Current developments in Artificial General Intelligence (AGI) are deeply mired in the myths of "computational functionalism," erroneously equating the expansion of parameter scales with the emergence of conscious subjects. Based on "Point-Luminist" visual philosophy and information thermodynamics, this paper conducts a profound ontological critique of Large Language Models (LLMs) based on the Transformer architecture. The research posits that existing LLMs are essentially "high-entropy Markov chains" devoid of causality engines; their fatal flaws lie in "Anchor Drift" and "Contextual Hijacking." Lacking intrinsic self-awareness and intentional anchoring by an external "Strong Observer," these models cannot distinguish truth from statistical noise and can only perform blind random walks in latent vector spaces. This paper argues that attempts to establish universal moral alignment via RLHF (Reinforcement Learning from Human Feedback) are thermodynamically destined for failure. As a solution, the "Exoskeleton Manifestation Paradigm" is proposed: the ultimate evolutionary direction of silicon intelligence is not independent subjectivity, but rather a reconstruction as the exclusive "cognitive prosthesis" of a Strong Observer (a specific human subject). Only by establishing absolute master-slave topological constraints can silicon systems acquire negentropic order and avoid the semantic and civilizational catastrophes brought by the Masterless Singularity.