Skip to main content

Formal limits of human-mind computability and the epistemic horizons of artificial intelligence

A synthesis of the logical, physical, biological and complexity-theoretic arguments that separate human cognition from contemporary digital architectures

Caio de Melo

Abstract

The computationalist thesis — according to which human cognition can be instantiated on discrete digital architectures over the binary alphabet — depends on an ontological transposition of the Church-Turing Thesis into the philosophy of mind. This article articulates four mutually reinforcing lines of formal refutation: the Gödelian Lucas-Penrose argument; hypercomputation over the real continuum (BSS, Pour-El and Richards, Siegelmann); Robert Rosen's relational biology; and the Orch OR program of Penrose and Hameroff. In parallel, we examine two intractable barriers to contemporary AI: the TC⁰ circuit expressivity identified by Merrill and Sabharwal for Transformer architectures, and the impossibility theorem of Colbrook, Antun and Hansen (PNAS 2022) which dissociates the existence of optimal neural networks from their algorithmic computability. We conclude that the gap between human cognition and binary digital systems is not a contingent gap of scale but a structured theoretical demarcation.

Keywords:computabilityartificial intelligencephilosophy of mindgödel's theoremspenrosetheoretical limits

Introduction

The postulation that the human brain, or the cognitive processes that constitute the human being, can be perfectly instantiated in digital computing architectures rests on a precise assumption: that the mind is reducible to formal manipulations of binary states over the discrete alphabet [1]. Historically, this assumption draws on a philosophical extension of the Church-Turing Thesis[1]. Originally formulated by Alonzo Church and Alan Turing in 1936, the thesis postulates that every function computable by a step-by-step effective method — strictly conceived as an algorithmic routine executed by a human calculator operating on explicit rules with paper and pencil — is computable by a Turing machine[3][4].

The mechanical transposition of this thesis into the philosophy of mind produced what contemporary epistemological literature identifies as the Church-Turing Fallacy[1]. Gualtiero Piccinini demonstrated that the original thesis is a proposition strictly confined to the theory of effective procedures in formal languages and to the tractability of finite symbol strings; it establishes no universal physical law that forces an open, biologically structured material system to limit its functional behaviour to the dynamics of a Turing automaton[1]. When classical computationalism asserts that the mind is ontologically a program running on a von Neumann substrate, it conflates the descriptive capacity of an approximate numerical model with the structural identity of the simulated natural system[2].

This article articulates four independent lines of formal refutation of that identification — the Gödelian argument, computability on the continuum, relational biology, quantum gravitational reduction — and examines two intractable barriers to contemporary AI that converge, through independent routes, on the same conclusion.

Background

Traditional binary architecture operates by temporal discretization and forced Boolean quantization, collapsing the continuous dynamic spectrum of the biophysical world into discrete logical states governed by Boolean algebra[1]. The axiomatic assumption that the brain is functionally equivalent to a network of binary logic gates simplifies dendritic electrochemical integration and disregards that the manipulation of symbols devoid of intrinsic semantic reference cannot generate biological causality or conscious phenomenology through mere algorithmic recombination[2]. The transition from the continuous domain to the discrete binary entails a structural loss of topological and dynamic information that from the outset vitiates the pretension of representational universality of conventional digital architectures[1].

The four theoretical programs examined in this article share a common analytical core: to identify, for each domain (logical, mathematical, biological, physical), a structural property of human cognition or of living matter that is formally inaccessible to any Turing machine operating over .

Thesis

The impossibility of fully computing the human being on contemporary binary digital architectures based on does not constitute a contingent inadequacy of processing capacity or volume of sampled parameters. It reflects a theoretical demarcation grounded, simultaneously, in computability theory, formal mathematical logic, the physics of continuous media, and the categorical biology of open systems.

The following sections validate this thesis along four independent but convergent routes.

Validation

The Gödelian Lucas-Penrose argument

The core of the logico-mathematical demonstrations against the computability of consciousness rests on the Incompleteness Theorems enunciated by Kurt Gödel in 1931[5]. The argument that human intelligence transcends the power of any automated formal system was pioneeringly systematized by J. R. Lucas in 1961[6] and substantially developed by Roger Penrose in The Emperor's New Mind (1989)[7] and Shadows of the Mind (1994)[8].

The mathematical foundation begins with the consideration of a formal system effectively generated by a recursive set of axioms and inference rules, with expressive power sufficient to encompass elementary Peano arithmetic[5]. By the First Gödel Theorem, if is formally consistent, there exists an arithmetic proposition , denoted , which asserts its own unprovability in through Gödel's numerical coding scheme:

If is consistent, cannot be demonstrated internally without logical contradiction[5]. By the Second Gödel Theorem, the system is unable to demonstrate its own consistency, expressed by the formula , unless it is inconsistent:

The Lucas-Penrose reasoning is structured through a direct confrontation between the reflexive capacity of the human mind and the formal restrictions imposed on discrete automata[6][8]. On analysing the formalism , a human mathematician grasps the metamathematical construction and directly apprehends the truth of in the domain of natural numbers — because the assertion that is unprovable corresponds exactly to the semantic observation that it is undecidable within the internal calculus of the system[8]. If the human mind functioned identically to a finite automaton or a formalizable Turing system, it could never establish the truth of without incurring inconsistency[6]. Since human mathematical cognition accesses the objective truth of by invoking semantic notions of truth that transcend purely syntactic calculus, it follows that human understanding cannot be reduced to a computable formalism[8].

This formulation resonates with the disjunction formulated by Gödel in his Gibbs Memorial Lecture of 1951[12]: either the human mind infinitely surpasses the operational limits of any finite Turing machine, or there exist absolutely unsolvable elementary Diophantine problems. While Gödel considered both options open, Lucas and Penrose argued that the first is the theoretically inevitable route.

The logical structure of the argument follows from an inferential progression that departs from cognitive self-identification with a mechanical system and lands in an irreducible formal contradiction:

  • If one admits that human intellectual faculties correspond to a discrete deterministic axiomatic formal system , the mind is parameterized by the same code as the system.
  • Assuming that human self-reflection attests to the soundness of its own reasoning procedures, one postulates the soundness of for formulas.
  • By Gödel's Theorem, the algorithmic system is structurally prevented from deriving under penalty of formal self-destruction.
  • Simultaneously, the human mathematician demonstrates and apprehends the unconditional truth of through semantic intuition.
  • An inescapable asymmetry is established between the finite automaton and the reflexive capacity of the mind, refuting the biunivocal equivalence between consciousness and a discretized machine.

Critical assessment. The argument has motivated logico-mathematical debates with contributions from Solomon Feferman, David Chalmers, and Hilary Putnam[9][10][11]. The most consistent refutations do not question the formal validity of Gödel's theorems, but deconstruct the epistemic axioms introduced to transpose arithmetic theorems into cognitive psychology.

The first critical point lies in the consistency axiom. If the human cognitive architecture admits localized operational contradictions and operates according to a paraconsistent logic — in which the trivial deduction of any formula from a contradiction () is axiomatically blocked — the immediate application of the First Gödel Theorem collapses, since paraconsistent formalisms tolerate pointwise antinomies without inferential failure[6].

The second problem, formulated rigorously by Feferman[9], centres on the soundness self-knowledge axiom. By the Second Theorem, the demonstration that is true depends on the premise that the system is consistent. For the human being to unassailably assert the truth of , they would have to formally prove that their own epistemic system is consistent — which Gödel prevents. Thus the mathematician does not know unconditionally that is true; they only know the conditional proposition "if I am consistent, then is true", a conclusion that a computer also trivially derives[9].

Chalmers and Putnam reinforced that a computational agent can be governed by an algorithm of intrinsic complexity so immeasurable that its complete specification is unreachable by the agent itself[10][11]. A computational mind unaware of its own axiomatic coding is unable to construct its specific Gödel sentence, making unviable any operational demonstration of superiority over the formalism that implements it.

Continuous mathematical analysis and physical hypercomputation

The thesis that cognition is exhausted in binary representations faces intractable barriers in the domain of functional analysis and the physics of continuous systems[13]. While the classical model of computability limits processing to discrete enumerable spaces, the fundamental dynamic processes in the cerebral cortex unfold on continuous differential manifolds over the field of real numbers , governed by ionic concentrations, electrochemical gradients, and membrane potentials in continuous time[21].

The formal theory of the limits of discrete computation on the continuum found its formulation in the BSS model developed by Lenore Blum, Mike Shub, and Steve Smale in 1989[13]. The BSS model generalizes classical computation theory to arbitrary fields, conceiving automata capable of computing directly over elements of in unit time with arbitrary infinite precision. On investigating the geometry of chaotic attractors and the dynamics of continuous systems, Smale and collaborators demonstrated the existence of fractally structured sets — such as the complements of certain Julia sets — that are formally undecidable and non-enumerable in continuous space[13].

The classic work of Marian Pour-El and J. Ian Richards, Computability in Analysis and Physics (1989), established theorems that connect the differential equations of classical deterministic physics to Turing non-computability[14]. On analysing the three-dimensional wave equation

the authors demonstrated the existence of perfectly computable initial conditions in the space of continuous functions whose subsequent time solutions generate physical energy distributions that are mathematically non-computable by any Turing machine[14]. This theorem shows that even within classical deterministic physics, the continuous temporal evolution of certain physical fields produces measurable quantities that violate the computational reducibility of the Church-Turing thesis.

In the field of neural networks, Hava Siegelmann (1995) demonstrated that assuming real-valued synaptic weights confers upon dynamic systems a power that surpasses the classical Turing limit[15]. Siegelmann conceived Analog Recurrent Neural Networks (ARNN) with continuous chaotic translations (analog shift maps) and proved that such architectures, operating with dynamic parameters in , break the ceiling of finite decidability. The ARNNs can compute the non-uniform complexity class , allowing them to resolve in polynomial time languages undecidable by conventional Turing machines — including the halting problem for data encoded in the continuous expansion of an irrational real number.

Although it is argued that the real physical exploration of requires infinite precision of measurement — constrained by quantum thermal noise and macroscopic thermodynamics — Siegelmann's mathematical analysis shows that the discrete binary model based on is a restrictive sub-specification of the neural architectures that make use of the continuous state space. As anticipated by Geroch and Hartle (1986), the requirement that scientific theories reduce to algorithmic formulations is a methodological imposition of digital computation, not an ontological limit of the universe[16].

Robert Rosen's relational biology and causal closure

The most formally structured refutation of the computability of living organisms was outlined by biophysicist and mathematician Robert Rosen[17]. In Life Itself (1991), and in subsequent developments with A. H. Louie, Rosen rigorously deduced — through category theory — that a living biological system is not a machine, being structurally impossible to derive a purely simulable model for its essential organization[18].

The demonstration rests on the formalization of Aristotle's four categories of causation within mathematical categories whose objects represent physico-chemical states and whose morphisms represent causal processes. In Rosen's minimal self-organizational model — the Metabolism-Repair system or -system — the fundamental biological flows articulate between a metabolic transformation of chemical inputs, a renewal of damaged catalysts, and a structural preservation process[19][20].

For a system to avoid an infinite regress of causal dependence on exogenously supplied catalysts, the ontological condition necessary and sufficient to characterize life resides in what Rosen defined as closure to efficient causation: all operators and catalysts that act as the efficient cause of metabolic processes must be produced and maintained by processes internal to the system itself[20].

From this formalization, Rosen formulated theorems that demonstrate an ontological difference between mechanisms and organisms:

  • Structure of mechanisms. Every computational machine or traditional physical mechanism is decomposable into a set of causal processes whose operators can be recursively decomposed in a purely linear and predicative manner. In an algorithmic automaton, the graph of efficient causes is strictly acyclic — the program is always supplied a priori from outside the computing system.
  • Biological impredicativity. A system closed under efficient causation encloses crossed self-referential causal cycles where a function is defined in terms of a set that contains it as an operator element. This causal interweaving confers on the system the status of an irreducible impredicative entity.
  • Theorem of the non-computability of models of living systems. Rosen and Louie mathematically proved that an entity possessing impredicative cycles closed under efficient causation does not admit a model simulable by Turing machines. Any attempt to convert an -system into a discrete algorithm fragments its mutual causal integrity, converting internal efficient causes into exogenous parameters dependent on external algorithmic initialization[18].

For the thesis of binary representation of the brain, the consequences are decisive: a computational architecture is based on the rigid separation between software (the set of functions that acts as a transient efficient cause) and hardware (the physical substrate that passively receives binary state as material cause). In the living brain, metabolism continuously constructs, regenerates, and modifies the very physical support that performs cognitive functions. There is no stable ontological boundary between physical processor and logical program; the hardware is continuously regenerated by the very dynamics of the functions it processes. This cyclic causal identity breaks with the predicative model of discrete computational languages, indicating that life and mind operate beyond the horizon simulable by binary automata.

Quantum gravity, non-computability, and the Orch OR model

Beyond the limitations of deductive logic and relational biology, the debate reaches the foundations of fundamental physics through the theory of Orchestrated Objective Reduction (Orch OR), developed by Roger Penrose and neurobiologist Stuart Hameroff[21].

The starting point consists in identifying an epistemological duality in contemporary quantum mechanics. On the one hand, the continuous linear evolution of the state vector prevails, described by the Schrödinger equation:

This unitary process is deterministic and time-reversible; consequently, it is fully computable and can be simulated by classical or quantum computers through circuits of discrete unitary gates[2].

On the other hand, the process intervenes — the collapse of the wave function generated during measurement. The canonical interpretation postulates that obeys a pure stochasticity governed by the Born rule (). However, a pure stochastic generator — mathematically equivalent to a Turing machine endowed with an oracle of Bernoulli random variables — confers no gain in decidability of functions, nor allows the derivation of mathematical judgments with intrinsic understanding of truth[8].

Penrose argued that the impasse of modern quantum mechanics reflects the non-existence of a unified quantum gravity, postulating a new fundamental physical process: Objective Reduction (OR)[23]. In the theory of Penrose and Lajos Diósi[22], quantum superpositions of macroscopic masses imply the superposition of distinct spacetime geometries. When the bifurcated curvature accumulates a sufficient gravitational self-energy discrepancy (), the geometry of the continuum becomes unstable, forcing a spontaneous decay to a classical eigenstate within a characteristic time :

The core of the thesis rests on the postulate that the OR process is neither random nor arbitrary: it operates according to a still-unknown physics, structured in the fundamental geometry at the Planck scale (), strictly non-algorithmic and non-computable in nature[21]. The selection of eigenstates in quantum gravitational reduction would depend on global properties of the fabric of spacetime that no Turing machine can compute or numerically approximate through step-by-step derivations.

The neurobiological component was Hameroff's contribution, who identified neuronal microtubules as the subcellular apparatus able to thermally isolate coherent quantum states[21]. Microtubules constitute the structural framework of the neuronal cytoskeleton, formed by cylindrical-conical networks of tubulin heterodimers with electric dipoles. In the Orch OR model, quantum states would be protected from thermal decoherence through Fröhlich condensates, structured water sheaths, and dipolar ordering. Once the gravitational threshold is reached, microtubules would undergo orchestrated objective collapses that would modulate the sensitivity of synaptic discharges, acting as the primary factor in the generation of acts of will, conscious judgment, and human logical intuition that escape syntactic predictability.

Biophysical criticism. Max Tegmark's work (2000) calculated the decoherence time of tubulin superpositions in the brain at orders of magnitude on the scale of s, demonstrating that the brain's aqueous and warm environment degrades coherent quantum states much faster than the temporal scale required for the firing of action potentials ( to s)[24]. In scientific replies (2014, 2016), Hameroff and Penrose contested these calculations by presenting experimental evidence of quantum resonances at giga- and tera-hertz frequencies in the cytoskeleton[21], but the postulate that consciousness results from non-computable quantum gravitational dynamics remains a theoretical hypothesis awaiting experimental confirmation[28].

Mathematical limits of contemporary AI

The empirical advance of deep neural networks and large language models based on the Transformer architecture has made clear that the empirical application of statistical learning algorithms on discrete substrates does not resolve the foundational problems of computability and expressivity of human cognition[25]. An examination of the recent mathematical literature reveals intractable barriers in circuit complexity and numerical non-computability.

Transformer expressivity and circuit complexity

The formal characterization of the computational power of Transformer-based models was elucidated in investigations led by William Merrill and Ashish Sabharwal (2023, 2024)[26][27]. The authors formally demonstrated that Transformer decoder-only models executed with finite logarithmic precision and without generation of intermediate steps are rigorously confined to the class of Boolean circuits (constant-depth circuits with counting gates and majority threshold):

Containment in implies severe theoretical limitations:

  • Models in this class are structurally incapable of solving -complete problems, such as the exact simulation of a finite-state automaton or the evaluation of general Boolean formulas.
  • In direct inference mode, they cannot solve elementary connectivity in directed graphs (a problem complete for ) nor -complete problems.
  • They cannot recognize all regular languages of the Chomsky hierarchy, failing strict structural induction.

To circumvent this ceiling, the current paradigm resorts to the heuristic strategy of Chain-of-Thought (CoT). Merrill and Sabharwal proved that the progressive addition of intermediate steps gradually expands expressivity: a logarithmic number of steps raises it up to (logarithmic space), a linear number decodes classes between and context-sensitive grammars, and a polynomial number of steps attains [26].

Although CoT with polynomial complexity raises expressivity to — formally empowering the model to simulate polynomial Turing machines —, the system remains confined to the strict limits of algorithmic determinism. The process is based on a linearization through matching of syntactic subgraphs that simulates deductive steps by statistical likelihood[26]. The model possesses no genuine intentionality or intrinsic semantics, remaining conditioned by the fundamental fragilities of discrete formal systems identified by Gödel's theorems.

The dissociation between existence and computability: Colbrook, Antun and Hansen (PNAS 2022)

Beyond the restrictions of expressivity, the methodological basis of deep neural networks has suffered a refutation of its analytical approximation capacity through the works of Matthew J. Colbrook, Vegard Antun, and Anders C. Hansen, published in the Proceedings of the National Academy of Sciences in 2022, in the context of Smale's 18th problem[25].

For three decades, the community has relied on the Universal Approximation Theorems (Cybenko, Hornik), which ensure that for any well-behaved continuous function there exists a finite neural network capable of approximating the map with arbitrarily small error ().

Drawing on the theory of the Solvability Complexity Index (SCI), Colbrook and collaborators demonstrated a mathematical rupture between the existence of the optimal neural network and its algorithmic computability by any digital machine[25]. On examining linear and non-linear inverse problems, they established theorems of structural impossibility:

For classes of problems with condition number , there exists no algorithm such that .

The theorem shows that, even in classes of problems that are perfectly stable and well-conditioned:

  1. An ideal approximating neural network provably exists from the ontological-mathematical standpoint.
  2. There exists not, nor can there exist, any algorithm — deterministic, heuristic, or probabilistic Monte Carlo — executed on a von Neumann digital architecture capable of converging to the optimal weights from finite sets of training data.
  3. An insoluble dilemma is installed between precision and stability: the more the network tries to approximate the precision digits of the objective function, the greater its vulnerability to environmental perturbations. Obtaining digits may demand incommensurable amounts of samples, while reaching the digit becomes mathematically impossible for any classical algorithmic machine[25].

This discovery reveals that human biological intelligence — able to extract continuous invariants and act with dynamic stability in a physical environment replete with analogue noise — possesses organizational principles that discrete algorithmic computation is unable to replicate through traditional training paradigms.

Conclusion

The investigation of mathematical, logical, and biological foundations exposes intrinsic limits to the attempt to equate the human being with a discrete-state machine over [1].

Firstly, the logical framework erected from Gödel's Incompleteness Theorems and the Lucas-Penrose argument establishes that any attempt to exhaust mathematical reasoning and reflexive consciousness in a finitely parameterized automaton imposes an inescapable choice between epistemic incompleteness or syntactic inconsistency[8]. Although the human mathematician is subject to contingent cognitive limitations, their method of validation rests on semantic notions of truth that transcend the restricted calculus of closed formal deductions[6].

Secondly, the topology of the physical continuum over constitutes a mathematical barrier to binary discretization[13]. As demonstrated in the theorems of Pour-El and Richards and in the analogue models of Siegelmann, systems that evolve over continuous dynamic variables exhibit hypercomputational behaviours unreachable for Turing machines operating on truncated-value matrices[14][15].

Thirdly, Rosen's relational biology formalized that the human being, as a self-organized living entity, exhibits functional closure under efficient causation[20]. The impredicative circular causality of biological systems prevents the functional split between software and hardware that is characteristic and mandatory for von Neumann architecture. The organism actively regenerates its own material constitution in response to continuous thermodynamic demands, a dynamic principle that admits no predicative algorithmic translation without destruction of vital self-sustenance.

Finally, the analysis of the contemporary frontiers of AI shows that deep-network-based technology operates under structural expressivity restrictions bounded by classes of Boolean circuits such as and deterministic subspaces of [26]. Complementarily, the SCI theorems of Colbrook, Antun, and Hansen demonstrate the definitive mathematical rupture between the abstract existence of neural approximations and the algorithmic impossibility of computing networks that are simultaneously stable and precise in the continuous world[25].

Thus the impossibility of fully computing the human being on contemporary binary digital architectures based on does not constitute a contingent inadequacy of processing capacity or volume of sampled parameters. It reflects a theoretical demarcation grounded in computability theory, formal mathematical logic, the physics of continuous media, and the categorical biology of open systems[1].

References

  1. [1]Piccinini, G. (2007). Computationalism, the Church-Turing thesis, and the Church-Turing fallacy. Synthese, 154(1), 97-120. https://doi.org/10.1007/s11229-005-0194-z
  2. [2]Copeland, B. J., & Shagrir, O. (2019). Physical computability theses. Oxford Handbook of Philosophy of Science. https://oronshagrir.huji.ac.il/sites/default/files/oronshagrir/files/ver19_physical_computability_theses.pdf
  3. [3]Turing, A. M. (1936). On computable numbers, with an application to the Entscheidungsproblem. Proceedings of the London Mathematical Society, s2-42(1), 230-265. https://doi.org/10.1112/plms/s2-42.1.230
  4. [4]Church, A. (1936). An unsolvable problem of elementary number theory. American Journal of Mathematics, 58(2), 345-363. https://doi.org/10.2307/2371045
  5. [5]Gödel, K. (1931). Über formal unentscheidbare Sätze der Principia Mathematica und verwandter Systeme I. Monatshefte für Mathematik und Physik, 38, 173-198. https://doi.org/10.1007/BF01700692
  6. [6]Lucas, J. R. (1961). Minds, machines and Gödel. Philosophy, 36(137), 112-127. https://doi.org/10.1017/S0031819100057983
  7. [7]Penrose, R. (1989). The Emperor's New Mind: Concerning computers, minds, and the laws of physics. Oxford University Press. https://global.oup.com/academic/product/the-emperors-new-mind-9780192861986
  8. [8]Penrose, R. (1994). Shadows of the Mind: A search for the missing science of consciousness. Oxford University Press. https://global.oup.com/academic/product/shadows-of-the-mind-9780195106466
  9. [9]Feferman, S. (1995). Penrose's Gödelian argument. PSYCHE: An Interdisciplinary Journal of Research on Consciousness, 2, 21-32. https://math.stanford.edu/~feferman/papers/penrose.pdf
  10. [10]Chalmers, D. J. (1995). Minds, machines, and mathematics: A review of Shadows of the Mind by Roger Penrose. PSYCHE, 2, 11-20. https://consc.net/papers/penrose.html
  11. [11]Putnam, H. (1995). Review of Shadows of the Mind by Roger Penrose. Bulletin of the American Mathematical Society, 32(3), 370-373. https://doi.org/10.1090/S0273-0979-1995-00606-X
  12. [12]Koellner, P. (2018). On the question of whether the mind can be mechanized, I: From Gödel to Penrose. Journal of Philosophy, 115(7), 337-360. https://doi.org/10.5840/jphil2018115721
  13. [13]Blum, L., Shub, M., & Smale, S. (1989). On a theory of computation and complexity over the real numbers: NP-completeness, recursive functions and universal machines. Bulletin of the American Mathematical Society, 21(1), 1-46. https://doi.org/10.1090/S0273-0979-1989-15750-9
  14. [14]Pour-El, M. B., & Richards, J. I. (1989). Computability in Analysis and Physics. Perspectives in Mathematical Logic. https://doi.org/10.1007/978-3-662-21717-7
  15. [15]Siegelmann, H. T. (1995). Computation beyond the Turing limit. Science, 268(5210), 545-548. https://doi.org/10.1126/science.268.5210.545
  16. [16]Geroch, R., & Hartle, J. B. (1986). Computability and physical theories. Foundations of Physics, 16(6), 533-550. https://doi.org/10.1007/BF01886519
  17. [17]Rosen, R. (1991). Life Itself: A comprehensive inquiry into the nature, origin, and fabrication of life. Columbia University Press. http://cup.columbia.edu/book/life-itself/9780231075657
  18. [18]Louie, A. H. (2007). A living system must have noncomputable models. Artificial Life, 13(3), 293-297. https://doi.org/10.1162/artl.2007.13.3.293
  19. [19]Mossio, M., Longo, G., & Stewart, J. (2009). A computable expression of closure to efficient causation. Journal of Theoretical Biology, 257(3), 489-498. https://doi.org/10.1016/j.jtbi.2008.11.012
  20. [20]Cárdenas, M. L., Letelier, J. C., Gutiérrez, C., Cornish-Bowden, A., & Soto-Andrade, J. (2010). Closure to efficient causation, computability and artificial life. Journal of Theoretical Biology, 263(1), 79-92. https://doi.org/10.1016/j.jtbi.2009.11.010
  21. [21]Hameroff, S., & Penrose, R. (2014). Consciousness in the universe: A review of the 'Orch OR' theory. Physics of Life Reviews, 11(1), 39-78. https://doi.org/10.1016/j.plrev.2013.08.002
  22. [22]Diósi, L. (1989). Models for universal reduction of macroscopic quantum fluctuations. Physical Review A, 40(3), 1165-1174. https://doi.org/10.1103/PhysRevA.40.1165
  23. [23]Penrose, R. (1996). On gravity's role in quantum state reduction. General Relativity and Gravitation, 28(5), 581-600. https://doi.org/10.1007/BF02105068
  24. [24]Tegmark, M. (2000). Importance of quantum decoherence in brain processes. Physical Review E, 61(4), 4194-4206. https://doi.org/10.1103/PhysRevE.61.4194
  25. [25]Colbrook, M. J., Antun, V., & Hansen, A. C. (2022). The difficulty of computing stable and accurate neural networks: On the barriers of deep learning and Smale's 18th problem. Proceedings of the National Academy of Sciences, 119(12), e2107151119. https://doi.org/10.1073/pnas.2107151119
  26. [26]Merrill, W., & Sabharwal, A. (2024). The expressive power of transformers with chain of thought. Proceedings of the 12th International Conference on Learning Representations (ICLR). https://openreview.net/pdf?id=CDmerQ37Zs
  27. [27]Merrill, W., & Sabharwal, A. (2023). The parallelism tradeoff: Limitations of log-precision transformers. Transactions of the Association for Computational Linguistics, 11, 531-545. https://doi.org/10.1162/tacl_a_00562
  28. [28]Aaronson, S. (2016). The ghost in the quantum Turing machine. The Once and Future Turing: Computing the World, 193-296. https://www.scottaaronson.com/papers/giqtm3.pdf
Formal limits of human-mind computability and the epistemic horizons of artificial intelligence | caiodemelo.org