Consciousness as Meaning in Latent Spaces

This essay introduces Latent Space Theory, a novel framework positing that consciousness emerges as a dynamical regime within a high-dimensional, amodal, semantically rich embedding space—the latent space—where specialized sensory and interoceptive manifolds converge to generate unified, meaningful representations. Drawing on recent neuroscience and computational findings—such as the superiority of contextual embeddings (e.g., from GPT-2) in predicting brain activity, multimodal convergence in conceptual spaces, and human-like object representations in large language models—the theory argues that phenomenal experience is not a passive or mysterious "what it’s like" property but the experiential perspective of efficient, compressed information processing in these embedding spaces. Qualia, including color sensations, pain, and emotions, are identified with relational positions and trajectories in domain-specific manifolds (e.g., opponent-process color space), rendering them identical to their functional geometry rather than epiphenomenal. The explanatory gap is reframed as perspectival: neurological computation viewed from within is phenomenology. Binding is achieved through representational coherence in latent space, without requiring additional mechanisms, while the sense of self emerges as a self-referential attractor in this semantic topology. Consciousness can be understood a sophisticated mode of meaning-making, evolutionarily favored for its dimensionality reduction and generalization advantages, enabling complex agency in biological and potentially artificial systems. Integrated with predictive processing and global neuronal workspace models, the theory predicts graded consciousness in animals and substrate-independent emergence in sufficiently sophisticated AI, while dissolving zombies and inverted qualia as incoherent. By treating experience as an optimal data format rather than an ontological add-on, Latent Space Theory offers a parsimonious, unified account that preserves phenomenology's richness while closing the hard problem through functional necessity.