REVIEW 4 major objections 5 minor 25 references
Toward the Axiomatization of Intelligence: Structure, Time, and Existence
T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash
Pith's one-line read The paper claims that intelligence can be defined axiomatically as a set whose element count grows, shrinks, and internally rearranges in prescribed ways over time, and uses this criterion to compare neural networks with a biological…
desk verdict A serious but flawed attempt to axiomatize intelligence; the cardinality-based I/O conditions are internally inconsistent and don't match the paper's own state-based examples. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the time-indexed set pair $(I, O)$ inside a finite universe $U$, together with the time mapping $T_i$ that carries $I_i$ and $O_i$ to $I_{i+1}$ and $O_{i+1}$ while keeping $|U|$ constant. Input and output structures are defined by strict cardinality inequalities on element transfers across the boundary; the processing structure is defined separately by cardinality changes between internal subsets $T$ and $V$, so no interaction with the exterior is required. This machinery carries the argument because it converts the verbal naive definition into a condition that can be checked against a temporal record, and it supplies the objects and morphisms that the categorical extension reuses.
What would settle it
Build an adaptive switch that changes its internal state in response to a signal but never adds or removes elements from its defining set: under the paper's criteria it has no input or output structure because $|I|$ and $|O|$ never change, even though it satisfies the naive input-processing-output description. A working instance would settle whether the cardinality conditions are necessary for intelligence.
Extended reading notes
Core claim
The central claim is that a system is an intelligence set $I$ exactly when, under a time mapping $T_i$, it possesses structures $C_i$, $C_p$, and $C_o$ obeying specific set-size conditions. Input is $|I_{i+1}| > |I_i|$ and $|O_{i+1}| < |O_i|$; output reverses the two inequalities; and processing is an internal rearrangement between subsets $T, V$ of $I$ with no element transfer to $O$. The paper then demonstrates the criterion on three systems and concludes from their time-step structures that the reflex and the local-learning network are closer to each other than either is to an error-optimizing network. A categorical reformulation treats each intelligence as a triple $(C_i, C_p, C_o)$, time evolution as a functor from a Time Category to an Intelligence Category, and imitation or translation between intelligences as functors between categories.
Load-bearing premise
The load-bearing premise is that every act of input, processing, and output can be represented as elements physically moving between finite sets, so that the act shows up as a change in set sizes.
Editorial extensions
If this is right
- The definition makes 'possesses input, processing, and output structures' a checkable property: from a system's time-evolution data one can decide whether it is an intelligence set.
- The time-step analysis of the three examples implies that backpropagation-based networks require more stages than input-processing-output, which under the definition marks them as less biologically plausible than the local-learning and reflex cases.
- Local, self-organizing learning and the biological reflex show the same short input-processing-output loop, so the definition provides a formal sense in which biologically inspired AI should avoid global error functions.
- The categorical extension makes learning, imitation, translation, and abstraction representable as morphisms and functors, so relationships between different intelligences can be composed and studied abstractly.
- The activity concept implies that a structurally capable but unpowered or inactive system is not intelligent during that interval; intelligence as defined is a property of temporal interaction, not of static structure.
Reading between the lines
- Beyond the paper: a digital computer whose program flips internal bit states without changing the number of active elements would be invisible to the cardinality conditions; extending the formalism to track state values would make it apply to ordinary software.
- Beyond the paper: the categorical machinery suggests a concrete comparison test for rival definitions of intelligence, namely check whether their functors into a common universe category classify the same three example systems identically.
- Beyond the paper: the paper's solution for simultaneous input and output still asks the exterior set to shrink and grow at the same time; a definition that splits the exterior into disjoint input and output channels would handle real-time bidirectional systems more cleanly.
- Beyond the paper: activity could be operationalized as the fraction of time steps in which at least one input or output inequality holds, giving a quantitative ranking of how persistently a system counts as intelligent.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a meta-framework for defining intelligence: first fix a set-theoretic depiction of the universe, then select a naive definition of intelligence, and finally formalize that naive definition in the chosen depiction. The naive definition selected is that intelligence is an entity with structures for externally inputting, internally processing, and externally outputting information or matter. The formalization represents an intelligence set I, its exterior O, and a time mapping T_i, and states conditions on cardinality changes of I and O for input and output, and on internal subsets for processing. The framework is then applied to three examples — Hebbian non-optimized neural networks, backpropagation-optimized neural networks, and the Aplysia gill-withdrawal reflex — and is extended to a category-theoretic setting with a Time Category and an Intelligence Category. The paper also introduces the informal concept of "activity" and argues that the same axiomatization procedure could be applied to consciousness and emotion.
Significance. If the formalization worked, it would provide a transparent template for turning informal definitions into formal criteria, and it would allow concrete systems to be classified as intelligent or not in a uniform way. The paper has real strengths: it is explicit about the meta-method, it does not hide the choice of a naive definition, it avoids circular empirical fitting, and it illustrates the framework on three concrete systems. However, the load-bearing formal steps contain type errors and inconsistent cardinality conditions, and the three stated examples are not actually classified by the proposed inequalities as written. Because the central axiomatic criterion is therefore not well-defined and not faithful to its own examples, the current contribution is more a research proposal than a working formalization.
major comments (4)
- [Set-Theoretic Depiction of the Universe; Axiomatic Formalization of the Naïve Definition] The definition of a structure is ill-typed. For an existence E={ε1,...,εj}, the paper writes C_E ⊆ ε_j = ε_1 × ε_2 × ... × ε_j, and Def 2-4 repeat the same pattern with ν_l. The symbol ε_j is used both as an element of E and as the Cartesian product of all elements, and a product of elements is not defined unless each ε_i is itself a set. Consequently C_i, C_p, and C_o are not well-defined subsets, and the central expression I(C_i,C_p,C_o) lacks a precise mathematical meaning. Because every later condition is a condition on possessing these structures, the axiomatic framework is not well-founded as written.
- [Axiomatic Formalization of the Naïve Definition of Intelligence] The cardinality conditions for simultaneous input and output are internally inconsistent. Input is defined by |I_{i+1}|>|I_i| and |O_{i+1}|<|O_i|, while output is defined by |I_{i+1}|<|I_i| and |O_{i+1}|>|O_i|. The proposed patch with subsets R,S⊆I replaces the I-inequalities but still requires both |O_{i+1}|<|O_i| and |O_{i+1}|>|O_i| to hold for the same set O in the same time interval. These two inequalities cannot be satisfied by any cardinal number. The added assumption that the same element is not simultaneously input and output does not remove the contradiction, because the inequalities on |O| are aggregate conditions on the whole set O, not conditions on the transferred elements.
- [Formalization and Reinterpretation of Specific Intelligences (Tables 1 and 2)] The three examples are not classified by the formal definition as stated. In Table 1 the intelligence set I is specified as nodes, program code, computers, or cells, but the input/output conditions require |I| and |O| to change between time steps. In Table 2, inputting an MNIST image is described as a response in the input and intermediate layers, and output is described as a response in the output layer: these are changes in node activations or states, not changes in the number of nodes, cells, or code elements. The inequalities |I_{i+1}|>|I_i| and |I_{i+1}|<|I_i| therefore do not fire for any of the three examples unless transient data elements are silently added to I, which would contradict Table 1 and make the intelligence set depend on the particular input data. The paper itself also concedes in the section on naive definitions that axiomatic reformulation does not resolve the "pile of sand" counterexample, further weakening the claimed classification power.
- [Discussions: Extension to Category Theory] The categorical extension is not formally coherent as written. Objects are written as I=(C_i,C_p,C_o), using I both for an object of the category and for the intelligence set, and the direct product structure is inherited from the ill-typed definitions of C_i, C_p, and C_o. More seriously, the claimed composite functor H=F∘G is ill-typed: G is a functor from ℐBio to ℐAI, while F is a functor from the Time Category 𝒯 to an Intelligence Category ℐ. Composition F∘G requires the codomain of G to equal the domain of F, which is not stated and generally does not hold. The category section is illustrative, but as formal mathematics it does not yet establish the promised functorial representation of intelligence changes and mimicry.
minor comments (5)
- [Set-Theoretic Depiction of the Universe] The notation conflates set membership and subset inclusion: E, O, and I are often written as E∈U, O∈U, I∈U, although they are defined as subsets of U. Please use E⊆U, O⊆U, and I⊆U consistently.
- [Axiomatic Formalization of the Naïve Definition of Intelligence] The notation for C_I is inconsistent: Def 2-4 use C_i∈C_I and also C_I⊆ν_l, but if C_I is the set of all structures of I, it cannot simultaneously be a subset of a Cartesian product set. The paper should define whether C_I is a set of structures or a structure itself.
- [Set-Theoretic Depiction of the Universe] The operation C/C_p is not well-defined: in C/C_p = ∏_{b_i∈C,b_i∉C_p} b_i, the elements of the Cartesian product C are tuples, not factors, so the condition b_i∉C_p is ill-typed. Please define the removal operation on structures using a well-defined indexing set.
- [Tables 1 and 2] The time-step labels in Table 2 reuse T_i at every row; this makes it appear that the same map is applied at every stage. Use T_i, T_{i+1}, T_{i+2}, ... consistently to reflect the sequential stages.
- [On the Activity of Intelligence] The concept of "activity" is described only informally as the extent to which a structure engages in interactions; if activity is intended to be a quantitative measure, a formal definition with a domain and unit of measurement is needed.
Circularity Check
No significant circularity: the paper is a transparent definitional formalization, not a derivation.
full rationale
The paper explicitly states that it is axiomatizing a chosen naive definition of intelligence: 'Axiomatic intelligence is not defined directly through axioms but by axiomatizing a chosen naïve definition of intelligence.' The set-theoretic conditions for input, processing, and output are stipulated formalizations of the naive definition, using cardinality changes as a representational choice, and the paper openly acknowledges the arbitrariness of this choice ('the definitions and methodologies proposed here represent just one possible approach'). No parameters are fitted, no empirical predictions are made, and no uniqueness theorem is imported from prior work. The self-citations to Itoh (2024a, 2024b) are used as a motivational example and as one of three case studies, but the axiomatic framework does not depend on those results; the classification of the three examples follows directly from the stipulated definitions plus the paper's own description of each system. Consequently, there is no circular step: the derivation chain does not reduce to its own inputs beyond the explicitly acknowledged act of formalizing a chosen definition.
Assumptions & free parameters
assumptions (6)
- domain assumption The universe is a finite set U of all elements, and time is a cardinality-preserving mapping T_i: U_i -> U_{i+1}.
- domain assumption Existence is any subset E of U, and the exterior is O = U \ E.
- ad hoc to paper A structure C_E is a subset of the Cartesian product of the elements of E.
- ad hoc to paper Input and output correspond to strict changes in the cardinalities of I and O, and processing to internal subset rebalancing.
- ad hoc to paper The same element of I or O is never both input and output within one time interval.
- standard math Category theory axioms, identity and composition preservation, hold for the time and intelligence functors.
invented entities (1)
-
Activity
Cite this review
Pith. "Pith review of Toward the Axiomatization of Intelligence: Structure, Time, and Existence." pith.science (2026). https://pith.science/paper/A6MUVHLY
@misc{pith2026250414596,
author = {Pith},
title = {Pith review of: Toward the Axiomatization of Intelligence: Structure, Time, and Existence},
year = {2026},
howpublished = {\url{https://pith.science/paper/A6MUVHLY}},
note = {Machine review of arXiv:2504.14596}
}
read the original abstract
This study aims to construct an axiomatic definition of intelligence within a meta-framework that defines the method of definition, addressing intelligence as an inherently naive and polysemous concept. Initially, we formalize a set-theoretic representation of the universe as the domain wherein intelligence exists and characterize intelligence as a structure that involves temporal evolution and interaction with other sets. Starting from a naive definition of intelligence as "an entity possessing structures for externally inputting, internally processing, and externally outputting information or matter," we axiomatically reformulate it within this set-theoretical depiction of the universe. Applying this axiomatic definition, we compare and interpret three examples -- Hebbian non-optimized neural networks (NNs), backpropagation-optimized NNs, and biological reflexive systems -- in terms of their intelligence, structural properties, and biological plausibility. Furthermore, by extending our definition into a categorical framework, we introduce two categories, "Time Category" and "Intelligence Category," along with the functorial relationships between them, demonstrating the potential to represent changes and mimicry relationships among intelligent systems abstractly. Additionally, since intelligence, as defined herein, functions effectively only when accompanied by temporal interactions, we introduce the concept of "activity" and explore how activity-based conditions influence classifications and interpretations of intelligence. Finally, we suggest that our definitional methodology is not limited to intelligence alone, but can be similarly applied to other concepts, such as consciousness and emotion, advocating for their formal reinterpretation through the same procedural steps: defining a universal representation, selecting naive definitions, and axiomatic formalization.
Reference graph
Works this paper leans on
-
[23]
知能“を持つ。 また、 人間以外の生物やコンピュータープログラムで構 成された人工知能、 果てはより単純な道具にもその内に知能が宿っているか、 もしくはそう 見える。 この
Wittgenstein, L. Tractatus Logico-Philosophicus. Routledge & Kegan Paul (1922). Proposition 7. Toward the Axiomatization of Intelligence: Structure, Time, and Existence 伊藤 慧* April 20, 2025 Abstract 本研究は、 知能という本質的に素朴かつ多義的な概念に対して、 その定義方法自体を定義す るというメタ的枠組みのもと、 公理的定義を構築することを目的とする。 まず、 知能が存在 する場としての宇宙を集合論的に形式化し、その中で時間発展や他集合との相互作用を考 慮した構造として知能を記述する。次に、”情報や物質を、外...
work page 1922
-
[24]
情報や物質を、外部から入力される構造、処理する構造、そして外部へ出力する構造のそ れぞれを有する存在
集合論的宇宙描像に基づく素朴的知能定義の公理的定義化 2 は前述の通り”知能”が元来素朴的概念であることから必要な過程である。公理的知能は公 理的知能定義から直接定義されるのではなく、ある素朴的知能定義を公理的定義化するこ とによって定義される。３では定義された公理的知能を基に具体的な知能を再解釈するこ とで知能の持つ能力の解釈の例をも示す。 具体的にはItoh(2024b)[13]によるヘブ学習型非最 適化 NNs と従来的な逆伝搬型最適化 NNs[e.g., 19]、そしてアメフラシの鰓開閉機構[e.g., 20] の 3 つの”知能”に対し本研究による公理的知能定義を適用する。 そしてそれらを比較するこ とによりそれぞれの知能性やその差異について解釈を行う。最後に本研究における知能定 義の任意性を集...
work page 2024
- [25]
- [26]
-
[27]
The Measurement of Adult Intelligence, 3rd ed
Wechsler, D. The Measurement of Adult Intelligence, 3rd ed. Williams & Wilkins, 1944
work page 1944
-
[28]
Turing, A. M. Computing machinery and intelligence. Mind, 59(236), 433–460 (1950)
work page 1950
-
[29]
What is Life? The Physical Aspect of the Living Cell
Schrödinger, E. What is Life? The Physical Aspect of the Living Cell. Cambridge University Press, 1944
work page 1944
-
[30]
A collection of definitions of intelligence
Legg, Shane, and Marcus Hutter. “A collection of definitions of intelligence. ” Frontiers in Artificial Intelligence and applications 157 (2007): 17
work page 2007
Show all 25 references
-
[31]
The law of accelerating returns
Kurzweil, R. The law of accelerating returns. In Alan Turing: Life and Legacy of a Great Thinker, pp. 381–416. Springer (2004). https://doi.org/10.1007/978-3-662-05642-4_16
2004 doi
-
[32]
Brown, T. B. et al. Language models are few -shot learners. arXiv preprint arXiv:2005.14165 (2020). https://arxiv.org/abs/2005.14165
2020 arXiv
-
[33]
Kaplan, J. et al. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020). https://arxiv.org/abs/2001.08361
2020 arXiv
-
[34]
Reward is enough
Silver, D., Singh, S., Precup, D., & Sutton, R. Reward is enough. Artificial Intelligence, 299, 103535 (2021). https://doi.org/10.1016/j.artint.2021.103535
2021
-
[35]
Inverse scaling: When bigger isn’t better
McKenzie, Ian R., et al. “Inverse scaling: When bigger isn’t better.” arXiv preprint arXiv:2306.09479 (2023). https://arxiv.org/abs/2306.09479
2023 arXiv
-
[36]
Rethinking Deep Learning: Propagating Information in Neural Networks without Backpropagation and Statistical Optimization
Itoh, K. Rethinking Deep Learning: Propagating Information in Neural Networks without Backpropagation and Statistical Optimization. arXiv preprint arXiv:2409.03760 (2024). https://arxiv.org/abs/2409.03760
2024 arXiv
-
[37]
Rethinking Deep Learning: Non -backpropagation and Non -optimization Machine Learning Approach Using Hebbian Neural Networks
Itoh, K. Rethinking Deep Learning: Non -backpropagation and Non -optimization Machine Learning Approach Using Hebbian Neural Networks. arXiv preprint arXiv:2411.05861 (2024). https://arxiv.org/abs/2411.05861
2024 arXiv
-
[38]
S., & Pitts, W
McCulloch, W. S., & Pitts, W. H. (1943). A logical calculus of the ideas immanent in nervous activity. The Bulletin of Mathematical Biophysics, 5(4), 115-133
1943
-
[39]
Rosenblatt, F. (1958). The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review, 65(6), 386-408
1958
-
[40]
Hebb, D. O. The Organization of Behavior: A Neuropsychological Theory. Wiley, 1949
1949
-
[41]
Large-scale machine learning with stochastic gradient descent
Bottou, Léon. “Large-scale machine learning with stochastic gradient descent. ” Proceedings of COMPSTAT’2010: 19th International Conference on Computational Statistics Paris France, August 22-27, 2010 Keynote, Invited and Contributed Papers. Physica-Verlag HD, 2010. https://do...
2010 doi
-
[42]
E., Hinton, G
Rumelhart, D. E., Hinton, G. E., & Williams, R. J. Learning representations by back-propagating errors. Nature, 323, 533–536 (1986). https://doi.org/10.1038/323533a0
1986 doi
-
[43]
Krizhevsky, A., Sutskever, I., & Hinton, G. E. ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, 1097–1105 (2012)
2012
-
[44]
Kandel, E. R. Cellular Basis of Behavior: An Introduction to Behavioral Neurobiology . W. H. Freeman, 1976
1976
-
[45]
LeCun, Y ., Cortes, C., & Burges, C. J. C. (1998). The MNIST database of handwritten digits. http://yann.lecun.com/exdb/mnist
1998
-
[46]
Critique of Pure Reason
Kant, I. Critique of Pure Reason. (1781/1787). A30/B45
-
[47]
Tractatus Logico-Philosophicus
Wittgenstein, L. Tractatus Logico-Philosophicus. Routledge & Kegan Paul (1922). Proposition 7
1922
Reviewed August 16, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.