**Links to**: [[Postulate]], [[Problems]], [[Question]], [[Philosophy]], [[Appendix]], [[Language]], [[Index]], [[Bibliography]], [[13 Conclusion]].
# THE _POLTERGEIST_ IN THE MACHINE
## _Sedimented, semantic, social and spectral perspectives on prediction in the artificial mind_
 
>Every intelligent ghost must contain a machine.^[“If human thought, feelings, sensations, decisions, emotions, moods etc. are amenable to any kind of scientific analysis and explanation then it must be because we are made of appropriate kinds of (virtual) machines: Every intelligent ghost must contain a machine. Alternatively, there may be inexplicable, magical, phenomena in the universe. So far I have seen no evidence for this, only evidence for our currently limited ability to understand and explain, and evidence that the frontiers are constantly being pushed forward.” Sloman 1990 p. 13.]
>
>A. Sloman, 1990, _AISB_ quarterly, p. 13.
 
>Experiment, never interpret. Make programmes, never make phantasms. [...] Programmes are not manifestos—still less are they phantasms—*but means of providing reference points for an experiment which exceeds our capacities to foresee*.^[Please note the original is reversed, where the third sentence appears slightly earlier in the text, before the first two. On “interpretation”: please also note that throughout this project it is often preferred _not_ to explain citations away, but to let them do their work. Without an explanation, more computation on the reader’s behalf, and less violation of the original grounds.]
>
>Parnet and Deleuze, 1987 (1977), *Dialogues II*, p. 36.
<div class="page-break" style="page-break-before: always;"></div>
# _The Poltergeist in the Machine_
This is a thesis on some philosophical considerations regarding the complex developments of so-called artificial intelligence (AI), with language-modelling as its main object of analysis. Arguments can and have been made about how it is (becoming increasingly) important to look at AI through a philosophical lens. An increasing number of figures of high influence in the public perception of AI have,^[To name but a few: Amanda Askell, François Chollet, Lex Fridman, Gary Marcus, Stuart Russell, Melanie Mitchell, Stephen Wolfram, Joscha Bach, Blaise Agüera & Arcas, Joshua Bengio, Jürgen Schmidhuber, and many others. Please note we need more influential _others_.] in the course of the last few years, emphatically promoted variations on the idea that we need philosophy to frame and fundament the development of AI. In the scholarly realm, expectable, perhaps: myriads more propose careful, slow philosophical analyses.^[E.g., van Rooij et al. 2023 suggest AI should be a “slow (computational cognitive) science.” p. 15. “[It may] seem excruciatingly slow compared to the apparent speed of progress in today’s machine learning approaches to AI. But to genuinely make progress we need to go this slowly, and in fact cannot go any faster (Adolfi & van Rooij, 2023; Rich et al., 2021.).” Ibid.] In thinking what a philosophy of AI might entail, the first challenge would be to define—or at least disentangle—*artificiality*, *intelligence* and *philosophy* to a satisfactory degree.^[This is why provisional definitions are given later on in this introduction, as well as in the appendix, not to limit the possible functions of these concepts but to focus their function and scope within _this_ project. Please note that we continually refer to “AI” or “GenAI”, as this signals the broad landscape of the “mind outside the brain” project, but the examples we draw from can be more aptly referred to as Machine Learning (ML).] This project _sort of_ proves this is impossible: what concepts do, as philosophical problem-machines, is present the possibility of _new_ concepts. Contemplating philosophy as the artificial _excess_ of mind—socially-, materially-, etc. distributed: outside itself, evolving through concepts—one can also claim that AI has _always_ been the preoccupation of philosophy, and vice versa. Indeed, if philosophy is understood as one of most complex evolutionary metalearning affairs: a historical endeavor where the mind evolves by recursively witnessing the mind, and AI has as its aim the understanding—and possible imitation—of (certain information-processing aspects of) the mind, then there is no separating one enterprise from the other. The proliferation of concepts stacking upon concepts—something one could call “_Geist_,” but also many other things: the human mind-type collectively understood and evolving through various structures—is a system which can now be philosophically recast in its new _artificial_ light (van Tuinen 2020). Given the unfolding technical infrastructure, it can now be understood as an _organizing inorganic_ (Hui 2019): a condition the organic has worked itself into, by doubling down—through externalization and the abstractive reflection this allows—on some of its own mechanisms.
This thesis is a program _for_ phantasms: primarily, it hopes to become the starting point of a conversation with linguistic agents, carbon or silicon-based. It is a series of conceptual proposals intended for the future modulation of _natural-language-programmable AI_. The hypothesis of this project is that controversial (philosophical) concepts^[Some of the ones at play in this work, to exemplify: “intuition”, “illness”, “scale”, “negation.”]—currently residing and evolving through intricate, distributed networks of assumptions, formalizations, sociocultural habits—will become, in their probing and recombinations, central objects of analysis for future language-model-like systems. The phenomenon we call _generative_ in “generative AI”—where the interest for us is in language-modeling—reveals all combinations of things we can _identify_ are **valid**, their generative relevance is whatever we consider salient at that moment in history. Stephen Wolfram speaks of _interconcept space_ (2023), as that which can be explored between current saliences. The drive behind this project is based on a simple question: **if current language models represent statistical mappings of linguistic plasticity, how can experimental philosophical contortion explore their possible operational futures?** If philosophy can largely be considered the coming to terms with the plasticity of language, the eternal reorganization of categories; saliences, and if we currently explore this plasticity through methods such as language models, how does the creation of new concepts interfere with these mappings? An additional question explored is the (anti)_sociality_ of concepts: if salience is observer-relative, are some concepts “better” than others at revealing the perspectivism inherent to their mechanics? Innovation from _outside_ language-modeling can induce statistical changes in future models, which will then come to operate through altered conceptual relationships, and therefore novel saliences.
In creating new philosophical concepts or reframing existing ones, we alter the probability distributions that language models learn. Mathematician and cryptographer Kristin E. Lauter, who is currently senior research director of Meta AI, says: “Machine learning is simply modeling from noisy data.”^[*Joint Mathematics Meetings* lecture, January 2025.] Deep learning scholar Simon Prince alerts us of the violent, volatile, capitalist and inherently unpredictable of trajectories AI, warranting care and caution (Prince 2023, p. 14). All “AI” modeling is subject to infinite types of noise: whether intentional or unintentional, all moves are inherently risky. Prince reminds us of the problems inherent to modeling from known noises: biases from the past inevitably creep into our unfolding worldviews (ibid., p. 13). Modeling with and from noise implies classic ontological distinctions between relevance and interference, which unavoidably leads to questions of the (chrono)politics of what patterns are to be considered salient, and therefore what accounts of _sameness_ we can narrate in order to cohere (biologically, politically, culturally, etc.).
Noise—the “polter” in our “Poltergeist”^[I only recently realized that the *psofos* (ψόφος) in my friend and colleague’s _Psofotopias_ (Raponi 2025) is perhaps the relative of _polter_ as noise. Some etymological speculations also relate the term onomatopeically as perhaps inheriting from the sound made by something literally dropping dead. Also noteworthy, on the topic of speculative metaphysics and AI, is that Simon Prince calls his own 2023 book the “spiritual successor to Deep Learning (Goodfellow et al., 2016)” (2023, p. 15). What is such a reference doing in such a technical book? These are important questions.]—will be a chief concern, taking various shapes in different versions of our arguments.^[Inspired by the people who have taught me all I’ve gathered around this concept: the Noise Research Union ([[NRU]], that is: Cécile Malaspina, Miguel Prado, Inigo Wilkins, Martina Raponi & Mattin.] Questioning the boundaries between information and noise (Shannon & Weaver 1948), or repetition and difference (Deleuze 1968), or recursivity and contingency (Hui 2019), or invariance and transformation (Shepard 1990), is a matter of accounting for notions of _equivalence_ versus _difference_ across domains, one which offers a schematic point of inquiry which allows for the consideration of _basic patterns_ as possibly translatable between different disciplines.^[Or, as Viveiros de Castro puts it, where the same could be said of the anthrophilosophy of AI: “...it could in all modesty be said that the future of the master concept of anthropology—relation—depends on how much attention the discipline will end up lending to the concepts of difference and multiplicity, becoming and disjunctive synthesis.” (2014 (2009), p. 170).] Insisting on the idea that new concepts (as events: transformations, salient differences) can be considered as introducing _noise_ into the statistical structures of language models (invariances, repetitions), this project seeks to explore how they unfold under current—and possible future—dialogical constraints. New concepts could be said to represent a kind of productive interference, as they elicit interpretation. Through traditional means (books, lectures, classes, conversations, etc.), the creation of conceptual novelties becomes subject to what we could call a kind of _philosophical denoising_: as they enter into dialogue with their substrate they undergo a process of _fitting_ into—or _falling_ out of—the canon. Philosophy, previously only evolving through conversations, papers and books, is now parsed through the machine (e.g., through language-modeling), a different kind of project which elicits different ways of understanding this fitting or _parsing_.
Please note that in our search for frictions, we use the word _parse_ technically but flexibly (sometimes interchanged with _chunk_), it can be applied to any system which filters (language) data, human or otherwise. To chunk is to, e.g., tokenize: to decide what counts as atomic or salient in a communication system. To parse is to _process a chunk_, to consume a segment, and (in most cases, therefore) to produce an output, to _compute_.^[For an incursion into tokenization and NLP in more technical terms please see Gastaldi et al., 2024.] In psycholinguistics, _parsing_ is terminologically used in this way, too: it is distinction by conceptual categorization as well as the incremental understanding of meaning as applied via syntax. In A. A. Cavia’s _Logiciel: Six Seminars on Computational Reason_ (2022) this is referred to as _encoding_; the assigning of a token to a type (Cavia 2022, p. 13), in our formulation: _chunking_ is contemplating a token (isolating an event, a singularity), and parsing is making it predictable under a type (a way to _parse_ reality).^[Cavia: “The interpretation of encoding can take the form of a program that outputs instances of [a] type or alternatively the construction of a class of paths in a continuous space.” (ibid.). See: [[Parse]] and [[Chunk]].]
Following Bernhard Rieder, our chunking and parsing takes much from his exploration of _interested_ **orderings** of reality (2020): where historical momentum, directionality: _bias_, is what makes salience _salient_.^[He notes conceptual *dis*-congruence of his view with David Weinberger’s, but congruence in this sentence: “‘How we choose to slice [information] up depends on why we’re slicing it up’ (Weinberger, 2008, p. 82).” (Rieder 2020, p. 32). More on _slicing_ and _cutting_ metaphors later, but already note that etymologically, it’s chunking all the way down, **to compute is to cut**: from the Latin _computare_ “to count, sum up, reckon together,” which resulted from the merging of _com_ “with, together” and _putare_ “to reckon,” which originally meant “to prune,” from the proto-indo-euro root _pau_- “to cut, strike, stamp.” Source: etymonline.com, accessed January 12 2023.] To chunk, question, analyze and exploit this parsing, a returning point of insistence throughout this project is that elusive concepts such as _ghost_, _soul_, _imagination_, and other traditionally fuzzy logics, should be continually contrasted with and tied down to what can be considered more concrete, crude things such as _machines_, _meat_ and _matter_. To cite Rieder again: “The way the computer serves as a[n epistemic] center should not be understood primarily through notions like quantification, formalization, or unification around a central paradigmatic reference like Shannon’s information theory or logic-based computationalism. These elements are far from irrelevant, but, as the history of information retrieval shows, the computer’s capacity to connect and to compel stems primarily from its capacity to _function_.” (2020, pp. 258-9.) This is the logic of “function” as in “we do not understand it unless we can build it”, resounding everywhere in ML. In this project we observe _function_ as something which remains invariant until time _x_ (see entry on “Grue and Bleen” in Appendix). We observe _function_ as a _transferrable_ quality or pattern between phenomena (e.g., _fluidity_ as a _function_ of different liquids, _balance_ is a function of many phenomena constrained by gravity, _filtering_ is a function of many feeding and porous systems, etc.). Once a function comes to recursively function on itself (from fluids in fluids to metacognition), novel functions evolve. This proposal is our version of advancing technodiversity, with language being the most obvious tool at mouth, “by inventing [yet] another recursive process” (Hui 2019, p. 224.). Our insistence on _function_ can be understood as inheriting from media and information philosophy traditions, where the focus on medium takes precedence over content. In our context, function represents both the idea of *process* as well as of *identity*.
This project takes both the poetic license (and risk) and the rigor of philosophy by questioning and, at the same time, applying the creatively literary to the formal, and vice versa. In a time of high specialization,^[Which is much needed, but cannot be only _that_, as this leads to incommunication.] it is crucial that translation between modes, fields and epochs becomes a point of attention: “advances in abstraction and the new conceptual formations that already exist or are emerging in interdisciplinary contexts should be made usable” (Luhmann 1995, p. 11).^[The quote ends: “in sociological research”, ibid. We apply it across the board here.] This project makes an attempt at abstractive usability, even if it is a failed one, because it refuses to say what something ought to be _for_.^[This is partly inspired by Deleuze following Whitehead: “the abstract does not explain, but must itself be explained” (Deleuze & Parnet 1977 (1987), p. vii), as well as by Sellars in seeking the way in which things hang together in the _most general_ (i.e., abstract) sense as the task of philosophy (Sellars 1962, p. 1). Given that Deleuze, in this same citation, also points to how “the aim is not to rediscover the eternal or the universal,” but to seek how it breaks, through what he calls _creativeness_ (ibid.), one may interpret these as polar opposites, as incompatible (particularly as they represent exemplary continental and analytic approaches), but what this thesis partly intends to show is how they can be seen as overlapping through a _functional_ explanation (see, e.g., [[04 Concepts as pre-dictions]].)]
The functional computation of information is the discovery that making things equivalent results in ever-diversifying, generative problems, making the organic come to terms with its own differential matter: systems which attempt to contain themselves always result in excessive paradoxes. This project therefore explores computation as an open-ended metaphor, and as a concrete chunk-parse activity. In this line, chasing after metaphors, Stengers notes that Whitehead suggested how “Plato’s ideas are those things that first of all erotically lure the human soul—or, we could say, ‘animate’ humans.” (Stengers 2012, p. 5). Our predicament being AI: the animation of the material could not be better (metaphorically)^[Following our interest in seeking the (predictive) functions in concepts and how they emerge from the metaphorical: “metaphorical style is the sign of a plenitude of life, just as “demonstrative” style indicates a poverty of life. The deliberate use of metaphors affirms life, just as the privileging of concepts reveals a will to nothingness, an adherence to the ascetic ideal” (Kofman, _Nietzsche and Metaphor_, p. 19).] represented as an inseparable union any more emphatically than in the 2024 _physics_ Nobel prize awarded to Geoffrey Hinton for the development of the perplexing methods of _deep learning_.^[The ground for scientific and academic integrity behind this award has been questioned by Jürgen Schmidhuber in his 2022 [“Annotated History of Modern AI and Deep Learning”](https://people.idsia.ch/~juergen/deep-learning-history.html), revealing another history of said developments which deserves missing credit, one ignored by Hinton (and Bengio and LeCun, with whom he shared the 2018 Turing prize).] As far as we can implement things computationally, or understand the lives and deaths of organisms: there is nothing more than inputs and outputs, and paths between them. However, saying this, is precisely what makes it possible to explore the possibility that there may be something more.^[Wittgenstein’s (generative) expression that we must be silent of that which cannot be expressed in language seems fair insofar as what that pronouncement can be understood as doing is demarcating the limit from which we may leap into the next metaperspective, or metacognitive phase. If anything because we need to stay with the trouble (Haraway) of what “nonsense” might entail, in many senses of the word: as not able to conceptually; i.e., predictively, track the unfolding of reality; or as able to _make_ sense, to become experiential; or as seeking novelty in or through the negative. Or because, related to the last, as Adorno has observed: Wittgenstein’s observation is (in our sense productively) anti-philosophical: “If philosophy could ever be defined, it would be as the effort to say that which cannot be spoken of” (cited and translated by McKeane 2018, n. pag.). If we read Wittgenstein with utmost charity in the _Tractatus_, as the Wittgenstein who became the writer of the _Philosophical Investigations_ (1953), we can reread: “What can be said at all can be said clearly; and whereof one cannot speak thereof one must be silent. The book will, therefore, draw a limit to thinking, or rather—**not to thinking**, **but to the expression of thoughts;** for, in order to draw a limit to thinking we should have to be able to think both sides of this limit (**we should therefore have to be able to think what cannot be thought**). The limit can, therefore, only be drawn in language and what lies on the other side of the limit will simply be nonsense.” (Foreword _Tractatus Logico-Philosophicus_, Ogden/Ramsey translation, our emphasis in bold), as indicating the possibility of another perspective, one which plunges into the productivity of _nonsense_. We cannot think what cannot be thought, but saying this does pry _something_ open. We follow Patricia Reed here, too: “If sensitivity to interference is a minimum criterion for the possible transformation of a world, under what conditions can given world configurations of insensitivity be suspended? How is it possible to localize conditions of otherworldly thought that are not absolutely beholden to the here and now of the local belonging to a given world configuration? In other words, how to localize speculative thought (which necessitates the construction of a world to embed it), and what are the consequences of such activity on an epistemological register?” In: “Thinking Beyond Experience: Prosthetic Agency in The Form of Not”, in _The Unmanned_, 2022, p. 5. More on this in [[12 Negintelligibility]].]
 
### One massive metaphor: proposing possible, coming abstractions
There are so many ghosts, phantasms, spirits, souls.^[Please see: [[Geist]]. Sometimes references are best left unsaid, but the reader should keep in mind the _hauntological_ (Brooke-Rose, Derrida, Fisher); lurking persistence of the past in everything that is supposedly new, as well as the opposite: the spectral being that which is absolutely elusive and non-graspable; (Wittgensteinian) nonsense; the distant future; the virtual (Deleuze); essences (as ultimately undefinable qualities), etc.] The practice of philosophy is a form of _textual animism_ (Stengers 2012, p. 1): we are bound to keep breathing life into that which already was or something which could be. Why _poltergeist?_ Initially, influenced by a recently renewed interest in the _extended/external sociality of mind_ in the work of Hegel and its possible relations to the concept-turned-project of artificial intelligence (e.g., R. Negarestani 2018),^[But also Honneth 2014, Pippin 2018, Brandom 2019, and more.] it was “_Geist_ in the machine.” This title seemed, to me, self-explanatory: the mind/animus/spirit *is* the machine (the power, the _maker_^[Etymonline: from PIE _magh-ana_- as that which _enables_, root: _magh_- “to be able, have power.” Interestingly, the word “mogen” in Dutch is of the same root, and the word “vermogen”, meaning _capacity_ or _ability_, can be literally translated to far- (as in distance, extension) ability. _Vergmogen_ is crucially also the word for _asset_, in finance. What these incursions reveal, to us, is the _predictive_, future-oriented force of concepts, an effect that will be explored and exploited throughout the project.] of sense in sense-_making_) and is _in_ and _of_ the machine, _Hegel ex machina_ style. That is, an emergent teleonomy we are, as of yet, unable to “fully” “explain,”^[I cannot retrieve the reference from notes, but Borges seems to have said of Schopenhauer that he only wrote _Die Welt als Wille und Vorstellung_ to explain the title to those who don’t [already] get it.] partly because we are both observer and observed, and because organic systems are “deeply interdependent with one another, simultaneously made and maker.” (Bongard and Levin 2021, p. 2). Thinking in terms of function: does the tape—the momentary chunk, the data, the pattern—drive the machine program—the parser—or does the machine program exert its function “on” the tape?
But the haunting had to be more absurd, because, well, _look around you_: we find ourselves in a rather ridiculous situation in which move-fast, capitalist _AI-as-engineering_ (van Rooij et al., 2023), rather than AI as a slow and careful tool, does quite some harm and in many cases _reduces_, rather than sustains, the amplitudes of this distributed cognition. In these and other ways, then, the _poltergeist_ haunts, and _mocks_—laughs _and_ imitates—and in the folklore: it does so with annoying sounds.^[This also being a hint at this project’s investigations into [[Noise]], again: much indebted and thanks to my friends, the [[NRU]].] You never see it, but can certainly feel it loom. It is omnipresent but evasive, characteristically impossible to capture because, perhaps, _it does not exist_. It is what _animates_,^[I.e., an _anima_; a soul.] and it is necessarily _spectral_. Social things evolve sociality as attractors that (convince them to) move beyond their carnally-bounded selves. As Winkelman (2024) argues in the context of interpreting the “natural” versus “cultural” status of “_super_-natural” beliefs: social environments do not determine whether supernatural beliefs exist, _all cultures have supernatural beliefs_. If this constant does not vary and all cultures see ghosts of one kind or another, this means, to Winkelman, that this might be rather _natural_, i.e., determined by evolutionary preferences. Is the cell-turned-self-turned-god-turned-machine a kind of _function_ striving for what is **_not_** there—meanwhile rendering concepts such as *nothingness*, *infinity* or *immortality*—now finding a new phase-change in what we call “AI”? A ghost that always was, propelling itself towards ever-renewed ghostly practices.
_The Poltergeist in the Machine._^[Invisible books, mere titles, and their promises are always more interesting than _actual_ proposals, Viveiros de Castro reflects on this ghostly effect in _Cannibal Metaphysics_: “fictional works (or, rather, invisible works) that Borges was the best at commenting on ... are often far more interesting than the visible works themselves” (2014 (2009), p. 39.)] How readily does the concept of the _mental_ jump out? Souls? Cartesian dualism? How present is Hegel as an interlocutor? Ryle? Why does this white mythology’s^[“What is white mythology? It is metaphysics which has effaced in itself that fabulous scene which brought it into being, and which yet remains, active and stirring, inscribed in white ink, an invisible drawing covered over in the palimpsest.” J. Derrida, _White Mythology_, 1971, p. 11. In this extended essay Derrida argues for attention to the importance of metaphor over metaphysics. More on this subject in the first chapter, [[03 Semantic noise]].] authority of incestuous pedigrees continue to structurally veil _and_ functionally haunt so much text? How much expectation is there of _actual_ talk of ghosts with such a title? If it is accepted that any exchange, any communication, harbors and delivers much more than what may be gleaned by a quick _literal_ assessment, then one could propose that the best course of action is to let a title run wild. What _does_ a reader _do_ with the thought(s) in a title? If a title provides initial semantic (at)traction, then what follows after the title is supposed to contain this. Contain it: i.e., restrain it _and_ give substance to it. Somehow, the text _explains_ the title because the title is not enough, and at the same time too much. Without a title, however, there is no indication of direction, so the title _does_ explain the text.^[A consideration which ultimately leads us down the search for sufficient origins: what explains _us_?] But if perfect propositional prose is impossible—its proposition is at least questionable—then it also seems logical to say that what follows the title is just a long series of additional title-like things that keep postponing the containment until it’s too late. Or perhaps too much. Which is not the same as _enough_, though we often say so. Philosophy always agrees to disagree with itself.^[This is what we will consider, in the last chapter, as rather _negintelligible_.]
The subtitle signals how these (dys)functional dimensions will be explored: the perspectives presented are understood to be **sedimented** because they are assumed to emerge from an unavoidable and irreversible material history (this is explored in, e.g., “Falling into Place”, “Phenomenology of Sick Spirit” and “C is for Communism and Constraint”); they are **semantic** because they contemplate the nature of meaning(s) and their possible teleologies (this is explored in, e.g., “Semantic Noise” and “Concepts as Pre-dictions”); they are **social** because they are inevitably dialogical and distributed (e.g., “Principle of Sufficient Interest”, “Post-Control Script Societies”), and they are **spectral** because they are metaphysical: transcendentally (and existentially) haunting (e.g., “Prediction” and “Negintelligibility”). They imply futures past and beyond our own, where the ghost is and the ghost is not.^[In our dissatisfied search for the _general_, especially in the context of artificial _general_ intelligence, we cannot but avoid citing Whitehead on the subject of elusive ghosts of generality, in the realm of mathematics: “The study of mathematics is apt to commence in disappointment. ... like the ghost of Hamlet’s father, this great science eludes the efforts of our mental weapons to grasp it—’Tis here, ’tis there, ’tis gone—and what we do see does not suggest the same excuse for illusiveness as sufficed for the ghost, that it is too noble for our gross methods.” A. N. Whitehead, _Introduction to Mathematics_ (1911), p. 7. We should not mystify mathematics any more than we mystify the elusive idea of AGI. Researchers such as Prince (cited earlier, 2023), or Emily Bender insist on abandoning the quest as one of “AI” altogether, and refer instead to concrete methods such as machine learning and even more specific terms (Hanna and Bender, 2023). Again, we stick with “AI” more often than not, because we speak of a general project, traversing Democritus (Aaronson 2013), Ada Lovelace and current trends in deep learning.] We move this way until we find something better: linguistic generalizing—a form of computing, as will be argued—in the cosmic washing machine, is what we currently have access to in the search for diverse complexification. As will be argued, concepts can be seen as dynamic _attractors_, and their fuzzy mechanics enable us to _chunk_ and _parse_,^[More on these terms in [[Chunk]] and [[Parse]].] predicting different variants on collectivity: we hook onto these structures as much as they hook onto us, such as in the weaving of this text and your reading of it. Otherwise we would calculate. But calculation—as relative, _chunking_ computation: measuring something up against something else, resulting in sometimes but not always either-or zero-sums—whatever it might imply (often: the chunking and parsing of relative perspectives) is entrenched within the confines of our current fleshy distribution: one which is understood primarily through concepts.^[In Sellarsian terms: manifest images guide the scientific ones, for now, or perhaps forever if complex technoscientific apparatuses inherit our concepts, which is what we question here.] Moreover, in the context of this project—much of which rests on the premises and promises of active inference (AIF)—if what we are often interested in computing are **probabilities**, which are _possible_ observables, never with fully certain guarantees, then our computations will inevitably be open-ended, mortal (Ororbia & Friston 2023), and particularly interested (Rieder 2020) by the milieu they exist within. Through AIF, a relativistic analysis becomes possible because what is on offer is analysis of predictive inference as probabilistically-perspectival, observer-dependent, which sets the agenda for an understanding of our given condition as one of **chance**, alleviating the castigation of cybernetics as a paradigm of accurate **control**.
Taking active inference as one of our starting points provides a modeling entry into chunks and parsings, where we can usefully compare and contrast computational limits with thermodynamic constraints. When viewed through the free energy principle, information changes in mortal computers can be read as inferences: updates to _prior_ saliences. When this is cast as a complex, multilayered process of surprise-minimization which _must_ share the energy-efficiency constraints of any and all thermodynamic processes “it then follows that the **statistical and thermodynamic efficiency of any system — to which the free energy principle applies — are two sides of the same coin**.” (Ororbia and Friston, 2023, p. 3, my emphasis in bold).^[In terms of “mortal computation”, Ororbia and Friston explain: “[D]ynamics are interpreted as a gradient flow on variational free energy or, equivalently, (the logarithm of) Bayesian model evidence. This stands in contrast with the simulation of inference in modern-day computers, e.g., von Neumann architectures. In this simulation setting, there are additional thermodynamic costs that are induced by reading and writing to computer memory, i.e., the power consumption of moving data between off-chip memory units and a computing processor is nearly 100 times greater than the floating-point operations on those data. These costs are sometimes associated with the ‘von Neumann bottleneck’ or the ‘memory wall’. This means that, in order to realize the potential thermodynamic efficiency of Bayesian computations, it is necessary to implement belief updating in memory, also known as in-memory processing or processing-in-memory.” ibid. More on this in [[Mortal computation]].] What offers promise, to us, is that AIF seeks to unfold a multilayered, 5E, naturalized understanding of the functions of life, through a practically modellable proposal, which ultimately rests on a highly panpsychist, perspectival materialism.^[Its constructively Kantian in that “[u]nlike ‘passive’, perceptual inference processes (e.g., inferring the presence of an external object) based on patterns of light impinging on the retina), the inferences underlying decision-making are ‘active’, in the sense that the agent infers the actions most likely to generate preferred sensory input (e.g., inferring that eating some food will reduce a feeling of hunger). Agents also infer the actions most likely to reduce uncertainty and facilitate learning (e.g., inferring that opening the fridge will reveal available food options). This leads decision-making to favor actions that optimize a trade-off between maximizing reward and information gain.” (Smith et al., 2022, p. 2).] Of course, the intricacies of this are many, and active inference is still an emerging, novel approach with many open-ended questions. Outlining these dynamics depends upon the dynamics of an already-existing complex web of unfolding language(s): of science, of politics, of life-reproduction in general. This project therefore embarks on investigations which seek symmetries and asymmetries between the (sometimes much too apolitical) neutrality of (active inference) models, and the harsh situatedness of much of what we call ‘life’, in order to attenuate the pace of grand narratives, such as AI (by creating novel f(r)ictions). A complex system such as an organelle, can hereby be compared to a complex system such as a concept, if they are both understood as structures effectuating a **generative model** in adaptive ‘conversation’ with their environment. In (unsupervised) machine learning, a generative model is the mechanism which ‘learns’ to synthesize examples that are statistically indistinguishable from its training data (Prince 2023, p. 7). In AIF, the concept is borrowed (and technically implemented) to refer to the inference model of any agent, as that which best represents its world, given observables. A generative model is therefore, for our purposes, a dynamic process of reality-abstraction where an agent (be it cell or society) actively generates reality-tracking abstractions; _syntheses_, about the hidden, latent causes of its sensory experiences. The generative model of something like ‘philosophy’, can be understood as that which constructs vast conceptual patterns that aim to track _thought_ through a distributed dialectical refinement of predictive abstractions (expectations) about unobservable states of reality (surprise), presenting a possible isomorphism between philosophical reason and the fundamental mechanisms of _anything_ that self-organizes.
If we desire to slow things down (Stengers 2005, 2018), we need to continue to deal with the conceptuality at hand. One which is generalistic, inevitably and polycomputationally-interestingly erroneous,^[See: [[Polycomputation]], [[Computation]], [[Interest]] and [[Error]].] and makes an effort to explain itself in the models of science, and in the sonic, haunting poetry of words. _Inspired_ by Margaret Masterman:^[Foundational yet often forgotten NLP scholar who challenged the idealized language conditions proposed by the British School of Linguistics, as well as the (also idealized) syntax-focus of the Chomsky school, (see: Masterman (ed. Wilks) 2005). “Inspired” is italicized to stress a focus of Masterman’s, she paid crucial attention to _breath_ as guiding the rhythms of language, and therefore meaning. Her ideas of how concepts are open-ended pointings into the future is briefly treated in [[04 Concepts as pre-dictions]]. Her work on this was strongly influenced by Wittgenstein, and by L.E.J. Brouwer.] the latter cannot continue to be ignored. Urged by the poetry of mathematics:^[See, e.g., Catren 2016, Grietzer 2017, on the _life_ between poetry and mathematics.] their union is obvious. The title, then, says: there is something, most likely an elusive and simple myth, which keeps this annoying promethean fire burning. The _illusory_ aspect of the ghostly plays an important role in this title, too, as **illusion** is one of the fundamental concepts of this project, where our proposal, following from the AIF view that all perception is hallucinatory is that—auditory, visual, etc.—illusions actually allow for metacognitive entry points into perception itself (see: [[Illusion]]).^[The concept of _seeing as_ or metacognitive perspectivism is best exemplified by Varela here: “I suppose this is why paradox appears over and over again in situations such as Zen training, where the learning is precisely that of leaping out to a larger domain where one can consider one's thoughts and values with detachment. To the extent that the student is fixed on one level or another, with one preference or judgment—good or bad, positive or negative, spiritual or mundane—the aim of the teaching is not achieved. A good teacher, I suppose, is one who can convey the unity or circularity, the tangledness of the situation, so vividly that the student is forced to leap out of it.” (Varela 1984, p. 314). More on this in [[04 Concepts as pre-dictions]].] Chasing ghosts is, therefore, all we do. Ghosts can be: the casting of shadows, the interpretation of pasts, the gaps created by absences, the breaks imposed by humiliations, or far-away wants and desires (see: [[Desire]]), etc. This work, in many senses, chases its own tail. Because it’s fun, many animals do it. Jokes aside—_lies_: there will be jokes _everywhere_—this project observes the metaphorical grounding of concepts as a philosophical experiment to observe what we mean, exactly, when we talk ‘generally.’^[Which is: _all the time_.] As Parisi notes: “To point to a transcendental dimension of the [computational] medium is important in order to ... argue that the machines’ function of using and being used implies more than a mere application or execution of instructions, or in other words, their sheer usability.” (2019, p. 39). We will pay close attention to this, by exploring the functioning of _function_ itself (more on this in the next section, and in various chapters).^[See also: [[Function]]. Please already note we work with a novel approach to the notion of function, one which is underdetermined and under construction, and certainly not with the foreclosing, restricted idea of teleological functionalism as treated by, e.g., Chalmers here: “What makes the hard problem hard and almost unique is that it goes beyond problems about the performance of functions. To see this, note that even when we have explained the performance of all the cognitive and behavioral functions in the vicinity of experience—perceptual discrimination, categorization, internal access, verbal report—there may still remain a further unanswered question: Why is the performance of these functions accompanied by experience? A simple explanation of the functions leaves this question open.” (1995, p. 5). The notion of function we explore is one which sees it as a meta-pattern evolving atop physical and evolutionary constraints, it is the coherence we witness in unfolding dynamical systems, where there is a _tendency_, but this tendency is _teleonomic_: where apparently purposeful or goal-directed structures and functions can only be understood retroactively, but arise through natural processes without originary intention, direction nor design. Our expanded understanding of function is therefore also not meant to imply the functionalism challenged by Hui here: “Transhumanists, on the other hand, take an opposite position and exploit technology to an extreme. They embrace functionalism (seeing the human as composed of functions that can be improved individually) and an interdisciplinary program for human enhancement, including information technology, computer science, cognitive science and neuroscience, neural-computer interface research, materials science, artificial intelligence, regenerative medicine and life extension, genetic engineering, and nanotechnology. They emphasize the importance of technology as a means to extropia (as opposed to the “static utopia”), an open-ended perfection of the human species.” (2019, p. 212.) We are not waving a flag here, we are watching the undulations. Which, we argue, is also an inevitable flag-waving, as seen in [[03 Semantic noise]], but one that, if taken as such: can provide a novel meta-perspective which modulates its condition through the conceptual. Also note that if “we are in [an] embarrassing situation with technical systems” and if [l]ines of flight can exist only as a refusal to engage with the system, as a self-marginalization or escape to occultism and sectlike communes. The question of system remains to be tackled, not only from the perspective of deconstruction, which was carried out in the twentieth century, but also to fragment the system by allowing diversity to emerge.” (Hui 2019, p. 2019). Here, we follow, only insofar as we can truly commit to the “[reappropriation of] cybernetics, ... and [if] to do so we must understand the inhuman beyond the human and nonhuman.” (ibid., p. 238). In this thesis, we propose that function will enable us to glean aspects of this.]
The overarching hypothesis _and_ method explored by this project is that once (some of our more controversial) concepts are understood under a predictive light,^[Which to us implies metaphysical, _transcendental_. This is primarily treated in [[05 Prediction]] and in [[04 Concepts as pre-dictions]], but resonates throughout the project in various presentations.] many of the sociopolitical problems that often disguise themselves in philosophical attire (and vice versa) become productively disarmed, re-(in)formed and can be explored as recursive functions. It is important to disarm the concept of prediction before continuing. Prediction is immanently possibilistic,^[We endorse the critiques of possibilism and risk presented by Louise Amoore, and her following of Deleuze here: the politics of possibility imply “‘a life of pure potentiality comprised only of “virtualities, events, singularities”’ (2013, p. 157).] it is _not_ predetermination. Concepts widely construed, as generative abstractions, represent open-ended predictive _continuity_. Importantly, they allow for communication between entities: to see another creature has eyes is to see it _sees_. This is already remarkably abstract. Conceptuality, as such, can help localize important aspects of perspectivism. To paint a banal human example, the concept of “freedom” can be understood to predict, for someone, that they expect little to no encounters with governmental intervention, while another human senses the concept to predict that agreed-upon constraints actually increase meaningful future flourishing. Do note that this is different than framing things through variants on reason or understanding: a naturalized concept of prediction implies the desire to self-evidence, to persist according to an already _sedimented_ logic of persistence.^[Persistence is different than survival in an evolutionary-determinist sense, this is explained through active inference, more later.] Different concepts—whether intuitive and dynamic, or projectibly “clear” and resting on _formalized_ definitions—as prediction _models_, explore different temporalities and with varying priorities (e.g., effects for the coming 5 or 500 years). Transgenerational knowledge systems might track environmental changes across large chunks of time, while some scientific approaches might prioritize more immediate, measurable results. Both are valid but operate at different depths.^[The concept of “sustainability” localizes different perspectives by revealing how an economist, an ecologist, and an indigenous community each focus on different horizons when discussing forest tempering. Localizing predictive perspectives through concepts, reveals how viewpoints might productively overlap or completely erase/forget about each other. The concept of “health” can localize where varying medicinal approaches might each contribute predictive wellbeing insights, allowing for the _integration_ of methods of perseverance (what do we want to go _on_? And for how long? Is living with prosperity and intensity for a shorter amount of time preferrable over living a longer, blander life? Etc.). Concepts, in this predictive-productive light, can offer a different approach to how and where perspectives communicate. Rather than thinking “we just have different understandings of a, b or c”, we can understand the localized, unavoidable, self-evidencing nature of a concept-network—comprised of beings, objects, etc.—as it effectuates itself into future states.]
Understanding how concepts—effectuated by and shared amongst partial perspectives—predictively track and have influence on different spatiotemporal depths, opens the philosophical arena up for generativity: conceptual disagreements are not to be resolved by “it’s all relative” **nor** “ground truth is X”, but rather expanded by exploring how different modes of thought resolve uncertainty _differently_, and how there is overlap but significant diversity in spatiotemporal approaches.^[This is treated in [[10 Bias, or Falling into Place]].] Sociopolitically, this is a question of control,^[And it can, sometimes, be accounted for in metrics, this is treated in [[11 Post-Control Script-Societies]].] philosophically, surprisingly, too: concepts and their definitions modulate attention and behavior,^[As treated in [[03 Semantic noise]].] as opposed to the idea that they are inert, innocent, or simply the fungibles of knowledge. The predictive terrain becomes interesting in areas, such as mathematics, where concepts such as infinity, substitution, neutrality or the possibility of multiply-realizable fungibility, play a prominent role.^[Quoting Wittgenstein, in this case mathematics really is “blind calculation” (_Tractatus Logico-Philosophicus_ 6.24, cited in Badiou 2011 (2009)): a system predictive only of itself.] This is where the confused prediction(s) of AI also step in: through technical structure, and as the prediction of something we cannot know^[Defining intelligence, i.e., predicting what it ought to be, would mean locking thought into a box-making-factory.] yet aim to construct nonetheless: “As the social consequences of computational machinery grow more extensive, the reciprocal influence of ideas about machines and people will presumably grow as well.” (Agre 1997, p. 315).
Prominent AI researcher and initiator of the ARC challenge,^[The _Abstraction and Reasoning Corpus_ (Chollet 2019), is treated in [[10 Bias, or Falling into Place]].] François Chollet, is of the opinion that in order to make progress^[On progress, it is also worth noting how: “‘[s]cience progresses one funeral at a time [Planck dixit]. The future depends on some graduate student who is deeply suspicious of everything I have said.’ Geoff Hinton, grandfather of deep learning September 15, 2017.” Cited in Marcus 2019. Hopefully: not _just_ graduate students.] towards “the promise of our field, we need precise, quantitative definitions and measures of intelligence—in particular human-like general intelligence.” (Chollet 2019, p. 3). He explains that these should be “precise, explanatory definitions meant to serve as a North Star, an objective function showing the way towards a clear target, capable of acting as a reliable measure of our progress and as a way to identify and highlight worthwhile new approaches that may not be immediately applicable, and would otherwise be discounted” (ibid.). Chollet seeks precision but his logic seems circular, as he explains how common-sense dictionary definitions of intelligence “may be useful to make sure we are talking about the same concepts,” but “they are not actionable, explanatory, or measurable” (ibid.). The proposal made by this thesis is that no concept is innocent, all are actionable, often explanatory, and measurable in how they provide affordative entries into reality. Cavia (2024) similarly criticizes the Sellarsian vision of a possible unification between the scientific and the manifest images,^[See: [“Turing Trauma,”](http://cavvia.net/media/AACavia-TuringTrauma.pdf), 2024.] as it splits sentience and sapience, which leads to the “problem [of] how we come to interpret normative acts and how these distinctions play out in mental representations”—a “Sellarsian blind spot”^[We treat aspects of this in [[06 Principle of Sufficient Interest]].] which Cavia diagnoses in the pursuit of a general theory of intelligence, too (pp. 3-4), and of which he concludes (traditional Marxist) _alienation_ is a fundamental aspect,^[We treat aspects of this in [[09 C is for Communism, and Constraint|09 C is for Communism, and Constraint]].] one which can help create a “coherent account of objectivity compatible with the context sensitivity of conceptual frameworks” (ibid. p. 8). The intent here is not dissimilar, and also follows in line with Denizhan’s (2023) criticisms that importing technological conceptions of intelligence into cognitive science and the humanities presents theoretical and ethical consequences. However, the intent in this project is precisely to experiment with concepts by observing the frictions that ensue from these imports. If we keep asking for ‘criticality’ as well as ‘creativity’ in AI (source: _too many sources_ and AI meetings I have attended), we treat concept-contortion here as an obvious road to novel domains of inquiry.
The original _raison d’être_ of this project is that we^[This project deals with the intricacies of language to a high degree of self-reference. [[Self-reference]] is assumed to be unavoidable and complex. The “we” that speaks in this text is language itself as mediator. Following proposals of distributed dialogism—as a matter of fact extending beyond the humanities, see also: [[Problems]]—it is assumed throughout this entire project that, despite the individualist nature of the PhD game, natural language is evidence to the communal, social nature of thought. This is why we/I (often) speak from the first-person plural, though it would be more appropriately-termed _no-person plural_, or just _plural_. This is treated in [[03 Semantic noise]] and in [[Modulations]]. Alternatives to pronouns, and related challenges, are also explored in [[Pronoun]].] do not know what **intelligence**^[Intelligence will be the catch-all concept used throughout, primarily because we are dealing with “AI” as our object of study. Other words such as _thought, reason, cognition, understanding, knowing, learning_, will be employed as well, but bear in mind that “intelligence” will be used to imply all these notions. Whenever other concepts appear, it will be explained why they appear as such. In general, we follow the observation by Matteo Pasquinelli that “older concerns of the relation between technology and reason re-emerge today as concerns of the relation between computation and cognition.” (2015, p. 8). Preemptively, it can already be mentioned that: a) “reason” will be used in the context of philosophy-specific parlance, where it is often contrasted with “understanding”, following Kant. These two categories are not opposed, however: _reason_ is that which supervenes over _understanding_, which tends to finite, temporary assessments, while reason is infinite as it must be fundamentally “open” and renewable. b) _Cognition_ will be used primarily in the context of cognitive science, and _thought_ will be used in the context of philosophy but also psychoanalysis, mostly in reference to “armchair” activities, such as remembering your youth or doing math in your head. c) _Knowing_ and _learning_ are both used to refer to phenomena of cognition/intelligence/thought/etc., such as in the first sentence of this text: “knowing” what intelligence is would imply cognizing/thinking/reasoning/etc., about intelligence.] is, yet “it” is conceptually ubiquitous: from IQ tests and AI, to natural selection and the politics and projections of longtermism. Knowing what something is/does—i.e., having a [[Model]] of it: i.e., knowing things that can be done with its [[Pattern]]—has the technical, speculative and generally entertaining advantage that novelty can be generated atop of the old. This kind of metacognition is the contemplation of the abstractive character of perception-cognition-action. Classically presented as seeing _as_: Kantian noumenality or his “as if”, Dewey/Sellars’ “seeing as”, Wittgensteinian _aspect perception_, Gibsonian _affordances_, and we could even overextend this to *free-wheeling* compatibilism, to name a few comparable proposals. _Comparable_ because they are absolutely **not** the same proposal: _seeing as_ is not strictly the same as _acting as if,_ but in our proposal, in crass predictive abstraction: they are. The abstract characteristic joining them is that of _abstraction_ itself, which is to say:^[As will be argued, see: [[04 Concepts as pre-dictions]].] a functionally **predictive** structure can be discerned beyond immediate practical ramifications, which can be schematically derived/applied and observed across a variety of experiences, leading to the contemplation of possibilities. The capacity to _see as_ conforms with the logic of _as if_, given they both share the conceptual structure of comparing and/or replacing one experience for another. Abstraction results in the productive-predictive experience of a sense of _distance_ from something, by not being subject *to* the experience but understanding the experience _as_ experience. Under the right conditions, metacognitive effect provides a local perspective the advantage to modulate contingency (see: [[05 Prediction]]), rather than be subject to it.^[Though these two are not mutually opposed, this is explored in [[09 C is for Communism, and Constraint|09 C is for Communism, and Constraint]].] The realms receiving plenty of attention where this is actively treated as a topic of exploration are philosophy (including mathematics, or vice versa), cognitive science and AI, because they intend to isolate, arrest and sometimes formalize supposed chunks of experience as aspects we can derive different parsing methods from and for. Presentations of different ways in which this effect ties the metaphysical to the concrete, the ghost to the machine, is what will be our tacit focus throughout everything that is proposed hereon.
This demands that we take a wide, dangerously blurry view to see what connects with what, to efface differences in favor of abstraction. Because of this approach, we will treat the possible functions of reason (and related notions: intelligence, intuition, common sense, etc.), as _orienting_ activities. But not in the sense that they require a North Star (to cite Chollet again). These activities are only possible given the right _sustenance_—from nutrition to stimulation to status—and they are currently most often associated with adult human beings (but more and more also with other beings).^[Exemplified by, e.g., Michael Levin’s work on diverse intelligences.] Which means: they clearly have a North Star, a rather prominent one, to begin with. In “general” these orientations are usually functions framed as leading to the optimization—where no definition of “optimal” may be given, it depends on the desiderata—of manipulations (which depend on capacities and/as constraints) of a spatiotemporal path. In other words: **reason^[Or intelligence.] may be defined as the capacity to modulate the computations of the future by deciding how they ought to be chunked and therefore parsed (i.e., divided up).** Starting from the _self-incurred problem of perspective_ that much of the modern turn has had as a conceptual result—from Descartes’ _cogito_, through Kant’s transcendentalism, to the current traps _and_ critiques of anthropocentrism and North Stars: we better be sure that we only see reason where we expect to find it (that is: where we observe what we care to parse, at that time, as _ratio_; measure).
This problematic perspectival framing allows not only for a critique of the narrow image of computation and its metaphors (see: Agre 1997) or _computational reason_ (see: Cavia 2022),^[“[C]omputational reason ungrounds thought by foregrounding realizability over the myth of the given qua any absolute notion of truth.” Cavia 2024, p. 17.] but particularly, for this project, for an understanding of many kinds of activities as computation navigating different spaces: from building mathematical structure, all the way to the ‘competent’ *or* interested *or* invested nourishing of another living being. These are all reasoning, intelligent, dynamic activities which further additional computations. We follow the path proposed by Cavia (2022, 2024), where the idea of computation as _strictly_ formal can be countered by endorsing a mathematically-intuitionistic view,^[To Cavia, but beyond our scope, “founded on a spatial theory of types first proposed by Voevodsky, interpreted under realizability semantics originating in the work of Kleene and further developed by Martin-Löf, as the best candidate framework for a unified account of computational reason.” 2024, p. 10). R. Negarestani also advances thoughts on Martin-Löf’s typing in relation to AGI, and suggests that types might be usefully understood as forms of judgment or Kantian categories: as “pure concepts”. E.g., existence, mentality, etc., can then be understood as functions which we can organize into highly specific and analyzable proofs, where the meaning of a program (or proposition) is constructively determined by an **example** of its instantiation, by the encounter of evidence (2018, pp. 414-422). This is perhaps as tautologically self-evidencing as the [[Free energy principle]].] where the sentient-sapient character of constructive cognition “is only ever imperfectly captured by symbolic formalism,” (2024, p. 10) and therefore by narrow interpretations of computation. In Cavia’s view, departing from an intuitionistic, constructive framework allows us to see “computation [as] synonymous with the forging of paths in continuous spaces” (ibid.).^[We partly treat this, framing intuition differently, though, in [[10 Bias, or Falling into Place]]. For Cavia, the preferred image of computation is one where: “computation rests on a constructive logic which proffers a treatment of identity grounded in an isomorphism expressed as topological invariance.” (Cavia 2024, p. 13).] This kind of reason—in our formulation: as self-evidencing and ever-changing _ratio_—leads to the malleable (indeed, topological) concept of computation, “not to be identified with an abstract universal machine so much as an epistemic theory of encoding” (ibid.). What Cavia refers to as _encoding_—assigning tokens to types—is, again, what the reader will often encounter as “chunking” (i.e., encoding as _creating_ types) or “parsing” (encoding as _assigning tokens_ to types) in this thesis, where the first denotes how patterns become identified, and the latter how they are digested, filtered, interpreted. In our framing, because of the possible circularities and paradoxes emerging, chunking will sometimes imply parsing (just as reading implies writing), and vice versa.^[It is also in our interest to exploit the literary effects of _parse_ and _chunk_, where parse is associated with abstraction: linguistics and computation, and chunking (though very present in computation, too) is often more denotative of handling ‘inert’ matter or clunky materials. Please see: [[Parse]] and [[Chunk]].]
Our trajectory is also inspired by—and our definition of intelligence not unlike—the definition of cognition promoted by experimental biologist Michael Levin:^[Who often cites William James as his inspiration in his definition of (a highly cybernetic) intelligence as the way to achieve the same _goal_ through different means. Perhaps overstretching, this can also be interpreted as the jumping between two planes: “In the type systems of Martin-Löf, type assignment resembles not so much an act of classification, as the construction of a proposition whose meaning is synonymous with all the ways we have of realizing it.” (Cavia 2024, p. 11).] “competent navigation in arbitrary spaces” (Fields and Levin 2022). Levin also proposes the idea of _polycomputation_ to explain-explore how organisms perform computations at multiple scales simultaneously, from subcellular components to tissues, organs, and entire cultural niches, all of which can be understood as effectuating different functions, depending on the interested perspective at hand. Unlike traditional computational models that focus on neurons or specific networks in isolation, polycomputation recognizes that information processing in biology happens across these different scales, and in parallel. Framing information-processing as a fundamental biological phenomenon represents much of the impetus behind Active Inference (AIF), too, a multidisciplinary proposal which will be guiding much of the speculation presented here. AIF is an emerging normative framework exploring fundamental aspects of behavior in living organisms. _Normative_, because it rests on the idea that _all facets of behavior_ in living organisms “follow a unique imperative: minimizing the surprise of their sensory observations” (Pezzulo, Parr, Friston 2022, p. 6), and because _surprise_, witnessing a state which does not correspond with what had been predicted, “has to be interpreted in a technical sense: it measures how much an agent’s current sensory observations differ from its preferred sensory observations” (ibid.) by comparing probability distributions (_chunks_ parsed in a probabilistic manner). Merging the phenomenal and the technical (in AIF distinguished as _surprise_ and _surprisal_, see Smith et al. 2022): it would be informationally-surprising and therefore *salient* to experience many unpredictable variables, possibly risking thermodynamic dissolution (in humans, the rather unpleasant ones: from not eating, to being out of work, becoming neglected, etc.). Crucially, “minimizing surprise is not something that can be done ... passively[:] agents must adaptively control their action-perception loops to solicit desired sensory observations. This is the active bit of Active Inference.” (ibid.). Active inference will allow us to frame the aspects of predictive perspectivism presented earlier, allowing insights into the mechanics (chunking) of how agents resort to the (parsing) strategies they do. For a reader not familiar with AIF but familiar with work on autopoiesis and enactivism or ecological (gestalt) psychology: these domains can be understood as principled by the (Bayesian) mechanics understood through AIF (Friston et al. 2022, Ramstead et al. 2020, though there are detractors of this conflation, too, see, e.g.: Di Paolo et al. 2022).
In the Levin citation, “arbitrary,” or variable, spaces, and in Pezzulo et al. the “desired” aspect is the very pressing _political parsing_ matter often in question. State-enclosures in the formalisms of AIF, Markov blankets separating one thing from another, do not pertain to individual somatic states alone. In expecting the future, we do not only modulate what will be perceived at the level of an inner world, or voluntarist imaginative volitions, we are also given to and construct a vast network of systems—conceptual, infrastructural, legal, habitual, etc.—in order for expectations to be possible at different scales, sometimes robustly, sometimes less so. What we will permanently set into question, when diving into AIF—particularly in the domain of the social—are the types of human _sameness_, or symmetry, that are often presented as background claims for furthering the theory.^[We can also think here of the generalized symmetry approach in Latour or Stengers, where we try to rid a perspective of ad-vantage or primacy. More on this in [[Vantagepointillism]].] Like Luhmann (1995 (1984)) showed with his examination of _double contingency_:^[Which we treat in [[06 Principle of Sufficient Interest]].] recursive dialogical loops of absolute difference, and ignorance, are what lead to interesting social interactions, _not_ the assumption of a tacit sameness or the expectation of consensus. Additionally, a large part of our argumentation will orient itself around the AIF claim that the process by which organisms persist and evolve, is by either adjusting *themselves* or their *niche* through action. Intelligence (or reason) can therefore be understood as a distributed process that cannot be said to seek an end-goal (representation), but continuously anticipates future sensory states in order not to dissolve, and in being distributed: it inevitably sheds and grows parts, all the time. Whenever we engage with _any_ patterns we are _always_ exploring their functions towards predictability, chunking and parsing, whether we realize it or not. Else, we would cease to exist: “_intelligent behavior, whether manifest in biological or artificial form, is inextricably intertwined with persistence in the face of an environment that could bring about its end._” (Ororbia and Friston 2023, p. 2, italics in original). Psychosociopolitically, chunking an environment in order to parse it, means modulating legitimation as we go.
Language is one of the, if not _the_, paradigmatic soft rule-based effectuator(s) of how we seem to frame the sociality of reason, and here it will be defined as **1)** a system: because a system is what we call an enduring, dynamic whole made up of simpler components which may change while the system, in “essence”, remains the same. And: **2)** as a collective endeavor which presents single or collective agents the possibility of diverse environmental explorations (which would not be possible, or drastically different, without it). However, to remind again: as we present it, these—and any other—concepts are to be considered working hypotheses, _semantic attractors_: explorative programs which verge between the chunked and the parsed. They are collectively composed _approximations_, estimations, variations. “Semantic attractor” builds on the concept of attractors from dynamical systems theory, where we understand them to represent states _toward_ which a system tends to evolve. Additionally, as Hinton and Shallice proposed in the original context (1991), to model concepts as represented by patterns of activity _across_ many neurons, implying that attractors form from distributed dynamical patterns rather than localized representations. In socially orienting ourselves through concepts, semantic attraction can be understood as “conceptual regions” in unfolding landscapes which exert attractive pull on attentional orientation. Throughout our text, there will be plenty of unusual concepts, often vandalized in this way: taking metaphorical, polycomputational risks. As we will explore: **conceptual closure is almost-always neither possible nor desirable.** Though we apologize for the overstretching of certain aspects of language, it is an attractive trap we fall into quite often.
Concepts track spacetime by predicting-producing future realities, in other words: they abstract from experience towards the possible production of experience in the future. Examples of this range from how the schematic concept of ‘apple’ reduces uncertainty about encounters with future spherical, edible fruit, all the way to how the concept of ‘ideology’ reduces uncertainty about large cultural structures which model behavior. It is the opinion of mainstream AI research that reasoning is high-level abstraction; conceptual generalization (from no or few examples), and that this “remains one of the most important ways in which human intelligence surpasses AI.” (Li et al., 2024, p.1). François Chollet has provided a formal definition of intelligence, describing it as “_skill-acquisition efficiency_ and highlighting the concepts of _scope_, _generalization difficulty_, _priors_, and _experience_, as critical pieces to be accounted for in characterizing intelligent systems.” (Chollet 2019, p. 1). His contention is that we have no formal methods for testing intelligence, though this claim might suspect if we think of all the standardized tests and rites of initiation into formal systems, performed by human (and other) beings on a daily basis.^[Not disregarding the complex histories of these metrics being highly problematic, a topic which has been explored by Stephen Jay Gould his 1981 _The Mismeasure of Man_, and more recently by Catharine Malabou in _Morphing Intelligence: From IQ measurement to artificial brains_, 2019. Additionally, as Shannon Vallor notes in her 2024 _The AI Mirror_: “Still, the lack of a reliable test or agreed-upon criteria for artificial intelligence has not impeded progress in the field [of AI]. Many of the systems and tools now commercially characterized as AI already perform tasks that would have astounded researchers at the 1956 Dartmouth conference.” (p. 21).] It’s just that these situations are elusive if we want “total abstraction”, precisely because they are highly distributed and subject to fuzzy, open-ended affairs effectuated by multiple perspectives in an attempt to coordinate (or destabilize) existing dynamics. While it is clearly intelligent to understand complex aspects of nuclear physics, or to help another person out, it is certainly not intelligent to drop an atomic bomb or enslave an entire population. Yet here we are. Classic philosophy of science knows to deal with these cases by advising for caution, and current philosophy of AI deals, primarily, with the ethics of matters in similar ways. Unfortunately, in the case of the latter, the predictive horizons are tuned by violent visions of what is and what is not supposed to be human(ly valuable).
 
### _Functionalism_: process philosophy meets enactive cognition and computation
The consequence of existing is that existence can hardly be denied. “That it happens ‘precedes’, so to speak, the question pertaining to what happens.” (Lyotard, 1991 (1988), p. 90). Following a tradition which is strongly constructive—and Kantian (see: Swanson 2016)—AIF presents us with the functionalist (Bayesian) question of: “if things exist, what must they do?” (Friston et al., 2022). Once we witness something, once we gather evidence and construct hypotheses, what does this imply? This turns the principle of sufficient reason on its head: from ‘what is the reason for this?’ to ‘what is this a reason _for_?’ Following in line with this, we aim to question the functions of phenomena. This sets the focus on constraints (Juarrero 2023),^[Briefly, we can cite Juarrero here to set the focus on an understanding of how constraint-enabled coherence displaying the effect of “control”, is to us revelatory of an interest in _function_: “Recognizing that systems of constraint are organized hierarchically is not news. However, because of the received understanding of mereological relations philosophers have rarely focused on relations between processes at different hierarchical levels. ... (1) relations between levels of hierarchical organization are **not energetic exchanges; they are relations of constraint.** This is particularly significant with respect to **top-down control from whole to parts.** (2) Levels of hierarchical organization ordered by contextual constraints cannot be extensionally defined in terms of their reference. ... History plays a leading role in these constraints, as do spatiotemporal constraints generally. The **properties and powers of contextually** generated dynamics must therefore be defined intensionally, that is, in context. (3) Finally, philosophers have long puzzled over the power of the content of mental events and intentions to bring about actions that satisfy that content (Juarrero 1999). ... those debates were framed by a flawed understanding of mereology and causation.” (2023, p. 179, our emphasis).] on collective reproductive conditions (Marx), on multidirectional evolutive dimensions (5E) and on what happens retroactively speaking when a new technology induces, degeneratively, novel functional dimensions (Simondon exploring transindividuation through technicity, or Deleuze asking ‘what is philosophy _after_ cinema?’, for example). If Joyce designed _Finnegans Wake_ for the reader to have to mouth the words as well as read them (Theall and Theall 1989, p. 52), dislodging the ‘problem’ of the aural and/versus the textual and making it productive, this thesis exploits the generative in the plasticity of concepts, searching for tacit or possible functions within them. What are these concepts _for_? The brief answer is: functions, all of which are open-ended.
It is also in hope of joining aspects of what is known as _functionalism_ (more traditionally discussed in the context of analytic philosophy, a key figure here being Putnam), with aspects of what is known as _process-philosophy_ (more often discussed in the context of continental work: a leading figure here being Whitehead), this thesis looks into _functions_ as the evolving phenomena resulting from contextual constraints (Juarrero 2023). Shifting from a _form_ or identity-based to a function-based ontologization, reveals the productive effects of predictive cognition. We can observe, rather than static entities interacting through processes, how, e.g., environmental constraints shape convergent solutions across phylogenetically distant species. Gravity—on Earth a pretty _universal_ constraint and subject matter of chapter [[10 Bias, or Falling into Place]]—has produced similar yet different locomotion strategies across beings (from swimming to flight to rocket launchings). In the dynamic interplay between organism-environment, between pure contingency and revelatory repetition, function ties them together, bringing our attention to their evolution and polycomputational possibilities, rather than to atomized subjects or objects undergoing (linear, causal) processes. Function, in this thesis, rings both timeless _and_ durational. Nested within the vantage point^[See: [[Vantagepointillism]].] of their Markov blanket,^[See: [[Markov blanket]].] organisms are functioning _responses_^[See: [[Autosemeiosis]].] to unfolding, ever-complexifying ecological problems rather than as predetermined forms. A river is, in this way, not a body of water nor a channel or vessel; but perhaps rather a nutrient-spreading function intimately coupled to a variety of other nutrient-filtering systems, such as mussels, birds and sewage systems. Understood this way, a river would, perhaps, be less likely the target of pollution.
These onto-epistemological implications also have effects on our methodologies, as we will see in the chapters: we are often after the question “what does X do?”, whether we are talking about concepts, constraints or other conditions. Arresting properties as functions can reveal how these are transferred across domains and modalities. Understanding how photosynthesis functionally connects seemingly disparate organisms (e.g., plants, algae, bacteria) opens up aspects of understanding these as resolving a particular domain of constraints (being subjects of the Sun, having to feed, etc.) at different _scales_ (chunks). Different scales of analysis looking at similar functions has drastic implications for how we understand cancer cells as coupling^[See: [[Structural coupling]].] or not to the larger morphological space that is the unfolding of a complex organism, for example (Levin 2019). As Levin’s work also shows, plenty of organisms often perform similar functions with equal or greater “efficiency” than humans; moving the “intelligence” dial in all sorts of directions; informing and reorienting understandings of the teloi of evolution. In the wake of post-anthropocentric desires, _function_ also allows us to set the focus outside human frameworks; putting ecological attention on the things we, humans, need to survive (which seems to be what matters to us, there’s no way around this, it seems).
Operational similarities across domains, at least in organic cases, appear to have the effect to increase _empathy_: once we witness how dolphins play, trees communicate via fungal networks, octopi nurture, elephants visit their dead relatives, orangutans use tools, gorillas learn sign language, etc., etc., etc.; we seem to care more for these beings. Particularly, our focus is also interested in moving away from agentic views of an all-powerful, voluntary self, and reveal how, just as digestion happens _to_ us, so does cognition. This shift is a philosophical _legitimacy_ switch: instead of thinking about who we are, what we want, what we ought to be, we can think about how functions pass through us, and _maybe_ about how these can be modulated. As we will see in [[09 C is for Communism, and Constraint]]: psychosocial problems require attention at the functional system level rather than merely observing atomized-aggregate individual behavior. Or, inheriting from Deleuze, as observed in [[11 Post-Control Script-Societies]]: social power operates through the functional modulation of distributed attentional frames, rather than through top-down discipline or authority.
 
### Unnatural language, limits and (dis)satisfactions
As we have so far framed it, mind, through language, has always been artificial or _unnatural_: unstable, unfinished and interrogative of its own given, or natural, condition. Pointing at deities or at life after death, a common concern for many cultures, presents us with rather complex sense/reference problems. The more radical, speculative proposal offered here is that new concepts, in their interplay with synthetic language-modulation, will become capable of producing futures currently unimaginable.^[An implication might be a speculative vision where, probing a highly advanced language-synthesizer, one might ask that it produces novel interpretations of abstraction itself (in an effort to, for example, communicate between different scientific disciplines, or expose an alternative model of social coordination not based on what current social coordination inherits from historical models).] According to Gregory Chaitin, mathematics only advances by changing its foundations and thereby proposing new concepts, creatively. He compares biological with mathematical change, underlining how in both the inherent (possible) creativity emerges through recombination and birth (2006, p. 40). We witness a paradigm in which conceptuality is exploring new shapes and dimensions: different structures lead to novel questions. What is the “shape” of the concept of evolution in a specific text corpus compared to another one? This question is now answerable in novel ways, given the technology. As technologies such as LLMs become more and more programmable through natural language, the possibilities seem promising. At the same time: careful attention to the supposed _natural_ language foundations is needed. We have already begun exploring the possibility to reduce the complexity of vast text data to coarser, *general* chunks, through methods such as unsupervised clustering, topic-modelling and concept-induction (Lam et al., 2024), and more recently the desire is to create prompts that create prompts that create prompts that create prompts, etc., seems to hold promise for arriving at deep(er) abstractive structures through querying language models.
What about the reverse? What about the generation of outputs based on novel concepts? “... to transform material inputs into spiritual outputs” (Catren 2023, pp. 415-6). In high-dimensional probabilistic spaces such as those navigated by LLMs, concepts may be thought of as manifesting through complex but _distinctly identifiable_ embedding structures. Humanly-relevant semantic (e.g., spiritual) contents gravitate around possibly identifiable regions, the structures of which can be compared or combined to yield novel relations,^[The famous and simple example is always that of king - man = queen, in word-embedding. We treat aspects of this in [[03 Semantic noise]].] always being mindful of our own hallucinatory tendencies. The “hallucination” capacities that we ascribe to LLMs and other generative technologies is revealing not only of possible _interconcepts_ (Wolfram 2023), and inferentially-structured conceptual embedding (Cavia 2023), but especially of our own interested access to their structure. It is in this way that conceptual exploration through LLM seems promising to this project. As these technologies evolve, their attractor landscapes might reveal unexpected semantic geometries that correspond to conceptual construction currently beyond imagination: controversial conceptual contemplation and formal computation have much to compare and contrast.
As Cavia suggests, the concept of computation itself “functions somewhat like a _motif_,”^[Cavia refers to Pierre Deligne on this: “providing a “system of realisations” for unifying disparate theories construed as structures, pursued by exposing certain topological invariances between them.” (Cavia 2024, p. 10).] or a “conceptual scheme” we can expand on to examine and expose novel patterns. As an open-ended and constructive phenomenon: it continues to evolve because its foundations can be challenged and transformed. Our _polycomputational_ grounding of (poly)computation_, framing computation as **different formal methods amplifying the capacity to witness _difference_**, parse it, and therefore organize perspectives around patterns emerging from differential exchanges,^[See expanded definition in the appendix.] aims to question some of the assumptions about how patterns are produced and communicated between linguistic beings. In light of the Schmidhuber quote that follows below, speculating at the very limits of what “interesting patterns” to AI might entail, we should not be surprised to be considering how the concepts in this and all other theses to be parsed by the machine, might have an effect on the types of concepts generated and explored in the future. Less surprising should it be that, since hypothesis-testing and experiment-design are now offered as possible ML products,^[By Google. The project does not have a name yet, but builds on Google’s Gemini LLMs. See: Le Page, February 19, 2025, _New Scientist_.] the perhaps rather unorthodox speculations entertained here might be explored through “virtual laboratories” sooner rather than later.
 
>AIs will be curious scientists, and they will be fascinated with life. Fascinated because life and civilization are such a rich source of interesting patterns, at least as long as they are not fully understood.
>
>Jürgen Schmidhuber, MLST interview, Jan 16, 2025.
 
The drive to witness as well as create increasingly complex systems, only to then seek even greater complexity once we have simplified the previous complexity into an overseeable pattern, frames mind as drawn to operate at the edges of its models. As Yağmur Denizhan has presented it: intelligence might be defined as the border activity between the modeled and the unmodeled (2023). Intelligence—or reason—might be defined, then, as _dissatisfaction_.^[Or even as asceticism: permanently in accepting the expectation that everything it “knows” is subject to possible destruction, doing away with any notion of progress. In his first seminar on _Technologies of the Self_ (series at the University of Vermont in 1982, published in print 1988), Foucault sets the stage for his inquiry (into the self, and following his research on the history of sexuality): “Max Weber posed the question: If one wants to behave rationally and regulate one’s action according to true principles, what part of one’s self should one renounce? What is the ascetic price of reason? To what kind of asceticism should one submit? I posed the opposite question: How have certain kinds of interdictions required the price of certain kinds of knowledge about oneself? What must one know about oneself in order to be willing to renounce anything?” (p. 17).] Science and philosophy, as intelligent, socially-distributed models which map possibilistic spaces, as explorative—even anarchic—affairs setting off from established patterns, are the disappointment or misalignment with given, current affairs. As ideas: science and philosophy are fundamentally unsettled with regard to their own sense of _self_-(evidencing). The logic of both science and philosophy as _institutions_, however, seems to be that of permanence or persistence: they ought to be relevant in the future. Strangely, the stage set by AI as an object of study for both, is that we are to study that which we characteristically witness as being in permanent transformation: thought itself. In the context of evolutionary biology, this is the paradox of the surviving organism: in order to survive, the organism must adapt, which means whatever it was before, it no longer is (Levin 2024, p. 1).
Concepts employed through structures such as language—socially-distributed memory patterns which allow multiple organisms to cohere into predictable projectibles—are the “cognitive glue” (ibid.) which ensure the salience of affordances (_not_ the fidelity of representation). Concepts and organisms are both adaptive phenomena because they entertain “more or less”, flexible functions against a contingent reality. Concepts and organisms have attractor states which make things apparent by inevitably realizing them through coupling. As all possible concepts are possible,^[See [[Semantic attractor]], and the notes there on Wolfram’s notion of _interconcept space_ (Wolfram 2023): “It’s as if in developing our civilization—and our human language—we’ve “colonized” only certain small islands in the space of all possible concepts, leaving vast amounts of interconcept space unexplored.” Do note the unfortunate evidence of the violence we just referred to.] we can indeed consider ourselves to be expectably, _reasonably dissatisfied_ with what is currently at hand. How might a deep dive into the structures of concepts help create a new dialogics, amenable to conceptual transfers between humans and humans and other types? Stated otherwise: how does our current use of concepts, reflected in activities such as learning and communicating, affect how conceptuality evolves in possible non-human systems? A more pressing question might also be: given the distribution of concepts across systems: can philosophy—including cognitive science and vice versa^[We include philosophy in cognitive science and vice versa given that both have the same object of study, namely the mind, and in a European-centric light: often with similar metaphysical commitments, inherited from the same (outlined and criticized, e.g., by Denise Ferreira Da Silva (2016)). Levin criticizes the cognitive science paradigm as one which doesn’t take the intelligent substrate that coheres as a subject: “Traditional cognitive science operates on a living Subject: we study the information it absorbs, computations it performs, memories it may have or form, behaviors it can deploy in different circumstances, goals it may pursue, etc. All those capacities are taken to be the properties of a fixed, embodied Agent ... the fact that it is a collection of cells or subcellular fragments, which proliferate and actively interact to build its body, is relegated to developmental biologists. That phase of a subject’s life is usually ignored as behind-the-scenes setup, after which real study can begin.” (Levin 2021, p. 114).]—claim to hold analytic legitimacy when it comes to witnessing the project of AI? Or, relatedly, how does the legitimacy of any field hold up against the game of catch-up with the technocapitalist complex that (currently) entrenches life? In the preface to the English edition of _Difference and Repetition_, Deleuze gives us his take on what philosophy ought to do in facing these challenges (in his presentation, in the case of philosophy versus art and science):^[Art, science, philosophy, religion, etc., are all forms of _revelation_, Catren, 2023.]
 
>Every philosophy must achieve its own manner of speaking about the arts and sciences, as though it established alliances with them. It is very difficult, since philosophy obviously cannot claim the least superiority, but also creates and expounds its own concepts only in relation to what it can grasp of scientific functions and artistic constructions. A philosophical concept can never be confused with a scientific function or an artistic construction, but finds itself in affinity with these in this or that domain of science or style of art. The scientific or artistic content of a philosophy may be very elementary, since it is not obliged to advance art or science, but it can advance itself only by forming properly philosophical concepts from a given function or construction, however elementary. Philosophy cannot be undertaken independently of science or art. It is in this sense that we tried to constitute a philosophical concept from the mathematical function of differentiation and the biological function of differenciation, in asking whether there was not a statable relation between these two concepts which could not appear at the level of their respective objects. (1994 (1968), p. xvi).
 
In our proposed philosophy of language-modulation, we look for functional conceptual bridges between AIF as a theory of cognition and the computational mechanisms that sustain linguistic prediction architectures. Such a philosophy doesn’t claim superiority over either, e.g., computational linguistics or cognitive neuroscience, but rather develops distinctive concepts by exploring their predictive functions in their possible applicability across biological evolution, neural prediction error-minimization, and transformer attention mechanisms. Taking this interlocuting risk might reveal something interesting about information-processing across different substrates. New events necessitate new concepts. A suspicion driving this project is that in the near future all silicon-based computing will most likely be effectuated by conceptual directives (as is evidenced by current LLM interfaces and their functionality/cross-domain applicability). It is therefore paramount that we continue to challenge the monocultural aspects of such as LLMs, with the exploration of dynamic processes such as (novel and controversial) concepts. This project’s neologistic approach proposes new concepts in order to disorganize, reconsider, and most likely irritate, a lot of the work being done in AI.
The main research question this work is after is therefore: **if philosophy aims at abstraction through concept-creation—both of which can be seen as different processes, or as the same, as will be shown—what kinds of abstractions can this project engineer, in order to predict the further development of abstraction in or of AI, as mind “outside” or “beyond” the organism, particularly given the dialectic that AI has inherited its “basic concepts” from philosophy?** In simpler sub-formulations: what concepts of abstraction might be useful for predicting “generality” in AI development? What philosophical concepts has AI inherited—transcendence, multiple-realizability, intuition, etc.—and how do these shape its development? What new dialectical relationships emerge when AI systems built on philosophical foundations begin generating their own abstractions?^[This question is particularly important, given form and function transform each other, through metaphors, as technologies evolve.] Based on our own abstracting condition, inherently based on prediction, how might we predict the evolution of abstraction capabilities in future systems? It is also important to mention again that we begin from the assumption that philosophy always was AI, and vice versa, and that the mediatized buzzword “AI” category-container-concept is just the disguised marketing of activities which explore-exploit hoarding, colonial realms. In other words: concepts narrate, constrain and shape _everything_, our inquiry explores the role of philosophy in shaping coming abstractions within/for/of AI, considering the emergence of distinctly new processes based on inherited concepts. The project seeks to question various ways in which abstractions create the future, and how future abstractions might bridge a possible gap between philosophical thought and technological development. We consider these to be aspects of the same thing: variants of seeking preservation in light of dissipation: where preservation is paradoxical, given no known system holds a stable identity. All things which persist are variants of seeking invariance, in other words.
As AI increases in complexity, its challenges bring philosophy to the foreground as a possible avenue for exploring issues pertaining to questions of futurology, survival and desire: “_all_ technologies, and thus all forms of AI, are mirrors of human thought and desire, [they] reflect what we think we want, what we think we need, what we think is important” (Vallor 2024, p. 35). What the challenges for AI as explored through philosophy reflect, is a novel moment for reconsidering ways of life: who are ‘we’? Can we even speak of a homogenized sameness through this ‘we’ or should we fully alienate ourselves from ourselves? And can be said of _want_ or desire when answers and solutions come at us faster than we can formulate questions or contemplate problems?^[I talk about this as playing eternal catch-up with Silicon Valley AI.] In light of the nonlinearities exposed by studies in complex systems:^[For an overview see, e.g., Alicia Juarrero’s 2023 _Context Changes Everything_.] how does the function of our basic analytical tools for understanding causality, such as the very word _why_, our access to principles of sufficient reason, change?^[I conducted an interview with Gerard de Zeeuw about this and other themes, right at the beginning of this thesis. I am indebted to his support and feedback throughout the PhD process, he provided my research with valuable insights and mentoring. Equally indebted to Caroline Nevejan for putting us in touch in 2019. See: [[Who is we and what is why]] for the interview. See [[Pronoun]], [[Question]] and [[Modulations]] for further considerations of these questions.] How does the quickly shifting perception of what AI is or could be,^[The infamous moving goalposts.] where achievements are no longer considered “AI” once accomplished, reflect both our evolving relationship with uncertainty as conscious curiosity, as well as the tendency for desire to eternally shift its telos?^[I asked this question, inspired by our eternally intransparent access to ourselves, not knowing what we want, the question of desire, to Stuart Russell. [[Stuart Russell DigitalFUTURES 2024 Doctoral Consortium on AGI 2, “AGI What if we succeed?”]], https://www.youtube.com/watch?v=SfWe_rDPbpM, streamed live on Sep 15, 2024.] Orientation, teleology, implies the political problem of _reproduction_ writ-large,^[Writ-large, as in: how to go on? Who gets to go on? What is “going on”?] which has been the preoccupation of myriads, arguably since the beginning of cognitive time. What of the political question of reproduction once we have ever more computational ways to approach it?^[I asked this question to Blaise Agüera y Arcas, in light of his definitions: “Life is the computational phase of matter, arising from evolutionary selection for dynamic stability. Intelligence is the ability to model, predict and influence the future, evolving in relation to other intelligences to create a larger symbiotic relation. Both definitions have multi-scale properties and are highly parallel.” [[DigitalFUTURES 2024 Doctoral Consortium on AGI 6, Blaise Agüera y Arcas, Are we there yet?]], https://www.youtube.com/watch?v=mza3Vxg0oQw, streamed live on Oct 18, 2024.] If our interest is in the diverse, the plural, the new, can we identify basic patterns sustaining the production of this experience? Relatedly: what is _sameness_? Can it be attributed to symmetries exhibited across multiple scales? Is symmetry (as equivalence, as translatability, as invariance, as reversibility, etc.) a basic common denominator across all computation? What of its definition of _invariance under transformation_ being a contradiction in terms?^[This question was directed at Karl Friston. [[DigitalFUTURES 2024 Doctoral Consortium on AGI 5, Karl Friston, The Physics of Sentience]], https://www.youtube.com/watch?v=GsRVcVVwnqc&t=4408s, streamed live on Oct 13, 2024.] We learn to reduce complex structures to simple rules so we don’t need to interpret patterns anew each time, stated otherwise: we cannot but perceive other than through patterns. At the level of language structures, this aspect of invariance is obvious. At the developmental level of the self: as we grow we need an image of _something_, a model, in order to continue to subsist. All these systems evolve tracking different spatiotemporal scales, making translation between them an eternally unfinished project: this is where philosophy has often made attempts at _witnessing_ (in formalizing computations through logic all the way to computing radical political novelty). The presentation of the ever-approaching _possibility_ of “actual” artificial intelligence presents us with the practical implications of all that has been philosophically considered before.
 
> LULA: The people accept you as a ghost of the future. And love you, that you might not kill them when you can.
>
> Baraka, _Dutchman_, 1964.
 
>The case for risk is strong enough that we should act now ... [but] it’s surprisingly hard to address these risks, because they’re not here today, they don’t exist, they’re like ghosts, but they’re coming at us so fast, because the models are improving so fast. So how do you deal with something that is not here but is coming at us very fast?
>
>Amodei, _Anthropic_ CEO, 11 Nov 2024, interviewed by Lex Fridman.
 
>‘[T]he soul never thinks (or understands) without a phantasm’ (431 a 16–17: οὐδέποτε νοεῖ ἄνευ φαντάσματος ἡ ψυχή).
>
>Aristotle, _De Anima_, (cited in Steel 2018, p. 185).
>
>“[To be understood as:] for the human soul no thinking is possible without being accompanied with some object in imagination, though it remains unclear how exactly this connection with imagination is to be understood” (Steel, ibid.).
 
Only surveying the quotes above, the following can be stated, in light of ghostly insistence: the twists of self-alienation are rather crude. The violently simple and simply violent pasts we seem to come from, as well as the intricate social dynamics exhumed from that past, tie in the chronopolitics we have been outlining with the fear of its own destruction, by its own means. The birthing of an alien force from the future already tormenting the present (as Amodei—and plenty of other voices in AI—expresses).^[Not unlike how, before AI: abduction stories can be seen to reflect the fears of a much too present colonial past in the US (Dean 1998).] All of this, even disregarding the question of slavery, brings us to Aristotle and the question of understanding as _having_ something, having an object, an image, a telos, etc. Perhaps we ought to stop projecting, stop having. But instead of stopping here, this thesis continues. Within the strange condition that envelops us, neither gods nor the death-drive are any longer at the center-stage of thought’s alienating turns.
As outlined in the preceding footnotes, I have asked these questions to some of the most prominent AI, cybernetics and complexity researchers, respectively, following the order of questions above: Gerard de Zeeuw, Stuart Russell, Blaise Agüera & Arcas, Karl Friston and Joscha Bach.^[Thanks to the doctoral consortium organized by the Digital Futures team (special thanks to Michael Just!), the interviews are on their Youtube channel. My only gripe with their series would be the emphatic focus on OpenAI’s definition of Artificial General Intelligence (AGI): “‘a highly autonomous system that can outperform humans at most economically valuable work.’”, cited during each intro. To frame our gripe, the following can be read along the definition: “Prima facie, the automation of cognitive labor by artificial intelligence (AI) seems to act as a canonical example of alienation in the Marxian tradition—the knowledge and skill of the worker concretised by machines that transform human labour into fixed capital.” (Cavia 2024, p. 1).] My insights into their responses exist in the chapters, an outline of which follows after this introduction. The eclectic theoretical framing that this project offers is a result of the undecidability in my estimations. I could not narrow my scope, because I was not capable. Before proceeding to an outline of the AI theories guiding the larger parts of this work, we present AIF in order to frame the aspects of perception as abstraction, and the question of the prediction of generality we referred to throughout this introduction.
 
### Active Inference: Hallucinations all the way down, for humans _and_ for GenAI
The process of perception _abstracts_ in order to be experienced as such: all percepts are an effect of syntheses of an otherwise unmanageable amount of sensorial complexity. This is true of what we normally refer to as more or properly abstract thought, too: concepts such as totality, infinity, emptiness, etc., are but an abstracting dream of predictive perception. None of these concepts “exist,” they are useful placeholders^[Literally and figuratively holding things in place, negating change: motion.] for abstractions to become useful points of comparison.^[Reza Negarestani points to the simplicity of the pointing point, as a starting point for thinking about the recursive dialectics of indexicality and pointing (2018, p. 40). How do concepts become point(er)s?] In order to frame these observations, this project follows many of the proposals made by the emerging _active inference_ framework (AIF): where the organic tracking of spacetime can understood as enactive^[5E enactivism: ecological, embedded, embodied, enactive and extended.] prediction. The framework presents many philosophical challenges but also offers the trans/interdisciplinary possibility to combine research in “neural networks, contemporary cognitive and computational neuroscience, Bayesian approaches to dealing with evidence and uncertainty, robotics, self-organization, and the study of the embodied environmentally situated mind.” (Clark 2016, p. 10). In our context, AIF offers an avenue for observing the unfolding of cultural mechanics through AI by understanding technical “progress” and organic adaptation as guided by the same principles, aiming at a naturalized understanding of these mechanics, while leaving room for possibilistic speculation: we are dealing with how systems estimate the future, which is productively and dangerously uncertain. The evolving complexity ensuing from combinatorial overlaps between systems inevitably leads to novelties which we can only speculate about from the perspective of surviving estimation. Abstractions—categorial percepts: affordances, concepts, paradigms, seeing the whole sky as “blue”—are estimations about what _could_ be there. AIF is therefore a strong challenge to any naïve realism, and is capable of producing a highly reflexive, epistemic humility as it urges understanding of how each perspective, with its own model of reality, necessarily operates through approximations rather than direct access to ‘truth,’ while simultaneously acknowledging the adaptive value and practical necessity of such estimations for any system’s survival and development as it interfaces with its contexts.
AIF, which might therefore be better termed _enactive inference_ (Ramstead, Kirchhoff & Friston 2020), in our context is able to reveal the evolution of concepts as extended cultural niche-formation, subject to coordination dynamics that can sometimes be modeled in terms of Bayesian approximation. The key philosophical implication the AIF sets forth is that all perception can be framed as an *inference*, or a process of belief-updating (“belief” taken in a Bayesian and an ample sense: a belief can be an embodied condition, it does not have to be a conscious thought). We will term this, largely: prediction.^[In earlier literature, often when conflated with predictive processing, the theory has also received the name of “predictivism”, Bruineberg 2017, p. 18.] The main—cybernetic—idea fundamenting the AIF “is that every action–overt or covert–is in the game of resolving uncertainty by sampling the sensorium to reduce expected surprise (or a variational bound on surprise called free energy).^[See also: [[Variational autoencoders]]. Variational Bayesian methods enter where things are effectively intractable for calculation. This is where sampling approaches such as Monte Carlo or perfect Bayesian inference is impossible (we always have too many variables determining an outcome).] This is known as active inference and requires us to commit to actions that have the greatest epistemic affordance or information gain.” (Mirza et al. 2019, p. 1). As will be seen, all of these concepts need be taken with a grain of salt, they survey technical and metaphysical terrains which sometimes appear incompatible.
Sustained by the *Free Energy Principle*, active inference is the ‘mechanism’ underlying theories of predictive processing, a field whose philosophy suggests that cognition can be understood most parsimoniously as a process which predictively produces experience by projecting expectations and adjusting these on the basis of what it encounters (Clark 2023, 2016). If the tendency of organic systems is to maintain themselves within a restricted set of states in the face of uncertainty (i.e., our context is entropic dissipation) while relying on a partially observed environment (Friston 2012, 2013), then anything in the service of persistence that is effectuated by such a system can be understood as a strategy towards the production of a more predictable environment: from cell-replication (making more of yourself is a sure strategy for predictability)^[As is _making_ yourself: “We start a new narrative from the point where we perceive ourselves as a person. It’s like creating a second spirit in the image of the first, built into the mind as this personal self that thinks of itself as a human being. I might be overgeneralizing, but I suspect there’s something to this theory of development.” Joscha Bach, September 8 2024. See: [[DigitalFUTURES 2024 Doctoral Consortium on AGI 1, Joscha Bach, Is consciousness a missing link to AGI?]].] to the producing-witnessing of affordances (Gibson), to technological ‘protensions’ (Stiegler) such as clothing or architecture (tempering the chaotic environment), or indeed: AI, i.e., as the emergence of a complex, distributed cognitive econiche. In the context of what a good life might entail, AIF frames “experiences of human flourishing to be the result of predictions and uncertainty estimations along many dimensions at multiple levels of neurobiological organization” (Kiverstein & Miller 2023, p. 1). None of this is without its problems, questions and possibilities, some of which we treat in [[05 Prediction]], [[04 Concepts as pre-dictions]], [[Free energy principle]] and [[11 Post-Control Script-Societies]].
Whatever it is we have been calling ‘the mind’ is always _partially_ implicated in what it tracks, and tracking is always abstracting. Difference is its inevitable a starting point: the *Free Energy Principle* in active inference dialectically grounds this proposition, as will be seen. The claim made by this project is that steps towards understanding the body-brain-extended-mind-complex can be better framed by understanding (philosophical) concepts as predictive abstraction(s): this is our claim about what perception-cognition as computation (as chunking and parsing) _does._ As far as we can tell right now, these abstractions are analyzable as information-processes which are subject to (cases and combinations of) indefinition, intractability, incalculability and, most importantly: _constructibility_. This much is agreed by various of our points of reference, in drastically different contexts: it is noted by Agüera & Arcas, the _Atlas of Experimental Politics_ (2021), and the constructive computation of Cavia (2022). This again, sets us off to consider notions of equivalence and difference, which at the level of humans learning from each other rests on a dangerous concept of _sameness_, which perhaps finds its basis on a general, malleable concept of symmetry.^[Which, we entertain (though this will be the focus of future research and appears rather incompletely here): in the effect of how perception reveals patterns as variations on symmetry (look around you and observe this), it could be that this is the after-effect or the smoking gun of dissipative structures seeking energy-minima _in general_.] One of the analyses explored ([[All things mirrored]]) will be the contradiction that is the very definition of symmetry: _invariance under transformation_, when tackled through the phenomenology of prediction. Comparably, notions of truth or _the absolute_ render equally problematic images of sameness. According to Cavia (2024): “a philosophy of intelligence must dispense with such a notion of the absolute if it is to faithfully render AI as a vector of reason” (p. 19). In pairing up computation as “the coming together of logic and matter” (ibid.), we are challenged by a reorientation around how matter links to logic, and vice versa. This means attention to situatedness, and an avoidance of pretensions to views from nowhere. “If I am to argue for a positive valence to the notion of alienation induced by AI, which I am inclined to, it must function outside of any _absolute_ frame of reference.” (Cavia 2024, p. 19). We absolutely agree.
All intelligence is distributed and collective^[“The second lesson we learn from embryogenesis is that all intelligence is collective intelligence. We are all made of parts – not just ant and bee colonies’ swarm intelligence but all of us. We are collections of competent parts (molecular networks, cells, tissues) which work together to implement an emergent higher-order mind with goals, preferences, and competencies that none of its parts have. This is as true of the collective of neurons that enables our cognition as it is of the cells that build the body itself; Turing understood this well, in his interest in both diverse embodiments of mind across substrates and in the origin of chemical order during embryogenesis. The story of the self-assembly of the body and its emergence from a formless chemical chaos shares a deep symmetry with the emergence of minds from a mindless void.” (Levin 2025, p. 4).] and therefore, the conversation between AI and philosophy emerging at the interface of language models is a crucial space of analysis. As Negarestani presents it in _Intelligence and Spirit_: contemplating the possibility of AGI—which he defines as “a thinking subject with physical substrate that is not biological, or one that is capable of using an artificial language that in every respect surpasses the syntactic and semantic richness of natural languages”—is neither technoscientific hysteria nor intellectual hubris but actually a rather transformative expression of a possible new phrase for (critical) self-consciousness (2018, p. 40). **How** _critical_ the Platonic butchery,^[On the topic of computation as the cutting of chunks, Chaitin says of LISP (the programming language) that it is “like a surgeon’s scalpel or a sharp-edged diamond cutting tool.” (2004, p. 38).] and therefore what ratios this implies, is the question.^[“My excuse [for violence done by way of a cybernetic analysis of religion] is that it is only through the knife of the anatomist that we have the science of anatomy, and that the knife of the anatomist is also an instrument which explores only by doing violence.” Norbert Wiener 1964, p. 9.] From the philosophical “outside” we may come at the products of commercially-driven AI with an arsenal of practically-inapplicable reproaches, which is not always effective nor desirable. What is the alternative? This project makes a modest attempt at _language-modulation_,^[As argued in [[06 Principle of Sufficient Interest]], coming to see how language has a hold of behavior much more than we tend to give it credit for, leads to a modulatory meta-awareness, which can sometimes be helpful.] it it predicts itself forward by hoping to become parsed and reinterpreted by future language agents— human, machine or both: as we are and have always been cyborgs (Haraway cannot be left unmentioned)—in order to become something else. Partial answers to these questions will be given by tracing the implications of the observations above, and by proposing a series of concepts which enable new perspectives on philosophy, the mind and AI, and the possible overlaps and differences between these concepts as predictive strategies, through AIF. These proposals will largely by framed by the creation of neologisms that aim to be productive; generative, by paying their respects to earlier concepts and challenging them with a new distinctive functionality.
 
### How AI is currently diagnosed, and some answers from this thesis
 
>What, moreover, when we have put the decision [of life and death] in the hands of an inexorable magic or an inexorable machine of which we must ask the right questions in advance, without fully understanding the operations of the process by which they will be answered?
>
>Wiener 1964, pp. 68-9.^[It continues: “No, the future offers very little hope for those who expect that our new mechanical slaves will offer us a world in which we may rest from thinking. Help us they may, but at the cost of supreme demands upon our honesty and our intelligence. The world of the future will be an ever more demanding struggle against the limitations of our intelligence, not a comfortable hammock in which we can lie down to be waited upon by our robot slaves.” Ibid., p. 69.]
 
Melanie Mitchell (2021), contemplating the diverging paths of (old) symbolic and (new) statistic orientations in the field, presents four possible anthropomorphizing/centering/biasing AI fallacies. _Fallacy 1:_ “Narrow intelligence is on a continuum with general intelligence.” This tendency of thought is most often apparent in the cases in which we witness a non-human machine do something remarkable (e.g., beating Kasparov at chess), and assume the field much further along toward general AI than it probably is. Essentially, this is the problem of mistaking narrow performance on specific benchmarks for much broader advancement. Mitchell also notes how Heideggerian-AI Dreyfus had already remarked—against e.g., the visionary predictions-productions by Herbert Simon^[Demands/desires for generality in AI is nothing new if we consider Shaw, Newell and Simon’s _General Problem Solver_ was already around in 1957.]—that the “unexpected obstacle” always eroding at AI is/was the problem of common sense. Philip Agre also treated important aspects of this in _Computation and Human Experience_ (1997), as we will explore later. We advance an exposition of problems with narrow benchmarks, and the problem with the “problem” of common sense in [[03 Semantic noise]].
_Fallacy 2_: “Easy things are easy and hard things are hard.” According to Mitchell, the aspects of cognition that AI has tended to focus on as salient, relevant or particularly important, have led to the misconception of mind as something de- and reconstructible, accessible, analyzable. She suggests considering how not only how what we call the _unconscious_, but also the gradual sedimentation of evolution resulting in a system such as the brain, are processes which fundament cognition in ways we _quite literally_ cannot access (to make the link to Dreygedder again: ‘what we forget is there because it works’). We focus on these problems pertaining to the realm of the tacit or implicit, in [[06 Principle of Sufficient Interest]], where an account of social learning is explored, and in [[10 Bias, or Falling into Place]], where an account of intuition is treated.
_Fallacy 3_: “The lure of wishful mnemonics.” Mitchell points to McDermott’s (1976) criticism that “wishful mnemonics” presenting shorthands for AI programs, implementations or structures (“GOAL” or “UNDERSTAND”, p. 5) distracts from these programs not actually doing these things or possessing these qualities. This criticism is, again, very similar to the critique-exploration of metaphors presented by Agre. Currently, as Mitchell points out, we observe these linguistic effects revealing technodesires in concepts-problems such as _neural_ networks or machine _learning_. We treat this effect throughout the project by paying attention to the large landscape of (AI/machine) metaphors, as our title is clearly suggestive of. Additionally, we employ a returning strategy to exploit the productive generativity that has been the “animal as machine” or “mind as machine” historical metaphor. Some of the ways we do this are through concepts such as [[Modulations]], [[Chunk]], [[Parse]] or, indeed, [[Machine]] itself. Inheriting from Douglas Hofstadter (Mitchell’s mentor), much of this project focuses on the fuzzy aspects of analogical reasoning, and how seeing this as predictive-productive leads to new concepts and therefore to new ways of interfacing with reality: reading and writing, parsing and chunking. We try, therefore, to be careful in our analogies, but analogize nevertheless to find frictions.
Finally, _fallacy 4_: “Intelligence is all in the brain.” Mitchell points to the brain-centric view of cognition, and of cognition as information-processing that can disregard the body (which would be otherwise understood, in the context she criticizes, as a mere sensory input/motor output effectuator). On this note, regarding critical takes on information-processing: we take information-processing, defined as the _witnessing of change_ or _difference_, very seriously in this work. Considering our thoughts on [[Pattern]], [[Difference]], [[Modulations]] and [[Computation]], it should be clearly stated that we—in this work—cannot but abide to thinking about everything as change, which right now we can access through languages between creatures, autosemeiotic _sign-processing_,^[Sign here taken to mean anything with meaning, directionality, pointing, bias. See: [[Autosemeiosis]].] colloquially and concretely or highly abstractly. A variation on this effect are technical implementations of meaningful structures through signal-processing in formal computation. The semeiotic strategies we orient ourselves by set the focus on arresting mobile aspects of reality, which in our account can appear as in the service of the minimization of uncertainty (across an entire body-landscape, social landscape, etc.). What becomes telling, to our account, is how this is therefore essentially revealing of the **constructive** and **extended** dimensions of—collective, distributed, multiply perspectival, etc.—meaning-making. That is to say, as it becomes more apparent that processes cannot be “captured” by representations, it also becomes apparent that they ways we capture them (through representations, or [[Model]]s, for example) are not indices nor maps but entry points, new perspectives: actual constructions of possible explorations in the service of predicting-producing future encounters with other, novel processes. This is where reality meets science fiction meets philosophical speculation meets scientific experimentation.^[“Any transformation of matter from entropy to information, from a million sleeping transistors into differences between electronic potentials, necessarily presupposes a material event called reset.” (Kittler p. 150). Paper or voltage-through-chips computation is tied to a series of _other_ computing things which enable it and witness it, for computation begin, there must be an action (a state reset) that establishes the initial conditions from which what we call information can emerge. “When meanings come down to sentences, sentences to words, and words to letters, there is no software at all. Rather, there would be no software if computer systems were not surrounded by an environment of everyday languages. This environment, however, ever since a famous and twofold Greek invention, has consisted of letters and coins, of books and bucks.” (ibid.). These are all instances of chunks and parsings.]
It seems likely that, because of the rampant implementation of language-tools such as LLMs across the board, signal-processing in the production of text, and what we could in fact call pure informatic *texture*,^[In a scenario where incoming email is automated, and outgoing email is, too: who/what/where is _reading_?] will change how we relate to the concept and experience of meaning; relevance. Because of the ongoing and incoming exponential increase in generative _stuff_^[This is a technical term in image recognition meaning undifferentiated background noise; irrelevance.] it is also likely that the need for (automated) curatorial, contextualizing, summarizing, etc., services will increase as well. We will probably witness massive amounts of text being produced from “basic units” of **currently relevant** meaning structures (e.g., as complex as: “produce an entire database of alternative science fiction”, or as innocent as: “write a friendly email to my coworkers that I will not make the deadline”)^[The latter, at the same time, on the other end, might become summarized back into these “basic units” because, as Silicon Valley sells it to us now: people don’t have time to parse through vast seas of “useless” grammatization (e.g.: “write me a summary of all the incoming email regarding the project deadline”).] One may also think, not just about whether there is any objective reality to some possible “basic units,”^[Which there could be, some ulterior, foundational pattern, but given perspectivism: it is hard to believe in the possibility.] but where the belief in an objectively summarizable—and, at the same time, expandable yet expendable—reality comes from, beyond dreams of immortality. This brings in important questions regarding _compressibility_ that this project only provides a small entry into, and will be better treated in the future.
From the (possible) perspective of this project is that there are as many (possible) worlds as there are (possible) perspectives. We should therefore pay close attention to the creation of new perspectives and the futures possibly ensuing. Perhaps the question is: can we veer the course of this imaginary exercise from exclusively taking place in the realm of coloniocapitalist tool-and-product-oriented obsessions? Probably not. It also seems the “bias” discussion on the partiality of systems seems not to have taken _actual_ critical hold, we continue to promote systems—such as academia—which supposedly parse reality in “neutral” and unbiased ways, producing universal conditions which are far from universally generous. This leaves one wondering about the evolution of meaning through time: the signal-processing structures of the future will be based on some awkward remnants from the past, which are no longer relevant but **will determine** relevance (such as how the university erratically deals with the hangover it has). As this project sees it now, this is an inevitable effect of (language) evolution in general. Philosophy is what deals with this: concepts passing through bodies who inherit them. This is the _Poltergeist_ joke: the **corpus** through the **corpse**. Because of this, throughout this project, we promote music above language, like Nietzsche did. Sounding and musicking must have existed before speech and articulation (Cross 2001). It is very difficult to say where the music and poetry of language begins, and the normal, even formal, life of language is supposed to suddenly take hold. All language is a poetic call to order.
It is paramount that we look into this unfolding musical structure of language, its dynamics and rootedness in matter, so as to understand what we want to ask of “AI.”^[See also: [[What is a question]].] Margaret Masterman’s (analysis of Guberina’s) gestalt-take on the chunks of language and the speeds of breathwork, is something to signal to for the reader’s interest (see: Masterman 2005, e.g., pp. 227-233), something we have taken inspiration from. In the chunking done by what we call data science, a system which deals with granular depths human brains can barely abstract upon: how to chunk in multidimensional spaces that exceed all possible contortions, all ways humans can parse? In thinking the sociopolitical consequences of such systems, the paradox here is that (as a consequence of data science) “AI” is a (dis)illusion statistically composed: neither harmony nor dissonance are the rule here, we all vocalize, move and adjust to each other’s improvisations. One system to fit all, where the scales are such that we do not only not know how we fit, but but also where the “one size” fits no one. Again, history repeats itself. Leading us to observations about the possible rhythms of thought (its chunks; parsing): who gets to dictate the pace, the order, the ends. Open-ended computation does not allow for any proof of “one-dimensionality” (one truth, one objectivity), so for now we’re bound to remain in challenges which only deal with the modulation, the _tempering_ of speeds, and the intractability which is a result of attempting to predictively arrest _anything_ indeterminately. Anything we present sequentially—and in communication we primarily only have sequentiality as human perceivers, again see: Masterman 2005, p. 233—after the fact, opens itself up to generative interpretation, contemplation, criticism. At the same time, it’s bound to remain only a symptom, a chunk that has occurred, becomes interpreted, and needs to be parsed.^[As explored in [[07 Phenomenology of Sick Spirit]], there is no depth to symptoms, symptoms are all there is. And AI is plastic. AI is _like_ plastic. AI is the very idea of plasticity. See also: [[AI is plastic, plastic is AI]].]
Stuart Russell’s view that AI has vast both potential and has unstoppable _momentum_, and that current approaches will inevitably lead to a loss of control, leads him to conclude that the only viable option which is provably beneficial is not making AI safe, but _making safe AI_. This is another call for a careful, slow approach, albeit completely different than that of, e.g., Isabel Stengers. On the other hand, it is also important to consider proposals by the likes of Joscha Bach and other techno-optimists. In their view, in order to protect “the human” from events such as natural disasters, we need AI approaches that might seem unorthodox from current perspectives.^[See, for example: [[DigitalFUTURES 2024 Doctoral Consortium on AGI 1, Joscha Bach, Is consciousness a missing link to AGI?]].] While I do agree that we need to challenge thought with extinction possibilities even beyond those of its own possible making, it is important to remember some foundational problems inherent in all risk-assessment: “Risk expertise[’s] authority is derived precisely from the capacity to decide the exceptions that make it possible to act even in the face of radical uncertainty.” (Amoore 2013, p. 8)^[Amoore continues: “Even where the limits of statistical risk knowledges of probability are breached by the exceptional case—a terrorist attack, flood event, hurricane, pandemic, or financial crisis—the more speculative knowledges of scenario planning, social network analysis, and predictive analytics reorient risk to action on the very basis of the imagination of those possibilities.” (ibid.).]
Such a “precautionary principle” (Ewald, cited in ibid.) becomes capable of authorizing _anything_, given high enough uncertainty bets. Besides, when humans are often the only risk, what also seems puzzling in light of our coming to terms with processing _change_ (changing minds, changing substrates, changing relevance): why do we care about the human being? We are, a lot of the time, terrible creatures: competitive, destructive, narcissistic (Prince (2023) grounds his thoughts on AI ethics in these observations about human beings, too, p. 13). Why not work towards a future **without** these characteristics as actual “safety”? Why not, perhaps, work towards abstractions such as conceptuality, or abstraction itself, as a project? Hoping instead for _loving semantic attractors_ to survive,^[This is a joke about _All watched over by machines of loving grace_, 1967 poem by Richard Brautigan, and title of the 2011 series by Adam Curtis.] rather than their particular instantiation through the vehicle of human beings. Perhaps I ought to be accused of misanthropism with such claims. But let this be clear, the case is exactly the contrary: loving that I love that humans love love, I would love for _love_ to endure, rather than for the chance that it is eaten up and destroyed.^[In _Pleromatica_ (2023) Catren sees love as a “meta-idea” (p. 465) that can be largely ascribed to the flourishing of life and, in human culture, to religious practices. This fundamental perseverance unfolding across fields (between living beings), Catren shows, lacks in other realms (science, philosophy, etc.) as a method of guidance. We would also note that religion is also the realm that enables plenty of war and destruction.]
From the perspective that we sustain cross-generational patterns-concepts such as love, it is not that far-fetched to suggest that we ought to pay attention to the pattern much more than the substrate. Separating message from channel is not just Shannon’s foundational observation: the medium really is the message (McLuhan), because it is the message that we care about. _Every intelligent ghost must contain a machine_: the software is transmittable, the substrate is brittle (and unreliable, and jealous, and hungry, etc.). But thoughts in this project are not settled on this, and if they were, that would be a rather stagnant sediment. On top of this sediment, currently, the economic system we rely on for ‘getting things done,’ is in the business of exacerbating of qualities which, as I see it, run counter to the patterns I find important to preserve. Money, the universal solvent, is an uncertainty minimizing structure that simply encourages the selfish gathering of its own abstract object in order to secure an unloving future: because hoarding, at the levels we currently witness, is simply the *possible promise* of something else, it is holding out for... what, exactly? It is a recursive function that seems to have gotten stuck. How do we start moving away from that? I asked Joscha Bach this question, in a convoluted way, but received no answer about the infancy of our home economics. The difficult political situation is that, attended to by another Russell (not Stewart): “in an increasingly networked world, through and beyond the digital, we find ourselves constantly navigating how to care for those we cannot see, whom we do not know, and negotiate what we perhaps feel entitled to take from them.” (Legacy Russell, 2021, n. pag.).
If the controversies inherent in concepts such as ‘intelligence’ emerge from their complex, metaphorical, multi-anecdotal nature,^[See: [[All knowledge is anecdotal]].] then this phenomenon should be considered a (socially) enjoyable opportunity and not an undesirable side-effect. This project will make attempts to witness how it is that _reasonable_ concepts—i.e. pertaining to reason, which reason, and are reasoned about—enjoy variable, divergent definitions precisely because their ‘meaning’ is as much in development as their technical sub- or superstrata are. A fact of engineering _and_ of cognition is that simplicity can be made complex, and complexity can be made simple.^[This is explored in [[12 Negintelligibility]].] The modulation of these two extremes is what is often referred to as _intelligence_: something which can compress vast dimensionality, as well as summarize it into coherent units. Attention to the collectively modulated, semantic metastability of concepts as functions of the complex and simple operations we entertain, can yield a critical interpretation of the technocapitalist abuse (and monodimensional neglect) of this fact, as well as propose the composition of a novel engagement with the nonlinear, multiply-realizable effects of concepts. “The identification of a generative metaphor is often a liberating experience, giving concise expression to a previously diffuse sense of being oppressed by unarticulated false assumptions. Unfortunately, the most common response to such a discovery is to place the blame on metaphors themselves, rather than on the unreflexive neglect of metaphors and their consequences.” (Agre 1997, p. 35).
 
### Writing about AI in the age of AI acceleration
 
“AI” research today is incredibly exhausting. I have been tempted to abandon fueling the fire altogether. It is not necessary to cite the myriads of frustrated AI voices chanting the fact that whatever is written about AI becomes outdated as soon as it is printed (to cite Mitchell again).^[Or Cameron Buckner: “Trying to write a book about the philosophical implications of deep learning today is a bit like trying to saddle a mustang at full gallop.” 2024, p. 9.] Others such as E. M. Bender, as already noted, suggest we should stop using the term “AI” altogether.^[Also, regarding frustration, AI, and being sidelined, I had to put this somewhere: “Oh the frustration of being found frustrating! Oh the difficulty of being assumed to be difficult! You might even begin to _sound like_ what they hear you as _being like_: you talk louder and fast as you can tell you are not getting through. The more they think you say the more you have to say. You have to repeat yourself when you keep coming up against the same thing. You become mouthy. Perhaps we are called mouthy when we say what others do not want to hear; to become mouthy is to become mouth, reduced to the speaking part as being reduced to the wrong part.” (Ahmed 2023, p. 23). This has happened way too much in the context of my PhD, to myself and others. Not only this, but also: being listened to dismissively, as if the listener _already knows_ what is being said, being looked at as if one doesn’t _look right_, being ignored because of not _appearing intelligible_. Not being taken seriously. The guy next to you saying _the same thing_ you just said, but _him_ they listen to. I thought some collegial parts of the world were beyond this, in many respects, but they’re far from it. Also: “Killjoy truth: if you have to shout to be heard, you are heard as shouting.” Ahmed 2023, p. 34). Terrible contemporary truths.] This much can be said about most emerging and quickly developing (“disruptive”) technologies,^[“Disruptive” because of the speed at which we never asked for them.] and perhaps about all published knowledge in our interconnected day and age. It is for this reason that this project will be an evolving book. It is meant to serve as a resource for others interested in these topics, and try to cope with an advancing landscape of thought.^[Thorough analysis of the topic of AI is pretty much an impossible task for a single work. This is why this project has taken the form of a living book, all thoughts presented here are subject to change. The project is hosted at: www.n-o.ooo. Previous versions of this “book” can be accessed thanks to the wonderful _wayback machine_. This also project takes advantage of the possibility of _hyperlinking_ (see also: [[Anticontinuous]]).]
Philosophy deals with the mobility of concepts, and this work of philosophy deals with a _rapidly_ moving concept: so-called ‘artificial intelligence’. These computational systems are “also philosophical systems—specifically, mathematized philosophical systems—and much can be learned by treating them in the same way.” (Agre 1997, p. 41). Cross-examining conceptual evolution in both is necessary, “[s]haring concepts does not suffice for understanding even when they are associated with the same words; the pragmatic ability to make sense of utterances in context is also important, and a challenge for AI.” (Butlin, 2023, p. 3080). AI, as a pragmatic concept, not only attempts to encompass the entirety of the transdisciplinary field that is cognitive science and its multi-level philosophical implications, but it is also—as opposed to related-concepts such as ‘reason’, ‘autopoiesis’ and ‘evolution’—very strongly driven by its highly problematic military-industrial-technofinancial context, the tacit and boring target of much of this work’s and others’ critique.^[As a hopeful and friendly reminder we can also think about how “The world of the capitalist subject—unfolding between the devastation of the earth and the ‘conquest of space’—is **only one world among others**, a sick world that—having lost its collective soul—becomes more and more detached from the living forces of the arche-soil.” (Catren 2023, p. 460, our emphasis).] Tracing lineages and legitimacies is, often, all we can offer in the humanities. Perhaps not unlike how like biology carves out species, playing catch-up with organic evolution: both of these are chunks and parsings observing different speeds. Regarding conditions of possibility, we can also recognize that the condition of self-incurred damnation presented by our current technocapitalist landscape is not only just _one_ of many conditions currently unfolding, but importantly that “our current situation is still an abstract image that depends upon the perspective defined by our available transcendental resources ... to explore the possible deformations of these regimes ... other affects, other percepts, other concepts ... other valves, other filters, other resolutions, other scales, other sieves, other categories, other bodies, other languages by means of which it is possible to constitute [other realities, without falling prey to misplaced universality].” (Catren 2023, pp. 460-1).
And if nowadays “automation can be dynamic and not dependent on a prescribed set of calculables” (Parisi 2015, cited in Chukhrov 2020, p. 68), then we could be tempted to ascribe to it the ability to produce its own conditions of evolution, since we certainly lack a level of _access_ to (the reproduction of) its possible prestatability conditions—irreversible _blackboxing_ problems across all fronts: from big data to all possible intractabilities emerging from our demands. This should urge us all the more to collectively seize the means of computation (Doctorow 2023). Practically, this is entrenched by current power structures. Speculatively, this is bothersome: humans create a pattern-digesting technology capable of exceeding the pattern-processing capacities of humans, but then: patterns become illegible, precisely because of this. And because of this, all reference to AI in this work will contain this irresolute, generative paradox. Of course, like with any other concept, we speak of something in the making. Something which currently seems to _speak back_ more than other technologies, and thus sets the ground for a reconsideration of _speaking_ and of the future evolutions of concepts which have been fundamental in historical self-conception(s) of all kinds. The gradual appearance of something (“AI”) which seems to enjoy a different relationship to concepts than “humans do”, makes humans anxious.
 
>The AI community has, by and large, found [philosophical critiques] incomprehensible. One difficulty has been AI practitioner’s habit, instilled as part of a technical training, of attempting to parse all descriptions of human experience as technical proposals—that is, as specifications of computing machinery.
>
>Agre, _Computation and Human Experience_, 1997, p. 21.
 
If the concept of intelligence remains largely undefined, be it in the field of cognitive science, AI or philosophy, then the main heuristic of this project is to elaborate some elucidations of the necessarily recursive ungroundedness of the concept as it repeatedly grounds itself on convenient (technical) premises, as well as pay close attention to how this concept fails at patching up unfinished definitions within itself. The work borrows from said fields, as well as from psychology, biology, and others: the employment of these various bodies of knowledge is not only meant to present an ample enough context for the themes pertaining to intelligence, but also to allow these fields to converse with each other in order to highlight their (dis)similarities. The discussion is also unavoidably selective, given the vastness of available literature on the topic of intelligence, ranging from molecular biology to political science and beyond. What is studied here is aimed at finding points where (a)symmetry, divergence and dissonance may be witnessed, particularly by observing the continuum between philosophical claims pertaining to semantics and reason, cognitive science models and recent AI developments pertaining to prediction and intelligence. From the perspective of this research, it can be quite unambiguously stated that philosophy needs to reflect on the technoscientific imaginary, and vice versa. ‘Imaginary’ here is not meant to invoke a critical undertone of (techno)science as relative and “socially-constructed.” That which is imaginative pertains to _all_ knowledge-seeking endeavors; it is the access to and retrieval from the combinatorial space so often attributed to what is termed _intelligence_: i.e. the capacity to observe-yet-think-otherwise, to dialectically interact with an interlocuted ghost and counter-_fact_. Throughout the work, different philosophical perspectives are brought to bear on particular scientific questions regarding intelligence: intentional behavior, the boundaries of the _mental_, the definitions of _explanation_ or _understanding_, and even the exploration of the origins of different types of motivational states, based on a personal account of experiences with the administration of corticosteroids.^[In [[07 Phenomenology of Sick Spirit]].]
The advancing and uncritical dominant approach to technology as a human “universal,” is countered by Hui’s proposal of technodiversity, where it is, as van Tuinen paraphrases: “[o]nly through the affirmation of “technodiversity” [that] technics become universal. This means that we do not deduce difference from sameness (as in the identity politics of multiculturalism) but induce sameness through the affirmation of differences.” (2020, p. 99). But perhaps we should not even affirm, and just let it go. Affirmation is impossible, as an _absolute_ it is inevitably _incomplete_ (always out of phase with itself (Simondon)), composed of a vastness of predictive dimensions which are all differently interested, and each of them as composed of/with inevitably contradictory interests (desires, initial conditions): all the way down. Your body’s mitochondria, my liver, tropical weather systems, memes on the internet, etc., all coupled, all predictive events in their own right, the _enduring_ qualities of which only become salient through the modulation of affirmative attention: affirmation is the starting point of control and domination. Attention for modern humans is primarily modulated through discourse, and we can only work at this with what we’ve got: fuzzy exchanges.^[See: [[E Pointing]], [[Vantagepointillism]].] Understanding dimensionally-complex attention begins by accepting possible interconnected communication between all parts, while inevitably, perpectivally knowing that synthesis is far from guaranteed, but always a possibility. “The very existence of the living insubstance relies on the finite local modes that express it in unexpected manners from a multiplicity of singular perspectives. ... what is ‘beyond’ our finite lives is life itself, the impersonal ‘immanent cause’ of our finite lives, the open multiplicity of heterogenous lifeforms, and the transpersonal living communities we are part of.” (Catren 2023, p. 368 & p. 380).
But what can we say, other than “proceed with caution”? The alternatives are sparse and, arguably, unrealistic. Not proceeding seems impossible, we’re marching down a one-way street already. Proceeding without caution rings ridiculous and careless (though, again, most “AI” nowadays is based exactly on the principle of moving fast and breaking things). And if, “the resistance to contingency is the signature of Western philosophy: the overcoming of the irrational through reason and control” (van Tuinen 2020, p. 7), and if we have always lived in a time when “[t]he privileged are processed by people, the masses by machines.” (O’Neill 2016, p. 8, O’Neill claims the time is now), and if the project of AI offers even more complications to these observations by revealing how we are far more soulless than we assumed to be, then what? The sad side of the story is that _poltergeist in the_ _machine_ is the—necessarily—mindless, intersubjective, disembodied ‘cultural algorithm’ on the loose: predictive strategies (“those are not human beings”) that perpetually eat at the margins (the marginalized; as well as the future).
 
### Terms: provisional definitions
A lot of the terminology in this project is explained in the [[Appendix]]. But before muddling things it would be convenient to set up a brief preliminary lexicon in order not to confuse a possible reader, as this project is highly eclectic: it takes risks by merging many areas of knowledge and experimenting with language (sometimes to an annoying, but hopefully amusing, degree).^[One of the project’s goals, after all, is that of [[Language-modulating]], so this cannot be avoided.] Since _intelligence_ is what we are after, and we have already presumed that it is currently unexplainable (as is any complex concept), let’s begin with aforementioned concepts that we think can be delineated to a degree of accuracy amenable to this project’s needs. These concepts, presented as follows, are central tempering elements of all the arguments/proposals explored in the project. The following definitions are thus neither final nor inflexible, they are meant to elucidate a specific space of thought for this project in particular.
**[[Philosophy]]**: the realm of abstract thought, where witnessing is witnessed. Abstract in many ways. In order to point we must ignore all else. *Ab-stract*.^[Catren: “... an _abstract modality of experience_ is a mode of experience resulting from a process of abstraction, i.e., from a process by means of which one abstracts certain dimensions of experience in order to focus attention on a single dimension. For instance, I can focus on the demonstration of a theorem by abstracting the perceptual field (e.g., the temperature of the room I am in, the visual field that surrounds me), the affective field (e.g., my current mood, the pulse of a symptomatic tic that affects me), and the social field (e.g., the actual human environment in the space where I am, the current political situation). “The Trans-Umweltic Express”, Glass Bead Journal, 2016.] Abstract, at least: both in the sense of abandonist^[Think the _armchair,_ again. See also: _abstrahere_.] and in the sense of something which, after much ado, results in conceptual X-rays,^[Jameson calls the schematism, in Kant, “the X-ray of a concept” (2023, p. 31). We take from this metaphor the reductive yet functional image of the X-ray as that which reveals an underlying, general structure behind something more dynamic and fleshy. Also, a further note on transparency and opacity: “For Leibniz, every text is only a pretext and every communication an initiation. “Those who know me only by my published works, do not know me” (Leibniz, _Opera Omnia_ 64–65). By moving in and out of the shadow, we protect what we communicate against the risk of overgeneralization and immediately implicate ourselves within a kind of shared intimacy.” (van Tuinen, 2020).”] that is: in the assigning of names to presumed patterns which apply at different spatiotemporal scales (e.g., the concept of “linearity” or, relatedly: “history”). Interference between scales is perhaps the most salient aspect of philosophical disagreement (this is treated in [[04 Concepts as pre-dictions]], and explained through the concepts of [[Chunk]] and [[Parse]]). Philosophy, for our purposes, includes aesthetics and mathematics. It includes aesthetics—in a Nietzschean fashion: _all_ is aesthetics, and eventually all aesthetics is (in)digestion—because the relentless metabol precedes and processes the conditions of thought, as it seems we, creatures, *sense* before we think. It includes mathematics, as mathematics deals with the possible shapes of language(s): it deals with the most abstract structures that can be made spatiotemporally durable. _All language is philosophy_, whether those who disregard or dismiss philosophy realize it or not. All words bottom-out at self-reference (Derrida and Wittgenstein, but also Peirce) and/or unprovable ontological commitments, therefore, language’s philosophical excess cannot be denied nor avoided. All philosophy gathers momentum from the negative (Hegel via Žižek), and can sometimes be understood to *hate* itself (Deleuze, Negarestani).
**Parse**: To parse is to _process chunks_, to consume segments, and (in most cases, therefore) to produce an output, to _compute_. At least, as we currently see it, because information flow has directionality and cannot be destroyed. However, both _chunk_ and _parse_ can sometimes be confused (e.g., when encounters with new patterns occur).
**Chunk**: We use this word, sometimes literally and sometimes metaphorically, to refer to the way(s) in which _unstructured or otherwise unmanageable_ (i.e., unpredictable) information is divided up, in order to make it _parsable_. The reason _chunk_ and _parse_ can be interchangeably used is because sometimes we may parse an already predicted chunk, and sometimes parsing implies novel chunking.
>By _chunk_ is meant here the smallest significant language unit that (1) can exist in more than one context and (2) that, for practical purposes, it pays to insert as an entry by itself in an [mechanical thesaurus] dictionary. Extensive linguistic data are often required to decide when it is, and when it is not, worthwhile to enter a language unit by itself as a separate chunk.
>Masterman ((1956) 2005, p. 83).
If, according to Masterman, science is the experimental creation of analogies ((1980) 2005, pp. 296-8), we take the liberty to expand on the concept of _chunk_—which she beautifully studied as the resulting from rhythms of how much can be said in one breath—and apply it beyond the context she refers to. All chunking a decidedly historicomaterial, evolutionary matter. For us, therefore, _chunk_ is significance; salience, in general. A poem is a chunk, an hour is a chunk, a person is a chunk. We also use the word _parse_ technically but flexibly (sometimes interchanged with _chunk_), it can be applied to any system which filters (language) data, human or otherwise. This means organizing tokens (e.g., as in the _chunking_ done through LLMs) and/or rendering a token (an event, a singularity) essentially predictable under a type (i.e., a conceptual category to parse reality). “The interpretation of encoding can take the form of a program that outputs instances of [a] type or alternatively the construction of a class of paths in a continuous space.” (Cavia 2024, p. 13).
The metaphor is particularly apt in our current scenario: both human systems and LLMs have **information capacity limitations** that make chunking necessary, and parsing possible. Humans, in the flesh, have memory and physiological limits (I suspect _the infinite_ does not (always) make us go insane because we eventually always fall asleep and are mortal) and LLMs have “context windows” and “attention mechanisms” to enable parsing. Both humans and machines segment continuous streams of information into discrete units allowing for processing, and although the underlying mechanisms may differ significantly, the metaphor is employed in order to run into their differential effects and explore, precisely, how it is that they are different. This strategy borrows from the productive generativity that has been the “animal as machine” or “mind as machine” historical metaphor. Saying how they are alike signals, often, how they differ, leading to new concepts and ways to interact with reality.
**Language**: the mark of _any_ language is the establishment of an arbitrary bond, which results in the preservation of that arbitrariness as a function. In general, the ensuing effect of this, leads to language as something that can be described as a system dynamically sustained by a collection of beings, which supports-controls the collective discovery-invention of affordances. Language can be understood as a life of its own (Wittgenstein), having the capacity to modulate predictive attention _through_ creatures. Language is what _enables_ a lot of chunking and parsing, and therefore gives us _meta_-physics: vast speculatively predictive scope. Language’s redundancies can, if observed; chunked, for periods long enough, be understood in analogical parallel with the lives of creatures: each word, like each individual, is an attempt which serves for the survival of an (initially always arbitrary and aleatorily-produced) function, a trait.
**Function**: the characteristic operation or effect that an individuated entity, system, or process produces and (temporarily) allows to persist and evolve within its context of interaction. Functions reveal themselves as crucial/salient/relevant/etc. at the time of observation, in its a role within a larger system. This take allows us to focus on what something does rather than what something is (on identity), which has theoretical, practical and therefore political implications. Thinking about the tape and the machine again, about whose agency are making salient—the written or the writing?—it could be said that we only have subject/object re(ve)lations, or functions, through a lot of our languages: paths from x to y. Function allows us to move beyond that: we observe the _path(s)_, rather than the subject/object.
**Model**: the “X-ray-like” (Jameson 2023) pattern in the practices named above (all practices are modeling practices) that becomes dynamically workable and, sometimes, applicable. E.g., we may observe the periodicity of the planets as a pattern, which can lead us to models such as the Ptolemaic or Copernican versions of the Solar system. Or: we may observe the pattern that is a plant resisting the seasons, leading to the modelling of aspects of plant nutrition such as photosynthesis. Note: concepts can be thought of as (sometimes rigid, sometimes loose) models. More on this in: [[Concept]], [[Model]], [[Pattern]].
**Pattern**: the central conceptual commitment of this project, the already differentiated structure of things which is chunked and parsed. Can sometimes be thought of as shadowing the concept of _information_, but this project prefers _pattern_ as it seems to unmuddle the field, while information is a highly contested notion with many definitions across many areas (see also: [[Information]]). A pattern can be defined as a _repetition of differences_.
**Difference**: the unavoidable starting point of anything, i.e., where the quality of any pattern, as pattern, begins (though we also need to think of _equivalence_ here, which enables repetition, more on this in [[Equivalence]]). Perceptual-conceptual salience _is_ difference, and vice versa, and anything repeatedly/ably salient is what makes a pattern (see also: [[Difference]], [[different-ciation]], [[Equivalence and difference]]).
**Equivalence**: where the notion of difference bottoms out. But also, relationally: difference is the notion where equivalence bottoms out. Both concepts are complementary and cannot be separated. Note: we opt for equivalence over identity or repetition, this is explained in [[Equivalence and difference]]. Much of what is treated around the concept of equivalence will pertain to the aforementioned (problems of) _sameness_ and [[Symmetry]].
**Prediction**: the basic _patternal_ activity we observe across all regimes of sense/thought/modeling/etc.^[The concept of _patternal_ is indebted to its etymology, pattern is _literally_ paternal, and re:modeling to observations such as this: “No history or sociology can camouflage the complacent paternalism of this thesis, which simply transfigures the so-called others into fictions of the Western imagination in which they lack a speaking part. Doubling this subjective phantasmagoria with the familiar appeal to the dialectic of the objective production of the Other by the colonial system simply piles insult upon injury, by proceeding as if every “European” discourse on peoples of non-European tradition(s) serves only to illumine our “representations of the other,” (Viveiros de Castro 2014 (2009), p. 40).] which allows for the sensing/thinking/modeling/etc. of salience (i.e. difference/equivalence interplay: chunking and parsing). It is very important to note that, despite all the criticisms it has seen as a concept (and rightly so in many domains which aspire to domination by way of _pre_-scription), prediction **does not imply providence or determinism**. Prediction is what allows for the senses to be synthesized coherently, for music to sound, for this text to be read, it permeates all action-perception-cognition and is an unavoidable feature of _life_: anything that tracks spacetime must be producing predictions about its future states (see also: [[05 Prediction]], [[04 Concepts as pre-dictions]], [[Free energy principle]]).
**Computation**: Following arguments made by Agre (1997), formalisms are only interesting if they couple to life, if life is understood as always _tending_ towards another state, and its formalisms as placeholders of this dynamism. As stated earlier, we frame computation as **different formal methods amplifying the capacity to witness _difference_**, parse it, and therefore organize perspectives around patterns emerging from differential exchanges.^[Gregory Chaitin, in _Meta Math!_ (2004) says: “the computer was (and still is!) a wonderful new **philosophical** and mathematical concept. The computer is even more revolutionary as an **idea**, than it is as a practical device that alters society ... the computer changes epistemology, it changes the meaning of “to understand.” To me, you understand something only if you can program it. ... Otherwise you don’t really understand it, you only **think** you understand it.” (p. vii). We would only dissent on the last part (the classic engineering trope), because, precisely: to question the meaning of computation on the basis of its appearance on the scene, to question a possibly digital physics or philosophy, is to constantly, dialectically reinvent the concepts constructively so as to arrive at new paths between existing, or possible, things. If Chaitin sides with intuitionism (ibid.), then open-ended programs are the only interesting ones, and even though we can program them, we certainly do not _understand_ a lot about them.] This process manifests in two interrelated ways: first in how patterns are identified and grouped together (what we call “chunking”), and second in how these patterns are processed and interpreted (“parsing”). These two processes often work together dialectically, as well as through retroactive interpretation (e.g., the parsing of the past changes as we make a chunk out of a pattern in the future). See also: “mortal computation” where “the definition of a mortal computer does not commit to implementation in any particular substrate/niche.” (Ororbia and Friston 2023, p. 22).
**Reason**: reason, as explained earlier, can be understood is an orienting activity in affordance manipulation.
**Cognitive science**: deals with philosophy in the flesh, and emerges from a (often war-driven) technico-medicinal impetus.
**AI**: deals with philosophy outside the flesh, and emerges from a war-driven, capitalist-technical impetus. Much of what we have to say about the concept is captured throughout the project.
These concepts are briefly delineated here, and certainly incomplete and only serving the purposes of this project, they can be further explored in the respective entries in the [[Appendix]].^[Please bear in mind that presenting the history of these concepts is the same/just as important as presenting their semantic-conceptual structure (as according to this project’s perspective/filtering). Therefore: both approaches will be mixed, sometimes implying the Carnapian difference between explication and explanation (_Meaning and Necessity_, 1947), where explanation draws out the _whence_, the history, while explication delineates the specific contours of a concept’s (constructive) structure and its (logical or “rational”, i.e., predictive; chunking) implications. We are not married to this idea but find it useful in the context of the interplay between pattern and model: explication is a positivist-constructivist proposal, it aims to define how a concept is constructed, in other words: how its pattern can be put to work. If the “work” metaphor does not sit nicely, please understand it broadly: as implying that something can move, transform, become, etc. While Carnap sought a certain forma/ulistic closure with the concept of explication, we would like to present it _inexactly_ here in order to allow for the emergence of new concepts, which is something Carnap also wanted (although in his proposal there is an incremental architecture one should aim to construct, rather than a house of flying cards. We are not sure where we stand here). N.B.: the entire project is littered with these kinds of jokes. Inspired by Nietzsche of whom Kaufmann writes that it seems as if through the entire writing in Zarathustra, he “prefers even a poor joke to no joke at all” (1968 (1954), p. 107). At the end of the day, humor is all. At least, it is the “deadliest of weapons against all that is pompous, staid, and holy”. Watson on Zhuangzi’s methods, 2013.]
 
### Short- and upcomings
The most obvious critique this work deserves is that concepts are overstretched (especially, e.g., _computation_). However, the retort is that this is precisely what the project argues: concepts are more flexible than they appear. Stretching them is what might lead to novel structures. Additionally, it must be said, part of the impetus here is to set computation on a continuum between meat and circuit boards. After all, electrical currents are continuous, and flesh, too, is electric. However, point taken: the danger in extending conceptual structure lies in losing analytical traction, creating confusing equivalences between, perhaps, fundamentally different processes.^[However, making the ontological claim that there exist fundamental differences might imply more unwanted problems than the general tendency towards _oneness_ that this project displays/prefers.] By treating everything as computation, we cloud some aspects distinguishing biological from other types of information-processing.^[Sorry for contrabanding this here.] Yet the drive toward total abstraction serves both computational and philosophical aims—the philosophical and computational pursuit of universal systems that apply “everywhere and nowhere,” detached from particular instances and material constraints is the problematic _generalizability_ we aim to study. In order to study it: it must be set center-stage. The perspectival-overstretching risk remains, though, this problem is treated in [[Vantagepointillism]].
On scales and speeds: prompted both by the speed of the field, and by my own unexplainable obsessions and lack of orientation, I tried to do too much: speak too many languages, please too many crowds, and at the same time rattle too many cages. This is the beginning of future, more concentrated research. With a topic such as AI it’s nearly impossible to not fall prey to doing too much. I tried to narrow my focus at times, e.g., in [[03 Semantic noise]], or in [[07 Phenomenology of Sick Spirit]], but other times I could not but take risks, be general and sometimes even rude, e.g., [[10 Bias, or Falling into Place]] or [[09 C is for Communism, and Constraint|09 C is for Communism, and Constraint]]. In these latter chapters, it will be clear that this is what I took the _poetic license_ of philosophy to mean, and it is an advantage as much as it is a shortcoming. I struggled with the impulse to want to dislodge from the canon, the formats, the methods. Inspired by scholars such as Catren, Eshun, Ferreira da Silva, Moten and Harney, fueled by the impetus of an ever-looming background Deleuze and a surreptitious Derrida. Challenged by the diagnoses of academic adaptation as overly concerned with its theoretical pedigree, a ‘reliance on illustrious names and specialization in them’,^[Luhmann 1995, xlvi, but more remarkably and earlier: Du Bois 1905.] and universalizing tendencies. Sometimes (often), because of this, I end up with ugly patchworks: pieces that are far too disoriented. In challenging that I did not know where I wanted to arrive, I produced some texts that verge between the mediocre-illegible and the melodic-comforting. Sometimes I loved this effect, sometimes I despised it, yet was compelled to engage it.^[I write like a dog, see: [[Anticontinuous]], I also _hide_ behind metaphors. “By moving in and out of the shadow, we protect what we communicate against the risk of overgeneralization.” van Tuinen on Leibniz (2019, p. 102).]
What I see as a benefit in philosophical poetic license is that it _polycomputes_ (Levin), allows for interconceptual exploration. Its text performs a variety of remarkable functions, all at the same time. Text, because it is not constrained by the speed of breath and speech,^[The speed that conversation requires, for example.] can be figurative and declarative, but also demonstrative and performative: it can do what it says by reflecting on itself.^[Conversation does this as well, as does thought, both _sui generis_, but is often too fast for us to realize this. Except, sometimes, through jokes, slips and trip-ups.] In considering this, it would be difficult in the case of this text for it to be **honest** and at the same time not be produced in the manner which follows. Text is often linear, but associative thought can hardly be said to follow a line.^[On the other hand, dialogical thought (two or more entities converging into yet another conversational entity) can carry one linearly, and text can invite non-linear readings and pauses.] Text invites a forced, or forked,^[It would be highly unlikely that if this text told you to start at reading this introduction again, you’d actually do so. However, if you did, you'd reread the first couple of sentences entirely differently.] association, which would otherwise not have been. In betraying specificity, I also produced something perhaps so wide that it is flat. But: flatness provides a certain kind of perspective. A map is ‘useless’ without the capacity to move through the territory it represents, but it remains an interesting object of contemplation. A painting elicits a response that has nothing to do with paint, a program directs behavior towards things other-than-programs.
To save myself a little more: it would be naïve to claim that all that is published is ultimately ‘finished’ and thus lacking in the associative potential which will follow from it being read.^[This is me asking you for mercy. But also another reason why the living book exists.] What happens when something written is thought about? Do thoughts crawl; fly out of these words? Are they ripped off? Copied or emulated? Teleported? What is the basic unit of (_a_) thought?^[**Note:** These questions are strong plays on words and represent moments in the history of philosophy, not sure if a footnote is needed or would off this.] A highly contested notion where chunking-parsing range anywhere from a couple of seconds or minutes, to variations in entropy, to paragraphs, to complex instances where thought is performed and/or experienced dialogically. We don’t know, this has urged me to write as I did. It would be dull and perhaps impossible to write a text about intelligence without reflecting on these thoughts happening here as they occur between us. _All knowledge is anecdotal until it isn’t._^[See: [[All knowledge is anecdotal]].]
It is important to once again warn the reader that some chapters are more explorative and risky in their style and level of speculation, while others play within safer margins. This thesis is highly speculative, this is the risk it allowed itself, and in light of the technical risks of the past: at least here we are only _suggesting_ modulations of thought. The authors which will be encountered here are from diverse and sometimes “incompatible” fields. These thinkers (most of them presented in the introduction) are contemporary authors who reflect on current AI developments and their histories. Deeper historical overviews will be traced in the chapters, involving compatible traditional and foundational AI authors such as A. Turing, J. McCarthy, C. Shannon and W. Weaver, H. Simon, N. Wiener, who are taken as structural sediments of our narrative.
If the above seems satisfactory enough, please continue reading.^[Also, in the spirit of relaxation, if one could imagine the formality of most finished theses as a celebratory dinner at a restaurant with thick napkins, resentful waiters and terrible acoustics, then see this as the cold pizza at your friend’s place, any day, any time. Norbert Wiener writes in the introduction to his essay on religion “God and Golem, Inc: A Comment on Certain Points where Cybernetics Impinges on Religion” (1964) that “The spirit in which [this work] is to be undertaken is that of the operating room, not of the ceremonial feast of weeping about a corpse.” (pp. 3-4). We, in this project, want both. PS: I am also very sorry if, in the future, any of the concepts/methods in this work become co-opted by antisocial assholes. This work is highly social and means no harm. I am equally sorry if I am unaware of using things that may make people think this is a work of bad magic, in bad faith. It is not.]
 
### Chapter outline
Since AI is our topic, concepts are our vector, and concepts effectuate themselves linguistically,^[Just as it is almost impossible to differentiate between perception and action, a returning theme around the actively-inferential corners of this thesis, it is equally difficult to distinguish between concepts as linguistically effectuated, and concepts as affordances; as the categories of perception we are bound to, in a Kantian sense. This is because of the simple observation that the abstracting, future-projecting quality of concepts, rests on a network of other, more “perceptual” concepts, such as object-permanence and the continuity of sensory appraisal. We will make sure to note the distinction between one and the other wherever necessary, but the reader is encouraged to read both elements into the concept of concept, whenever the concept appears.] we are bound to an analysis of language-modelling in AI. Our first step of inquiry is thus into _Natural Language Processing_ (NLP).^[As well as _Natural Language Understanding_.] In [[03 Semantic noise]], we analyze some basic (generative AI) problems: the main argument made here is that if one thing can mean many, and that this is what enables conceptual generativity ([[Polycomputation]], [[Language-modulating]], [[Meta-language]]), then how come language-modeling cannot do justice to this basic, fundamental linguistic fact and its dialogical effects?
_Semantic noise_ was the type of noise Shannon and Weaver diagnosed as the effect ensuing from this denotative variability exhibited by different terms in different contexts. It is a common concern in NLP and language-modeling, but its crucial functionality remains a side-concern, or worse: is even considered a bug to be eliminated. While unarguably problematic in specific applications (e.g., certain translation tasks where we _need_ to disambiguate), the main argument of this chapter is that failing to observe this linguistic matter of fact as a generative effect rather than as an obstacle, leads to _actual_ obstacles in instances where language model outputs are presented as ‘neutral’ outputs. Given that a common and long-standing challenge in NLP is the interpretation of ambiguous—i.e., semantically noisy—cases, this chapter focuses on an exemplar ambiguity-resolution task in NLP: the problem of anaphora in _Winograd schemas_. The main question considered is: to what extent is the standard approach to disambiguation in NLP subject to a stagnant “image of language”? And, can a transdisciplinary, dynamic approach combining linguistics and philosophy elucidate new perspectives on these possible conceptual shortcomings?
In order to answer these questions we explore the term and concept of _noise_, particularly in its presentation as _semantic noise_. Owing to its definitional plurality, and sometimes even desirable unspecificity, the term _noise_ is thus used as proof of concept for semantic generativity being an inherent characteristic in linguistic representation, and its _concept_ is used to interrogate assumptions admitted in the resolution of Winograd schemas. Meaning is collectively (de)generated by indeterminate, combinatorial processes involving massive amounts of (semantic) noise. These exchanges cannot be thought of as voluntary and/or unpublic under _any_ circumstances. While the banality of this point screams loud enough, it cannot be stressed enough that we still _speak_ as if this was not the case, particularly with and through contemporary language models. The problems investigated here do not dive into a proper scholarly exposition of the philosophical background, but certainly pertain to language issues brought up by Derrida in his exploration of anti-accuracy and deferǎnce.^[Please mind this différance joke.] But also, of course Wittgenstein—as what we could jokingly call, in our context: “what language still can’t do.” What I could have done but did not, was also look at these problems through the lens of criticisms against Chomsky. However, my intention with this piece^[Which was published as intended for a wide audience] was to speak to the broader NLP community, and not to philosophers or linguists.
Again, to remark, the future of AI as we can intuit it today will primarily be guided by _prompts_, i.e., natural language _programming_, leading to an even more opaque relationship with back-end scripts than what we experience today. The way language is composed and delivered in the present will yield the combinatorial realities of tomorrow. The motivation that launched this project was the observation that even though the definition of a concept or specific word (such as _artificial intelligence_) may be highly contested and even fully absent, this does not deter publics and experts from employing it extensively. We are all stochastic parrots. When employing terms, we might make attempts at delineating the contours of one aspect or another that pertains to the topic, but the core of the term, its situatedness in the landscape of semantic possibilities, remains mostly unspoken for: whoever has legitimacy moves the pieces. This phenomenon is what this work eventually interprets as the strength of the conceptual, even though it is now entrenched within the technocapitalist framework. It is precisely in the fertile eventfulness of an unstable _semantic attractor_ that meaning resides; in the _promise_ of meaning (again, Derrida). In this opening chapter, the first ghost brought to the stage is therefore the assumed _normality_ proposed by the early foundations of NLP.^[It should be noted that Terry Winograd, especially through his work with Flores, transformed himself _a lot_ since the proposal of the schemas.] This first step allows us to witness the project of AI as a major metaphoric horizon against which a certain proposal of intelligence compares itself (based on earlier proposals: eternal/incomprehensible deities, specific images of enlightened _Man_, etc.).
 
>I want to see LLMs say “no” or “but” like children might.
>
>François Chollet, _MLST_ 2024 interview.
 
Moving forward, into a more structural analysis of what is happening as concepts “apply” to reality, in order to analyze how concepts function as semantic attractors, how the effectuate themselves by tracking differences, we move onto “[[04 Concepts as pre-dictions]].” This chapter presents an intervention which schematizes the concept of ‘negation’ by combining questions in philosophy of language(s) (what is ‘negation’?) with those of phenomenology and psychology (can we experience ‘nothing’?), and concentrating our efforts in the claim that concepts are actively-inferential: the main proposal advanced is that ‘negation’ can be understood as a _hyperprior_ (a “proto-concept”): a prior which tunes predictions at a rather fundamental level,^[See: [[Hyperprior]].] revealing itself at the level of the generative structures of natural language. Analyzing this as such reveals that framing these functions as predictive has implications for formal languages, for computation. Chunking and parsing therefore result in variable coordinations of perception and communication, which sometimes present as _paradox_. As explored in other chapters, too, in order to “experimentally” reveal this effect on paper (or screen), we show how this paradoxical effect is comparable to experiences in non-linguistic phenomena, such as sensory illusions. This chapter thus highlights the _perception-cognition-action_ continuum promoted by AIF, and is a first attempt to merge philosophy and physiology through AIF (the second attempt is [[10 Bias, or Falling into Place]], and the third [[07 Phenomenology of Sick Spirit]]).
Following closely from what came before, moving into the presumptuous claim that everything is prediction, we follow with: “What is the future? Prophecy, providence, preference, pattern, planning, prognosis, persistence, procreation: [[05 Prediction]]”. Here, we take an even deeper dive into the concept of prediction and analyze how it is that the technoconceptual structure supporting the possible future of mind—AI _and_ philosophy; both as abstraction, metaphysics and/or transcendence—can be understood as the dream of _absolute generality_, i.e., what we could also frame as total prediction. If one were to summarize the impetus driving contemporary machine learning (ML) and 20th/21st century science under one heading, it would be difficult to argue against this being “prediction.” Vernacular conceptualizations of prediction have a long history: as premonition, providence, prophecy, divination, fate, destiny, etc. In comparison, its scientific conceptualization is more recent. From early cryptologist al-Kindi (801–873), through Pascal and Fermat, to today: measuring/tracking phenomena in terms of their predictability indices (by way of statistical inference and probability) is now a prominent conceptual tool spanning all scientific-modeling. In recent decades, the concept has become even more prominent by the hand of AI technologies, many of them potentially useful for augmenting humanity’s capacity to risk-assess impending—sociopolitical, technoclimatological, etc.—(self-made) crises. Given the increasing implementation of ML-techniques across scientific-modeling, it will be argued here that a technical-philosophical conceptualization of “prediction” in the _broadest_ sense (Sellars 1962), in a “general” sense, can clarify some of the driving factors behind the construction of the future. This chapter tries to bring generality down on itself.
The dominant future, as of now, is linguistically parsed by interested perspectives, this prompts us to look into the concept of _interest_, and how this concept is sociolinguistically effectuated: learned and trained. We thus then proceed to the following chapter, which focuses on language-learning in interested humans: “The [[06 Principle of Sufficient Interest]].” Interest is a concept with many names across different areas (motivation, drive, and concern), each tied to distinct theoretical frameworks (such as relevance, psychology, and care), and all frameworks have different interests. This banal concept permeates all aspects of sociocultural life, appearing in everyday comparisons (“this is more interesting than that”), ethical considerations (conflicts of interest), and personal dilemmas (questions of self-interest). In this chapter, we connect both everyday and theoretical understandings of interest to demonstrate how it functions as an essential and inescapable element in language learning and, by extension, reasoning processes. Interest, which indicates a person’s specific focus or inherent subjective orientation, is frequently—though not always explicitly—positioned in opposition to reason, which is traditionally viewed as transcendentally detached or disinterested. Philosophy has extensively explored reason’s many possible interests, however, most modern theories of reason or rationality have tried do away with “interest”, in order to consolidate rationality’s procedural, inferential, and logical aspects into a reproducible model, seeking to establish foundations for a supposed scientific objectivity or in AI: _generality_, which instead elevate various forms of realism into predictive, self-legitimizing frameworks. Historically, reason’s most notable self-reflective attempt at this is the “principle of sufficient reason,” which gives this chapter its title.
This chapter, in obtuse and profuse twists of confusion, modestly aims to show that reason, despite its aspirations toward generalistic detachment, is unavoidably inherited from a process of “interested learning” as a procedural, iterable social activity. Inspired by the work on cultural learning by Tomasello (2016), this chapter highlights how language encodes interests, and how what we often forget to notice in transparent images of “reason” is that _learning itself_ is what is transferred across beings, as a **function**. This process, shaped by the vastly different paths of those who engage in it, cannot be said to converge toward achieving a common transcendental reasoning goal , or a human “sameness”. The challenges in developing genuine “artificial intelligence” demonstrate this impossibility, as challenges like those involving issues of “explainability,” “perspectival frame,” “common sense,” or even seemingly straightforward natural language understanding, present the same returning obstacles when pursuing universal solutions. We propose that by examining the concept of interest in learning processes (which combine biological mimicry, psychological imitation, historical-material contexts, relevance dynamics), a new understanding of reason’s processual nature can emerge. This approach informs a different concept of “intelligence” in AI development—one that emphasizes learning as interest-encoding. This argument strengthens existing theories about intelligence’s socially-distributed nature, its perpetually incomplete and unstable character, and its inevitable aesthetic dimensions. Various authors are discussed for context and examples, particularly Wilfrid Sellars’ ideas around “language games,” which connects learning, reasoning, AI and interest into a single argument (which we partially support but also deconstruct).
In order to provide depth into the conceptual assumptions made by this chapter, some of the entries in the [[Appendix]] may be elucidating, the reader is advised, if they wish, to see the entries [[Modulations]], and [[Neologistics]] which briefly treat aspects of how the polycomputation of pronouns, moving verbs and other linguistic phenomena can be modulated, and are themselves modulators of experience. [[Xpectator]] & [[Autosemeiosis]] explain how agency or agents are better termed xpectators, given they are predictive perspectives, and how the semeiotic communication between parts—whether this be between agents, neurons, ecosystems—_is_ inevitably **autosemeiosis**. [[Metalearning]] & [[Collective intentionality]] explain these two concepts in more depth, metalearning as the _seeing as_ capacities of distributed, interested minds, and _collective intentionality_ as the production of an affirmative (this not being without its possibly top-down fascist problems), constructive learning space.
In the following chapter, “[[07 Phenomenology of Sick Spirit]]”, we take a rather deep dive into the infrapersonal and megasubjective: a first person account of mind-body dualism in the trap(s) of a sick body. This chapter might seem beyond scope, but the consequences for the topic of AI and LLMs are vast if we consider the medical and commercial drive to develop systems which _interpret_ symptoms and _interact_ with patients. The paper “Sharing Our Concepts with Machines” by Patrick Butlin (2023) opens precisely with this challenge, by pointing out how “Babylon Health, a London-based private healthcare company, claims that their AI can ‘understand and recognise the unique way that humans express their symptoms’ (babylonhealth.com/ai; accessed 26 June 2020).” (p. 3079). The parameters which determine a state of (mental) ‘health’ are debatable, if not altogether objectively impossible to determine (Fava et al. 2019, Sterling 2020). In psychosocial allostasis, that is: our collective coping with stress, these parameters are especially difficult to delineate because they pertain to the sociocommunicative frameworks in which a so-called _person_ is embedded. As we will have seen through previous chapters in our retelling of AIF: predictive bidirectional dynamics between an agent and their ecological embedding result in the preservation of very interested states—_chunks_—which are “negotiated” as (collective) health and therefore survival. What we explore in this chapter, through a first-person account, is an insight missing in neuroendocrinology as it currently stands: health, of course, is relative, but how do concepts tune hormones, and how are they in turn tuned by hormones? Despite decades of research on the psychiatric symptoms associated with corticosteroids (Sirois 2003), showing how corticosteroid-excess and deficiency induces neuropsychological changes, even in short-duration therapy (Russell et al., 2023), an account of how they affect semantics remains elusive. In order to explore explanatory links between the phenomenology of semantic allostasis and neuropsychiatric approaches to psychopathology, this chapter exposes aspects of allostatic regulation through an account of corticosteroid self-administration. I will attempt to frame and narrate aspects of my own self-prediction when living with adrenal insufficiency,^[A condition that was given to me as a working diagnosis basically coinciding with the beginning of the PhD.] and this account will be theoretically informed by an allostatic account of active inference and neuroendocrinology. Philosophical introspection will ensue, in every sense of the expression, something which is much needed for updating our understanding of the (predictive) self and (allostatic) health (Cavel 2016, Gallagher 2024).
Symptoms are, of course, all there is. These explorations lead us to understanding the entire planet as _one sick being_, in the best sense of the word: a persevering, striving, surviving-cycling of entities. Switching the magnification back to the level of human coordination, (and based on an article written together with Inigo Wilkins and Jimena Clavel Vázquez): “[[08 Active ignorance]]” turns back to a social question of survival: how to organize ourselves once we know how we are inevitably at fault? This article points to some of the issues with the “homo economicus” presentation of the (voluntarist) human in the active inference framework. Active inference can be said to explain many of the effects that link perception, cognition, and action, by fundamentally framing these processes as in the service of uncertainty reduction, achieved through hierarchical processes of prediction. Cognitive systems apparently instantiate approximate Bayesian hierarchical architectures whose dynamics ensue from the interactions between the system’s generative model of the world, and a changing, relatively unpredictable world. In AIF, cognitive systems are therefore generally presented as optimizing systems: they optimize the trade-off between predicting with precision and acquiring novel information. Its roots partly lie in a Kantian philosophical framework, this is noticeable in the role played by prior knowledge in the organization of incoming information (Hohwy, 2013, p. 5). This imprint is also apparent in the individualistic perspective of the cognitive system that results from AIF as essentially isolated from its environment and from other agents. This methodological individualism not only has consequences for how we think about cognitive systems, but especially for the way we think about social and political structures as uncertainty-reducing systems. In this chapter, we—Clavel, de Jager, Wilkins—want to make active inference more explicitly enactive. We propose an approach to cognition that focuses on processes of _ignorance_ rather than _inference_, if only for the terminological point. Many have already emphasized how the body and the environment are involved in reducing uncertainty (e.g., Clark, 2016; Gallagher & Allen, 2018; Kirchhoff & Kiverstein, 2019), and how active inference is _enactive inference_ (Ramstead et al., 2019). The political question is _how_. This is implicit in the notion of active inference (Friston, 2010), but rarely expressed in terms of the organization of life at the sociopolitical scale.^[We also treat this in a collaborative paper on [[Power]] (with Mahault Albarracin and David Hyland). My contribution here is presented in [[11 Post-Control Script-Societies]].]
The argument here is that cognitive systems are not only engaged in (long-term, average) uncertainty reduction, but also in the complex and permanent restructuring of the _conditions_ of uncertainty reduction. These conditions are sedimented in the sociocultural environment in which cognitive systems are embedded. Drawing on post-Kantian critiques of the individualism of modern political philosophy (e.g., Mills 2022), we argue that this sociocultural environment is always unfortunately reflective of the ideological inclinations of the histories which have shaped it. We thus advance the concept of _active ignorance_. This is the process by which a system **neither confirms (accurately predicts) nor denies (model or hypothesis-adaptation), but simply _ignores_ evidence**, bending-constructing the world to meet its expectations. While, as mentioned, this idea is already implicit within the account of active inference, making it explicit adds a necessary critical dimension. It allows us to show that uncertainty reduction is unequally socially structured, and always socio-historically located relative to the (theoretical) framework and (practical) abilities of the system in question. We show that by simply focusing on single ‘agents’ and processes of ‘uncertainty reduction’, we often fail to reflect multiply-instantiated perspectives of sociocognitive systems. In the chapter, we illustrate this with the case of so-called maladaptive predictions.
The following chapter seeks to explore, in more sociopolitical dimensions, how it is that coordinated adaptation through (language) decisions happens. This chapter plays with the poetry of a Marxist philosophy, and is titled “[[09 C is for Communism, and Constraint]]”. This incursion assumes that its main idea is very, very simple and has been repeated throughout history _ad nauseam_ (in flows of the Dao, or more recently by Nietzsche via Schopenhauer), yet we still fail to see its real consequences: we are unfree. Why do we narrate ourselves as free? Chiming in with the oft-told historical chant of anti-freedom, this chapter is tuned to and by this: a *verbal*, and thus conceptual; *dialogical* and thus political move needs to be made from **choices to chances**: from isolated “decisions,” to distributed conditions and the layered implications of their perspectival [[Vantagepointillism]]. This is a move from the vestiges of domination and control (a presentation of Deleuze’s _reactive_ memory), to the variations of chaos and contingency (perhaps his _active_ forgetting). It is assumed throughout this chapter that this is a simple argument and in its implication lies true communism, where communism is understood as rather simple, too: as the cooperative construction of living conditions which seeks to rid itself of alienation, through inevitable alienations. Getting rid of the (alienating) concept of *choice* gets rid of the bullshit, free-agent, future-capturing, colonially-regurgitating, meritocratic condition we seem to inhabit under the various guises of desiring-legitimacy (fearful hoarding; capitalism). An encounter with *chance*, instead, actually awakens the distributed, modulatory power that is held by the tuning of a collective historicomaterialist process. Chances are based on what is given (not that anything strictly _is_ given, clearly), on conditions, while choices (and ensuing “decisions”) are surreptitiously presented as commandeering eagle-eye outlashings. It seems much of what we do rests on the question: _why did **x** do **y**_? But, given our dis- and re-orienting narratives throughout (deep) time: there is no **x**, only **y**. **x** had no choice, there were only chances that **y** would happen (and that this could be contemplated as such provided a nourishing substrate). The suggestion made is that (natural) language can be more attuned to chances.
This chapter might be helped by the following entries in the [[Appendix]], in order to ground itself more substantially. [[Agency]] deals with this concept, and its problems. The entry on [[Vantagepointillism]], explores Viveiros de Castro’s perspectivism, and how any and all of our meanings, as distributed, extended and co-constructed (via interindividuations through constraints) between perspectives and their environments, which are also perspectives, give rise to a _psychedelic pandemonium_. This pandemonium^[See [[NRU]] issue nr. 2 on contestation, Nietzsche, pandemonium and noise.] is explored in [[Music as permanent revolution]]. Here, we treat an understanding of musicality in speech and text as a fundamental driver of communication. If speech began as, and can continue to be, _song_, then all *generative* and *generous* reading-writing, chunking-parsing—in the amplest sense: all languaging—is modulating for the sake of making an opening for others to join in. That way: conversations, fiction, speculation, and even mathematical modelling, can be seen as musical proposals that invite others to join. Crucially: we pick up the rhythm where others left off. Rhythm and its modulation is resistance, is labor, is exhausting, is fascinating. In line with this, as Isabel Stengers muses: “Perhaps I became a philosopher because writing such footnotes [to Plato, following Whitehead] implies feeling the text as an animating power—inviting participation, beckoning to me and suggesting the writing of another footnote that will make a bridge to the past, that will give ideas from the past the power to affect the present.” (2012, p. 1).
If, as we have now attempted to show, _control_ and _agency_ are highly suspicious images of voluntarist, liberal legitimacy dreams, and we should modulate our verbosity away from them, then how to better understand our constraints, in order to do this? This leads us to “Falling into place”: here we explore our thoughts under gravity, and how our concepts reflect this. In this much-too-long and convoluted chapter we take Herbert Simon, some current AI themes, Hegel and AIF as our stepping stones towards the analysis of gravitational metaphors. We attempt to highlight several things, among them, how: a) intuition is an often-employed and philosophically-overlooked concept in the field of AI, whereas in philosophy and psychology it is a contested faculty which is more often than not posited against rationality. I try to show that it is a form of inevitable prediction as agents participate in their environments, and how this inevitability can be allegorized as a process of _falling_. b) Intuition is the _connection_ between two or more spatially-predictive models. Intuitive prediction is driven by the combination and recombination of spatial speculations. Even at the level of highly abstract concepts: these originate spatially and become contracted; summarized estimations, whose energetically-demanding spatial aspect we lose track of, as a result of free-energy minimization. c) Mind thus becomes space, gravitationally-bound, as mind _falls into its own predictions_. This as an aspect of the evolution of social intelligence; the actualization of perception-action as _beyond_ the singular agent, is excellently captured by Hegel’s concept of _Geist_, we therefore revisit Hegel in the context of free-energy minimization. Gravity is a basic affect: be where you are, reading this, and observe the pull. You dangle, people walk, falling step by step. Frictions, everywhere, ensue. If our concepts reflect our gravitational boundedness by way of the spatiotemporal metaphors they emerge from, then our thesis is that all concepts do is propose ever different (speculative) relationships to gravity. If, at the very least, _conservation of momentum_ is what matter does, then it is gravity which makes meaning happen, on Earth, by allowing things to fall into place. Because of this, different relationships to gravity (e.g., in outer space) might in the future give way to drastically novel types of linguistic coordination. Especially so if spatial reasoning is granted new perspectives by a body’s ability to float and move in a new dimensional axis. We only know what the body can do, under gravitational constraints. An entry that can be explored together with this chapter is [[Disorientation]].
The long entry on [[Assembly and assemblage]], compares two apparently drastically different approaches to **tagging reality with concepts**: Deleuze and Guattari’s _assemblages_, and Marshall, Sharma, Cronin, Walker, et al.’s _assembly theory_. As we have established in the previous chapters, we can say that life reveals and sustains itself through patterns by chunking and parsing these. Patterns (regularities which necessarily exist _before_ life, and which can be latched onto and/or created by it) are the very conditions of spatiotemporal experience. The search for metapatterns in these patterns, a common goal for philosophy and much of science, can be said to reveal a desire for a synthesis between biology (that which sustains meat-cognition and philosophy) and engineering (that which sustains silicon-cognition and technoscience). The metapattern of these research patterns, therefore, seems geared towards the construction of something capable of revealing-engendering novel life patterns. This is where it becomes interesting to compare two historically differing approaches to organisms and their interwoven levels of complexity: _assemblages_ and _assembly theory_. As will be explored, it is no surprise they share an almost identical name. An assemblage—_agencement_ in Deleuze and Guattari’s original context—is a philosophical proposition for the systemic acknowledgement of unavoidable multiplicity: dynamic, interlocked and interlocuted complexity. Assemblages are saying-doings. Assembly theory is a recent proposal for a new understanding of complexity in the organization of things that make themselves (i.e., life). Assembly theory puts the concept of assemblage to work in order to figure out the rhythms—chunks—and possible logic—parsing—of complex systems such as organic life. In this proposal we expose a comparison between assembly-theoretic foundations, and assemblages as a fact of life. Notions of difference and equivalence temper the entire sequence that will follow, and chase after the question: _which came first: difference or equivalence?_
In “[[11 Post-Control Script-Societies]]” we move towards the current conditions of language modulation through regimes of attention. “Algorithmic techniques are ways of making computers do things, of creating function, and their history is characterized to a greater extent by accumulation and sedimentation than by paradigm shifts or radical breaks.” (Rieder 2020, p. 16). If, for Deleuze, philosophy creates **concepts** and science creates **functions**, then perhaps control, in general, as a power over attention that pervades multiple disciplines-behaviors, creates _scripts_. As we noted in [[06 Principle of Sufficient Interest]], interest is guided by linguistic scripts, and scripts are guided by interest. What comes _after_ Deleuze’s modulatory control? How has power transformed from a condition of material and territorial domination, from enclosure and capture, to information and its modulation? These questions, partly answered by Deleuze already, are further explored by looking at how is power currently modulated, and by asking the question: can we analyze social power through formal methods in order to reveal novel modulations, new technics? In the social realm, the acquisition and maintenance of power depends on (changing) established habits: chunks and parsings. The _exertion_ of social power results in organizing mechanisms such as (new) chunks; as e.g., scripts, which come to determine habits of communication and behavior. Those in power are able to modulate the landscape of behavior and therefore material conditions by leveraging attentional scripts, all of which can be understood as *flows* of information, after Deleuze. Therefore, it is argued that power manifests in the capacity to influence information-processing, and therefore material conditions, through scripts that organize social attention and behavior. Observing the effects of scripts directing behavior as the basis of power and its material effects, makes power amenable to analysis as (effective) information-processing. Empowerment grants agents more information-processing opportunities, as they can rely on others to bring about desired outcomes, and therefore shape future trajectories. Again, starting from AIF, social power is conceptualized as a function of variables including attentional-attraction capabilities and information-processing efficiency, which together enable both script-creation and/or maintenance. This examination of how power relations materialize through attention and information processing, reveals connections between habitual structures, cognitive-behavioral patterns, and material realities. The final analysis offers a novel methodological approach for examining the habits of power by bridging social theory with cognitive science paradigms, the formal-modelling contributions are by Mahault Albarracin and David Hyland, and are excluded here, as this chapter bases itself off of a collaboration with them, which is forthcoming in print. The version presented here is an expanded account of what I contributed to the collaboration.
Switching our lens back to as “universal”, or if you will, as _blurry_ as possible: in our entry on The _[[Free Energy Principle]] and its Tautologies_, we explore what the FEP implies, as a tautology. Here we rely on the work of Chris Fields and Karl Friston. Very simply put, the FEP makes the claim that anything that persists _persists_. This necessitates that, although overlaps are unavoidable, a “thing” is something other than its environment because it must persist with regard to/in the face of it. The FEP is proposed as a grounding, organizing principle delineating the border between what we can and cannot call _a thing_. It has often been remarked that the tautology of what the FEP explains can be analogically compared to that of natural selection or survival of the fittest (Ruse 2008, pp. 44-45, Sims 2017, p. 8): why do some organisms survive? Because they are fit: i.e., because they survive. Keeping in line with the idea that _being is doing_—and maybe the distinction between those two verbs is not that distinct, after all—the FEP tries to get at the basic common denominator for what a thing needs to be in order for it to do what it does: be a self-containing coherence. Perhaps in awe of this tautology, and taking an additional a step back, what ask we would actually need to consider to have the concept of an _organizing principle_ in itself: what do we mean by this? Why do we seek to reduce complexity? It is even more tautologically interesting to consider that the desire—or perhaps _need_—to recognize, understand and/or employ an organizing principle is, tautologically, the defining characteristic of organized entities in general under the FEP. What is meant by this? Organic life, if we follow active inference as driven by the FEP, reduces predictive uncertainty by relying on induction strategies; summarizing patterns out of otherwise _indifferently_ ostensive, undecided series (i.e., creating chunks, indices out of opaque disorientation). By restricting this complex representation, _or_ recognition, of life to a simpler principle, what we observe, ultimately, is pattern-complicity. We explore whether it can be as simple as either the making of patterns or the exploiting of patterns which exist outside it: some lives search, others create, and most of them do both.
After all these incursions, we try to get at the mechanics of thought itself. The imagination, ‘detached’ thought, often presented as access *to* and retrieval *from* a space of possibilities which is _separate_, and experientially different from the inevitable state experienced through the senses implies, paradoxically, our unavoidable **lack of access** to what appears as given and contingent (Kant). The ability to know (metacognize) that one doesn’t ever _really_ know, is the cliché origin of all philosophy—whether presented as nothingness, infinity, doubt, negation, self-hatred, change, difference, the _other_, the past, a subconscious, or whichever other challenge. Not knowing is the basic requirement for knowledge-seeking, and the intentionality or volition that drives this process is undergirded by a multiplicity of factors, including the predictive drive behind abstract representation in scientific models, or, perhaps, hormonal homeostasis, all ultimately driven by gravitational balancing acts. The challenge lies in understanding the relations between these by observing them in ever-changing hierarchies, which depend on knowledge’s situated ambitions. After all these navel-gazing twists we are ready to conclude with an anticonclusion: _Negintelligiblity_, framed by the question: What is it that _drives_ intelligence? A possible answer: transitions between simplicity and complexity.
This chapter examines intelligence through active inference, as the dynamic interplay between model complexity and performed simplification (a.k.a.: predictive accuracy) in adaptive systems. This analysis is motivated by a desire to challenge contemporary AI discourse that dismisses “mere” pattern recognition as a subordinate cognitive function, as well as challenge proposals that simplification is the trademark of scientific explanation. We propose that intelligence can be understood as the systemic, organized capacity to adapt by _traversing_ pattern hierarchies: intelligence not only resolves complex patterns into simple ones (exploitation), but, crucially: also _explodes_ simplicity into complexity (exploration). We frame simplicity as predictive spatiotemporal compression (summarization, reduction, etc.), and complexity as (the designation of) pattern intractability and/or unpredictability. We define patterns as sensitivities to difference that constitute the fundamental substrate of perception. The intelligible domain is thus circumscribed by patternability—that which can be recursively, iteratively mapped—while that which ‘resists’ patterning remains outside the scope of the intelligible. To frame this effect, we introduce the concept of _negintelligibility_: a meta-pattern that can be speculatively understood as an attractor driving intelligent systems toward perpetual pattern-restructuring. Drawing an analogy with negentropy, we argue that negintelligibility operates _within_ negentropic systems, compelling intelligence to continually remodel established patterns against persistent epistemic barriers (these can be understood as presentations of novelty, noise, excesses, negativities, absences, catastrophic limits, symmetry-breakings, noumenal limitations, and other phenomena). This theoretical speculation suggests that intelligence fundamentally operates through bidirectional translations between simplicity and complexity, a metastable condition driven by what we term “negintelligible”: that which resists final pattern stabilization. Through various examples, expositions and comparisons to similar proposals (e.g.: irruption theory, entropic brain), we demonstrate how this conceptualization can offer novel insights into the possible radical naturalization of intelligence, learning, and the limits of knowability, through the lens of AIF.
Essentially, this conclusion says nothing new. It is a simple entropic challenge: we are bound by uncertainty. But of what measures (chunks) and for what purposes (parsings)? We might need an _immanental (perspectival) phenoumenodelics_ (Catren 2016, 2023) to slow things down. As proposed earlier, linguistic musicality allows us to play all this by different names. Negintelligibility tries to give another name to many things, for example to what Deleuze explains through the “dark precursor.”^[“We need something like a “differenciator” of difference, which relates the different to the different. This is the role played by what is called the _dark precursor._” (Deleuze 1994 (1968), p. 94).] Whatever we are able to apprehend, even the idea of the inapprehensible, if it is cognized, it exists in a series of cognitions that preceded it. But their virtuality is what is of interest because we know the future is larger than the past (entropy dictates unpredictability). This condition renders predictability as tuned by absolutely unidentifiable attractors. In the case of articulating a language of cognition (things being _made_ intelligible, through perception, science, language, and e.g., systems like LLMs), all arresting attempts are confronted with the very thing drawing what we call cognition forward: thinking pulls away from established ground toward something that it cannot capture. That which is made intelligible tends _against_ its own structures, abducting away from the already established. This is not an argument for the ineffable in thought, but its confrontation.
In other words, if the future didn’t seem incomplete, we wouldn’t be able to think.^[This is an effect of energetic dissipation: we remember the past because the future is larger. This asymmetry in time, its irreversibility, is what gives entropy its perceived direction and allows memories to form in one direction but not the other.] But this is, in so many ways, a very dangerous proposal. Paraphrasing Agre: conceptions of computation _now_ will determine the worlds of tomorrow (1997, p. 315), in ways we can barely grasp. In light of all of the above, the main problem this thesis sees itself and other—philosophical and more widely: all—theses as implicated within, is the question: _how to reconcile commonality with absolute difference, in the socioconstructive realm_?^[A.k.a., why can’t we be friends? Or, relatedly: how is science about consensus _and_ eternal destruction?] We do not answer this question.
<div class="page-break" style="page-break-before: always;"></div>
### Footnotes