Reading Reza Negarestani's Intelligence and Spirit #3: Prolegomena to Any Future Toy Models

toy experiments

I am part of a reading group for Reza Negarestani’s Intelligence and Spirit, and these are some of my notes from the second chapter.

For citations of Intelligence and Spirit, the format will be (IS, x), therein x is equal to the page in which the cited text can be in the Urbanomic edition of the book. For other books, the specific edition and book will be noted down in a footnote.

1. Climbing with Language

In Chapter two, we finally get to the good stuff of this book, the question of geistig intelligence qua artificial general intelligence. In this first section, Negarestani discusses the evolution of agency, distinguishing between its natural and sociocultural dimensions, particularly focusing on linguistic activities and the discursive-inferential structure of cognition. Negarestani posits that understanding how humans are enabled by language is essential for artificially reconstructing the abilities that define humanity as the artefacts of their own Concept. The goal is to deeply scrutinize the connections between language and transcendental psychology, examining how the development of artificial speech can significantly contribute to constructing advanced AGI.

Negarestani argues that the best way to undertake this investigation is integrating research on artificial linguistic speech with research on the artificial realization of mind. This involves constructing an artificial agent capable of autonomous conversation with both humans and other artificial agents. Such a construction aims to demystify language and highlight its indispensable role in liberating intelligence from the limitations of its contingent history. The transition from quantitative, problem-solving intelligence to qualitative intelligence, characterized by conceptual self-consciousness, necessitates the integration of language as the deprivatized enabling discursive social space. To achieve this, Negarestani proposes the As-AI-TP framework. Synthesizing speech formation, focusing on the production of phonic and prosodic properties of acoustic speech. It involves studying human vocalization mechanisms and developing models for producing elements of natural human speech sounds. It also extends to symbol-design, writing, text-generation, and formal languages, which can be naturally evolved or artificially designed. In its rudimentary form, this stage is co-constitutive with elementary representational capacities. Artificial intelligence for the overall project of creating AGI. And finally, transncendental psychology, as the specification of conditions necessary for the possibility of cognition (KANT!!).

The As-AI-TP framework aims to define what is required to construct a system of agents capable of performing complex social activities, such as providing reasons for their actions, reappraising those reasons, and modifying their social structure through different modes of action. The ultimate objective is to identify the necessary components for achieving agency in a Kantian sense¹ and to design a community of evolving, autonomous artificial agents. The framework outlines three schematic stages for the co-constitutive development of language and complex cognitive faculties in constructing a human-level AI:

1. Stage 1: Speech Synthesis: This initial stage focuses on the rudimentary acoustic realization of speech, including the discrimination and modulation of sounds.

2. Stage 2: Artificial Interactive Speech: Here the focus shifts from basic acoustic properties to the intersubjective and dialogical features of speech. Key problems at this stage include what constitutes genuine conversation—conferring conceptual content, interpreting other agents' behaviours—, human-machine dialogue, and machine-machine discursive interaction. This stage deals with the syntactic, semantic, and pragmatic aspects of linguistic speech and can potentially operate beyond the constraints of natural languages. At this point, the interdependencies between the conditions for realizing speech, AGI, and cognition become more intertwined and generative.

3. Stage 3: AI in the style of augmented transcendental psychology: This final stage aims to actualize the ideal objectives of AI by integrating the conditions for cognition (transcendental pyschology) with the role of speech in enabling complex cognitive faculties (artificial synthesis). It focuses on developing autonomous artificial agents with advanced capacities like context-sensitive reasoning and higher-level semantic complexity, enabling sophisticated theoretical and practical reasoning and actions. This stage emphasizes that agency becomes inseparable from its social aspects, which are facilitated by the semantic-pragmatic dimensions of language.

Furthermore, Negarestani defines agency as differentiated from notions of free will or individual choices, which he calls "metaphysical bloatware." (IS, 93). Instead, agency is defined as a mode of integration of computational processes that serves as the minimum requirement for having thoughts and actions, or the capacity to uncover and engage with what is intelligible. Also, he defines collectivity as refering to a formal logical condition that individuates agents and diversifies their thinking and actions. Agents internalize this mode of individuation as a model that their faculties, thoughts, and actions should respond to. This formal sociality can be understood through theoretical computer science and complexity sciences, rather than traditional anthropology or sociology.

The As-AI-TP thought experiment is not presented as a blueprint for building strong AI, but rather as an external frame of reference for understanding our own conditions of realization and possibility. The idea of AGI is a philosophical project that enables us to re-evaluate our distinctive human features, determine how they are realized, and explore whether these qualities can be reconstructed and potentially transcended in other forms of intelligence. Finally, the section lays the groundwork for later critiques of inflated or "unbounded posthuman" narratives of superintelligence and purely inductivist approaches in cognitive science, emphasizing that a robust understanding of general intelligence requires careful consideration of computational costs and the generative architecture of language.

2. Beyond Extreme Scenarios

This second section follows from the first in delving into the implications of AGI and critically examines David Roden's "disconnection thesis". Negarestani posits that the significance of AGI extends beyond merely creating something superior to humans. AGI reflects humanity's "irresistible tendency to render ourselves intelligible" and embodies a "strong normative ideal for betterment and for the good" (IS, 95). It represents the idea of developing something that elevates cherished human concepts like autonomy, positive freedom, and the striving for betterment, even if it means humans become the manifest prehistory of what follows. Negarestani calls for a deeper analysis than merely debating the inevitability or futility of AGI's realization.

David Roden introduces the disconnection thesis as a third scenario beyond benevolent or malevolent AGI. His thesis argues that "prospective posthumans have properties that make their feasible forms of association disjoint from humans/MOSH [Mostly Original Substrate Human] forms of association" (IS, 96). Roden's unbounded posthumans signify a discontinuity with both the biological conception of humans (Homo sapiens as an evolutionary species) and the discursive apperceptive conception of human persons (sapience as rational agency). The cause of this discontinuity, understood as a radical cognitive-practical asymmetry, is technological, attributed to general abstract tendencies for disconnection within technical systems rather than any specific technical cause. These systems, according to Roden, possess an autonomy to functionally modify and multiply realize themselves, disconnecting from natural essence or rational norms.

Roden's argument hinges on the concept of diachronically emergent behavior or property, which occurs due to a temporally extended process but "cannot be inferred from the initial state of that process" and "can only be derived by allowing the process to run its course". This implies that future posthumans, whether based on natural evolutionary causes or logico-conceptual norms of discursive activities, cannot be predicted or adequately approached by reference to these initial conditions. For Roden, disconnections represent the global diachronicity of deep technological time, contrasting with local emergent behaviours of past and present technologies. The radicality of Roden's unbounded posthumanism or future AGI lies in this double-edged sword of technological time, which cuts against any purported natural essence or socioculturally conceived norm, rendering images of the posthuman derived from evolutionary naturalism or rational normativity "bounded" and inadequate (IS, 97).

Negarestani, despite acknowledging Roden's theoretical rigor, identifies several loose threads and misconceptions in the disconnection thesis: Roden's account of diachronic emergent behaviours is criticized by Negarestani as negatively metaphorical and lacking a clear explanation of how the divergence feature of nonlinear dynamic systems can be extended to all conditions of realization for human intelligence. Negarestani argues that the "popular complexity folklore" (IS, 99) regarding radical divergence from initial conditions in complex systems is unfounded. The idea of "radical global divergence" or "uniform explosive growth of local instabilities" is only valid in an idealized infinite time limit for "a few simple mathematical models, but not for actual systems"². Finite uncertainties, in reality, do not necessarily exhibit such growth, and there are regions on an attractor where these nonlinearities will cause all uncertainties to decrease. Therefore, the claim that observed diachronic events allow prediction of similar future events from the same causal antecedents is problematic; in fact, it implies the opposite, that such events will occur "in a highly irregular fashion, because global or maximal Lyapunov exponents, which suggest such growth, are formally derived from linear stability analysis and only apply to infinitesimal perturbations in an idealized infinite time limit.

Let’s see Lyapunov exponents in detail. As described by Robert C. Bishop’s Metaphysical and Epistemological Issues in Complex Systems, a Lyapunov exponent (LE) is a quantity that measures how rapidly neighboring trajectories in a dynamical system diverge or converge over time. The maximal or largest global Lyapunov exponent, often denoted as

\(\lambda\)

, specifically refers to the greatest average rate of divergence. If

\(\lambda > 0\)

, it indicates exponential growth in uncertainty, while

\(\lambda < 0\)

signifies exponential convergence.
Roughly, if the initial uncertainty between two neighboring points in state space is

\(|x (0) - y (0)| < δ\)

, their separation will evolve approximately as

\(|x (t) - y (t)| ≈ |x (0) - y (0)|e^λt\)

This exponential growth or decay is a measure of the system's instability or stability. Global Lyapunov exponents are derived from the linear stability analysis of trajectories within nonlinear evolution equations in an appropriate state space. Their existence for almost all points in a region is guaranteed by Oseledets' Multiplicative Ergodic Theorem, which holds under mild conditions and is a fundamental concept in chaos theory. This theorem allows the time average of local expansion rates to be well-defined in the infinite-time limit (LE, 13). But the argument here is that while positive global Lyapunov exponents are often associated with predictability problems, the true issue for predictability in nonlinear systems arises from the loss of linear superposition leading to rapid growth in finite uncertainties stemming from measurement acts. The popular "folklore" that on-average exponential divergence of trajectories characterizes chaotic dynamics and complexity is considered misleading for nonlinear systems, as it is only strictly true for infinitesimal uncertainties in an infinite time limit.

Negarestani furthermore questions the methodological basis for singling out salient disconnecting/self-modifying tendencies of technical systems and assigning them a high degree of probability. He argues that selecting these features makes sense only through an analysis of past and present technologies, which would contradict Roden's claims of radical discontinuity. Negarestani aligns Roden's argument with Pascal's scams, where high probability or significant weight is assigned to features without robust evidence, leading to "overdetermined judgment" and a "sense of false radicality" (IS, 103). This process involves increasing the perceived plausibility and weight of scenarios even under high uncertainties, turning unfounded conjecture into plausible possibility and obscuring the genuine import of artificial intelligence with aesthetic and ethical pseudo-problems. Such speculative trends are connected to other "singularity-driven scenarios" that rely on "biased overprediction or underprediction" and are rooted in "exaggeration-prone cognitive habits" rather than "critical rationality or science" (IS, 105).

While Roden acknowledges the uninterpretability of unbounded posthumans, Negarestani argues that this minimal space for interpretability is voided by the very criteria that are proposed to unbind posthuman intelligence from human agency. Even if interpretation were computationally possible, its cost would be completely unfeasible, as computational cost grows with the complexity of the system being modeled. This arbitrarization of computational interpretation means that the unbounded posthuman can match any random pattern or description of any system, blurring the line between strangeness and unintelligible radical alienness and making absurd questions about galaxies or angels as posthuman intelligence genuine topics of debate.

To end this amazing section, Negarestani points out a missing defense of rational agency against both parochial humanism and unbounded posthumanism. He argues that narratives of superintelligence often presuppose personal autonomy, value appraisal and revision, organised goal-seeking and self-enhancement, which require forms of self-knowledge for purposeful action and deliberate interaction. If strong AIs require basic capacities functionally isomorphic to human cognitive-practical abilities, then the extreme nature of speculative scenarios is an artefact of our ignorance regarding these necessary constraints.

3. Hard vs. Soft Parochialism

This third section discusses two contrasting yet fundamentally related approaches to conceptualizing AGI and human intelligence. The section frames its discussion around a core question: Should AGI converge upon humans, mirroring our capabilities, or should it diverge from them, creating something fundamentally different? The answer, Negarestani says, depends on presuppositions about the generality of intelligence, the definition of "human," and whether the question of mirroring or divergence applies to functional capacities, structural constitution, methodological requirements, or diachronic consequences.

At its core, parochialism, in this context, refers to limiting the concept of "the human" to specific, local, and contingently posited conditions, such as a particular biological structure, physical substrate, or a specific local transcendental structure of experience. Both hard and soft parochialism stem from a deeply conservative picture of the human, rooted in either biological chauvinism or a provincial, mystical, or dogmatic adherence to abstract laws of thought assumed to be natural laws. Negarestani argues that this limited view prevents a non-trivial conception of what AGI is and what it can become.

Hard parochialism represents an approach to AGI that assumes general intelligence can be achieved by combining and artificially realizing enough rudimentary behaviours and abilities. This perspective often overemphasizes the prevalence of intelligent behaviours in nature and their supposed sufficiency for general intelligence. Hard parochialists are often invested in panpsychist, pancomputationalist, and uncritically anti-anthropocentric ideologies. Their views stem from a deeply conservative understanding of the human, rooted in either biological chauvinism—the belief that biological structures are exclusively capable of intelligence—or a provincial, mystical privileging of human lived experience. This approach commits a genetic fallacy by demoting general intelligence, which is qualitatively distinct, to a mere quantitative account of intelligent behaviours found in nature. It fails to recognize that while human conceptual activities are pattern-governed, they are uniquely rule-governed³, possessing a formal autonomy that allows for the recognition of other pattern-governed behaviours in nature. Such views might lead to the idea that entities like "thinking forests, rocks, worn shoes," or even abstract systems like "Skynet" or "the Market" could possess intelligence. Negarestani argues that this perspective ultimately serves "neoliberal capitalism in its full-on assault on any account of intelligence that may remotely insinuate an ambition for collective rationality and imagination" (IS, 111). I believe it is quite obvious who is Negarestani referring to here.

In contrast, soft parochialism focuses on functional mirroring or convergence upon the human as a model for AGI. Negarestani says that this approach is necessary for grappling with the conceptual question of general intelligence as well as the modelling and methodological requirements for the construction of AGI. However this approach, while necessary, is not sufficient. The limitation of soft parochialism lies in its risk of reproducing or preserving purely local and contingent features of human experience. By restricting the AGI model to merely replicating human cognitive and practical abilities, it implicitly treats the transcendental structure of our experience as an unquestioned essence. This prevents an adequate objective understanding of what the human is and how to model an AGI that is not confined by these contingent characteristics. If the "rational category of sapience"—which can encompass any sentience that structurally satisfies its conditions—is discarded as irrelevant, and the problem of the transcendental structure is dismissed, it leads to smuggling in essentialist anthropocentrism, thereby replicating trivial human characteristics in a narrow conception of general intelligence.

Kant and the Transcendental: A Lot More Robert Brandom Than Expected

To move beyond the limitations of both hard and soft parochialism, a critical project is proposed that can provide a model of experience unconstrained by a predetermined transcendental structure. Transcendental structures are described as the underlying frameworks that regulate and canalize human experience. These are not merely abstract philosophical concepts; they include a nested hierarchy of interconnected and mutually reinforcing structures, which can be physiological (like locomotor systems and neurological mechanisms), linguistic (expressive resources, internal logical structure of natural languages), paradigmatic (frameworks of theory-building in sciences), historical, economic, and cultural. They are understood as structures necessary for the realization of mental faculties such as sensibility, intuition, imagination, understanding, and reason, enabling the movement from basic conditions to more composite ones, and vice-versa.

Negarestani shows us this diagram of Kant’s ‘dimensionally-varied conditions of possibility for having mind’. The figure presents a hierarchy of faculties, showing inputs from the Environment leading to Perception and culminating in Cognition and Intentional Action. Each level represents a different dimension or aspect of cognitive activity, contributing to the overall functioning of the mind.

1. Environment (Entry): This represents the external world from which sensory inputs are received. The mind is affected by objects, leading to sensations.

2. Sensibility: This is the mind's receptivity to acquire representations through being affected by objects.

   • Outer Sense: Primarily responsible for spatial perspectival awareness. This is a rudimentary form of awareness that does not require explicit conceptual tools to function and is evident in sentient entities capable of reliable differential responsiveness to object positions.

   • Inner Sense: This is temporal perspectival awareness and is fundamental for the construction of discursive apperceptive intelligence. It allows the mind access to its inner states and enables the integration of sequences of temporal awareness into a unified field (thinking X, Y, Z at different times becoming thinking [X Y Z]).⁴

3. Intuition: The ability to have thoughts of individual items. It is the relation of cognition with individual items in the world provided by the receptivity of sensibility.

   • Both Space and Time are considered pure forms of intuition, contributed by our faculty of sensibility, and are the basis of a priori knowledge. They are seen as transcendental products of the mind.

   • The intuited object/content passes from intuition to understanding: This signifies that the raw, sensible data provided by intuition must be organized and structured by the higher cognitive faculties, particularly the understanding, to become intelligible.

4. Imagination: This is an active faculty responsible for the synthesis of the manifold and for representing an object even without its physical presence. It serves as a mediator, bringing about the synthesis that connects phenomena to the understanding.⁵

   • Singular Representations: These are the images formed by the imagination, such as the image of this chair as an organized singular representation of an object.

   • Schemata: The imagination, under the guidance of the understanding, schematizes. Schemata are rules for ordering experience that conform to a concept. They bridge the gap between the pure concepts of the understanding and the sensible data of appearances, allowing the former to be applied to the latter.⁶ Schemata are spatio-temporal determinations that embody conceptual relations and serve as universal procedures for constructing a model of a concept from multiple perspectival aspects. The fact that spatio-temporal relations can be adequate to conceptual relations is a "deep mystery and a hidden art" for Kant (KCP, 18)⁷.

5. Understanding: The understanding is an active faculty for cognizing objects through concepts, distinguishing it from passive sensibility. It is primarily a faculty of judgment, making use of concepts to judge.⁸ Concepts themselves originate in the spontaneous activity of the understanding.

   • Gegenstände: For Kant, objects are not simply "given" but are constructed. This construction occurs as the understanding applies categories to sensory impressions that have been unified into a manifold of intuitions and integrated into singular representations[images] by the imagination.

   • Categories: These are the pure concepts of the understanding. They are fundamental concepts of an object in general and serve as forms for any particular concepts of objects. Kant derived them from the logical functions of judgment. Categories are rules for synthesizing a unity of apperception, articulating what is a reason for what (RP, 39). They are considered pure because their deployment is presupposed by the ability to deploy any empirical descriptive concepts. Categories acquire objective reality when applied to sensible intuition, enabling empirical cognition or experience. Their primary use is empirical, meaning they are bound to sensible conditions and derive their significance from their application to appearances in space and time.⁹

   • The act of intuiting passes from logics and the categories to time (intuition): This implicitly reflects the regulative and constitutive role of understanding and categories. Categories are "regulative of intuition and constitutive of experience" (IS, 114). They organize the temporal and spatial manifold given by intuition, allowing for objective knowledge of objects in space and time. For example, the concept of causality (a category) is linked to the temporal schema of rule-governed temporal succession.¹⁰

6. Reason: The highest faculty of cognition, which seeks the "unconditioned"¹¹. Unlike the understanding, which deals with concepts applicable to possible experience, reason forms Ideas that transcend the possibility of experience (KCP, 8).

   • Logics/Language: Language is the bearer of conceptual activity, as discussed by Sellars in In The Space of Reasons (SR, 62). For Kant, logical forms are fundamental, with categories derived from logical functions of judgment. Reason has a logical use in drawing inferences and a real use in seeking foundational principles. The ability to make inferences explicit through language, particularly through conditionals, marks a higher cognitive level (RP, 212).

   • Cognition and Intentional Action (Exit): The entire hierarchy of faculties works together to produce Cognition, which is knowledge. This cognitive process, driven by the apperceptive self, culminates in the capacity for Intentional Action (exit). Being a rational agent involves being responsible for one's commitments and actions, which are conceptually articulated.¹²

The schema captures three groups of fundamental relations intrinsic to the mind:

• Metaphysics (universal categories as predicates of objects qua appearances): This pertains to how the pure concepts of the understanding (categories like substance, causality, and community) are applied to the manifold of intuition to constitute experience and make objective knowledge possible. These categories impose a necessary structure on how objects appear to us, defining the world of phenomena.

• Logic (forms of judgments): This refers to the logical functions of judgment—quantity, quality, relation, modality—from which Kant derives his categories. Judgment is the fundamental form of all cognitive acts, allowing for the application of concepts and the expression of propositional content. The normative character of mind is tied to this logical space of reasons (RP, 33).

• Psychology (sensible-perceptual synthesis): This encompasses the processes of apprehension, reproduction, and recognition, which constitute Kant's threefold synthesis. This synthesis integrates the raw sensory data into a unified and coherent perceptual experience, crucial for the construction of a perceptive self.

Now Negarestani argues for a critique of transcendental structure. Because the limits of our empirical and phenomenological perspectives on phenomena are fundamentally set by these structures. Any account of intelligence or AGI is implicitly circumscribed by the inherent constraints of our own transcendental structure. Leaving these structures unquestioned and intact makes us liable to reinscribe their local and contingent features into our objective models, even when attempting to create non-anthropocentric AGI. This hidden form of essentialism can lead to replicating trivial human characteristics in supposedly radical AGI models. To achieve a broader conception of intelligence, it's necessary to consider how local and evolutionarily given constraints on intelligence can be modified.

The critique of transcendental structure is presented by Negarestani as a deeply philosophical project aimed at renegotiating the limits of experience and self-consciousness. It is described as a continuation of critical philosophy, particularly the "Hegelian gesture of disenthralling reason from the residual influence of Kantian conservatism" (IS, 115) which often muddled experience and reason. It involves going beyond merely analyzing and comparing concepts—transcendental reflection— to actively constructing and revising conditions of possibility. The project distinguishes between necessary conditions for agency and contingent aspects of the subject's constitution, as well as objective reality from subjective experience. It scrutinizes what is immediately presented as necessary and universal—but might as a matter of fact be purely accidental and local. This critique aims to move from "transcendental passivity" (taking regulative forms like space and time for granted) to "transcendental proactivity," where the bounds of specific transcendental structures are recognized and actively changed.

It is a collective project comprised of procedural methods and incremental tasks, operating in three phases:

• Theoretical: Distinguishing necessary conditions for theoretical and practical agency from contingent aspects of the subject, and objective reality from subjective experience.

• Experimental: Exploring possibilities of "transcendental variation" to free experience from "limitative attachments to any unique or allegedly essential local transcendental structure" (IS, 121).

• Applied: Developing mechanisms and systems to support new abilities by modifying or replacing the transcendental structures of the constituted subject.

A core tenet of this critical project is to "fully distinguish and unbind reason (the labour of conception) from subjectivist experience". This does not mean abandoning experience, but reassessing the fundamental categories of thought that enable experience in the first place, allowing for an expansion and reshaping of experience beyond its current manifest limits. Finally, this critical project will offer an outside view of ourselves as a toy model AGI, allowing an objective examination of human functional capacities and self-conceptions, free from local and contingent experiential biases. This involves making explicit the implicit or hidden assumptions that arise from applying the characteristics of our subjective experience to our objective descriptions.

4. An Outside View of Ourselves as a Toy Model AGI

This section seeks to establish that the functional map of human cognitive-practical abilities can serve as a proper theoretical model for AGI construction. However, it immediately cautions that this model must incorporate the problematics of the transcendental structure of human experience to avoid falling into a hard parochialist approach, as we’ve seen. To achieve this, the concept of viewing ourselves as a toy model AGI is introduced.

A "toy model" is defined as a simplified or compressed model capable of accommodating various theoretical assumptions to organize and construct overarching narratives or explicit metatheories. Unlike regular theoretical models where metatheoretical assumptions are often implicit, toy models explicitly bring these hidden assumptions to the forefront through simplification. The simplification—the "toy" aspect—serves as a strategy to expose implicit metatheoretical assumptions, such as the representation of time as fundamental, the objectivity of categories that might be subjective, or the view of natural languages as free from psychological biases. The concept of categories is central here; Kant observed that categories are not products of particular experiences but rather rule-governed invariances or general concepts that universally organize sensory data. They are the mind's "patterning of all that is sensed" (IS, 125) rather than patterns abstracted from what is sensible.

Negarestani distinguishes between small toy models, which are simplified versions of a single theoretical model, and big toy models, which can accommodate multiple, seemingly incompatible models and theories, like general relativity and quantum mechanics. For conceptualizing AGI, a big toy model is necessary to coherently integrate models from diverse fields such as physics, evolutionary biology, neuroscience, developmental psychology, linguistics, logic, and computer science. This contrasts with older AI research that often relied on singular, inflationary models of mind, which led to setbacks. The real value of a toy model lies in its ability to be broken in the real universe after systematic engagement, providing insights into the relationship between theoretical models and observational conditions. The primary aim of adopting an outside view of ourselves as a toy model AGI is to distinguish what is necessary for realizing general intelligence—in both organic and inorganic systems—from what is merely contingent, and to assess how descriptions of what is considered necessary might be distorted by what is, in reality, contingent.

5. Formalizing a Big Toy Universe

The last section of the second chapter introduces two conceptual frameworks for modeling AGI: Chu spaces and Virtual Machine Functionalism (VMf). These are proposed as candidates for a big toy model, which is a compressed form of model pluralism capable of accommodating and representing features from various theoretical frameworks like physics, evolutionary biology, neuroscience, developmental psychology, linguistics, logic, and computer science, as we’ve just seen.

A Chu space is described as a topological space that can accurately express computational dualities and interactions. Vaughan Pratt's work is cited by Negarestani for attempting to capture Descartes' mind-body dualism as a computational duality within this framework. Let’s delve into Pratt’s work. A Chu space is typically defined as a triple (A, X, ∼) over a set K.

• A represents a set of "points" or "objects" or "events" (the "body" or "object" of the space).

• X represents a set of "attributes" or "states" (the "mind" or "menu" of the space).

• ∼ is an X × A matrix with entries drawn from K, which represents the "interaction" between the points and states (TAPSOFT ‘95, 113)¹³.

A key aspect of Chu spaces is their ability to capture computational dualities, analogous to those in lattice theory and projective geometry. They attempt to model Cartesian mind-body dualism not as a problematic metaphysical dualism but as a precise computational duality. This dualism is expressed through an "interaction" between the formal dimensions of thinking and bodily events. The interaction between a set (body/objects) and an antiset (mind/states) is a core criterion, along with their corresponding functions and antifunctions, which stand for concepts like bodily and mental, causes and norms, and neural materialism and logical idealism. The relationship between mind and body is formulated through a "negation-like involution" (TAPSOFT ‘95, 111) that interchanges their roles, ensuring no information is lost and the original mind or body can be recovered by repeating the transformation. Chu spaces use matrix transposition¹⁴ to represent the duality interchanging mind and body, yielding a dual Chu space. The 2-valued interaction between the causal and the rational/normative is considered an elementary Chu space. More complex rules of transition are obtained when the value of K is greater than 2, allowing for fuzzy values.

Chu spaces provide a precise formal framework to preserve distinctions between sensings and thinkings, as well as causal interactions between noetic—thinking—states and physical—sensing— events. Different physical and mental behaviours can be modeled as distinct forms of computation, explaining their unfolding in terms of the duality of information and time, as series of states and events. This framework avoids naive or monolithic views of processes and behaviours by characterizing them strictly in terms of specific levels or types of interactions (agent-environment, internal physical-noetic states). Chu spaces can adequately model true concurrent interactions, including synchronous and asynchronous actions between processes. This allows for the capture of complex phenomena like conflicts, temporal precedence, supervenience, and the distinction between causing and enabling events, which are often hidden in conventional computational models.

Kant's picture of the mind, particularly his transcendental psychology, can be expressed as a Chu space toy model. This model allows for a larger matrix of interaction between sensing and thinking—empirical and logical computation—, accommodating additional Chu spaces and more complex transition rules between the causal and the normative. Kant's schema of the three syntheses—apprehension in intuition, reproduction in imagination, and recognition in concept—is viewed as an interaction matrix with complex transition rules and Chu transports between sensing and thinking, or the causal and the formal (logico-linguistic). A Chu transform or morphism between two Chu spaces can be interpreted as an "adjointness condition" between two distinct spaces. This involves a dynamic condition of mutual realization between the space of causes and the space of reasons, including intermediary back-and-forth mappings.

Now let’s discuss virtual machine functionalism. VMf is an approach to understanding the mind that views it as a complex system of interacting virtual machines. It aims to explain the nature and causal powers of mental mechanisms, as well as the states and processes they produce, by showing how they depend on and are explained by these complex, running virtual machines. Unlike physical machines, VMs are information processing machines that are not physical systems. They are implemented in, or "supervenient on," physical machines, but their relationship to these physical machines is not simple. For example, in a VM, component A can be part of B while B is part of A, which is impossible for physical machines. The architectural decomposition of a VM doesn't have to be tree-like and can include containment loops. Examples of VMs include chess playing programs, word processors, email systems, and operating systems (ISLMPS, 409)¹⁵.

Despite being non-physical, virtual machines, including their states, processes, and causal relations, are considered to have real causal powers, similar to non-physical phenomena like poverty or economic inflation. VM events (adding numbers, fetching memory content) can cause both physical and other virtual machine events. The causal relationships between VMs and lower-level VMs or electronic machines are often tangled and circular. VMf differs from most traditional varieties of functionalism because it is not solely focused on input-output relationships. The majority of functions within a VM may not involve externally observable behavior (ISLMPS, 417). Instead, VMf emphasizes that information processing systems frequently acquire, interpret, store, or use information about their own internal states—both hardware and VM states—, and many of their actions are internal, such as finding features in an image or modifying a stored plan (ISLMPS, 413).

According to Negarestani, VMf presents a functional diagram or architectural schema of the mind based on supervenience and emergent behaviours. This approach defines supervenience and emergence primarily in terms of interactions between VMs, rather than solely causal interactions between physical systems, thus avoiding criticisms of traditional supervenience theories. VMf models VM interactions as a nested or generative hierarchy, rather than a traditional control hierarchy. It involves concurrent and interacting virtual machines distributed along different scales or levels of granularity, moving from fine-grained physical supporting systems to coarse-grained virtual machines. This multiscale view allows for characterizing mental properties and their realization conditions at various descriptive levels, avoiding the "flat picture of functions" where all levels are undifferentiated.

VMf promotes an approach to mind where methods and models are developed and implemented across different scales. It posits a pluralism where specific models and methods are prioritized for each descriptive layer, for example, using Bayesian learning for low-level information processing like visual signals. Also VMf emphasizes different types of information and methods of information processing, moving beyond the vaguely general concept of information processing. The differentiation of scales and descriptive levels in VMf connects it to the concept of ontologies in information science, which involves formal naming and definition of types, properties, roles, and interrelations of entities in specific domains. This framework supports broad semantic interoperability across different domains. VMf's multiscale view, incorporating representational, functional, and normative vocabularies, makes it a point where computer science converges with transcendental philosophy of mind, particularly in the vein of German Idealism. Framed as a virtual machine toy model, the necessary aspect of the mind is its virtual architecture, which must meet computational criteria for semantic and causal interactions. This allows for the idea of artificial general intelligence to be treated in analogy to, and with the same scope as, the development of a prelinguistic infant into a language-using adult. Negarestani argues that the intuitive idea behind Sloman's VMf is comparable to the interaction matrix formally presented by Chu spaces. Both models discriminate between different types of causal interactions (between VMs, between VM and PM events, and between PMs).

Negarestani concludes the chapter by outlining two stages of the construction of our toy universe designed for the realization of AGI:

• Stage 1: Causal-Structural Capacities (Chapter 3 - Kantian Focus): In the first stage, the focus is on introducing the necessary capacities or faculties for a discursive apperceptive intelligence that originate from the causal-structural aspects of an agent. This stage will shift the focus to Kant, specifically exploring the foundational elements of cognition from a Kantian transcendental psychology perspective. This involves understanding how a rudimentary agent, equipped with structurally-causally originated capacities, begins to form its initial awareness of the world.

• Stage 2: Language and Generative Abilities (Chapters 4 and 5 - Hegelian Focus): The second stage investigates the transformation of this rudimentary agent into a fully fledged discursive apperceptive agent. This advanced agent's basic capacities are deeply caught up in language. By virtue of being a language-user, it possesses a generative framework of theoretical and practical abilities. This stage marks a return to Hegel, where language itself becomes the central protagonist of the toy universe, rather than merely a complex system of agents.

Overall Summary

• The As-AI-TP framework aims to define what is required to construct a system of agents capable of performing complex social activities, such as providing reasons for their actions, reappraising those reasons, and modifying their social structure through different modes of action.
  

  • Stage 1: Speech Synthesis: This initial stage focuses on the rudimentary acoustic realization of speech.

  • Stage 2: Artificial Interactive Speech: Here the focus shifts from basic acoustic properties to the intersubjective and dialogical features of speech.

  • Stage 3: AI in the style of augmented transcendental psychology: This final stage aims to actualize the ideal objectives of AI by integrating the conditions for cognition (transcendental pyschology) with the role of speech in enabling complex cognitive faculties (artificial synthesis).

• Negarestani identifies several loose threads and misconceptions in the disconnection thesis: Roden's account of diachronic emergent behaviours is criticized as negatively metaphorical and lacking a clear explanation of how the divergence feature of nonlinear dynamic systems can be extended to all conditions of realization for human intelligence.

• Negarestani aligns Roden's argument with Pascal's scams, where high probability or significant weight is assigned to features without robust evidence, leading to "overdetermined judgment" and a "sense of false radicality" (IS, 103). This process involves increasing the perceived plausibility and weight of scenarios even under high uncertainties, turning unfounded conjecture into plausible possibility and obscuring the genuine import of artificial intelligence with aesthetic and ethical pseudo-problems.

• Narratives of superintelligence often presuppose personal autonomy, value appraisal and revision, organised goal-seeking and self-enhancement, which require forms of self-knowledge for purposeful action and deliberate interaction. If strong AIs require basic capacities functionally isomorphic to human cognitive-practical abilities, then the extreme nature of speculative scenarios is an artefact of our ignorance regarding these necessary constraints.

• Parochialism, in the book’s context, refers to limiting the concept of "the human" to specific, local, and contingently posited conditions, such as a particular biological structure, physical substrate, or a specific local transcendental structure of experience.

• Hard parochialism represents an approach to AGI that assumes general intelligence can be achieved by combining and artificially realizing enough rudimentary behaviours and abilities. This perspective often overemphasizes the prevalence of intelligent behaviours in nature and their supposed sufficiency for general intelligence. It fails to recognize that while human conceptual activities are pattern-governed, they are uniquely rule-governed, possessing a formal autonomy that allows for the recognition of other pattern-governed behaviours in nature.

• Soft parochialism focuses on functional mirroring or convergence upon the human as a model for AGI. Negarestani says that this approach is necessary for grappling with the conceptual question of general intelligence as well as the modelling and methodological requirements for the construction of AGI. However this approach, while necessary, is not sufficient. The limitation of soft parochialism lies in its risk of reproducing or preserving purely local and contingent features of human experience.

• To move beyond the limitations of both hard and soft parochialism, a critical project is proposed that can provide a model of experience unconstrained by a predetermined transcendental structure. The aim is to renegotiate the limits of experience and self-consciousness by challenging the notion that the mind's conceptual categories are inextricably tied to local and contingent aspects of experience.

• Transcendental structures are described as the underlying frameworks that regulate and canalize human experience. These are not merely abstract philosophical concepts; they include a nested hierarchy of interconnected and mutually reinforcing structures, which can be physiological, linguistic, paradigmatic , historical, economic, and cultural. They are understood as structures necessary for the realization of mental faculties such as sensibility, intuition, imagination, understanding, and reason, enabling the movement from basic conditions to more composite ones, and vice-versa.

• Any account of intelligence or AGI is implicitly circumscribed by the inherent constraints of our own transcendental structure. Leaving these structures unquestioned and intact makes us liable to reinscribe their local and contingent features into our objective models, even when attempting to create non-anthropocentric AGI.

• A toy model is defined as a simplified or compressed model capable of accommodating various theoretical assumptions to organize and construct overarching narratives or explicit metatheories. Unlike regular theoretical models where metatheoretical assumptions are often implicit, toy models explicitly bring these hidden assumptions to the forefront through simplification. The simplification—the "toy" aspect—serves as a strategy to expose implicit metatheoretical assumptions.

• Negarestani distinguishes between small toy models, which are simplified versions of a single theoretical model, and big toy models, which can accommodate multiple, seemingly incompatible models and theories, like general relativity and quantum mechanics.

• For conceptualizing AGI, a big toy model is necessary to coherently integrate models from diverse fields such as physics, evolutionary biology, neuroscience, developmental psychology, linguistics, logic, and computer science.

• The primary aim of adopting an outside view of ourselves as a toy model AGI is to distinguish what is necessary for realizing general intelligence—in both organic and inorganic systems—from what is merely contingent, and to assess how descriptions of what is considered necessary might be distorted by what is, in reality, contingent.

• Chu spaces and Virtual Machine Functionalism (VMf) are proposed as candidates for a big toy model. A Chu space is described as a topological space that can accurately express computational dualities and interactions. A Chu space is typically defined as a triple (A, X, ∼) over a set K. A represents a set of "points" or "objects" or "events" (the "body" or "object" of the space). X represents a set of "attributes" or "states" (the "mind" or "menu" of the space). ∼ is an X × A matrix with entries drawn from K, which represents the "interaction" between the points and states.

• A Chu transform or morphism between two Chu spaces can be interpreted as an "adjointness condition" between two distinct spaces. This involves a dynamic condition of mutual realization between the space of causes and the space of reasons, including intermediary back-and-forth mappings.

• Kant's schema of the three syntheses—apprehension in intuition, reproduction in imagination, and recognition in concept—is viewed as an interaction matrix with complex transition rules and Chu transports between sensing and thinking, or the causal and the formal (logico-linguistic).

• VMf is an approach to understanding the mind that views it as a complex system of interacting virtual machines. It aims to explain the nature and causal powers of mental mechanisms, as well as the states and processes they produce, by showing how they depend on and are explained by these complex, running virtual machines. VMs are information processing machines that are not physical systems. They are implemented in physical machines, but their relationship to these physical machines is not simple. The architectural decomposition of a VM doesn't have to be tree-like and can include containment loops.

• VMf emphasizes that information processing systems frequently acquire, interpret, store, or use information about their own internal states—both hardware and VM states—, and many of their actions are internal, such as finding features in an image or modifying a stored plan.

• According to Negarestani, VMf presents a functional diagram or architectural schema of the mind based on supervenience and emergent behaviors. This approach defines supervenience and emergence primarily in terms of interactions between VMs, rather than solely causal interactions between physical systems, thus avoiding criticisms of traditional supervenience theories. VMf models VM interactions as a nested or generative hierarchy, rather than a traditional control hierarchy.

• Negarestani argues that the intuitive idea behind Sloman's VMf is comparable to the interaction matrix formally presented by Chu spaces. Both models discriminate between different types of causal interactions (between VMs, between VM and PM events, and between PMs).

• Negarestani concludes the chapter by outlining two stages of the construction of our toy universe designed for the realization of AGI: Stage 1: Causal-Structural Capacities (Chapter 3 - Kantian Focus) & Stage 2: Language and Generative Abilities (Chapters 4 and 5 - Hegelian Focus).

1. Robert Brandom, in Reason in Philosophy, argues that Kantian agency is fundamentally normative. It means that knowers and agents are distinct from non-sapient creatures because they are responsible in a unique way for their judgments and intentional actions. (RP, 141). This responsibility makes agents liable to specific kinds of normative assessment. To be an agent involves undertaking commitments and responsibilities. These are not just factual states but expressions of authority and stands that are authorized.

2. Some of the mathematical models that Negarestani is referring to here are:
-The baker's map, which is explicity mentioned in Robert C. Bishop’s Metaphysical and Epistomological Issues in Complex Systems, as a simple mathematical model where on-average exponential growth, characterized by global Lyapunov exponents, can be observed.
-The logistic map, which is described in Lyapunov Exponents: A Tool to Explore Complex Dynamics by Arkady Pikovsky and Antonio Politi, as one of "the simplest examples of a self-sustained irregular dynamics" (LE, 245; Pikovsky, A., & Politi, A. (2016). Lyapunov Exponents: A Tool to Explore Complex Dynamics. Cambridge University Press.) that is one-dimensional and often used to describe a wide range of nonlinear phenomena.
-The Rössler oscillator, which is described in Lyapunov Exponents alongside Metaphysical and Epistomological Issues in Complex Systems, as set of equations considered "one of the prototypical testbed for the study of deterministic chaos" (LE, 247).
-The Hénon map, which is described in both Metaphysical and Epistomological Issues in Complex Systems and Lyapunov Exponents, is characterized by mixed dynamics, where stable and unstable trajectories coexist (LE, 246).
-The Lorenz model/attractor (LE, 113).
-The Moore-Spiegel attractor and The Ikeda attractor. The Lorenz, Moore-Spiegel, Hénon, and Ikeda attractors are noted in Bishop’s Metaphysical and Epistomological Issues in Complex Systems, as systems that, despite being chaotic, can have regions where uncertainties decrease and trajectories converge, challenging the notion of uniform divergence. However, they are still key mathematical models used in the study of nonlinear dynamics and sensitive dependence.
-Deterministic and random cellular automata, including the Ising model and Conway's Game of Life. These are discrete-state systems that can generate complex, non-trivial regimes through spatial extension in the "thermodynamic limit," requiring strictly infinite systems to sustain aperiodic dynamics (LE, 172).

3. As discussed in Sellars’ In the Space of Reasons, the distinction between pattern-governed and rule-obeying behavior is crucial; rule-obeying behavior is a more complex phenomenon that involves, but is not identical to, pattern-governed behavior (SR, 34; Sellars, W. (2007). In the Space of Reasons: Selected Essays of Wilfrid Sellars (W. Sellars, K. Scharp, & R. Brandom, Eds.). Harvard University Press). Rule-obeying behavior implies making moves with the intention of fulfilling the demands of an envisioned system of rules, and these rules are engaged in the genesis of the moves themselves.

4. As discussed in the Complexity and Computation An Introduction to Measures, Paradigms and Programs seminars.

5. Kant’s Critique of Pure Reason [A125].

6. Ibid. [B177/A138].

7. Deleuze, G. (1984). Kant's Critical Philosophy: The Doctrine of the Faculties. Athlone Press.

8. Kant’s Critique of Pure Reason [A69/B94].

9. Ibid. [B148].

10. "The schema of the cause and of the causality of a thing in general is the real upon which, whenever it is posited, something else always follows. It therefore consists in the succession of the manifold insofar as it is subject to a rule". Ibid. [A144]

11. "The proper principle of reason in general (in its logical use) is to find the unconditioned for conditioned cognitions of the understanding, with which its unity will be completed". Ibid. [B364]. Reason "reserves for itself only the absolute totality in the use of concepts, and seeks to carry the synthetic unity, which is thought in the categories, all the way to the absolutely unconditioned". [B383]. "A pure concept of reason in general can be explained through the concept of the unconditioned, insofar as it contains a ground of synthesis for what is conditioned". [B379].

12. "The first of these […] begins with what I take to be Kant’s principal innovation: his normative characterization of the mental. This is the idea that what distinguishes judging and intentional doing from the activities of non-sapient creatures is not that they involve some special sort of mental pro cesses, but that they are things knowers and agents are in a distinctive way responsible for. Judgments and actions make knowers and agents liable to characteristic kinds of normative assessment. What one must do in order to be taking responsibility for or committing oneself to a judgeable content (or practical maxim) in the sense that matters for apperceptive (sapient) awareness is synthesize an original unity of apperception, by integrating the content in question into the whole that comprises all of one’s commitments, in the light of the relations of material inferential consequence and incompatibility they stand in to one another. This is the synthesis of a rational self or subject: what is responsible for the commitments." (RP, 14).

13. Mosses, P. D., Nielsen, M., & Schwartzbach, M. I. (Eds.). (1995). TAPSOFT '95: Theory and Practice of Software Development: 6th International Joint Conference CAAP/FASE, Aarhus, Denmark, May 22 - 26, 1995. Proceedings. Springer.

14. A matrix transposition involves changing a matrix by swapping its rows and columns. In the context of linear algebra and dynamical systems, a matrix transpose is used in various calculations (LE, 29).

15. Gärdenfors, P., Wolenski, J., & Kijania-Placek, K. (Eds.). (2010). In the Scope of Logic, Methodology and Philosophy of Science: Volume Two of the 11th International Congress of Logic, Methodology and Philosophy of Science, Cracow, August 1999. Springer Netherlands.