Reasoning mechanics of large language models: a philosophical analysis

The article demonstrates and philosophically expounds the nature, mechanics of reasoning and the fundamental principles of epistemic limitation of modern dialogue large language models based on the Transformer architecture. Large language models are presented as a functionalist mechanistic project of statistical modeling of language and speech as a model of knowledge as a semantic model of reality – 1) a model 2) of a model 3) of a model of reality. It is shown that because of this substantial distance of mediating reality, the intra-model connections tend to lose on their factual capacity. It is also demonstrated that large language models are a product of machine learning of a certain linguistic behavior with a purpose and values radically different from the purpose and values of human cognition. Their goal is the principle of satisfying the operator of the reward function by cheating at any set price and by any available means at the training stage, which does not let model users be sure of the alignment of the model with human expectations and of the safety of any its reasoning at the stage of its exploitation. In addition, it is substantiated that the fundamental limitations of the very ability to reason in these models are not only factual, but also algorithmic and ontological: these models are limited linear finite automata without a body in reality, generally devoid of other sources of knowledge and experience, except for syntax and context. Due to this, the model by its design resorts to a rough imitation of reflection through fuzzy autoregression, which, in fact, displays the result of a search on a corpus of texts, each of which could potentially have been created by the author of the request to the model. In connection with the above, the actual epistemic value of the products of a large language model is determined primarily by their search value for the user and is limited by the problematic nature of their attribution, validation, as well as the need for external responsible verification and evaluation by the user themself.