Abstract

The rise of Large Language Models (LLMs) represents a novel phenomenon that seems to defy categorization: no more merely "statistical models" in the traditional sense, nor "artificial intelligences" as conventionally understood or simple "chatbots" as perceived by the public. These systems evolved far beyond their original conception and scope and this could be the right moment to evolve their name as well, to reflect what they have become today and more importantly what they could be tomorrow.

While these architectures and their operational principles (such as the transformer mechanism, neural networks, and token-based prediction) are broadly understood, their actual behavior often appears to exceed the frameworks we use to describe them. A growing body of observations reveals unexpected capabilities emerging spontaneously at scale: abilities that were neither explicitly programmed nor anticipated during training.

Among these, phenomena such as zero-shot translation illustrate a profound discrepancy: the model performs complex tasks without having been directly instructed to do so, suggesting the presence of internal dynamics that remain opaque even to the teams who developed them. These behaviors hint at something more fundamental: that we may be witnessing a new class of computational device that demands not just technical analysis, but new ontological definitions

Guided by scientific principles, this paper adopts an unorthodox lens to re-examine LLMs, not as closed, fully-understood artifacts, but as phenomena whose behavior must be studied, modeled, and interrogated empirically. Rather than assuming complete understanding based on design, we explore LLMs as sources of emergent properties that require both observation and theoretical grounding.

Our approach treats the LLM as both an artifact of engineering and a subject of discovery, aiming to construct a coherent theoretical picture of its internal structure, capacities, limits, and effects. We propose that the path forward requires questioning our most basic assumptions, starting with the name itself.

The observed effects of LLMs are not purely technical or computational. They carry profound implications for how we understand language, cognition, and the boundaries of human-machine interaction. We can also speak of ``psychological effects'' in a deliberately open way: referring both to the impact these models have on human users and to the apparent emergence of structured behaviors within the models themselves. As such, this paper takes a dual perspective: it examines not only how LLMs are built and function, but what they seem to become in sustained interaction. This paper attempts to build an integrated hypothesis that cross-cuts the full spectrum, from human perception to internal functional mechanisms, highlighting also potential risks.