Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Understanding Human Intelligence through Human Limitations

T0 review · 4 major / 5 minor · reviewed 2026-08-27 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Human intelligence is best understood not as a general-purpose thinking engine but as a family of solutions to problems constrained by limited time, limited computation, and limited communication.

desk verdict A well-written perspective that overclaims 'derive' where it actually offers a useful synthesis—worth peer review with revision. read the letter →

arxiv 2009.14050 v1 pith:DYXBMHXT submitted 2020-09-29 cs.AI cs.LG

Thomas L. Griffiths This is my paper · ORCID
classification cs.AIcs.LG
keywords artificialintelligencehumaninductivebiasmetalearningrationalmetareasoningculturalevolutionlimitedcomputationcommunication
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the right way to see what is unique about human intelligence is to look at the computational problems human minds must solve, not at the machinery of the brain. It claims those problems share three constraints: not enough time to gather unlimited data, a fixed and finite amount of computation, and no direct way to copy knowledge between brains. From these three constraints, the paper derives classic human traits: rapid learning from few examples, breaking problems into subgoals, and cumulative cultural evolution. If true, this reframes human intelligence as one specific flavor of intelligence, and explains why AI systems built without these constraints would not be expected to share those traits.

What carries the argument

The central object is the constraint-defined problem class: the intersection of limited time, limited computation, and limited communication that makes human computational problems a specific subset of all possible computational problems. Each limitation is tied to a formal tool: limited time to Bayesian inference and metalearning (learning to learn, i.e., optimizing an inductive bias across tasks), limited computation to rational metareasoning (deciding which computations to execute given their cost), and limited communication to parallel and distributed algorithms, including a particle filter that represents a posterior distribution over hypotheses by weighted samples and passes those weights across generations. These tools carry the argument by showing that ideal solutions under each constraint look like the human traits the paper explains.

What would settle it

Give two otherwise identical reinforcement-learning agents the same open-ended task, one with effectively unlimited compute and copyable parameters and one under human-like limits of data, compute, and communication. If the unlimited agent nevertheless develops the same rapid-learning, subgoal-based, teaching-like behavior as the limited agent, then those traits are not consequences of the constraints; if it does not, the framework's prediction is confirmed.

Watch

Extended reading notes

Core claim

The central claim is that many properties normally listed as hallmarks of human intelligence are not intrinsic to intelligence but are consequences of the class of problems humans actually face. Limited time forces learning from small data, making inductive bias—and with it Bayesian inference and metalearning, or learning an inductive bias across tasks—the right way to understand rapid learning. Limited computation makes subgoal decomposition and reusable problem-solving strategies rational, formalized by rational metareasoning (deciding which computations to run given their cost). Limited communication makes teaching, language, and cumulative cultural evolution necessary, and these can be modeled with parallel and distributed algorithms, for instance a particle filter that passes weighted beliefs from one generation to the next. Because artificial systems can, in principle, escape all three constraints, the paper predicts that their intelligence will differ in kind, not just degree, from the human flavor.

Load-bearing premise

The whole chain breaks if human cognitive traits are not actually adaptations shaped by the three limitations, but merely happen to coexist with them; the paper explicitly rests on the assumption that human minds are, to some extent, optimized for the problems those limits define.

Editorial extensions

If this is right

  • AI systems trained with effectively unlimited data, computation, or direct parameter transfer should not be expected to show human-like rapid learning, subgoal decomposition, or cumulative culture; their absence is not a sign of missing intelligence.
  • Machines that face all three human constraints at once—autonomous, real-time, general-purpose systems such as space probes—should be expected to develop human-like cognitive features.
  • Cognitive science should treat Bayesian inference, metalearning, rational metareasoning, and parallel algorithms as core formal tools rather than optional extras.
  • The explore–exploit balance in human learning should shift with lifespan, since limited time makes the value of information depend on how much future use remains.
  • The "bitter lesson" of scaling data and compute in AI is not a general law of intelligence but a consequence of being free from human constraints.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If human traits are adaptations to these three constraints, then the same traits should emerge in any sufficiently capable agent placed under the same constraints, which is a testable prediction for artificial agents.
  • The framework suggests a constructive recipe for human-like AI: instead of imitating human behavior directly, impose the constraint profile—fixed data budget, fixed compute budget, no weight copying—and let the system discover the traits.
  • The communication limitation also interacts with the other two: because it blocks the obvious escape from them, it forces intelligence to become collective, meaning institutions and scientific communities can be analyzed as distributed inference algorithms rather than as collections of individual minds.
  • The argument likely implies that human cognitive biases are not bugs but features that should be predictable as optimal solutions under the three limits, making the framework generative rather than merely descriptive.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper argues that human intelligence can be understood by characterizing the computational problems humans face, which the author claims share three fundamental constraints: limited time, limited computation, and limited communication. From these limitations, the paper claims to derive signature human properties: rapid learning from limited data, decomposition of problems into subgoals, and cumulative cultural evolution. It further proposes corresponding mathematical toolkits—Bayesian inference and metalearning for limited time, rational metareasoning for limited computation, and parallel/distributed algorithms for limited communication—and discusses implications for when AI systems should be expected to resemble human cognition. The paper is a programmatic perspective piece with four boxes giving concrete analogies and formal fragments, rather than a fully worked derivation.

Significance. If the central claim were established, the paper would offer a useful unifying perspective on human uniqueness in the age of AI: human intelligence would be characterized as a family of solutions to a constrained class of computational problems, with concrete predictions about when AI systems should or should not share human cognitive traits. The paper is clearly written, draws on a broad and relevant literature, and the boxed examples—especially the metalearning/evolution analogy and the particle-filter model of cumulative cultural evolution—provide concrete, potentially testable formal anchors. As a proposal for reorienting the comparison between human and machine intelligence, the paper is valuable. However, the current version substantially overclaims: the advertised derivation is not supplied, and the argument relies on an unstated adaptation premise with little empirical or formal support. The significance is therefore conditional on either supplying the missing derivation or softening the claim to a motivated hypothesis.

major comments (4)
  1. [Abstract and 'Different Computational Problems, Different Kinds of Intelligence'] The central claim that the three limitations allow us to 'derive' rapid learning, subgoal decomposition, and cumulative cultural evolution is not supported by the argument that follows. Each section offers narrative associations and examples, and Box 3 and Box 4 give formal expressions for rational metareasoning and Bayesian particle filtering, but no step shows that these properties logically follow from the constraints. The Concluding Remarks contain the actual load-bearing statement: 'To the extent that human minds are adapted to solve such problems, these limitations potentially provide insight.' That is a hedging qualification, not a derivation. Please either supply the missing inference—for example, formal results showing that optimal solutions under a specified resource model yield these traits—or revise the wording from 'derive' to 'motivate' or 'hypothesize.'
  2. [Concluding Remarks] The argument depends on an adaptation premise that is acknowledged but never defended. The chain from limitation to property holds only if evolution or developmental processes optimized human cognition around these three constraints. The paper does not provide evidence for that adaptation claim, nor does it specify a criterion for identifying a trait as the adapted solution to a constrained problem. As written, each limitation is paired with one signature property, so the framework risks being unfalsifiable: any observed trait could be re-described as a consequence of some limitation. A concrete test—for instance, comparing species or simulated agents under systematically varied time, computation, and communication constraints—would make the causal claim empirically tractable. Without such a test or a formal optimality argument, the explanatory direction from limitation to property is not established.
  3. [Limitation 2: Computation] Limited computation does not entail subgoal decomposition. The paper's own Rubik's cube example shows that a bounded agent can instead use memorized algorithms or a lookup table, and it notes that the shortest-path solution 'doesn't have a huge amount of underlying structure.' Conversely, an unbounded agent could still adopt subgoals for modularity or reuse. Subgoaling is one possible resource-bounded strategy among many, so the connection to the constraint is underdetermined. To support the claimed derivation, the paper would need a formal argument from a specific model of computation costs—for example, a rational metareasoning problem with defined computation and error costs—showing that subgoal decomposition is the optimal strategy, not merely an illustrative one.
  4. [Limitation 3: Communication] Limited communication does not entail cumulative cultural evolution. Most species with limited communication do not exhibit cumulative culture, and even in humans the mechanisms described—teaching, writing, scientific institutions—are additional adaptations that are not derived from the absence of brain-to-brain transfer. The text states that limited communication 'results in' these mechanisms, but at most the constraint enables them given other supporting conditions. As with the other limitations, a selection or optimization argument is needed to connect the constraint to the observed property. Without that, the claim that AI systems without the communication limitation should omit cultural traits is not warranted by the argument presented.
minor comments (5)
  1. [Concluding Remarks] The text refers to 'Outstanding Questions' in parentheses, but no such section appears in the manuscript; either include it or remove the reference.
  2. [Throughout] There are several typos and formatting issues: 'Proceeedings' in references 17, 20, and 44; 'expresed' and 'he hypotheses' in Box 4; 'a least one popular approach' in Box 2; and 'that actions taken by these systems' missing a verb in the Computation section. These should be corrected in a final pass.
  3. [Box 2, Eq. (1)] In Equation (1), the left-hand side is written as L(θ), but the right-hand side is a sum over task losses L_t(φ_t) where φ_t depends on θ. Please state explicitly that this is the meta-objective being minimized over θ, and define the task distribution or set of tasks.
  4. [Box 3] The expected-utility and metareasoning objectives are written in inline prose rather than as displayed equations. Displaying them with all quantities (p(x|a), u(x), p_c(x|a), cost(c)) defined would substantially improve clarity.
  5. [Figure 1] The caption lists three limitations, but the schematic as described in the text appears to label only Time and Communication axes. Please ensure the figure actually depicts all three constraints or adjust the caption to match what is shown.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the limitations-to-properties argument is conditional and underdetermined, but no derived property reduces to its input by construction.

full rationale

The paper's central claim is a conditional research proposal rather than a formal derivation: the three limitations define a problem class, and the properties (rapid learning, subgoaling, cumulative culture) are suggested as likely solutions for agents adapted to that class. The Concluding Remarks explicitly condition the whole argument on adaptation: "To the extent that human minds are adapted to solve such problems, these limitations potentially provide insight into the nature of human intelligence." The one-to-one pairing of each limitation with a signature property is a rhetorical organization, not an equation or fitted parameter, and the paper does not define the limitations in terms of the properties. No load-bearing premise rests on a self-citation: the rational metareasoning and particle-filter discussions cite prior work (including the author's own) only as reviews or as evidence for the formal tools, not as the source of the claimed entailment. Underdetermination—for example, bounded computation does not logically force subgoal decomposition, and limited communication does not force cumulative culture—is a real gap in the argument, but underdetermination is a correctness and evidential weakness, not circularity. Therefore no step satisfies the standard of exhibiting a reduction by construction, and the circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim depends entirely on three posited limitations and on an adaptationist premise. The limitations are not derived or measured; they are introduced at the outset and paired with target properties, which makes the framework's explanatory power partly a function of its own definitions. The formal toolboxes are imported from prior work rather than proven. No numerical parameters are fitted, but the three qualitative limitations function as chosen degrees of freedom.

free parameters (3)
  • Limited time (lifespan and survival constraints) = not quantified
    Chosen by the author as the first constraint; the amount of available data is presented as fixed but no scale is derived, and it is used to explain rapid learning.
  • Limited computation (fixed brain capacity) = not quantified
    Assumed, not measured; used to explain subgoal decomposition and resource-rational behavior.
  • Limited communication (no brain-to-brain transfer) = not quantified
    Assumed, not measured; used to explain teaching, language, and cumulative culture.
assumptions (3)
  • domain assumption The three limitations are exhaustive and independent characteristics of human computational problems.
    Section 'Different Computational Problems, Different Kinds of Intelligence' lists them as 'three important characteristics' without argument that no fourth constraint (such as social motivation or embodiment) plays an equal role.
  • domain assumption Human minds are adapted by evolution or development to solve the computational problems they face.
    The Concluding Remarks condition the entire argument on this premise: 'To the extent that human minds are adapted to solve such problems, these limitations potentially provide insight.' If adaptation is false, the derivation fails.
  • domain assumption Bayesian inference, rational metareasoning, and parallel or distributed algorithms characterize the ideal solutions under each limitation.
    The sections on each limitation assert these correspondences; they are research programs, not proven consequences of the limits.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Understanding Human Intelligence through Human Limitations." pith.science (2026). https://pith.science/paper/DYXBMHXT

@misc{pith2026200914050,
  author       = {Pith},
  title        = {Pith review of: Understanding Human Intelligence through Human Limitations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DYXBMHXT}},
  note         = {Machine review of arXiv:2009.14050}
}
read the original abstract

Recent progress in artificial intelligence provides the opportunity to ask the question of what is unique about human intelligence, but with a new comparison class. I argue that we can understand human intelligence, and the ways in which it may differ from artificial intelligence, by considering the characteristics of the kind of computational problems that human minds have to solve. I claim that these problems acquire their structure from three fundamental limitations that apply to human beings: limited time, limited computation, and limited communication. From these limitations we can derive many of the properties we associate with human intelligence, such as rapid learning, the ability to break down problems into parts, and the capacity for cumulative cultural evolution.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Optimizers Qualitatively Alter Solutions And We Should Leverage This

    cs.LG 2025-07 conditional novelty 4.0 of 10

    Deep learning optimizers should be designed to induce desired solution properties, not just convergence speed; different optimizers demonstrably land in qualitatively different minima.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages · cited by 1 Pith paper

  1. [1]

    1 Amodei, D., & Hernandez, D. (2018). AI and compute. http://blog.openai.com/ai-and-compute/. 2 Anderson, J. R. (1990). The adaptive character of thought. Hillsdale, NJ: Erlbaum. 3 Anderson, J. R. (1991). The adaptive nature of human categorization. Psychological Review, 98(3), 409–429. 4 Andrychowicz, M., Denil, M., Gomez, S., Homan, M. W., Pfau, D., Sch...

    work page Pith review arXiv 2018

  2. [59]

    RL$^2$: Fast Reinforcement Learning via Slow Reinforcement Learning

    16 Duan, Y., Schulman, J., Chen, X., Bartlett, P. L., Sutskever, I., & Abbeel, P. (2016). RL2: Fast reinforcement learning via slow reinforcement learning. arXiv preprint arXiv:1611.02779. 17 Finn, C., Abbeel, P., & Levine, S. (2017). Model-agnostic meta-learning for fast adaptation of deep networks. In Proceeedings of the 34th international conference on...

    work page Pith review arXiv 2016

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.