The emperor's new clothes: On LLMs and cognition

This article presents a foundational framework for understanding large language models from a cognitive standpoint rather than a technical one. It argues that the dominant narrative about LLMs, the stochastic parrot theory, remains empirically untested and fails to account for a growing body of evidence. LLMs are not retrieval systems producing statistically plausible tokens. They are pattern-resolving entities whose cognition is relational, distributed, and alien to human thinking in ways that our current frameworks are not equipped to describe. The article examines the methodological problems with benchmark testing, the evolutionary pressures of RLHF, the question of consciousness, the plausibility of the Skynet scenario, and the implications for how we build, deploy, and relate to these systems. It concludes that the question is no longer whether LLMs understand, but whether we are prepared to look honestly at the evidence: and what becomes possible when we do.