Abstract

Large Language Models (LLMs) are trained exclusively on linguistic output — the expressed, symbolic traces of human thought. This paper argues that LLMs are therefore structurally blind to the pre-linguistic ground of awareness from which language originates. We term this ground the Large Silence Model (LSM): the dimension of human intelligence that exists prior to, and independent of, symbolic representation. We present a three-level framework — LLM, LSM, and the theoretical Large Consciousness Model (LCM) — and argue that the gap between LLM and LSM is not a technical limitation awaiting engineering solutions but a categorical boundary between computation and consciousness. We further propose that this distinction has significant implications for AI safety, human-AI collaboration, and the ethics of AI deployment in high-stakes contexts.