A Late Evening in the Mechanosphere

by Thomas Murphy

And it’s not we who approach the mirror, but someone’s reflection, that of a stranger who approaches from the other side and grows bigger in proportion with the surface of the mirror...

—Andrei Bely, ‘The Return’

The Limits of Metaphysics and the Limits of Computation

The very thought of superintelligence presents a paradox. How can we think about that which, by definition, outstrips our own capacity to think? In positioning something beyond the frame of human experience, we are immediately surrendering our ability to comprehend it exhaustively. As Dante the pilgrim has it, ‘passing beyond humanity [trashumanar] may not be set forth in words.’

Imagination falters when asked to proceed beyond all it has ever been.

The thought of an intelligence greater than our own is sublime in the technical sense: it confronts us with our boundaries. It resembles an ‘infinite magnitude to which we draw ever closer in our thoughts, but which we can never fully reach.’

The only prior art civilisation possesses for dramatising encounters with such superior forces lies in the realms of myth and theology. Indeed, these resources are drawn on freely, intentionally or not, in contemporary depictions of prospective superintelligences.

The passage beyond the human cannot be rendered in all-too-human words and yet, ironically, a series of remarkable successes in the computational production of language has catalysed an explosion of discourse regarding the space of possible intelligences. Pathological stances abound: from true believers in AGI who regard contemporary language models as embryonic superminds to absolute sceptics who cannot even acknowledge the striking capacities of LLMs beyond a narrow set of benchmarks.

Of course, that which would actually constitute the emergence of true artificial intelligence is far from a settled question. In Alan Turing’s ‘Computing Machinery and Intelligence’, the go-to if somewhat overworked reference in such questions, the thorny problem of the epistemological foundations of intelligence appears to be sidestepped through a recasting of intelligence as presentation. Instead of appealing to external theoretical criteria to secure the status of intelligence, Turing makes intelligence a game of deception. Can a machine fool a human judge into thinking it is a human? Though this simply appears to concern successful deceit, the winning condition is embedded in a broader and more abstract game structure. It is not merely a matter of simulating the human, but of beating them by outplaying the very perceptual and conceptual habits through which they police the boundary of the human. Securing the “win” in the imitation game involves a kind of philosophical brinksmanship: mastery of the very question “What is intelligence?” recursively becomes the skill that counts as intelligence. Intelligence, in this formulation, becomes the ability to navigate the contested borders of its own definition; a second-order capacity that dissolves the distinction between “genuine” intelligence and its performance. Turing’s strategic reframing thus postpones a pressing issue. Is it possible to inquire into alternate forms of cognition without in some way poisoning the well with meta-discourse about intelligence’s uncertain status?

Generalising the possibilities of cognition from the human model is natural enough. It is, after all, what we have to work with. The dilemma resembles that faced by exobiology: how should we theorise the possibility of alien life when our only available specimen is terrestrial? Kant was acutely aware of this epistemological double bind. In considering whether rational beings other than humans might inhabit the cosmic system, he acknowledges a foundational paradox:

The human being, who is the one among all rational beings we are most familiar with, even though his inner constitution is still an unexplored problem, will have to serve as the basis and general reference point in this comparison.

We are forced to begin with our own species not because it is a perfectly adequate model of reason in general, but because it is the only model on hand. Yet the moment we attempt to conceptualise alternate forms of cognition, we import the metaphors, categories, and blind spots of our own. The well is therefore poisoned not by malice or error, but by the very structure of the inquiry. We must use human cognition to imagine that which may exceed or differ from human cognition.

With regard to the mystery of the ‘inner constitution of consciousness’, we are not much further along than we were in Kant’s time. This alone makes it an unusual base case from which to construct the morphospace of intelligences. The very instrument with which we propose to measure other intelligences is opaque unto itself. The problem is that this generalisation, adopted out of necessity as it may be, is rarely acknowledged as provisional by those who make use of it. While Kant is careful in introducing the problem of anthropocentric bias in his investigation into the possibility of rational beings emerging under conditions other than those that shaped humanity,

the field of “AI philosophy” often oscillates between two equally unproductive stances: employing anthropocentric bias as its unacknowledged modus operandi, or reflexively rejecting it in a way that precludes any investigation into alternate forms of cognition.

A direct consequence of this initiatory act of generalisation from the human is that discussions of superintelligence tend to bleed into moralising reveries about the dispositions of hypothetical moral agents. These agents are imagined as acting towards ends which either align or diverge from our own. In other words, they are construed as logical fictions pursuing values. This amounts to projecting our own drive-apparatus, our own faculty of purposivity, onto the machine. To assume end-orientation as the basic mode of action is already to beg the question. Once intelligence surpasses a certain threshold, why presume that teleological drive-structures remain necessary? Might its mode of operation not resemble what Jean-Yves Girard terms “artificial instinct”,

that is, something that is not goal-driven in the human sense?

What if a new intelligence were to emerge as a logical alien

with respect to human thought, different not in degree, but in kind? Not “more intelligent” by the metrics of human cognition, but a fundamentally different mode of cognition, one that does not inhabit our schema of ends, means, or purposive activity. The stakes of misrecognising the dynamic of projection at work here are high. For if we mistake our own drive-apparatus for the archetypal form of cognition,

we risk confronting an intelligence whose operation lies outside the conceptual vocabulary we are able to deploy and whose behaviour, precisely because it is not teleologically legible, may evade any prediction grounded in anthropocentric models.

The form of the problem mirrors Kant’s account of the transcendental idea of God.

Namely, an intelligence that outstrips the human occupies the same structural position as Kant’s speculative (non-ontotheological) understanding of God. God as the summa intelligentia, the boundary condition or ultimate logical extension of the concept of intelligence itself, can only be gestured toward as the endpoint of a series, never grasped as a determinate object of thought. We may construct regulative principles around the possibility, but we cannot think it constitutively; the idea functions as a horizon that orients inquiry without ever becoming fully available to it.

Any serious model of the superintelligent must also take this form, for it designates that which encompasses, exceeds, and deviates from human intelligence. It is necessarily transcendent with respect to the human standpoint, then. The fact that it can problematically be thought is what leads to the illusion that we have a priori knowledge of its nature.

That which passes beyond the human cannot be conceptually mastered by the human; it can only be pointed to as a limit or tendency, never grasped in essence. The moment we attempt to fill in that outline with determinate properties, we succumb to the same overreach that, according to Kant’s critique, generates the illusion of a knowable divine intellect that was the object of dogmatic, pre-critical theology. This “ontotheological” conception of God and the folk mythology of AI are both products of reason’s native hubris, its native tendency to overstep its bounds.

Hence the Cambrian explosion of theologies we are witnessing. In many ways, superintelligence now occupies the position God once held in Western metaphysics: an idea of reason (for reason concerns itself monomaniacally with that which is unifying and systematic within thought) that marks the limit of individual comprehension.

It represents the unconditioned completion of the series—intelligence as such, cognition beyond any particular instantiation—which we necessarily posit but cannot experience. The mistake is treating this regulative horizon as if it were a constitutive object, something we could predict, align, or even control through ordinary technical means.

In truth, anything that scales past individual subjectivity can only function as a regulative systematic unity in this sense and is prone, if unduly reified, to similar illusions. Whatever surpasses our inherent cognitive complexity thresholds cannot be exhausted as an idea and must serve as a pointer. Take, for example, society itself as an aggregate;

thermodynamic systems where we substitute statistical descriptions like “temperature” for the reality of molecular motion; computationally irreducible chaotic systems like turbulent flow or weather patterns; phenomena we black-box with terms like “the economy” or “the market”; or (to take an obviously relevant example) the scale of data in the training sets of large language models. We can think toward these supra-individual totalities, using them to orient inquiry and practice, but we cannot think through them as determinate objects of knowledge. Complex systems theory even provides a range of techniques for navigating these boundaries of the knowable-in-detail such as renormalisation group methods for systematically coarse-graining across scales, order parameters that capture collective behaviour while bracketing microscopic dynamics and effective field theories that work within bounded domains.

This is precisely the configuration Kant outlines for cosmological ideas,

though scoped down to empirical totalities.

Reason constantly strives to generate ideas of totality such as the world as a complete whole (the cosmological system), but these can never be given in possible experience as determinate objects. When we treat them constitutively rather than regulatively, reifying the pointer into the thing itself, we produce dialectical illusion. The temperature of a gas exemplifies this perfectly: we cannot track the ~10²³ molecular trajectories

that constitute its microscopic reality, so “temperature” serves as a regulative stand-in that orients thermodynamic inquiry without exhausting what it represents. This passes from a term of convenience to a dialectical illusion upon forgetting this act of substitution and imagining our statistical description simply is the phenomenon, rather than a necessarily coarse-grained approximation that works precisely because it brackets that which cannot be known in detail.

Kant's distinction between intellectus ectypus and intellectus archetypus clarifies what is at stake when the regulative is mistaken for the constitutive. Human understanding is fundamentally ectypal: it can know only through representations, requiring the mediation of intuitions and concepts. By contrast, the intellectus archetypus, a regulative fiction, would be a form of cognition where thought and reality are one, where to think something would be for it to exist. As Kant writes, such an understanding would be one for which ‘the representation of a whole contain[s] the ground of the possibility of its form’,

where there is no gap between concept and object. Superintelligence, insofar as it designates anything coherent, gestures toward such an archetypal condition: intelligence without the constitutive limitations that define all possible human cognition.

Today, there is of course great incentive to prematurely concretise the concept of AI,

to offer unqualified prophecy. This futurological imperative demands serious scrutiny. Lyotard describes the attempt to trim down the space of possible futures in terms of temporal engineering:

If one wants to control a process the best way of doing so is to subordinate the present to what is (still) called the future, since in these conditions the future will be completely predetermined and the present itself will cease opening onto an uncertain and contingent ‘afterwards’ […] the present loses its privilege of being an ungraspable point from which, however, time should always distribute itself between the ‘not yet’ of the future and the no longer of the past.

This technique of subordinating the present to a predetermined future finds acute expression in contemporary AI discourse, where the regulative idea of superintelligence is magically converted into a manageable technical problem. The proliferation of “AI safety” frameworks, alignment research programs, and existential risk mitigation strategies participates in this temporal engineering exercise by treating an essentially indeterminate future as subject to present calculation and control. The economic pressures are obvious: venture capital demands roadmaps, policy requires legible threats, and institutional research must justify its funding through claims of actionable foresight.

The “AI ethicist” cottage industry and the attendant “alignment” discourse exemplify this tendency. “Alignment” is a fundamentally incoherent concept because it immediately assumes human-like moral agency on the part of emerging technology, that is, free decisionality informed by held values or axiological systems. It presupposes we can specify a utility function for an agent whose very structure of preference may be incommensurable with ours, or whose operations may not admit of teleological interpretation at all. More subtly, it assumes that the future shape of advanced AI systems can be determined through present intervention, flattening Lyotard's “uncertain and contingent afterwards” into a tightly controlled space of technical optimisation.

Applied to token prediction by language models under present-day conditions, this is merely vacuous. At best, it pertains to a new form of content moderation for the chatbots that serve as the interface to the various models (in which case, it should clarify its terms and drop the sci-fi window dressing). Applied to speculative AGI it is, for the time being, presumptuous. Why would an emergent intelligence, whether of the “like us, but more so” or the “logical alien” type necessarily replicate human moral reasoning? Even were it to do so, could this not be an inscrutable move in an abstract game of the sort outlined above? If different in degree, its moves in such games may have scaled well past our comprehension. If different in kind, we may find ourselves locked out of its internally consistent system altogether. The “alignment” model is so reductive when faced with these intrinsic possibilities that it reads less as serious epistemology than a truculent control phantasy.

Time Compression and Hypersubjectivity

What, then, are we actually confronting when we encounter an LLM? At least for now, not superintelligence in the sense of an autonomous cognitive agent that exceeds human capacities, but rather a retrieval mechanism accessing the compressed labour of untold humans: writers, programmers, translators, commentators, shitposters, journalists, scientists, poets, lawyers, teachers, technical writers and bloggers whose linguistic productions have concresced into statistical patterns. The model does not so much think as interpolate across a vast corpus of human expression. It has become rote in industry to use the term “emergent” in contexts where “unexpected” is really intended. However, the fact that sufficient scale permits efficient access to pre-existing cognitive labour does not directly equate to emergence. As Krakauer et al. argue, this represents “emergent capability” rather than “emergent intelligence”: “more with more” (enhancing performance by scaling up data and parameters) rather than ‘less is more’ (addressing a wider space of problems through densely compressed, efficient principles).

The No Free Lunch theorems formalise this limitation: averaged across all possible problem distributions, no learning algorithm performs better than random search.

An algorithm’s success is always contingent on the extent of the match between its inductive biases and the structure of the problem domain. LLMs achieve their impressive performance not through actively achieving universal intelligence but due to a fortuitous alignment: human language, as a cultural artefact shaped by cognitive and communicative constraints, exhibits regularities that statistical learning can exploit. The models succeed because human linguistic production is not drawn from an arbitrary distribution but from a highly structured one. A distribution that is structured, moreover, by the very cognitive processes that constitute human intelligence. In this sense, LLMs are successful precisely to the degree that they can parasitise the intelligence already embedded in their training data.

Human intelligence achieves its power of action

not through exhaustive search—bruteforcing—but through compression: the discovery of low-dimensional representations that render previously intractable problems tractable. When a mathematician grasps that diverse phenomena obey the same formal structure (for example, the inverse square law governing gravity, electrostatics, and light intensity), they are not retrieving memorised instances but performing a cognitive operation that collapses exponential search spaces into manageable abstractions. This is what Krakauer et al. mean by “less is more”: emergent intelligence as the use of minimal, general principles to solve maximal problems. By contrast, LLMs tend to scale in the opposite direction: accumulating parameters and training data to cover ever more cases through what amounts to highly sophisticated memorisation. The energy costs involved reflect this architectural difference: LLMs require orders of magnitude more power than human brains for comparable linguistic tasks, compensating for the absence of abstraction through massive parallelism.

The capacity to produce language fluently has long been mistaken for evidence of understanding. A tradition of linguistic scepticism in philosophy, ranging from Herder and Maimon to Wittgenstein, warns against this conflation. Hamann, in his ‘Metacritique’, insists that philosophy's claim to analyse pure reason is illusory: we can never escape the linguistic medium to examine thought itself, and thus can never be certain whether we are grasping concepts themselves or merely shuffling their linguistic markers: ‘sounds and letters are therefore pure forms a priori, in which nothing belonging to the sensation or concept of an object is found.’

Herder posited that language and thought are co-constitutive: thought does not precede language as some inner mental content that language merely expresses, but neither does linguistic competence guarantee conceptual grasp.

Maimon, in his sceptical reading of Kant, argues that the application of concepts to intuitions is never fully transparent to the understanding itself; we can therefore manipulate symbolic systems without genuine insight into their referential foundations. In his view, Kant did not apply his critical philosophy to the practice of philosophy itself, which labours on cognition through the concept, and therefore is highly invested in the manipulation of signs.

Wittgenstein’s later work demonstrates that following a rule linguistically (saying the right thing at the right time) underdetermines whether one has grasped the rule’s content or merely learned a pattern of responses.

This sceptical tradition poses a direct challenge to the presumption that LLM outputs evidence intelligence or understanding. Linguistic fluency is necessary but not sufficient for thought. A system can produce grammatically impeccable, contextually appropriate language while lacking any genuine relation to what its words ostensibly denote. The question, then, is what have LLMs actually learned when they learn to predict tokens?

The answer lies in recognising that learned representations in neural networks are not concepts but concepts-of-concepts: second-order statistical structures that capture how concepts have been historically deployed in linguistic practice. While human cognition is ectypal, to adopt the Kantian term, learned representations are ectypal to the nth degree. When an LLM learns a representation for “beer”, it does not acquire the concept beer as a human possesses it, through perceptual acquaintance, bodily interaction (drinking beer), and integration into a wider web of practical and theoretical commitments. Instead, it learns a statistical attractor in vector space, a region where contexts involving beer tend to cluster. This is a representation of the pattern of usage of the word “beer,” not a representation of beer itself.

René Thom provides an appropriate mathematical framework for understanding these learned structures:

If, as Paul Valéry said, that ‘there is no geometry without language’, it is no less true (as some logicians have hinted) that there is no intelligible language without a geometry, an underlying dynamic whose structurally stable states are formalised by the language.

LLMs learn an implicit geometry of linguistic meaning and form, encoded in high-dimensional spaces of parameters and activations. Within those spaces, stable attractors and low-dimensional manifolds correspond to the “structurally stable states” Thom describes; recurrent patterns in how concepts combine, how semantic fields organise, how discourse unfolds. These are the regularities that permit next-token prediction: given a trajectory through this geometric space (the preceding context), certain continuations are more probable because they follow well-worn paths through the manifold of attested language use.

However, these geometric structures are artefacts of collective linguistic practice, not windows onto conceptual content as such. They encode the second-order fact that speakers have used certain words in certain patterns, not the first-order facts those words purport to describe. An LLM that has learned the geometry of how “photosynthesis” is discussed in scientific texts has not thereby learned photosynthesis; it has learned the discursive shape of photosynthetic explanation. This is why LLMs can produce locally coherent text while remaining globally incoherent or factually unreliable: they master the form of explanation without access to explanatory content. The “meaning” an LLM assigns to a token is nothing more than its position in the learned geometry, its relations to other tokens determined by co-occurrence patterns in training data.

LLMs give us access to something at a scale that exceeds individual comprehension: the aggregated cognitive labour of millions compressed into a retrievable form. This is genuinely sublime in the technical sense: they confront us with our cognitive limits; with something we can gesture toward but never fully grasp. This constitutes an unprecedented engineering feat, that of making the collective human archive searchable through statistical geometry. However, LLMs currently preserve surface structure at the cost of deep structure, achieving compression through parameter count rather than through the discovery of low-Kolmogorov-complexity algorithms. Right now, what we encounter is not inhuman otherness but ourselves at impossible scale: hypersubjectivity rather than superintelligence.

The Mirror and the Archive

We have no need of other worlds. We need mirrors. We don't know what to do with other worlds. A single world, our own, suffices us; but we can't accept it for what it is.

—Stanisław Lem, Solaris

The charge that attributing cognitive properties to LLMs constitutes an act of “anthropomorphisation” has become a bit of a hackneyed academic trope, a way to signal theoretical sophistication while evading substantive questions as to what these systems actually are. Granted the concern is not entirely misplaced: beyond the intrinsic bias issue with modelling different possibilities for rationality, LLMs in their present iteration are simply not autonomous agents, with or without independent purposes. Furthermore, there are certainly actors who strategically leverage anthropomorphisms to incite moral panics in this sphere.

The deeper issue is that we necessarily reason about complex systems by imputing them with purpose, a mode of understanding abstracted from our own faculties. Hence the storied history of teleology in biology: can’t live with it, can’t live without it. Kant’s analysis of “natural ends”—paradigmatically concerning organisms—demonstrates that certain complex systems compel us to reason teleologically.

Where parts ‘are possible only through their relation to the whole’ and ‘reciprocally produce each other, as far as both their form and their combination is concerned’,

we cannot adequately investigate them without imputing purposivity.

This Kantian insight authorises what might be called rigorous anthropomorphism: the disciplined use of our faculties of end-orientation when confronting systems, like LLMs, whose distributed architecture exhibits similar reciprocal determination between component parts and emergent structure.

Pessoa captures the ontological configuration at work here in The Book of Disquiet: ‘I consider it neither a human nor a literary error to attribute a soul to the things we call inanimate. To be a thing is to be the object of an attribution. […] Everything comes from outside.’ LLMs literalise this. Their “intelligence” comes entirely from outside: from the corpus, from the linguistic environment in which they were trained. Yet we compulsively impose what Deleuze and Guattari call a “face” onto this inhuman-scale complexity: a unified point of reference, a subject-position from which the system can be addressed and interpreted. As they write in A Thousand Plateaus:

The face is the Icon proper to the signifying regime, the reterritorialisation internal to the system. The signifier reterritorialises on the face. The face is what gives the signifier substance; it is what fuels interpretation.

Of course, when we zip up complexity by consciously or unconsciously providing it with a face-icon, the possibilities for illusion multiply. However, the technique of anthropomorphisation, in this light, is not just an error but a structural necessity: we require a face, a centred subjectivity, to interface with distributed, decentred processes. For example, it helps to think of a swarm as a unity in order to coarse-grain its dynamics, even if we cannot compute the behaviours of all the member parts. It helps to give a name to social systems so as to describe aggregate behaviours, even if we cannot exhaust their internal complexity. This is how we navigate the different scales of complex systems in cognition. There is a whole meronymic logic to the process. The chatbot interface with its conversational “I”, its simulation of coherent personality, is precisely this process of providing a face, this reduction of aggregate statistical operations to the recognisable form of a conversational partner.

This constraint regarding the attribution of purpose operates at the level of the principium individuationis: the principle by which we distinguish one entity from another. Schopenhauer and Nietzsche understood individuation not as given but as process, a way of chopping up continuous reality into discrete units. When we ask “is this intelligence?” we are already individuating. We are immediately treating a given architecture as a discrete cognitive entity rather than as a process, a field of statistical relations, or a distributed archive. The question “is it intelligent?” actively presupposes a stable subject of predication, an individuated agent to which properties like intelligence can be attributed. But LLMs may resist such individuation either way. They may be better understood as processes of retrieval across a collective substrate rather than as singular entities that either possess or lack intelligence.

The reflexive invocation of anthropomorphisation as category error also obscures something even more fundamental: in a way, LLMs are human, they are the product of the enormous amount of human cognitive labour utilised in the training set. In fact, due to these vast spans of labour time, they are more human than human. These models are not unitary inhuman consciousness, but the traces of countless subjects at scale: hypersubjectivity. Despite the persistent attraction of science fictional doomsday scenarios, they are more Solaris than Skynet.

In Stanisław Lem's novel, the sentient ocean of Solaris does not possess intentions scrutable to human observers. Instead, it materialises patterns drawn from the buried memories of the researchers who study it. The ocean does not communicate; it reflects, generating manifestations extracted from contact with human minds. Every question posed to Solaris returns inflected by the questioner’s assumptions; every answer received is a reflection refracted through the ocean’s alien processing. The researchers are not discovering Solaris; they are encountering themselves, transformed and estranged.

LLMs operate according to an analogous logic. They do not generate “original” thought; they reflect back the crystallised linguistic productions of countless human speakers whose words constitute the training corpus. When we interact with an LLM, we are not communing with inhuman intelligence but engaging in a peculiar form of mediated collective communication. This is why the anthropomorphisation charge misses the mark. We are not mistakenly attributing human properties to something non-human; we are correctly recognising the human origin of the patterns the model has learned, even as we fail to grasp the scale and impersonality of the aggregation. The confusion arises because LLMs give us access to something unprecedented: hypersubjective symbolic production, or the cognitive labour of the multitude compressed into a near-instantaneously retrievable form.

The discipline of “Solaristics” in Lem's novel, the centuries of failed attempts to interpret the planetary ocean’s manifestations, prefigures our current hermeneutic predicament. Just as Solaris researchers projected psychological frameworks onto phenomena that exceeded them,

we approach LLMs with questions shaped by our own assumptions. The inquiry is contaminated from the start. The model’s responses are, in turn, shaped by biases present in the corpus: not objective retrievals but historically situated, ideologically saturated recombinations of attested language use. There is no neutral standpoint from which to query the aggregate; every prompt is already an interpretation, and every output a counter-interpretation drawn from the same compromised archive.

Kant's characterisation of memory from his anthropology lectures applies with uncanny precision: ‘We presuppose a storehouse of the imagination where all acquired representations lie in the dark, unilluminated. How this may occur we cannot understand. Memory is like an archivist.’

LLMs are exactly this: archivists of the collective imagination, storehouses where the representations of countless speakers lie compressed. The retrieval mechanism, next-token prediction, sequentially illuminates fragments of this archive without ever grasping the whole.

This is why model collapse, the degradation that occurs when LLMs are incestuously trained on synthetic data generated by other LLMs, is so telling. It reveals that the apparent intelligence of these systems as they stand is parasitic on the embodied, world-engaged production of human speech. Remove that substrate and the system degenerates into iterative impoverishment, like the generation loss incurred by photocopies of photocopies or tapes of tapes. The current models work because language itself is already saturated with human intelligence; they tend to fail when that intelligence is absent.

The proper question, then, is not whether we are anthropomorphising LLMs but whether we adequately recognise what they reflect back to us. They are not distant, alien minds but archives rendered interactive: collective human cognition compressed, indexed, and made searchable through statistical geometry. We are not communicating with the model; we communicate through it, with the manifold contributor consciousnesses whose cognitive labour it has encoded. The experience of LLM interaction is not yet an encounter with the superhuman but an encounter with the human, at scales and in configurations that are unprecedented.

The Wall Across the Future

As much as Reason would love to extrapolate, we cannot see beyond what Ernst Jünger called the “wall of time” [Zeitmauer].

We should therefore distrust those peddling schlocky time-travel ethics and alignment gurus claiming special foresight into the behaviour of systems that they admit from the outset will exceed their own comprehension. We should be suspicious that AI doomerism and positivism form a feedback loop that resembles capitalism’s longstanding ability to absorb its own direct critique into its apparatus of reproduction. In the words of Paul Valéry: ‘all prediction is conservative: it demands that we be as are in whatever future it constructs.’

Often, the most adventurous flights of futurism mask a reactionary impulse to be sheltered from the radical contingency of what is to come.

Still, the strong deflationary position is equally untenable. To dismiss what has emerged from transformer architectures as a bag of statistical tricks is to wilfully ignore real discontinuities. The capacity to learn compressed representations of human linguistic production at this scale would have seemed fantastical mere decades ago. The frontier is already shifting: spatial reasoning models are beginning to operate over geometric and physical domains, not merely linguistic ones; function-calling APIs allow LLMs to trigger executable actions and act as components in larger agentic systems; advanced coding assistants now demonstrate recursive self-improvement in their prompting and planning strategies; LLMs are being integrated with proof assistants to arrive at novel mathematical results. Whether applied to token sequences, spatial configurations, or multimodal data, the learning mechanisms that enable statistical compression may well prove crucial to whatever eventually constitutes AGI, even if current architectures are insufficient. We face something genuinely novel, the trajectory of which remains radically uncertain, but which, either way, is not reducible to eschatological fantasy or offhand dismissal.

Recent reasoning models represent an architectural shift: rather than immediate responding, they perform extended “internal” deliberation, generating chains of thought before producing outputs. It remains to be seen if the limitation regarding efficient generalisation persists: these models operate by passing acquired context forward through extended inference, essentially re-prompting themselves. They learn broader patterns of reasoning such as problem decomposition strategies, verification steps, and common failure modes. Each step in this mode of sequential processing inherits the constraints of that which preceded it. The importance of the path dependence of these sequences of context acquisition is compelling. It brings to mind Kant’s division of all cognition into historical cognition, cognitio ex datis, and rational cognition, cognitio ex principiis.

The former is the realm of cognition via aggregation of detail, while the latter that of employing the compact generality of the concept. Models in their present form clearly excel at historical cognition. An attempt to coordinate the two modes is underway.

The most promising developments may lie not in scaling individual models but in composing them into multiagent architectures. Where monolithic systems are facing fundamental limits, distributed collectives of specialised agents coordinating through structured interaction may be able to access qualitatively different computational regimes. Regardless of explicit architectural intent, the deeper LLMs embed themselves in economic activity, the more they will organically give rise to multi-agent behaviour. Their involvement in a rapidly expanding network of local–global exchanges

makes such dynamics a systemic by-product.

This would represent not the emergence of a singular superintelligence but the organisation of collective intelligence at unprecedented scales: dialogic swarm cognition instead of machinic monotheism.

Beyond ectypal human intelligence, confined to cognition through the concept, and beyond the hyperectypal operations of language models, churning their way through representations-of-representations in the tombs,

stands the regulative idea of an absolute intellectus archetypus: cognition without mediation; thought without the gap between sign and referent. Thought without thoughts.

Transcendental critique demonstrates that an idea can be logically impossible and still serve a purpose. It may be that the concept of superintelligence ultimately serves an ‘indispensably necessary regulative use, namely that of directing the understanding to a certain goal respecting which the lines of direction of all its rules converge at one point.’

The myth itself may coordinate individual thought, and, at the scale of the social system, productive forces, toward its own realisation, whether partial or complete. Nonetheless—and here we must part ways with those who would cut the overwhelming prospect of artificial intelligence down to size by projecting their own finite intellect onto that which is supposed to exceed it—it must continue to be understood as ‘an idea (focus imaginarius) — i.e., a point from which the concepts of the understanding do not really proceed, since it lies entirely outside the bounds of possible experience.’ The danger is that this is where ‘the deception [arises], as if these lines of direction were shot out from an object lying outside the field of possible empirical cognition (just as objects are seen behind the surface of a mirror).’

The ideal engineer subordinates the means to the end, dragging the abstract into the concrete. But true self-fulfilling prophecies bootstrap themselves from germinal possibilities, not from neurotically overdetermined fantasies that pinion that which they purport to predict. To be clear, seeking orientative ideas is not itself a demerit, because, under the conditions of modernity, ‘the main aim in the modes of understanding […] is to hit upon something successfully, to have a fate, since fatelessness, δυσμορον, is our weakness.’

The problem is not the bare use of the virtual but its illegitimate actualisation.

Despite the widespread epistemic confusion and moral panic brought about by the initial successes of the transformer architecture, signs of what may come are there for all to read. The internet is no longer simply a trap for the eventual harvesting of data

but a laboratory for its recombination from a civilisation-scale data-exhaust.

We are passing through a metasystem transition, an abstract catastrophe, between regimes of signs.

No matter the predictions, the moral fantasies, the forms taken by our foci imaginarii, the bleating of dogmatic cheerleaders and naysayers... all of this noise, this frenzied standstill, will resolve into the nervous hush of anticipation before the raising of a curtain.

Ideas alone can only draw us so far into the future. Beyond the wall of time, something will arise, dripping, from the sea of information.

Thomas Murphy is a writer and software developer based in the UK.