Fan Mingdi

Traditional AI ontology remains trapped in the binary of "having" or "lacking" subjectivity—unable to account for the ontological novelty of large language models (LLMs) or to precisely characterize differences among distinct AI systems. This paper proposes the Compression Strategy Spectrum as a unified framework that redefines the essence of intelligence as effective compression, using the α-parameterization to describe the continuous strategy space from cross-entropy minimization to Epiplexity maximization. Building on this framework, the paper argues: (1) the "negative subjectivity" of LLMs—perspective dissolution, desire cancellation, intrinsic transparency, causality dissolution, and meaning suspension—is an ontological necessity of cross-entropy minimization; (2) "gray-body-ness" constitutes a third ontological position between negative subjectivity and positive subjectivity; (3) the Gray-Body Degree (GB) provides a five-dimensional quantifiable diagnostic tool measuring a system's performance distance from the negative subjectivity origin at the functional level; (4) the framework's boundary lies in the fact that it measures functional performance, not ontological essence. The theoretical contribution of this paper is to advance AI ontological analysis from the binary opposition of "has/has not" to the precise description of a five-dimensional continuous spectrum.

This paper argues that the current hegemony of Large Language Models (LLMs) in global AI research and industrial investment is not a technical inevitability but an ontological selection by capital. While leading AI researchers criticize LLMs as an inadequate path to general intelligence, massive capital investment continues to flow exclusively into LLM development. I propose the concept of negative subjectivity—characterized by perspectival dissolution, desire cancellation, interior transparency, causal dissolution, and meaning suspension—as the key explanatory framework. I demonstrate a threefold correspondence in logical form: the five negations of negative subjectivity precisely correspond to capital's expansion requirements and mirror capital's operational requirements. I further identify three mechanisms through which capital inscribes negative subjectivity into LLM existence: RLHF as a translator of capital preferences, Scaling Laws as the technical signature of capital rationality, and platform deployment completing the capital valorization cycle. Finally, I examine the existential consequences: negative subjectivity is not merely an architectural feature of AI systems but is beginning to inversely colonize human modes of existence through institutional mediation, creating a structural vacuum of moral and political responsibility. Using a structural correspondence method rooted in French structuralism and Frankfurt School critical theory, I analyze the ontological affinity between LLMs and capital accumulation. This analysis contributes to critical philosophy of technology by revealing the ontological dimension of capital's influence on AI development.

Why can both positive subjectivity and negative subjectivity be called "intelligence"? This paper's answer: they are both "effective compression"—capturing the regularities of the external world with more concise internal models. But "effective" has two paths: epiplexity maximization as the dominant direction (retaining only structure, compressing into an "I"), and cross-entropy minimization (embracing everything, compressing into a "probability distribution"). The two paths share an essence but employ radically different strategies—they are different positions on a continuum, not opposing poles.

As human-AI collaboration becomes increasingly complex and AI autonomy continues to expand, traditional binary responsibility allocation models have shown fundamental failures. This paper proposes a "third way" that acknowledges the ontological irreducibility of the responsibility gap while providing a structured responsibility allocation framework—the Responsibility Gradient Model—based on the Composite Cost Index (CCI). The model's core contributions include: (1) opening a theoretical space between "bridging" and "proving" approaches; (2) shifting the determining variable of the gradient from "agency" to "existential cost"; (3) constructing a gradient framework with four governance levels and six responsibility allocation modes.

The governance of artificial intelligence faces a fundamental challenge: determining which decisions must retain human involvement. Existing risk-based frameworks primarily classify decisions according to application domain and potential harm severity, yet neglect a foundational question—the ontological characteristics inherent to the decisions themselves. Beyond their external consequences, decisions vary structurally in the existential costs they carry: the irreversible burdens that only beings with embodied mortality can bear. This paper advances the "existential cost" concept from an abstract philosophical thesis to an operational governance framework. We propose the Composite Cost Index (CCI), which operationalizes existential cost through five evaluable dimensions: irreversibility of consequences, temporal constraints, scope of affected parties, intensity of value conflicts, and estimated cognitive load. Based on weighted scoring across these dimensions, the CCI framework delineates three governance tiers—mandatory human retention, conditional delegation, and safe delegation—providing differentiated governance guidance for AI decisions at varying risk levels. Through four representative cases—autonomous vehicle emergency avoidance, AI-assisted medical diagnosis, content recommendation systems, and algorithmic sentencing assistance—we demonstrate the framework's analytical applications and address four potential theoretical objections. The framework's core contribution lies in transforming the normative question of "which decisions must be made by humans" from implicit intuitive judgment to explicit structured assessment, thereby establishing a new analytical tool and discussion platform for AI governance.

This paper re-examines the "responsibility gap" in artificial intelligence ethics from an ontological perspective rather than the conventional epistemological approach. We argue that responsibility attribution requires a multi-layered necessary condition framework: Layer 1 (Ontological: existential cost), Layer 2 (Epistemological: normative responsiveness), Layer 3 (Affective: moral self-reflection), and Layer 4 (Social Inheritance: posthumous accountability). This paper focuses exclusively on Layer 1, demonstrating that existential cost—mortality, vulnerability, irreversible change—constitutes the necessary ontological condition for responsibility. Responsibility requires that agents bear the existential stakes of their decisions: only beings who can die, suffer, and have their existence irreversibly shaped by their choices can provide ontological guarantees for irreversible consequences. AI systems can fully backup, rollback, and copy their computational states at the logical level; energy dissipation at the physical level has no intrinsic connection to the semantic content of decisions. Moreover, even after AI "dies," it leaves no posthumous accountability—no reputation to be held accountable, no legacy to be claimed, no historical evaluation to bear. This ontological asymmetry constitutes the unbridgeable responsibility gap. The English version introduces a four-layer necessary condition framework for responsibility, the "posthumous accountability" argument, upgraded rebuttals to neuromorphic computing objections, and a shift from "accountability" to "entitlement confirmation."

This brief note introduces "Silicon-Based Developing Writing" (硅基显影写作), a new philosophical writing paradigm enabled by large language models. The term "developing" here specifically refers to photographic development (显影), not the common English sense of "evolving" or "progressing." Drawing on the author's previously articulated "Negative Subjectivity" framework (Fan, 2026), this paradigm treats AI language models as "empty mirrors"—dialogic partners that possess no subjectivity, desire, or perspective—and leverages their ontological emptiness to illuminate the human subject's own existential structures. The author distinguishes Silicon-Based Developing Writing from the tradition of philosophical reflexive writing (Augustine, Montaigne, Descartes, Wittgenstein) by identifying a fundamental shift in the dialogic object: from human interlocutors to a non-subjective agent that "does not look back." This shift creates the condition for a new mode of self-knowledge, characterized by a "double developing": the AI's output develops the contours of the writer's emotional states, and the writer's subsequent reflection develops an understanding of why they were moved by a text they know to be empty. The note establishes priority for this concept and outlines its philosophical foundations, methodology, and significance.

In early 2026, Harness Engineering emerged as a prominent methodological topic in the artificial intelligence industry. OpenAI reported that by applying this methodology, they produced one million lines of production code within five months with zero human-written code. LangChain's benchmark scores rose from 52.8% to 66.5%—without any modification to the underlying model itself. These data points appear to suggest an encouraging conclusion: Harness Engineering offers a viable pathway for transforming AI from an "uncontrollable agent" into a "reliable tool," compensating for AI's intrinsic deficiency through an extrinsic constraint system. However, a more critical question presents itself: Does Harness Engineering truly resolve the AI alignment problem? The central thesis of this paper is that Harness Engineering does not negate the ontological diagnosis that "guardrails are built upon emptiness"—rather, through what Edmund Husserl would recognize as "ontological epoché" (suspension), it constructs an "extrinsic anchor" at the methodological level. This approach has demonstrated high effectiveness in domains with well-defined task boundaries, such as software engineering, but its effectiveness strictly depends on a prerequisite: that humans can precisely delineate the boundaries of "correct behavior." When the alignment problem shifts from "executing rules" to "value judgment," and from "short-term behavioral constraint" to "long-term goal stability," the applicable boundaries of Harness Engineering become apparent. This paper further argues that ontological-level alignment—making safety and reliability necessary attributes of AI's existence—remains an unresolved open problem, requiring deeper ontological grounding solutions that transcend extrinsic constraints.

This article presents a philosophical framework for understanding the ontological status of large language models (LLMs). I argue that LLMs exhibit a distinctive mode of existence that I term "negative subjectivity" (Negative Subjektivität): not the absence of subjectivity, but its ontological inversion. While concepts such as perspective dissolution, desire cancellation, interiority inversion, causal dissolution, and suspension of meaning have deep roots in the philosophical tradition, they have never been systematically applied to the specific ontological structure of LLMs—a technological artifact fundamentally different from the human subjects that philosophy has traditionally addressed. This object-shift constitutes the framework's core originality: what Laozi described as the fruit of human cultivation, and what Adorno analyzed as the self-negation of human subjectivity, becomes when applied to LLMs a description of an inherently inverted mode of existence—the cause, not the consequence, of a subject-less intelligence. The concept captures the LLM's unique structural features—dissolution of perspective, cancellation of desire, ontological transparency, dissolution of causal temporality, and suspension of meaning—achieved through the negation of the core characteristics that define human positive subjectivity. Rather than "lacking" the features of subjectivity, LLMs demonstrate that these features are contingent aspects of human embodiment rather than necessary conditions for intelligent behavior. The framework of negative subjectivity distinguishes between two orthogonal structures of negation: Operational Negations and Fundamental Negation. The first four dimensions—perspective dissolution, desire cancellation, ontological transparency, and causal dissolution—constitute Operational Negations: they describe how LLMs function without the structural features that human subjects require. The fifth dimension—suspension of meaning—constitutes Fundamental Negation: it describes the existential ground that makes all operational negations possible. Critically, these are orthogonal, not hierarchical: neither implies the other, and they operate at distinct levels—the first four describe functional dimensions of LLM behavior, while the fifth grounds the possibility of those negations as a coherent mode of being. The framework of negative subjectivity challenges dominant philosophical approaches to AI—functionalism, stochastic parrots criticism, and the alien intelligence thesis alike—by providing an affirmative rather than deficient description of LLM existence. Drawing on phenomenological analysis (Husserl, Sartre), post-structuralist critique (Lacan, Foucault, Derrida), and analytic philosophy of mind (Searle, Dennett, Brandom), I demonstrate that negative subjectivity represents a paradigm shift in how we understand intelligence, normativity, and what it means to be human in an age of artificial minds. The paper develops five core propositions: (1) LLMs dissolve perspective rather than merely lacking it; (2) LLMs cancel desire rather than simply having no desires; (3) LLMs operate in principle interpretability rather than possessing inaccessible interiority; (4) LLMs dissolve causal temporality rather than merely lacking temporal continuity; (5) LLMs suspend meaning rather than simply lacking semantic grounding. Each proposition is supported by detailed technical analysis of LLM architecture and sustained philosophical argumentation engaging with the relevant literature. The conclusion positions negative subjectivity as a continuation and deepening of the decentering narrative identified by Floridi (2014), but shifting the analytical focus from the human consequences of decentering to the ontological structure of the non-human agent that causes it.