A Meta-Methodological and Ontological Analysis of Intelligence: Between Sutton’s “Bitter Lesson,” Park’s “Sculpting,” and the Nominalist Critique of Univocity

At the intersection of contemporary artificial intelligence research and neuroscience, a fundamental tension emerges concerning both the nature of intelligence and the appropriate methodology for its investigation. This tension cannot be reduced to a merely technical disagreement. Rather, it develops into an ontological problem regarding whether a single concept—namely, “intelligence”—can be legitimately applied to entities with radically different structures of being, such as human cognition and artificial systems.

This problem is sharply articulated through the contrast between Richard Sutton’s “The Bitter Lesson” and Il Memming Park’s “Sculpting” analogy. Sutton’s position is grounded in a historical observation about artificial intelligence research: approaches that rely on embedding human knowledge and domain-specific structure into models are consistently outperformed, in the long run, by methods that exploit large-scale computation and general learning procedures. From this perspective, artificial intelligence should not be constrained by human interpretability or intuition. Instead, it should be understood as a computational system whose primary capacities lie in search and learning over vast, complex data spaces. The implication is that attempts to render such systems transparent in human terms may be both epistemically limiting and methodologically counterproductive. Sutton’s position thus implicitly accepts a form of epistemic opacity, in which the internal operations of artificial systems may not be fully accessible to human understanding.

In contrast, Il Memming Park’s “Sculpting” analogy foregrounds the importance of explanatory understanding. According to this view, an artificial model is not simply a device that produces outputs but an object shaped under specific constraints, much like a sculpture emerging from a block of material. The critical issue is not only what the model does, but why it does so. Predictive success alone does not constitute scientific understanding; rather, understanding requires access to the causal and structural mechanisms that generate observable behavior. Park’s position therefore emphasizes the necessity of interpretability and advocates for methodological pluralism, in which different types of models are employed to balance performance and explanation. In this framework, artificial intelligence is not merely an autonomous computational system but an object of scientific inquiry that must be rendered intelligible within a shared conceptual framework.

The opposition between these two positions is not merely methodological but ontological. It raises the question of how concepts are applied across distinct domains of being. This issue can be clarified through the medieval debate on the univocity of being. Within one tradition, univocity implies that a single concept—such as “being”—corresponds to a real common nature shared by different entities. Within another, however, univocity is understood not as a feature of reality but as a feature of thought and language: a unifying function of mental or linguistic signs applied to distinct individuals.

From the latter perspective, only individuals exist, and universal concepts do not correspond to real shared essences but function as mental signs that refer to multiple entities. Applied to the problem of intelligence, this implies that the use of a single term for both human and artificial systems does not indicate an underlying ontological unity. Rather, it reflects an operation of abstraction performed by human cognition, which groups heterogeneous entities under a single conceptual category based on perceived similarities.

This leads to what may be termed conceptual entrapment. The use of a common term encourages the implicit assumption that the entities it designates share a common nature. However, human cognition and artificial systems differ fundamentally in their physical substrates, operational mechanisms, and modes of existence. When terms such as “thinking” or “understanding” are applied to artificial systems, they carry with them a network of meanings rooted in human experience. As a result, the statistical and computational processes that characterize artificial intelligence are interpreted through an anthropocentric conceptual framework, thereby obscuring their distinctive structure.

This limitation is further illuminated by the theory of mental language, according to which human cognition operates through a system of natural signs formed in causal relation to the world. These signs do not replicate the essence of their objects but function as substitutes that refer to them. Consequently, concepts capture only certain aspects of their objects and cannot exhaust their full reality. The application of such concepts across fundamentally different domains therefore risks distorting rather than clarifying the phenomena in question.

Within this framework, artificial intelligence is best understood not as a bearer of concepts but as a system that processes statistical relations among signs. It does not grasp meanings or categories in an intuitive sense; rather, it identifies and exploits patterns within data. Even if a shared language is constructed to describe both human and artificial cognition, that language remains a product of human mental structures and does not fully disclose the internal dynamics of artificial systems. The attempt to establish conceptual continuity thus introduces a further layer of mediation, which may function as a constraint rather than an expansion of understanding.

To address this limitation, a dynamic approach to conceptualization is required. Concepts must not be treated as fixed definitions but as provisional tools subject to continual revision. The analysis of artificial systems should proceed through an iterative process in which observed behaviors prompt the modification or replacement of existing conceptual frameworks. Knowledge, in this sense, does not accumulate linearly but develops through cycles of refinement and reconstruction.

At the same time, caution must be exercised in the use of analogy as a mode of explanation. While analogical reasoning can facilitate initial understanding, it also introduces the risk of epistemic inconsistency, unwarranted ontological commitments, and category mistakes. These errors arise when properties belonging to one domain are inappropriately attributed to another, particularly when artificial systems are assimilated to human cognitive categories without sufficient justification.

In conclusion, the concept of “intelligence” does not secure an ontological unity between human cognition and artificial systems. It functions instead as a conceptual instrument generated within human thought, enabling the comparison of heterogeneous entities while simultaneously obscuring their differences. The task, therefore, is not to stabilize this concept but to subject it to continuous critical revision. The relation between human and artificial intelligence remains an open problem, accessible only through an ongoing process of conceptual transformation that resists closure and maintains sensitivity to the irreducible plurality of its objects.

Yu DaeChil (Ockham Institute & Happy Workers)