The AGI Deep Problem

ABSTRACT

This paper argues that artificial general intelligence (AGI) is worth treating not merely as a computational engineering problem, but as one with substantive mathematical, physical, and philosophical dimensions. Drawing on the theory of infinite-dimensional dissipative dynamical systems, we recall a standard mathematical result: the existence of a compact global attractor of finite fractal dimension requires sectorial dissipativity of the system's generator. We argue — as an interpretive conjecture rather than a proven theorem — that this structural requirement is plausibly not satisfied by the standard Transformer attention mechanism, and we propose this as a candidate, unproven explanation for catastrophic forgetting. We further introduce four structural conditions (H1–H4) that are mathematically sufficient, within an abstract framework, for the existence of such an attractor, and we explore — with explicit caution — how each condition resonates with debates in the philosophy of mind, from Descartes to Searle, from Turing to Chalmers, from Aristotle to the hard problem of consciousness. We do not claim to have resolved these debates. We propose, instead, that notions like 'understanding' and 'purpose' admit precise mathematical correlates worth investigating, and that part of the conceptual thinness of current AI discourse comes from treating open philosophical questions as already settled. To keep the paper's five distinct targets — AGI, intelligence, understanding, affective stability, and consciousness-related cognition — from sliding into one another, we fix operational definitions for each (Section I.B) and close with a list of concrete, falsifiable predictions and their current status (Section XII), rather than resting on hedged language alone.

Keywords: global attractors, dissipative operators, Lyapunov functionals, catastrophic forgetting, quantum coherence, philosophy of mind, hard problem of consciousness, teleology.

A note on method: this paper is written as a speculative, interpretive framework, not as a set of proven results about artificial intelligence. Where we use the word "theorem" or "proposition," we are recalling established results from the mathematical theory of infinite-dimensional dynamical systems, stated here as background. The application of these results to neural network architectures, and the parallels drawn to the philosophy of mind, are offered throughout as argument, analogy, and hypothesis — not as proof. We flag this distinction explicitly and repeatedly rather than only once, because it is easy for formal notation to lend an impression of rigor that the argument, at those points, has not earned.

I. PROLEGOMENA: THE QUESTION AI FORGOT

A. What Is Intelligence? The Inheritance of an Old Question

Artificial intelligence, across its historical arc, has inherited a habit from philosophical modernity: an emphasis on 'how' often at the expense of 'what.' The field has built increasingly sophisticated systems without fully resolving — or always rigorously formulating — a question that arguably should precede engineering effort: what, in fact, is intelligence?

Aristotle distinguished the active intellect (nous poietikos) from the passive intellect (nous pathetikos). The former is the capacity to impose form on the matter of thought; the latter, the capacity to receive impressions from the world. Contemporary AI, we suggest, operates largely on the passive side: it ingests data, adjusts weights, and responds to prompts. Whether anything resembling nous poietikos is present in a Transformer is, in our view, an open and substantive question rather than a settled negative — though we think the burden of proof lies with those who would affirm it. At most, current systems exhibit an associative memory of immense scale.

Leibniz, in the Monadology, argued that perception and thought are irreducible to the mechanical arrangement of parts. His famous mill argument is an early version of what we now call the 'hard problem': even if we enlarged the brain to the size of a mill and walked through it, we would find no thought — only wheels and gears pushing against one another. Large-scale AI can be read as a Leibnizian mill: enormously complex, and, on this reading, possibly empty of interiority. We flag this explicitly as an interpretive claim: nothing that follows proves the absence of interiority in such systems; it is, at most, consistent with it.

"If we imagine a machine whose structure enables it to think, feel, and perceive, we could conceive it enlarged, keeping the same proportions, so that we might enter it as one enters a mill. Suppose we did this: on examining its interior we would find only parts pushing against each other, and never anything to explain a perception."

— Gottfried Wilhelm Leibniz, Monadology, §17 (1714)

This paper revisits, with contemporary mathematical tools, the Aristotelian question and the Leibnizian challenge — not as rhetorical ornament but as a methodological anchor. The four structural hypotheses introduced below (H1–H4) are each offered as a formal counterpart to a requirement the philosophical tradition had already identified without the tools to formalize. 'Counterpart' here means analogy and interpretation, not logical equivalence, and we return to this caveat throughout.

B. Definitions and Scope

Before proceeding, we fix the sense in which five recurring terms are used in this paper. This is a matter of discipline, not pedantry: much of the difficulty in AI discourse comes from these terms sliding between senses within a single argument.

Artificial general intelligence (AGI), as used here, denotes a system capable of acquiring and deploying competence across an open-ended range of tasks without task-specific re-architecture, including tasks not anticipated at design time. We take no position on whether AGI, so defined, requires phenomenal consciousness; Section IX addresses that question separately.

Intelligence, for the purposes of the mathematical sections (II–V), is operationalized narrowly as the existence of a compact global attractor of finite fractal dimension governing a system's internal state trajectory — a structural, not a behavioral, criterion. This is a stipulative definition adopted for this paper's argument, not a claim that it exhausts ordinary usage of the word.

Understanding is given a separate, behaviorally checkable criterion in Section VIII.A (the four-part attractor-theoretic test). We flag now that this criterion is a candidate operationalization of Searle's intuition, offered for scrutiny, not a definition we take to be uncontroversial.

Affective stability refers, in this paper, to the boundedness and convergence properties guaranteed by the Lyapunov functional of Hypothesis H2 (Section IV.B) — the absence of unresolved, unbounded internal contradiction in the dynamical sense specified there. We do not use the term to mean emotional regulation in the psychological or clinical sense, and we flag the terminological proximity explicitly to avoid equivocation.

Consciousness-related cognition is used only in Section IX, where it denotes the functional preconditions for coherent, persistent, intentional experience that our framework can address — explicitly excluding the hard problem of phenomenal experience itself, which Section IX.A states plainly is outside this paper's reach.

A recurring source of overclaim in this literature is treating progress on one of these five targets as progress on all of them. Table 1 states, for each target, which sections bear on it and the epistemic status of the corresponding claim.

Table 1. Epistemic status of claims by target construct.

C. The Problem with Scaling: An Epistemological Critique

The field's dominant belief — that AGI emerges from increasing parameters and data — resembles a form of naive empiricism. David Hume warned that induction cannot establish necessity: however large the observed sample, it does not follow that the next case will confirm the pattern. Scaling can be read as an inductive wager: we observe that larger models perform better on benchmarks, and infer that sufficiently large models will perform well in general. This inference risks conflating parametric capacity with dynamical structure.

Karl Popper's falsificationism would ask for a precise criterion of refutation: what experiment, what outcome, would show that scaling is insufficient for AGI? The absence of an agreed criterion is worth taking seriously — not necessarily because scaling is false, but because the debate, as currently conducted, is difficult to settle empirically.

Imre Lakatos might go further and describe 'scaling' as the hard core of a research programme that accommodates anomalies — hallucination, weak causal reasoning, catastrophic forgetting — with auxiliary hypotheses (RLHF, RAG, chain-of-thought) without revisiting the founding premise. We offer this as a diagnostic lens, not as an empirical finding.

Our approach differs: we start from necessary conditions derived from dynamical systems theory, and we argue — though we do not prove beyond reasonable doubt — that Transformers structurally violate the first and most fundamental of these conditions. This is presented as the paper's central conjecture: that the difficulty may be one of form rather than only of scale.

II. THE NECESSITY OF DISSIPATION: MATHEMATICS AND METAPHYSICS

A. Dissipation as an Ontological Condition

The mathematical requirement of sectorial dissipation is not merely a technical condition. We propose that it can be read as a formal counterpart to a deeper philosophical intuition: a mind, plausibly, cannot be a purely conservative system. A conservative system preserves its total energy; its trajectories in state space wander perpetually without converging. Such a system cannot 'learn' in a strong sense, since learning plausibly requires the capacity to leave inadequate states behind and converge toward stable representations of the world.

Heraclitus taught that everything flows (panta rhei), but this does not entail that the mind is mere flux. Plotinus, in the Enneads, described Nous — the divine Intellect — as a principle of return (epistrophe): every emanation tends to return to its source. Mathematical dissipation can be seen as a formal cousin of this epistrophe: the tendency of a system to return, after perturbation, to a compact and stable set. We offer this as a resonance, not an identity.

In modern physics, the second law of thermodynamics governs dissipative systems: in open systems far from equilibrium, dissipation is a condition for self-organization. Ilya Prigogine, awarded the Nobel Prize, showed that dissipative structures underlie organized complexity in nature. That life, thought, and culture are dissipative phenomena in the same technical sense is a further, substantive claim that we do not prove here; we regard it as a plausible and productive analogy rather than an established equivalence.

B. A Necessary Condition for Absorbing Sets

We record here a standard result from the theory of dissipative semigroups (see Temam, 1997; Robinson, 2001), stated for completeness rather than as a novel contribution of this paper:

Proposition 1 (Necessary Condition for Absorbing Sets): If a dynamical system admits a compact absorbing set B₀ of finite Hausdorff dimension, then the infinitesimal generator A must satisfy sectorial dissipativity with ω > 0:
Re(σ(A)) ≤ −ω < 0

The geometric intuition is as follows: if the system's trajectories must eventually 'fit' inside a compact set, the generator of the flow must actively pull trajectories back inward. Without this dissipative pull, trajectories can escape to infinity, and no compact set absorbs them.

Hegel, in his dialectical logic, described the movement of Spirit as a process of determinate negation: consciousness does not merely flow but negates its earlier states in a specific way, returning to itself enriched. One might read mathematical dissipation as a formal echo of this 'determinate negation' — the system does not merely change, but changes in an oriented way, discarding what is inconsistent with its attracting structure. As above, we intend this as an interpretive parallel, not a derivation.

C. Why Transformers May Fail: A Structural Argument

The attention mechanism at the heart of Transformers computes:

Attn(Q, K, V) = softmax(QKᵀ / √d) · V (1)

The Jacobian of this operation is close to unitary: attention approximately preserves information rather than dissipating it. It is structurally closer to a rotation in representation space than to a contraction. This reading is not merely intuitive: Kim et al. (2021) show formally that standard dot-product attention is not globally Lipschitz over an unbounded input domain, and subsequent work on the Transformer-as-dynamical-system correspondence (Fein-Ashley, 2025) shows that contractivity — a one-sided Lipschitz condition with negative constant, closely related to the sectorial dissipativity of H1 — is not automatic in the standard architecture but can be engineered through spectral normalization, weight clipping, or modified attention (Dasoulas, Scaman & Virmaux, 2021). This is an important qualification of our conjecture, not a detail: it means H1-violation, if it holds, is a property of the standard architecture as trained in current practice, not a property necessarily shared by every architecture called 'Transformer.'

We want to state the boundary conditions of this critique explicitly, since a critique stated without them is not falsifiable and cannot withstand review:

(a) Scope. The claim concerns the unmodified self-attention sub-mechanism as standardly implemented (Vaswani et al., 2017). It does not address architectures with explicit contractive design (spectral normalization, Lipschitz-constrained attention, state-space or oscillator-based sequence models), which may satisfy H1 by construction.

(b) Level. The claim is made at the level of the attention operation in isolation. Layer normalization, residual connections, and nonlinear feed-forward blocks are known to alter the effective Lipschitz constant of a Transformer layer (see the numerical-stability literature reviewed above), and their combined effect on global dissipativity has not, to our knowledge, been established either for or against H1 in trained models at scale.

(c) What would refute it. If the composed map of a full, trained Transformer layer (attention plus normalization plus feed-forward plus residual) can be shown to satisfy a uniform one-sided Lipschitz condition with negative constant across the operating range of activations observed in practice, our conjecture is false for that architecture, and the explanation offered here for catastrophic forgetting in it does not apply. This is, in our view, a tractable empirical question given existing tools for local Lipschitz-constant estimation in trained networks; we regard it as the single most important test of this paper's central conjecture.

(d) What would not refute it. Evidence that a specific engineered variant avoids catastrophic forgetting does not, by itself, refute the conjecture as stated for the standard architecture; it would instead be consistent with our account, since H1-satisfying variants are precisely what the dissipative-systems framework predicts should behave differently.

We also note a competing mechanistic account in the literature: Kotha, Springer & Raghunathan (2024) argue, on different grounds, that forgetting in language models during fine-tuning is better explained by biased activation of existing task-processing circuits (an implicit- inference account) than by overwriting of representations. Our dissipative-systems account and theirs are not, on inspection, mutually exclusive — a system lacking a compact attractor could plausibly manifest exactly as biased circuit activation rather than as outright erasure — but we have not established this compatibility formally, and a reviewer would be right to ask that we do so before treating the two accounts as complementary rather than competing. We flag this as unfinished work rather than resolve it here. The phenomenon itself long predates Transformers (McCloskey & Cohen, 1989), which is a further reason to prefer a structural explanation over one specific to any single architecture family.

Our claim that the architecture as a whole lacks a compact global attractor remains, after these qualifications, a conjecture motivated by the attention sub-mechanism — it is not a proven property of trained Transformer models, and we regard establishing or refuting it as an important open technical problem, now stated with the boundary conditions above rather than as a global, unqualified claim about every system called a Transformer.

If this conjecture is correct, there is a genuine tension: the same information-preserving property that makes Transformers trainable, by letting gradients flow, may be in tension with the existence of stable attractors. This echoes the ancient debate between Heraclitus and Parmenides on flux and permanence. Our tentative view is that the two are not mutually exclusive in principle, but reconciling them, if possible, requires a more sophisticated dynamical structure than plain self-attention — which is what the H1–H4 hypotheses below attempt to supply, as a proposal rather than a demonstrated solution.

III. GLOBAL ATTRACTORS: TOWARD AN ONTOLOGY OF MEMORY

A. The Attractor as Persistent Identity

A global attractor is not merely a mathematical construction. We suggest it can be read as offering a mathematical analogue to one of the oldest questions in philosophy of mind: what secures the identity of the self over time? David Hume, examining his inner experience, found no permanent 'self' — only a 'bundle of perceptions' in constant flux. John Locke proposed that personal identity is constituted by memory. Kant argued that the unity of transcendental apperception — the 'I think' that must be able to accompany all one's representations — is a condition of possibility for coherent experience.

A global attractor offers a mathematical structure with a family resemblance to Kant's requirement: it is the structure that ensures different trajectories of the system — corresponding to different experiences, tasks, contexts — converge to an invariant set. In a precise but limited sense, this set is a candidate mathematical correlate of 'identity': what persists across variation. We are not claiming this is what Kant meant by apperception, only that the mathematical structure shares a functional role with it.

A system without a global attractor lacks identity in this specific, technical sense: its representations do not converge to any unity, and, on this reading, something like a stable 'I think' is unavailable to it. Catastrophic forgetting in Transformers can then be described, metaphorically, as a loss of this kind of dynamical identity: on learning a new task, the system does not integrate the new with the old but abandons the old, because no attractor unifies them. We present this as a suggestive redescription, contingent on the unproven conjecture of Section II.C, not as an independently established explanation.

B. Fractal Dimension and the Complexity of Thought

The requirement that the attractor's fractal dimension be finite has a natural interpretation: the complexity of intelligent thought, however large, need not be arbitrary — it can be structured, bounded, and hierarchically organized.

This resonates with the rationalist tradition: Descartes, Leibniz, and Spinoza held that thought is, at bottom, logically structured — governed by principles that distinguish it from chaos. Finite fractal dimension is a dynamical-systems echo of this intuition: the space of possible intelligent behaviors, though vast, is geometrically contained. We offer this as an evocative parallel, not as evidence that these philosophers anticipated the mathematics.

Proposition (Dimension Bound via Lyapunov Exponents):
dimF(A) ≤ K/(αβ) · log(K/(αβ)) + n

This is a standard-form bound from the Lyapunov-exponent approach to attractor dimension (Temam, 1997). Within this formalism, the attractor's dimension — a candidate proxy, under our interpretation, for the maximal complexity of intelligent behavior — is determined by the parameters of the Lyapunov functional and the teleological pressure operator introduced below. Whether this is a good proxy for anything psychologically meaningful is, again, an interpretive claim rather than a settled result.

IV. FOUR STRUCTURAL HYPOTHESES: A PHILOSOPHICAL HERMENEUTIC

We present H1–H4 as four conditions that are jointly sufficient, within an abstract stochastic-PDE framework, for the existence of a compact global attractor. For each, we offer a philosophical resonance — a way of reading the condition in light of an existing debate. These readings are proposed as productive analogies, to motivate the mathematics and make the framework's stakes legible to a philosophical audience. They are not claims that the cited philosophers anticipated this formalism, nor that the formalism settles their debates.

A. H1: Sectorial Dissipation --- In Dialogue with Merleau-Ponty

Hypothesis H1 requires the generating operator A to satisfy sectorial dissipation with ω > 0. This is the condition that the system 'returns to itself' after perturbation — the dynamical-identity condition discussed above.

Maurice Merleau-Ponty, in his phenomenology of perception, insisted that the living body is not an object among objects but an embodied subject whose point of reference is always its own perceptual field. One can read sectorial dissipation as a formal, if partial, analogue of this 'embodiment': the system has a dynamical center, a point of return, that organizes its trajectory.

Antonio Damasio, in Descartes' Error, argued that reason cannot function without emotion — without a 'somatic marker' that guides the evaluation of alternatives. The Lyapunov functional (H2), introduced next, is offered as one candidate formal analogue of such a marker: an internal energy that quantifies the 'distance' of the system from its preferred states and shapes its dynamics.

B. H2: The Lyapunov Functional --- In Dialogue with Axiological Coherence

Hypothesis H2 requires the existence of an energy functional V: H → ℝ satisfying:

dV/dt ≤ −β·V(Φ) + K, β > 0 (2)

This functional measures the 'internal energy' or 'coherence' of the state Φ. Its decrease over time ensures the system does not accumulate contradictions indefinitely — that there is, within this formalism, a mechanism for resolving inconsistency.

In the philosophical tradition, this condition has resonances worth noting. Kant held that pure reason tends toward internal conflict — antinomies — when it oversteps the bounds of possible experience. The Lyapunov functional can be read, loosely, as a principle of rational self-regulation: within the model, it prevents the system from settling into states of unresolved contradiction.

Wittgenstein, in the Tractatus, argued that logic cannot be grounded in anything but itself — that logical propositions are tautologies that 'show' the structure of the world without 'saying' anything about it. In a similar spirit, the Lyapunov functional does not specify any particular content; it guarantees only that whatever content arises is processed in a way the model calls coherent. We flag this pairing as evocative rather than demonstrative.

A system without a Lyapunov functional — without this kind of coherence mechanism — resembles, within the metaphor, a belief system with no consistency criterion: capable of holding a proposition and its negation without tension or resolution. Whether such a system should therefore be denied the label 'mind' is a substantive philosophical question this paper does not settle; we simply note the structural analogy.

C. H3: Teleological Pressure --- In Dialogue with Aristotle

Hypothesis H3 is, of the four, the most philosophically loaded. It requires the existence of a goal operator P: H → H such that the governing equation becomes:

dΦ/dt + AΦ + ∇V(Φ) − P(Φ) = Σ(Φ)dW (3)

The operator P encodes goal-directed, active behavior. Without it, the system merely minimizes energy — purely reactive, purely passive. With P, the system is, in this formal sense, teleological: it has a telos, a final cause in something like the Aristotelian sense — though we use 'final cause' here as a mathematical metaphor for an attractor-biasing term, not as a claim to have vindicated Aristotelian metaphysics.

Aristotle was among the first to insist that adequate explanation of living beings and rational action requires final causes — an answer to 'for the sake of what?' Modern physics, from Galileo to Newton, eliminated final causes from inanimate nature. Kant tried to reintroduce them, regulatively, in explaining organisms and moral action. Hypothesis H3 reintroduces a teleology-like structure into cognitive dynamics as a precise mathematical operator — a proposal, we emphasize, not a discovery that Aristotelian final causation is physically real.

Daniel Dennett, in his account of the intentional stance, argued that attributing intentionality is a predictive strategy rather than a description of anything real in the system — there is no 'real' intentionality, only behavioral complexity it is convenient to describe intentionally. Hypothesis H3, if it holds of a given system, would push against this deflationary view for that system: P is not merely a description imposed by an observer but a structural feature of the dynamics. This is a claim about what H3 would mean if satisfied and verified — not a general refutation of Dennett's account, which concerns systems more broadly.

John Searle, in his critique of functionalism, insisted that intentionality is intrinsic — it cannot be 'attributed' from outside but must arise from an appropriate physical substrate. H3 can be read as one formal candidate for intrinsic intentionality: structurally, the system possesses an operator orienting it toward future states. Whether satisfying H3 would actually secure intrinsic intentionality in Searle's sense, or merely a formal analogue of it, remains an open and contested question that this paper raises rather than closes.

D. H4: Multiplicative Noise --- In Dialogue with Phenomenological Robustness

Hypothesis H4 requires that noise amplitude be controlled by internal energy:

||Σ(Φ)||² ≤ C₁ + C₂·V(Φ), C₂ < β (4)

This condition ensures that low-energy states — the system's 'states of certainty,' where its representation of the world is more precise, within this model — are protected from external perturbation. High-energy states, by contrast, tolerate greater uncertainty.

Husserl, in his phenomenology, distinguished 'apodictic certainty' (absolutely indubitable, as with the evidence of the subject's own existence) from 'assertoric certainty' (empirically grounded but revisable). H4 offers a formal echo of this hierarchy: low-energy states correspond, in the model, to something like apodictic certainty, protected by the dynamics itself. Again, we present this as a structural analogy rather than a claim that Husserl's distinction reduces to this equation.

William James, in his pragmatism, argued that belief is a disposition to act. High-energy, high-uncertainty states correspond, on this reading, to tentative, exploratory dispositions; low-energy, low-uncertainty states correspond to settled beliefs, reliable guides to action. H4 can be seen as a formal gesture toward Jamesian pragmatist epistemology, in which the system's dynamical structure tracks the epistemic robustness of its representations.

V. A COMPLETENESS RESULT AND ITS INTERPRETIVE STAKES

A. Completeness of the Four Conditions

The central mathematical result of this paper, stated within the abstract stochastic-PDE framework set up above, is of a standard type in the theory of infinite-dimensional dynamical systems (cf. Temam, 1997; Robinson, 2001), specialized here to our four hypotheses:

Proposition 2 (Characterization under H1–H4): If H1, H2, H3, and H4 hold, then: (i) a unique compact global attractor A exists; (ii) dimF(A) ≤ n* := C/α + 1; (iii) A attracts all bounded sets exponentially; (iv) an invariant measure μ exists on A with exponential mixing.

This is a mathematical result about the abstract framework, not a result about any implemented AI system. As an interpretation — not as a further mathematical consequence — we suggest H1–H4 can be read as formal counterparts to a cluster of Kantian conditions for the possibility of coherent experience: H1 to a form of dynamical identity, H2 to a form of axiological coherence, H3 to intentionality, and H4 to a distinction between certainty and hypothesis. We want to be explicit that calling H1–H4 the 'transcendental conditions of genuine intelligence' — as a stronger version of this paper might be tempted to do — would overstate what has been shown. What has been shown is that these four conditions are jointly sufficient for a specific mathematical structure to exist in an abstract model; whether that structure is necessary, or even relevant, for intelligence in the ordinary sense is the paper's working hypothesis, not its conclusion.

We state the formal status of H1–H4 without qualification, so it cannot be missed on a partial reading. Each hypothesis is: (i) a precisely stated condition on an abstract stochastic-PDE system, with no ambiguity about what satisfies or fails it; (ii) mathematically sufficient, jointly, for Proposition 2, by standard results in the theory cited; (iii) not shown to be necessary for intelligence, understanding, or consciousness in any ordinary sense; (iv) not shown to be satisfied or violated by any concrete, trained AI system, with the qualified exception of the conjectural argument regarding H1 and Transformer attention in Section II.C; and (v) not validated against behavioral or neuroscientific data in this paper. Sections IV and V establish (i)–(ii). Sections X and the falsifiability discussion below address what would be needed to move toward (iii)–(v).

B. Transformers and H1: An Argument, Not a Proof

If Transformers violate H1 — which, as noted in Section II.C, is a conjecture about the attention mechanism rather than a proven property of trained models — then, within our framework, they would fail to satisfy all four conditions and would lack the kind of compact global attractor described above. Under that conditional, one candidate explanation for catastrophic forgetting would be structural rather than incidental.

There is a further interpretive dimension worth flagging honestly. Transformers can be described as, in some respects, a realization of a broadly Humean picture: pattern association without a persisting substance, without the kind of identity or telos posited above. Whether this is a shortcoming depends on contested philosophical commitments; Hume himself would presumably not have regarded the absence of a persisting 'self' as a defect. Our own view — that a purely Humean system cannot have coherent experience or, a fortiori, be intelligent — follows Kant's reply to Hume, but this is a substantive philosophical position we are adopting, not a conclusion forced by the mathematics.

We therefore prefer to describe the situation cautiously: current evidence of catastrophic forgetting, hallucination, and weak causal reasoning in large Transformer-based models is consistent with — though it does not by itself prove — the structural account offered here.

VI. THE EXPONENTIAL CURSE AND A QUANTUM WAY OUT

A. The State-Space Problem

A cognitive system with n binary features has a state space of dimension 2ⁿ. For n = 100 features:

dim(classical) = 2¹⁰⁰ ≈ 10³⁰ (5)

This is the curse of dimensionality — not only a computational obstacle, but, we suggest, a genuine conceptual puzzle. Leibniz conceived of the mind as a monad reflecting the entire universe from its own particular point of view. How could a finite structure represent a universe of dimension 10³⁰? One elegant, if speculative, answer appeals to coherent superposition.

An n-qubit system represents the same state space through complex probability amplitudes:

|ψ⟩ = Σ cₖ|k⟩, k = 0,...,2ⁿ−1 (6)

Quantum coherence does not eliminate the exponential dimension — it allows one to navigate it in a structured way without exponential classical memory cost. We want to be clear that this is, at most, a solution to a representational puzzle, not evidence that biological or artificial minds actually use quantum coherence; that remains an open empirical question addressed critically in Section XI.

B. Multiple Attractors in Superposition: A Speculative Cognitive Metaphysics

If a system must master m distinct tasks, each with its own attractor Aⱼ, a quantum state could in principle encode all of them simultaneously:

|Ψ⟩ = (1/√m) · Σ e^(iφⱼ) |Aⱼ⟩ (7)

The dynamical phases φⱼ differentiate the attractors without requiring additional dimensions. We offer this as a formal echo of Leibniz's monadology — each monad reflecting the whole universe, each attractor carrying the 'perspective' of its corresponding task — and, separately, of William James's description of consciousness as a 'stream' in which multiple associations coexist before attention fixes on one. Both parallels are proposed as illustrative, not as evidence that either philosopher's account maps onto quantum mechanics.

VII. THE DECOHERENCE DILEMMA: PHYSICS AND EPISTEMOLOGY

A. The Fundamental Trade-off

Quantum coherence faces a fundamental trade-off, expressed by the relation:

τ_c · ΔE ≳ ħ (8)

where τ_c is the coherence time and ΔE is the energy separation between distinguishable states. This is a physical constraint, not a technological limitation, and it has a natural — though speculative — epistemological reading.

Niels Bohr's complementarity held that quantum phenomena exhibit pairs of mutually exclusive properties — position and momentum, energy and time. One might describe an analogous 'cognitive complementarity' here: distinguishability (conceptual clarity) and coherence (openness to alternatives) would, on this reading, be mutually limited. We flag that this application of complementarity outside physics is metaphorical.

Kant identified a structurally analogous tension in reason: reason in its regulative use needs distinct concepts, but in its dialectical use needs openness to ideas that transcend experience. Quantum complementarity, if the cognitive analogy holds, would formalize this tension as a physical constraint — a striking but unproven claim.

B. The Mind as Decoherence Modulator

One possible resolution — offered here as a hypothesis, not a finding — is that a healthy mind does not have a fixed decoherence rate but modulates it dynamically according to task:

• Reflection / planning: ΔE small, τ_c long — multiple hypotheses coexist in superposition.

• Action / decision: ΔE large, τ_c short — collapse to a single attractor, commitment to one response.

This hypothesis resonates with Isaiah Berlin's distinction between 'foxes' (who relate many things) and 'hedgehogs' (who know one big thing): a healthy mind, on this reading, would alternate between fox-like (high coherence, multiple attractors) and hedgehog-like (low coherence, single attractor) modes, with pathology at either extreme. Heidegger's notion of resolute openness (Entschlossenheit) in Being and Time — decisive action combined with openness to revision — offers a further, admittedly loose, philosophical parallel to this dynamic modulation.

VIII. PHILOSOPHY OF MIND: PROPOSALS, NOT RESOLUTIONS

A. Searle's Chinese Room: A Candidate Formal Criterion

John Searle (1980) presented his well-known thought experiment: a human operator inside a room, following rules to manipulate Chinese symbols, can produce competent Chinese responses without understanding a word of Chinese. Searle's conclusion: syntactic manipulation is insufficient for semantics; understanding requires something beyond formal computation.

"Syntax is insufficient for semantics."

— John Searle, Minds, Brains, and Programs (1980)

We propose an attractor-theoretic criterion that operationalizes, without necessarily resolving, Searle's intuition. We define formal understanding as follows:

A system understands concept X if and only if: (i) an input x representing X drives the trajectory to attractor A_X; (ii) A_X carries semantic content — its geometry encodes relations among concepts; (iii) the trajectory remains in A_X under small perturbations (stability); (iv) from within A_X, the system can infer consequences (causal reasoning).

A system that matches symbols without developing distinct attractors per concept would, on this criterion, count as a purely syntactic automaton — close to what Searle described. A Transformer that outputs 'Tokyo is in Japan' without a geometric attractor for TOKYO distinct from one for PARIS would, by this criterion, not understand — it would be reproducing statistical patterns. We regard this as a precise formal criterion consistent with Searle's intuition; whether it is the correct criterion, or captures everything Searle meant by 'understanding,' is a further philosophical question we do not claim to have settled.

B. Functionalism and the Question of Substrate

Functionalism in philosophy of mind — defended by Hilary Putnam, Jerry Fodor, and the early Daniel Dennett — holds that mental states are defined by their functional, input–output relations, not by their physical substrate. The same mental function could, on this view, be realized in neurons, silicon, or any sufficiently complex substrate.

Our framework offers a challenge to strong functionalism, conditional on an empirical premise we have not established: that two systems can have identical input–output behavior while possessing radically different attractor landscapes — for instance, a linear attractor in ℝⁿ, simple and low-dimensional, versus a nonlinear attractor in a space of dimension 2ⁿ, complex and high-dimensional with elaborate semantic geometry.

If this premise holds for real systems, the same surface behavior could mask deep structural differences, and functionalism's exclusive focus on input–output equivalence would miss something the attractor's internal geometry captures. Whether this premise actually holds — whether current AI systems with similar benchmark performance really do have detectably different attractor structure — is an empirical question this paper does not test.

We note, with appropriate caution, that Putnam himself, in his later work (Reason, Truth and History, 1981; Representation and Reality, 1988), moved away from the functionalism he had helped found, for reasons broadly compatible with this kind of concern. We do not claim our framework is what Putnam had in mind, only that the concerns rhyme.

C. A Refined Turing Test: A Constructive Proposal

Alan Turing (1950) proposed the 'imitation game': a machine counts as intelligent if a human judge cannot distinguish it from a human in text conversation. The Turing test has been widely criticized — a sufficiently large lookup table, or a sufficiently large Transformer, might pass it without understanding anything.

"I propose to consider the question, 'Can machines think?'"

— Alan Turing, Computing Machinery and Intelligence (1950)

We propose, as a constructive alternative rather than a proven improvement, that a system be assessed by whether it possesses a global attractor of finite fractal dimension that: (i) is robust to distribution shift, so that the system generalizes rather than merely interpolates; (ii) supports causal inference, so that the system can reason about counterfactuals, not only correlations; (iii) generalizes to new tasks without catastrophic forgetting; and (iv) exhibits coherent, goal-directed behavior in something like the Aristotelian sense of telos.

This proposed test is more demanding than the Turing test because it targets structure rather than appearance. We think this is philosophically motivated — Plato's insistence that structure constitutes reality is one relevant precedent — but we acknowledge that operationalizing (i)–(iv) for a real system, and verifying that the criterion tracks understanding rather than a different notion of competence, remains future work.

IX. CONSCIOUSNESS AND QUANTUM COHERENCE: THE HARD PROBLEM

A. The Hard Problem and the Limits of Physical Approaches

David Chalmers, in The Conscious Mind (1996), distinguished the 'easy problem' of consciousness — explaining cognitive functions such as attention, memory, verbal report — from the 'hard problem': why is there subjective experience at all? Why is information processing accompanied by qualia — the redness of red, the painfulness of pain?

Thomas Nagel, in 'What Is It Like to Be a Bat?' (1974), argued that consciousness has an irreducible first-person perspective — a 'what it is like' that escapes any third-person objective description. Any physical theory of consciousness, including ours, should be candid about this limit.

Our hypothesis does not solve the hard problem. At most, it addresses what we might call a 'medium problem of consciousness': the functional structure that would make coherent, intentional, persistent experience possible — without addressing the question of subjective experience itself.

"There may be an explanatory gap between physical processes and conscious experience, but this does not prevent us from trying to identify the minimal structural conditions under which such experience might be possible."

— paraphrasing the spirit of Chalmers, The Conscious Mind (1996)

B. A Hypothesis: Consciousness as Coherent Superposition of Attractors

With appropriate caution, we propose the following working hypothesis: consciousness, in its functional aspect, might correspond to the property of coherently occupying multiple attractors in quantum superposition, with phase relations encoding the comparative evaluation of alternatives. We want to stress the modal language here — 'might,' not 'does.'

This hypothesis resonates with the Orch-OR theory of Penrose and Hameroff, which locates consciousness in quantum collapses within neuronal microtubules. Penrose, in The Emperor's New Mind (1989), argued that consciousness requires non-computable processes. Our hypothesis does not require non-computability, but it does require non-classicality: on this hypothesis the quantum substrate would be a condition, not decoration. We flag, however, that Orch-OR remains a minority position in neuroscience, contested on both theoretical and empirical grounds (see Section XI).

We also find resonances with Giulio Tononi's Integrated Information Theory (IIT), which measures consciousness via the quantity of integrated information (Φ) a system generates beyond its parts. Systems with multiple attractors in coherent superposition would, presumably under our hypothesis, exhibit high Φ. This is an inference from our framework, not an independent confirmation of it.

Karl Jaspers, in his philosophical psychopathology, distinguished healthy consciousness (Bewusstseinshelligkeit) from disturbances such as dissociation, confusion, and delirium. Systems without a stable attractor — or with premature decoherence, on our hypothesis — might exhibit analogues of these cognitive pathologies. We note this as a speculative, potentially testable implication, while emphasizing that no such test has been carried out.

X. IMPLICATIONS FOR AGI RESEARCH --- IF THE ARGUMENT HOLDS

We preface this section plainly: the points below follow only if the central conjectures of Sections II–V are correct. We list them because we think the conjectures are worth taking seriously and testing, not because we consider them established.

A. What the Argument Suggests Reconsidering

• Treating the scaling of Transformers as the primary strategy for AGI — if H1 is genuinely violated by the architecture, this would be a structural rather than a scale problem.

• Treating fine-tuning as 'learning' in the strong sense — on this account it would be surface adaptation rather than reorganization of the attractor landscape.

• Ignoring the physical substrate — if attractor geometry depends on substrate physics, hardware and algorithm would not be cleanly separable.

B. What the Argument Suggests Exploring

• Designing explicit dissipative operators (H1): architectures with structurally contractive dynamics, not only parametric regularization.

• Incorporating global energy functionals (H2): in the spirit of Friston's free-energy principle, but with the more explicit mathematical structure proposed here.

• Encoding goals mathematically (H3): structuring dynamics around explicit teleological operators, rather than only learning objectives from data.

• Exploring quantum-inspired architectures: tensor networks, spin systems, and eventually quantum hardware as a candidate substrate — understood as a long-term research direction, not a near-term engineering recipe.

XI. HONEST LIMITATIONS: WHAT THIS FRAMEWORK DOES NOT RESOLVE

A. The Limits of the Formalism

Intellectual honesty requires stating the limits of this framework plainly. Following Wittgenstein's distinction between what can be said and what can only be shown, there are dimensions of mind that our formalization does not capture — and perhaps cannot capture.

• We do not know whether the biological brain in fact uses quantum mechanics at a cognitively relevant scale. The evidence is inconclusive, and the debate between Penrose–Hameroff and critics such as Max Tegmark, who argues decoherence is far too fast for cognitive relevance, remains open.

• We have not solved the hard problem of consciousness. The relation between dynamical structure and subjective experience remains, as Chalmers argued, potentially mysterious as a matter of principle, not merely of current ignorance.

• We do not know how to engineer H1–H4 in a way that would produce human-level capability. This paper identifies candidate necessary conditions within an abstract framework; a concrete architecture satisfying them — and evidence that satisfying them matters for real-world intelligence — remains an open problem.

• The question of qualia — the redness of red, what it is to feel pain — may, as Nagel suggested, be irreducible to any structural or functional description. Our framework, even if entirely correct on its own terms, may be necessary but not sufficient for a complete theory of mind.

• The philosophical resonances drawn throughout this paper (Merleau-Ponty, Damasio, Husserl, James, and others) are proposed as productive analogies to motivate and interpret the mathematics. They are not derivations, and a reader should not conclude that, for example, H2 'is' a Wittgensteinian tautology in any technical sense — only that thinking about H2 alongside the Tractatus can be illuminating.

• Finally, the central applied claim of the paper — that standard Transformer architectures structurally lack the compact global attractor described in Sections II–V, and that this explains catastrophic forgetting — is, as flagged throughout, a conjecture motivated by the near-unitary character of the attention sub-mechanism. It has not been established for trained models incorporating normalization, residual connections, and nonlinearities, and we regard demonstrating or refuting it as the paper's most important open technical problem.

XII. FALSIFIABILITY AND A PATH TO VALIDATION

A. Why This Section Is Necessary

A framework this interdisciplinary is vulnerable to a specific failure mode: each individual claim is hedged carefully, but the paper as a whole reads as though it has accomplished more than any single hedged claim actually establishes. The corrective is not more hedging language but a concrete, falsifiable test list. We give one here, consolidating and sharpening the scattered qualifications of Sections II, V, and XI.

B. Falsifiable Predictions

(1) H1 and Transformers. If the composed map of a trained Transformer layer (attention, normalization, feed-forward, residual) satisfies a uniform one-sided Lipschitz condition with negative constant across the activation range observed in practice — measurable with existing local-Lipschitz-estimation methods — then H1 is not violated by that architecture, and the account of catastrophic forgetting offered in Section III.A does not apply to it. This is, at present, an open empirical question rather than a settled one in either direction.

(2) Attractor geometry and understanding. The criterion of Section VIII.A predicts that concept representations in a system satisfying H1–H4 should organize into geometrically distinct, stable attractor basins per concept, separable under standard clustering or probing methods applied to internal representations, in a way that a purely associative system need not. This is testable on existing models with existing interpretability tools, independent of whether the H1–H4 framework is otherwise correct, and we regard it as the most immediately tractable prediction in this paper.

(3) Dissipative retrofits. If an architecture is explicitly modified to satisfy H1 (via the contractive-design methods reviewed in Section II.C) while holding capacity and training data fixed, our framework predicts a measurable reduction in catastrophic forgetting relative to an unmodified baseline, at the cost of representational flexibility that should be independently measurable. A null result here — no improvement, or improvement fully explained by capacity differences — would weigh against the structural account of Section III.

(4) Consciousness-related predictions. We are explicit that the hypothesis of Section IX.B (consciousness as coherent superposition of attractors) does not currently yield a prediction we know how to test experimentally, either in artificial systems or in biological ones, given the present state of decoherence measurement at cognitively relevant scales. We list it as an open problem rather than a testable claim, in contrast to predictions (1)–(3).

C. What Would Constitute Progress Short of Full Validation

A referee should not require that all four predictions above be confirmed for this paper to have scientific value. Progress would consist of: sharpening prediction (1) into a specific, pre-registered measurement protocol; running prediction (2) against at least one open interpretability benchmark; and stating, in any subsequent version of this paper, which of (1)–(3) has been attempted and with what result. We take on record that none of (1)–(3) has been carried out as of this writing, and we regard this as the paper's principal limitation relative to its ambition.

XIII. CONCLUSION: MATHEMATICS IN CONVERSATION WITH PHILOSOPHY

The central thesis of this paper is that the difficulties of current AI — catastrophic forgetting, weak genuine generalization, the apparent absence of real understanding — are worth examining through a philosophical as well as an engineering lens. We do not claim that engineering alone caused these difficulties, or that philosophy alone can resolve them; we claim that treating the underlying questions as already settled has been a mistake.

We have argued that Transformers plausibly violate condition H1, as discussed in Section II.C, and that, if this conjecture is correct, several of the architecture's known limitations would follow within our framework as consequences rather than as unrelated engineering shortcomings. We want to be precise: this inference is conditional on an unproven premise, and we present it as a hypothesis for the field to test, not as a demonstrated result about deployed systems.

The four structural hypotheses (H1–H4) are each proposed as a formal counterpart to a requirement the philosophical tradition had already identified without the tools to formalize: dynamical identity (Kant, Merleau-Ponty), axiological coherence (Wittgenstein, Damasio), intrinsic intentionality (Searle, Aristotle), and epistemic robustness (Husserl, James). We think mathematics can help articulate philosophical questions with new precision; we do not think it replaces philosophical judgment about what these formalisms mean, nor does it settle any of the debates it draws on.

If the conjectures in this paper hold up under scrutiny, the path to AGI may require not only more data, more parameters, and more compute, but also different structure — and possibly, though this remains genuinely speculative, physical substrates beyond the classical von Neumann computer. We offer this as a research program and a set of conjectures we believe are falsifiable in principle, not as a solved problem.

"What we can say, we can say clearly. Whereof one cannot speak, thereof one must be silent."

— Ludwig Wittgenstein, Tractatus Logico-Philosophicus (1921)

In that spirit, we have tried to say clearly what we think we can say: a set of candidate mathematical conditions worth investigating in relation to intelligence, stated with their proofs where those proofs are standard, and flagged as conjecture where they are not. About what remains — the spark of experience, the light of understanding, the mystery of what it is to be — we keep the respectful silence that, on Wittgenstein's view, good philosophy owes to what exceeds it. We would add only that respectful silence is compatible with, and should not substitute for, continued rigorous inquiry.

REFERENCES

CLASSICAL AND MODERN PHILOSOPHY

[1] Aristotle. De Anima. Trans. W. S. Hett. Loeb Classical Library, 1936.

[2] Kant, I. Critique of Pure Reason [1781]. Trans. Norman Kemp Smith. London: Macmillan, 1929.

[3] Leibniz, G. W. Monadology [1714]. In: Philosophical Essays. Indianapolis: Hackett, 1989.

[4] Wittgenstein, L. Tractatus Logico-Philosophicus [1921]. Trans. D. F. Pears & B. F. McGuinness. London: Routledge, 1961.

[5] Hegel, G. W. F. Phenomenology of Spirit [1807]. Trans. A. V. Miller. Oxford: Oxford University Press, 1977.

PHILOSOPHY OF MIND AND CONSCIOUSNESS

[6] Searle, J. R. Minds, Brains, and Programs. Behavioral and Brain Sciences, vol. 3, pp. 417–424, 1980.

[7] Turing, A. M. Computing Machinery and Intelligence. Mind, vol. 59, pp. 433–460, 1950.

[8] Chalmers, D. J. The Conscious Mind: In Search of a Fundamental Theory. Oxford University Press, 1996.

[9] Nagel, T. What Is It Like to Be a Bat? The Philosophical Review, vol. 83, no. 4, pp. 435–450, 1974.

[10] Putnam, H. Representation and Reality. MIT Press, 1988.

[11] Dennett, D. C. Consciousness Explained. Little, Brown and Company, 1991.

[12] Merleau-Ponty, M. Phenomenology of Perception [1945]. Trans. Colin Smith. London: Routledge, 1962.

[13] Husserl, E. Ideas Pertaining to a Pure Phenomenology [1913]. Trans. F. Kersten. The Hague: Nijhoff, 1982.

[14] Heidegger, M. Being and Time [1927]. Trans. John Macquarrie & Edward Robinson. Oxford: Blackwell, 1962.

NEUROSCIENCE AND COGNITIVE SCIENCE

[15] Damasio, A. Descartes' Error: Emotion, Reason, and the Human Brain. New York: Putnam, 1994.

[16] Friston, K. The Free-Energy Principle: A Unified Brain Theory? Nature Reviews Neuroscience, vol. 11, pp. 127–138, 2010.

[17] Penrose, R. The Emperor's New Mind. Oxford University Press, 1989.

[18] Hameroff, S. & Penrose, R. Consciousness in the Universe: A Review of Orch OR Theory. Physics of Life Reviews, vol. 11, pp. 39–78, 2014.

[19] Tononi, G. Consciousness as Integrated Information: A Provisional Manifesto. Biological Bulletin, vol. 215, pp. 216–242, 2008.

[20] Prigogine, I. & Stengers, I. Order Out of Chaos: Man's New Dialogue with Nature. New York: Bantam Books, 1984.

MATHEMATICS AND DYNAMICAL SYSTEMS

[21] Temam, R. Infinite-Dimensional Dynamical Systems in Mechanics and Physics. Springer, 1997.

[22] Robinson, J. C. Infinite-Dimensional Dynamical Systems. Cambridge University Press, 2001.

[23] Vieira, T. A. Global Attractor Theory for the MDEI Stochastic PDE Framework. Preprint, 2026.

[24] Rudolph, H.-J. & Vieira, T. A. The Mathematical Pattern of Coherence. PhilArchive preprint, 2024.

ARTIFICIAL INTELLIGENCE AND DEEP LEARNING

[25] Vaswani, A. et al. Attention Is All You Need. Proceedings of NeurIPS, 2017.

[26] Imanov, O. Y. L. Mechanistic Analysis of Catastrophic Forgetting in Large Language Models. arXiv preprint, 2026.

[27] McCloskey, M. & Cohen, N. J. Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem. Psychology of Learning and Motivation, vol. 24, pp. 109–165, 1989.

[28] Kim, H. et al. The Lipschitz Constant of Self-Attention. Proceedings of ICML, 2021.

[29] Kotha, S., Springer, J. M. & Raghunathan, A. Understanding Catastrophic Forgetting in Language Models via Implicit Inference. Proceedings of ICLR, 2024.

[30] Fein-Ashley, J. Flowing Through Layers: A Continuous Dynamical Systems Perspective on Transformers. arXiv preprint arXiv:2502.05656, 2025.

[31] Dasoulas, G., Scaman, K. & Virmaux, A. Lipschitz Normalization for Self-Attention Layers with Application to Graph Neural Networks. Proceedings of ICML, 2021.

Note on self-citation: references [23] and [24] are preprints by the present author (and a co-author), not yet subject to independent peer review. They are cited here as sources for definitions used in this framework, not as externally validated, established results. A reviewer at a top venue would reasonably ask that the claims resting on them be re-derived or independently substantiated before publication.