Source-linked AI summary
Why We Care About Understanding: Competence through Predictive Compression
Matthieu Queloz, Pierre Beckmann
TL;DR
The paper asks how understanding relates to compression and why human understanding is so heavily compressed. It develops the CPC framework, linking robust competence to predictive mental models and social pressures for demonstrability and transmission. It concludes that compression is understanding’s representational shadow, while predictive accuracy alone can produce opaque machine competence.
Problem
Research variously identifies understanding with compression or characterizes it through connections, explanations, and novelty, leaving their relation unresolved.
Method
The paper develops the CPC framework through theses about understanding as a proxy for robust competence, mental models enabling prediction, and social pressures shaping compressed understanding.
Results
The paper concludes that compression is not identical with comprehension but is its representational shadow, and that predictive success alone does not establish understanding.
Takeaways & Limitations
Human understanding is socially disciplined predictive compression that can be displayed, challenged, and taught, whereas accuracy-dominated systems may develop opaque understanding.
Takeaways & Limitations
The paper does not depend on the stronger claim that shared understanding is prior to individual understanding.
Abstract
from arXiv · showhide
What is the relation between understanding and compression, and why does human understanding take such a heavily compressed form? Across information theory, machine learning, and AI research, a substantial tradition identifies understanding with compression-a thought captured in Gregory Chaitin's dictum that "comprehension is compression." Philosophers, by contrast, have characterized understanding in terms of grasping connections, giving explanations, and handling novelty. This paper bridges the two pictures through three interlocking theses. The first concerns the concept of understanding: it serves as an efficient proxy for a distinctive form of robust competence, enabling us to identify whom to trust and whom to learn from. The second concerns the state of understanding: to understand a domain is to possess a mental model of its relational structure that enables prediction, and what enables prediction enables compression, because what becomes predictable need not be stored separately. Compression is therefore not identical with comprehension, but its representational shadow. The third concerns the characteristically human form of understanding: the fiduciary and transmission functions highlighted by the first thesis impose pressures of demonstrability and transmissibility that drive human understanding toward principled simplicity. The resulting framework explains both the appeal and the limits of compressionist accounts of understanding while shedding light on the inscrutability of AI systems.
1 Introduction
The paper integrates philosophical and computational accounts by treating compression as the consequence of predictive mental models rather than as identical with understanding. It argues that social demands for trust, demonstration, and transmission drive human understanding toward principled simplicity.
- The problem: Compressionist research often identifies understanding with compression, while philosophical accounts emphasize connections, explanations, and novelty.This asymmetry motivates an integrated account spanning information theory, machine learning, AI, and philosophy.
- The Proxy Hypothesis: The paper treats understanding as an efficient proxy for robust competence, helping people identify whom to trust and learn from before competence can be directly observed.The proxy tracks a ground of competence that can be probed more efficiently than competence itself.
- Mental models and compression: Understanding consists in a mental model of relational structure that enables prediction, making the domain less surprising and thereby supporting compression.Predictable information need not be stored separately, but compression remains a representational consequence rather than an identity with comprehension.
- Sociality and principled simplicity: The paper argues that human understanding becomes highly compressed because social practices require it to be demonstrable and transmissible.These pressures favor compact principles or systems that can be displayed in conversation and passed through communication bottlenecks.
- Concept and state: The concept and cognitive state of understanding form a co-evolutionary feedback loop: attainable cognitive states shape the concept, which in turn shapes cognitive aspirations and pedagogy.The paper notes that its argument does not depend on the stronger view that shared understanding is prior to individual understanding.
- Contribution: The framework connects philosophical and computational traditions while using social demands to explain why human understanding differs from predictive competence unconstrained by communication.The paper organizes the account through the Proxy Hypothesis, predictive compression, objections, and implications for machine understanding.
2 The Functions of the Concept of Understanding
Understanding functions as a proxy for robust competence: it helps identify whom to trust and learn from, while tracking a cognitive organization that supports prediction beyond familiar cases.
- Understanding helps identify people whose competence is likely to survive unfamiliar and difficult conditions, not merely routine performance.The concept supports warranted deference by distinguishing robust competence from competence that collapses outside familiar conditions.
- Understanding also identifies effective teachers because principled explanations can transmit understanding and robust competence between agents.This inference is defeasible, but explanations can transfer the principles exhibited by the explainer to the learner.
- The concept’s fiduciary and transmission functions arise from the shared social need to locate robust competence for reliance and acquisition.These functions also shape the cognitive state of understanding through practical social pressures.
- The Proxy Hypothesis holds that understanding efficiently tracks robust competence because exhaustive direct assessment is infeasible and routine track records poorly predict novelty.Understanding can be probed through summaries, explanations, counterfactuals, and particular applications, complementing past-performance evidence.
- Understanding is a cognitive property distinct from competence, consisting in a structure-sensitive mental model that supports prediction, explanation, and counterfactual reasoning.Such models preserve enough relational structure to generate insights and extrapolate beyond the cases through which they were acquired.
- Mental models ground one form of robust competence, but robust competence is polygenic and may instead rely on strength, reflexes, muscle memory, or other properties.Kepler’s astronomical competence contrasts with Achilles’s primarily bodily and motor-grounded competence.
3 How Sociality Compresses Comprehension
Human understanding is shaped by four pressures: predictivity, cognitive manageability, demonstrability, and transmissibility. Together, these pressures favor predictive compression—models that capture relational structure while remaining economical, communicable, and useful for robust competence.
- Four pressures: Predictivity favors mental models that anticipate outcomes, variations, and effective interventions by making domains less surprising.Structure-sensitive models support successful action, diagnosis, explanation, and intervention.
- Four pressures: Finite cognition favors models that remain storable and manipulable rather than becoming too complex to retain or use.This pressure limits complexity even when additional complexity might improve accuracy.
- Four pressures: Demonstrability pressures understanding toward explicit, coherent principles that can efficiently signal competence and justify trust.Principled organization enables concise explanation, while fewer and more general principles yield more elegant displays.
- Four pressures: Transmissibility pressures understanding toward models that can pass through communication and reproduce competence in learners.Teaching favors distilling domains into basic principles arranged in coherent, memorable systems.
- Predictive compression: These pressures converge on predictive compression: models that reduce surprise by capturing relational structure while remaining economical in representation.The CPC framework treats this organization as the functional core through which understanding grounds robust competence.
- Predictive compression: Compression is not identical with understanding: structure-sensitive models enable data compression, while model compression concerns the model’s own description length.The model encodes reusable structure, leaving residuals as the data description; compression can therefore reflect understanding without constituting it.
- Limits: Unchecked predictive accuracy can produce opaque understanding when large models lack storability, demonstrability, and transmissibility.The paper places such systems in the high-predictivity, low-simplicity region of its conceptual map.
4 Objections and Replies
The objections distinguish compression from comprehension by asking what structure compression captures and what task it serves. The replies extend the framework to formal logic, tacit practical wisdom, and graded forms of principled understanding.
- Compression Without Comprehension?: A .zip compressor detects repetitions and symbol frequencies, exhibiting limited structure-sensitivity without thereby understanding what the compressed text is about.The reply treats file compression as predictive in a modest sense because it forms expectations about what is likely to come next.
- Compression Without Comprehension?: Compression alone underdetermines what is understood: it may capture structure in the target domain or only regularities in its representation.Kepler’s laws make planetary motion intelligible, whereas shorthand compresses Brahe’s records by exploiting notation.
- Compression Without Comprehension?: The same distinction applies to LLMs: their compression may reflect shallow corpus regularities, deeper linguistic structure, or the relational structure of the target domain.The question is therefore not whether LLMs merely compress, but what kind of understanding their compression reflects.
- Compression Without Comprehension?: Compression can reveal deep understanding of a representational scheme: the Sheffer stroke compresses logic’s operators and axioms, though it is poor for ordinary reasoning.Its epistemic value depends on the target domain and task family: it is false economy for practical reasoning but a metalogical insight about logic’s structure.
- Tacit Understanding and Graded Compression: Tacit practical understanding also involves compression, because compressed representations help agents recognize salient features, anticipate outcomes, and respond appropriately in novel cases.The framework’s notion of prediction includes reducing surprise about relevant considerations and effective interventions, not only forecasting explicit data.
- Tacit Understanding and Graded Compression: Human understanding is not always elegantly principled; it tends toward principled simplicity where pressures for predictivity, manipulability, demonstrability, and transmissibility are strongest.The framework accommodates both tacit cognition and the social demand to teach, assess, coordinate, and justify understanding.
5 Conclusion
The conclusion presents understanding as a socially useful proxy for robust competence and asks how predictive systems achieve their generalization. It argues that compression is informative only when we identify the structure it captures, while highly predictive systems may remain opaque when demonstrability and transmission are absent.
- Conclusion: Understanding functions as an efficient proxy for robust competence, helping people identify whom to trust and learn from.The framework derives pressures toward predictivity, storability-cum-manipulability, demonstrability, and transmissibility.
- Conclusion: AI understanding cannot be inferred from predictive success or generalization alone; the relevant question is what structure the system’s compression reflects.Candidate structures include shallow corpus regularities, deeper linguistic structure, and the relational structure of the target domain.
- Conclusion: When predictive accuracy dominates, systems may exhibit opaque understanding: high predictivity coupled with complexity that cannot be readily communicated.The paper contrasts this with human understanding as socially disciplined predictive compression that can be displayed, challenged, and taught.
List of Works Cited
The works-cited section assembles references spanning epistemology, philosophy of science, information theory, machine learning, artificial intelligence, and theories of understanding.
- Works Cited: The bibliography includes philosophical work on understanding, explanation, knowledge, reason, practical wisdom, and social epistemology.Examples include works by Baumberger, De Regt, Grimm, Hannon, Kvanvig, and Khalifa.
- Works Cited: The list further covers AI and language-model research on predictive compression, explainability, representation geometry, and understanding in large language models.Cited works include Delétang et al., Mitchell and Krakauer, Li et al., and related AI research.
- Works Cited: It also cites information-theoretic and compression-oriented research, including minimum description length, algorithmic information, and compressionism.References include Grünwald, Kolmogorov, Chaitin, Maguire and colleagues, and Li and colleagues.