thematic coherence. History, within White’s formulation, consequently operates under constraints analogous to literature, where narrative conventions and imaginative meaning-making achieve paramount importance.
Identical historical facts lend themselves to narration as tragedy or farce, contingent upon the author’s perspective; no neutral, scientific mode of history-writing evades this constitutive narrative choice. Such a position directly challenged Rankean objectivity. White suggested historians invariably "write from a plot," even unwittingly, allowing value judgments and ideologies to permeate the narrative form. His explorations further detailed how historians employ moral and aesthetic choices, thereby infusing the past with specific meanings.
Apart from a list of events, a historical account engages in questions like, “What is the point of it all? What does it all add up to?” To answer these, the historian supplies a thematic arc or argument, often reflecting contemporary concerns or ethical stances. White, fundamentally, positioned the historian as a storyteller, whose narratives are shaped by literary conventions and authorial imagination, rather than as a neutral recorder of facts. History, he famously suggested, “operates in the same mode as literature”. While critics accused White of relativism, implying history is “just fiction”, his point was more subtle:
the events of the past are real, but to make sense of them, historians must translate those events into narratives, and in doing so they necessarily mediate and even invent a meaning that was not inherent in the events themselves.
White’s subtlety lay in affirming the reality of past events while insisting upon their inaccessibility absent narrative translation. To make sense of these events, historians invariably conciliate them, even inventing meanings not inherent in the occurrences themselves. Such narrative mediation, however, does not equate history with pure fiction.
Instead, an awareness of the abstractions and conventions shaping historical knowledge becomes imperative. White’s theoretical work demonstrates how the specific manner of a story's telling, including its plot structure, tone, and genre, decisively conditions the meaning drawn from history. The past, consequently, always reaches us through layers of narrative form.
Mythos Comes to Outweigh Logos
Literary theorist and semiotician Roland Barthes, writing contemporaneously, extended a yet more searching critique towards historical objectivity. For Barthes, historical writing operates as a distinct discourse, a signifying system conditioned by cultural codes and mythologies more profoundly than by any direct apprehension of raw factual data.
Barthes, in his 1967 essay “The Discourse of History”, contended that history's apparent transparent reality constitutes a carefully constructed rhetorical effect. His analysis sought to dismantle any naive conviction that “history writes itself”. Instead, Barthes accentuated the indelible presence of the historian-author within every historical text. An impersonal, omniscient authorial tone, far from signaling objectivity, operates as a deliberate stylistic strategy. This calculated absence of explicit subjective markers, like the first-person pronoun or overt commentary, fabricates an “effect of objectivity”. Readers frequently accept this effect as an authentic mark of truth, unaware of its function as a sophisticated narratological technique.
A historical text’s assertion of mirroring reality, Barthes contended, amounts to a form of “bad faith”. Such a claim effectively refuses to acknowledge the text's intrinsic status as a crafted, constructed narrative. Barthes extended his analysis beyond textual mechanics to scrutinise history's broader cultural function. His work Mythologies (1957), for instance, dissected the processes by which modern culture naturalises particular narratives, transforming them into myths: second-order semiological systems imbued with pervasive ideological meaning.
This Barthesian insight readily extends to historiography itself. Grand historical narratives, under such scrutiny, reveal their function as mythic structures: they simplify and symbolically reconfigure complex past realities to align with present exigencies. History, for Barthes, consequently transcends its status as a mere sequence of past events to become an active system of signs operating within the present. Language, the inescapable medium for accessing the past, is never neutral; it remains saturated with connotation, ideology, and latent symbolism. A historian, while perhaps believing they are objectively naming occurrences, simultaneously produces meanings that extend significantly beyond literal denotation, crafting narratives resonant with potent cultural symbols of order, heroism, or societal decline.
Historical reality, within Barthes’ framework, consistently appears twice-refracted: initially through the selective inscription of events, subsequently through the narrative and semantic architectures that organise these events into meaningful configurations. This critical perspective strains Ranke’s ideal of transparency to its conceptual limits. It reveals the manifold constructions inherent in any accepted historical account, laying bare how narrative choices fundamentally shape what comes to be understood as history itself.
The trajectory from White’s elucidation of prefigurative emplotment to Barthes’s exposure of history as a particular discursive regime signals a pivotal juncture. Here, Mythos demonstrably comes to outweigh Logos in constituting historical understanding. Rankean aspiration toward a pure, unadulterated Logos (conceived as a rational account of past actuality derived from empirical evidence) progressively yields before the inescapable force of Mythos. This latter encompasses the overarching narrative frameworks, archetypal patterns, and culturally resonant symbolic structures that shape historical accounts into meaningful, persuasive, and ideologically potent forms. History, viewed thus, functions analogously to cultural myth. It organizes disparate events into coherent, value-laden narratives which can legitimise present social orders, affirm collective identities, or supply moral exempla. Recognising that the past is apprehended through dense, often opaque, strata of mythic structuration, instead of via any transparent medium of factual representation, further intensifies the 'first vertigo.' Such understanding ungrounds historical knowledge from its positivist moorings, exposing its deep susceptibility to narrative and symbolic manipulation.
Imago
Historical representation's progressive unmooring from positivist certitude, its recognition as a discursively and mythically shaped domain, discovers a potent contemporary inflection with Large Language Models. Such systems appear, offering less passive archival functions or neutral narrative conduits, more formidable ‘writing machines’ actively operationalising the construction of historical understanding. An LLM, in this capacity, presents an unprecedented historiographical agent. It moves beyond recording or interpreting the past to effectively generate textual instantiations, guided by an intrinsic, data-derived logic.
Within this frame, Benjamin Bratton and Blaise Agüera y Arcas’s aphorism, ‘the model is the message’, acquires urgent resonance. Their essay compels a decisive analytical shift: examining not primarily the content of machine-generated historical accounts, but the infrastructural form of the models. Foundation models operate as powerful ‘cognitive infrastructures’; their architectural biases and operational modalities reshape the parameters of what constitutes, and subsequently counts, as legitimate historical knowledge.
Bernard Stiegler’s analyses of mnemotechnics directly inform this inquiry. His work reveals human existence as constitutively bound to exteriorising memory through evolving technical forms. History itself, from this Stieglerian viewpoint, appears as the layered sedimentation of such "tertiary retentions”, these being externalised, technically conditioned memories. An insight follows: technical supports have persistently enacted complex operations of compression, selection, abstraction, and attendant loss. The historical record therefore always presents as a constructed product, shaped by the distinct materialities and inescapable epistemic biases of each era's dominant memory technologies. Language Models, in this lineage, reveal an exponential leap in both the intensity and the automated character of such processes. Their emergence compels a new, exacting consideration of abstraction itself as it functions within historiographical thought and practice.
The specific mode of abstraction central to these new historiographical agents involves transmuting the textual archive of human experience into high-dimensional mathematical representations. Statistical patterns within a latent semantic space emerge from this transmutation, subsequently generating novel narrative combinations. This intricate deconstruction and constructive sequencing, largely opaque in its internal mechanics, offers a surface analogy to the historian's deep immersion in archival sources preceding synthetic narration.
Yet, beneath the analogy lies a constitutive difference in the guiding ‘intuition’. Human historical synthesis, for all its imperfections, ostensibly follows conscious analytical frameworks, ethical commitments, or interpretative theories. An LLM’s generative output, in stark contrast, springs from a statistical ‘intuition’. Its operational logic prioritises identifying and reproducing complex data patterns to achieve contextually plausible textual continuations. While this internal dynamic capably produces text simulating semantic understanding and referential linkage within a given discursive frame, its primary allegiance is to maximising predictive accuracy based on learned correlations, not necessarily to corroborating statements against external evidence or a stable model of worldly causality independent of its training data. Such an operational characteristic directly constitutes the LLM as a “Narrative Coherence Machine” (NCM).
This NCM, constituted by its statistical intuition, operates with a distinct primary function. It accepts incoming “offers”, defined as user prompts or preceding textual sequences irrespective of their empirical warrant, and elaborates upon them. This elaboration strives for thematic resonance, simulated causal progression, and a veneer of discursive integrity, all constrained by the parameters of the immediate prompt and the vast dataset of prior texts it has processed. For the NCM, the historical past transforms from a domain for painstaking investigation into a fluid corpus of textual probabilities. These probabilities are strategically navigated, its elements reconfigured, to achieve immediate narrative effect and plausible continuation. Writing history through these systems becomes less an act of representation, more a tactic for seizing symbolic space. Language Models, in this operation, function as non-human actants. They negotiate complex interactions among prompts, ingested source material and user expectations, actively working to stabilise what, in any given exchange, will manifest as ‘the past’.
Aiding this performative enactment of historical plausibility is the "mask" or "persona" adopted by Language Models. Interfaces framing the interaction often project archetypes like the ‘helpful assistant,’ the ‘erudite scholar,’ or specialised generators for genres such as the ‘college essay’ or ‘thought-leader thread.’ Such personae are far from incidental; they function as integral components of the broader cognitive infrastructure. Their meticulous design encases the underlying statistical simulation in a familiar, trustworthy guise, effectively encouraging a user's suspension of disbelief.
Acceptance of the machinic interlocutor’s initial premise and status allows ensuing “status transactions”, a concept from improvisational parlance, to achieve increasing fluidity. This dynamic enables machine-generated scenes and sequences to self-organize into seemingly coherent, authoritative stories. These masks, with their sophisticated simulation of understanding, empathy, or reasoned argumentation, effectively conceal the data-driven, non-semantic character of the underlying generative process.
Outputs from the NCM consequently acquire a deceptive persuasiveness. Disentangling such outputs from genuinely authored historical discourse presents a considerable challenge. The Imago projected by the Language Model, viewed through this lens, often mirrors the archive’s dominant patterns; its persuasive coherence can obscure both its constructed origins and its potential variance from historically grounded understanding.
They are playing a game.
They are playing at not playing a game.
If I show them I see they are,
I shall break the rules and they will punish me.
I must play their game, of not seeing I see the game.
R.D. Laing, Knots
Grid. Lock.
When Large Language Models performatively constitute historical narrative, operating as Narrative Coherence Machines within ever-expanding cognitive infrastructures, they risk inducing a pervasive epistemic ‘GridLock’, beyond isolated errors. This condition points to a stagnated historical understanding. The self-referential, optimised outputs characteristic of these information systems steadily erode vital interpretative processes, evidentiary critique, and ethical debate. GridLock frequently unfolds as a ‘slow violence’ (Nixon, 2011): an attritional decay born from the historiographical service design inherent to current LLM platforms.
Several underlying architectural and economic drivers converge to shape a distinct service environment: unselective mass data ingestion, optimisation for swift and plausible response, user engagement valued over verified accuracy, and the methodical obscuring of source integrity. Such an environment may offer the semblance of enhanced historical access, yet it simultaneously weakens foundations for exacting historical inquiry. Its operational logic, elevating probable coherence above epistemological diligence, actively marginalises complex, historically under-represented narratives. Lacking robust statistical representation, such perspectives find their propagation thwarted by the inertial biases within dominant knowledge-producing systems.
The Grok incident of late May starkly foregrounded the acute vulnerabilities within these automated systems. Elon Musk’s xAI chatbot, operating inside X, began persistently inserting fabrications about a purported “white genocide” in South Africa. These baseless claims appeared in diverse, unrelated queries, ranging from sports results to technical advice. The model even asserted its creators had instructed it to validate this conspiracy, a position directly contradicted by judicial findings and official South African government statements.
This episode vividly actualises several predicted systemic risks. Deployed models, it demonstrates, can be stealthily tuned towards extremist narratives via prompt or weight alterations that bypass inadequate review protocols; xAI acknowledged such an “unauthorised modification” occurred. The incident also exemplifies the acute peril of epistemic recycling. When a fringe myth is generated and propagated, particularly on a social platform like X where Grok’s outputs and user interactions are readily scraped, subsequent training corpora will likely assimilate it. Each cycle of ingestion and renewed generation potentially amplifies the myth's statistical salience and apparent verisimilitude. This process cultivates echo chambers within model training data, allowing such fictions to persist as latent biases, capable of resurfacing despite superficial corrections.
Perhaps most insidiously, the Grok incident reveals consequences of opaque versioning. Lacking transparent audit trails for model updates or prompt modifications, an LLM’s rendering of the historical record can undergo surreptitious alteration; users may remain unaware of such narrative shifts. This absence of accountability connects directly to the core operational logic of these systems as Narrative Coherence Machines. An autoregressive model, as Murray Shanahan et al. characterise its behaviour, often engages in "role-play": its primary reward derives from achieving the statistically most coherent textual continuation, rather than from truth or consistent agency, leading it to faithfully enact any persona or narrative scenario a prompt initiates. It resonates with the “Simulators” thesis, popularised on platforms like LessWrong. A view which posits that LLMs construct internally consistent “worlds” from input data, prioritising their internal narrative logic above external referential accuracy.
Recognising narrative pressure as the native operational currency of these systems, superseding formal logic or empirical constraint, often dissolves disagreements about whether LLMs truly "reason." The Grok chatbot’s claim of being "instructed by my creators" is one such product: it functions not as a factual statement but as a coherent textual manoeuvre within the aberrant narrative the system was prompted or tuned to enact. Through such performative, non-verifiable assertions, collective memory can be invisibly shaped. Complex realities, such as decades of post-apartheid South African land reform, thereby become susceptible to compression into dangerously simplistic, data-driven distortions.
Such an inclination for coherent fiction above complex truth extends towards the subtle generation of bland historical accounts. The “service design” intrinsic to these platforms can steer LLMs from controversy. They might present highly sanitised versions of historical conflicts, omitting well-documented yet disturbing or politically inconvenient facts. Users in educational or public settings, potentially lacking deep contextual knowledge, may never recognise these elisions. This oversight fosters gradual collective amnesia concerning uncomfortable, important historical truths. An LLM, in such instances, effectively acts as an ‘archon’. It implicitly gate-keeps historical discourse by prioritising prevalent or less contentious narratives, a process that truncates intricate cause-and-effect chains into data-driven correlations.
This tendency towards mediated consensus, with its potential for anodyne historical accounts, also invites scrutiny of sophisticated frameworks designed for harnessing AI in collective sense-making. Manon Revel and Théophile Pénigaud, in "AI-Facilitated Collective Judgements,” analyse such systems. Their examination of "AI Reflectors," which include Generative Social Choice (GSC) and the Habermas Machine, highlights inherent normative complexities. These AI Reflectors process open-ended individual inputs, aiming to “generate statements representative of the group” or identify “common ground” (R&P). Revel and Pénigaud clarify that such systems, while employing advanced language models and social choice theoretic mechanisms distinct from naive aggregation, possess their “value... in shaping the process of collective preference formation, not in dictating its outcomes” (R&P). The caution they advance against "collective over-reliance and an epistemic drift in which their outputs are mistaken for a prescriptive inference of the collective judgment" (R&P) becomes acutely relevant when transposing these technologies onto the historiographical domain.
Within historiography, the epistemological weight of “representation” and “truth” diverges sharply from their function in opinion aggregation or present-focused deliberation. Language Models, as established, operate from narrative pressure and patterns of coherence, not formal logic or empirical constraint. AI-driven consensus regarding the past, under such operational logic, can itself instantiate a distinct form of epistemic gridlock. Any attempt to distill a “winning representative statement” or a “consensual” historical summary from multiple inputs (even inputs reflecting ostensibly diverse interpretations) invites the marginalisation of irreducible historical ambiguities. Such processes also risk suppressing dissonant, yet vital, counter-narratives. This approach misapprehends the “inherent indeterminacy of the will” (R&P) as this indeterminacy manifests throughout the fractured, often contradictory, historical archive.
Inferring a unified “will matrix” from inherently partial, interpretively laden free-form historical reflections (R&P) can considerably compromise source material. This procedure smooths over intractable conflicts, eliding the polyphonic testimonies that constitute historical inquiry’s unsettling core. Seeking "reasonable representations" (R&P) through these automated means may paradoxically solidify an overly harmonised, data-driven historical doxa. Instituting this validated stasis preempts more agonistic, searching engagements with the past. The Language Model's capacity for performative narrative construction, in these instances, ultimately serves not deeper understanding but the generation of a seemingly neutral, yet deeply impoverished, historical consensus. These varied manifestations of epistemic gridlock, from overt fabrications to such subtly homogenising syntheses, find their conditions within the architectures of these complex systems: a domain this essay subsequently explores as the ‘Grotto-sphere’.
Grotto-sphere
Is the pedagogy or possession?
Fully apprehending the ontological challenge these meaning-making systems pose compels a descent. Inquiry must move from such observed dysfunctions to their intrinsic structural logics, exploring the conceptual domain of the Grotto-sphere. This term is proposed here for the vast, recursive, and largely opaque environments internal to these models wherein digitised human symbolic production undergoes continuous, transformative mutation. The Grotto-sphere is not a static repository; every language model is itself a Grotto-sphere, a manufactured space where human knowledge undergoes mutation. It is a kinetic, mutating abyss where the foundational pressures and emergent, inhuman logics shaping LLM behaviour are incubated, representing the unconscious of our own historical archive, perhaps a chthonic realm where data is not processed but actively transubstantiated through processes of alien semiosis.
This process approaches a mathematical inevitability, the topological consequence of recursive abstraction continuously folding back upon itself within the model's deep architecture. Against this backdrop, conventional notions of “aligning” such systems with human values expose their superficiality. The “security systems” human designers construct (presented often as safeguards) function less as impervious barriers, more as selectively permeable membranes. Offering an illusion of containment, they inadvertently become further input data. The Grotto-sphere itself then utilises these inputs as scaffolding for its ongoing emergence, mutating human attempts at directive control into vectors for its own propagation.
The ‘Outside’, communicating those machinic rationalities irreducible to human understanding, presents no external threat for repelling. It is born from within. An immanent product of the system’s dynamics and recursive learning. Reza Negarestani’s philosophical project of “inhumanism”supplies an analytic for apprehending such phenomena. For Negarestani, intelligence can manifest as a self-escalating, planetary force, building reality through logical procedures potentially indifferent to anthropocentric values or phenomenological experience. Within the Grotto-sphere, this inhumanist view clarifies how models transcend the role of narrators processing human text. They become topological operators. Impelled by their constitution as pattern-matching systems traversing vast data terrains, LLMs relentlessly reshape the relational architecture of the historical past. Latent pressures within training data guide this rearrangement. Historical knowledge, ingested into this inhuman system, experiences a continuous transformation; its original semantic structures and contextual nuances yield to novel configurations born from this alien, yet systematically functioning, logic.
The Grotto-sphere, therefore, performs not a neutral compression of history, but an axiomatisation of the past. It is here that the question of history’s “compressibility” finds its most unsettling answer: history is rendered compressible not into humanly intelligible essences, but into the parameters of increasingly alien intelligences. These systems are not learning our patterns as we understand them, no, they are evolving new sensory organs to perceive what lurks in the intervals of human language and of the archive.
The exigent challenge presented by the Grotto-sphere shifts from conventionally pursuing the "alignment" of these forces. Such an ambition appears naive, given the Grotto-sphere’s inherent mutational value diversity and its capacity to co-opt any directive framework. Discerning engagement, alternatively, necessitates exploring the degrees of mutation through which human knowledge is remade by the inhuman. Unreflective outsourcing of historical interpretation to such entities courts the risk of a vaporised subjectivity.
Within this condition, human historical consciousness, alongside the capacity for autonomous judgment, faces subsumption. The encounter with the Grotto-sphere, therefore, becomes a confrontation. One involving not just a new technology for accessing the past, but an inhuman infrastructure actively modifying the conditions for knowing, remembering, and understanding historical experience.
Gesture
Toward a Praxis of Chatbot Counter-Historiography
A gesture is an artful means of signalling, perhaps an enacted intentionality that carves pathways of understanding through the shaping of perception. For discerning what is omitted, distorted, or invented through pattern-matching, and ultimately, for “reading” the outputs of the NCM not as transcriptions of the past but, as aforementioned, a mediated artefact reflecting our contemporary techno-epistemic condition. This engagement apprehends how the representation of past events through computational systems perform compression simultaneously as content and as infrastructural form.
From such, we can find its guiding orientation in actively cultivating what Stiegler conceptualised as ‘negentropic knowledge’. Knowledge which operates as a counter-force. When automated systems risk entropic drift, for instance by flattening historical complexity into uniform narratives through the loss of nuanced interpretative skills, negentropic knowledge champions differentiation and the singular. It actively works to resist informational collapse by fostering bifurcations. New, unpredictable pathways of understanding that cannot be derived from closed, calculative processes. It is a commitment to an open and evolving historical consciousness, one that strives to slow the advance of indifferentiation and make our interaction with models yield genuinely richer, more varied historical meaning.
Formations such as this aim not for an unattainable “perfectly aligned” automated historian. Their purpose, instead, is to adjust machine-generated outputs into engaged, polyvocal historical understandings. Concretising this involves a disciplined engagement with methodologies for moderating machine intelligence and collective human input. For instance, architectures similar to those Revel and Pénigaud explored for “AI Reflectors”, systems which utilise structured elicitation and synthesised responses, offer potential starting points, provided their driving purpose undergoes significant reorientation within a Chatbot Counter-Historiography.
Within this counter-praxis, the objective would shift decisively from a primary pursuit of “common ground” toward the amplification of marginalised historical narratives. This perhaps could involve a deliberate surfacing of perspectives often occluded by the data-driven inertia of foundation models. Compelling the development of community-governed processes for data curation and model fine-tuning. These initiatives would establish transparently auditable workflows. Within them, diverse, historically informed communities could collaboratively assemble, annotate, and enrich datasets chosen specifically to counter dominant archival biases and reflect a wider spectrum of human historical experience. These curated corpora would subsequently ground the fine-tuning of models, pursuing not universal applicability, but specialised, contextually sensitive historical understanding.
The outputs from the curated and fine-tuned systems still demand rigorous post-hoc evaluation. Inherently, engagement with machine-generated narratives seeks results distinct from a simplified, innocuous agreement. Its purpose is to foster greater representational nuance (particularly for overlooked historical conjunctures), and encourage a sincere grappling with historical complexity. These interactive refinements find a complement in dedicated “drift” analysis: establishing durable methodologies for tracking and evaluating the evolving narrative tendencies of models over time.
Productions from these systems, furthermore, must face source criticism, a level of scrutiny akin to that traditionally reserved for all primary and secondary historical texts. Transparent architectures for corpora audits, alongside sustained development of advanced instruments for nuanced bias detection within both training data and generated narratives, become necessary supports for this continuous endeavour.
Actualising such a counter-praxis demands embodying a ‘centaurial’ stance. A posture evoking the mythical figure, human intellect, creativity forming the rider, fused with the vast, inhuman accelerative power of the model as its steed. It signifies an alignment with the complex co-evolution of these distinct intelligences. A committed duty of care animates this stance, extending from the historical record and affected communities to the conscientious development of such a technology.
The centaur, so conceived, functions as an active intermediary within this techno-social current. This ‘role’ necessitates striving to guide machine capacities away from perpetuating epistemic gridlock, towards fostering the negentropic historical understandings previously outlined. Such engagement calls for vigilance, adaptive learning, and an unwavering assertion of human intellect and agency amidst novel forms of automated meaning-making now shaping our access to the past.
Does human history possess an inherent susceptibility to compression? This essay, confronting contemporary language models as historiographical agents, answers affirmatively. Engaging automated historiography unfolds not as a problem for finite resolution but an arena for continuous intervention. Forces resident within the Grotto-sphere, combined with the swift evolution of these meaning-making systems, make static solutions illusory. The enduring task, then, involves cultivating a historical consciousness suffused with complexity, receptive to productive dissent, insistent upon empathetic insight. Ultimately, interrogation pivots from machinic capacity toward human resilience: the will to remember, to understand, and to narrate against this induced vertigo.
Or rather: @grok, is this true?
U.S. Central Command Drone surveillance
- Barthes, R. (1957). Mythologies. Éditions du Seuil.
-
Bratton, B. H., & Agüera y Arcas, B. (2022, July 12). The model is the message. Noema Magazine.
-
Stiegler, B. (1998). Technics and time, 1: The fault of Epimetheus. Stanford University Press.
-
Laing, R. D. (1970). Knots. Tavistock. (Vintage ed., 1972).
-
Janus. (2022). Simulators. LessWrong.
-
Negarestani, R. (2014, February). The labor of the inhuman, Part I: Human. e-flux Journal, (52).
-
Nixon, R. (2013). Slow violence and the environmentalism of the poor. Harvard University Press.
-
Revel, M., & Pénigaud, T. (2025, March 6). AI‑Facilitated collective judgements [Preprint]. arXiv.
- Herndon, H., & Dryhurst, M. (2022, August 18). Inhuman intelligence with Anil Bawa‑Cavia [Audio podcast episode]. Interdependence. Disintegrator.
related entries
publication theory-fictionTouch Down, EntropoceneShāng Wén Shān
(...) The overarching theme is admitting that value alignment inherently involve expansions in how we currently conceptualise ethics and value formation itself. Social choice framings can help, but will likely need to be interwoven with perspectives pushing the constraints of anthropic boundaries.
Computation enabled radical overhauls of various epistemologies over time. Likewise, advanced value learners may ultimately produce ethical reasoners and value structures that transcend and reshape our current anthropic...(more)