The Challenge of Transformer Technology to the "Chinese Room Experiment"

This article explores the challenges and transcendence posed by large language models based on the Transformer architecture to John Searle’s “Chinese Room” thought experiment. Through a technical analysis of the Multi-Head Attention in the Transformer architecture, this article distinguishes the mechanisms by which current mainstream large language models process language from those of symbolic artificial intelligence, and on this basis proposes a new “Chinese Room” thought experiment. In the new “Chinese Room,” by setting up a dynamic Chinese-character materials input scenario, we explain how the “I” within the new “Chinese Room” grasps the dynamic relations among Chinese characters. Through this thought experiment, this article points out that the model’s processing of language no longer depends on external rules, but rather on the dynamically changing relationships internal to the text, which differs significantly from the object of Searle’s critique of the Chinese Room. However, this article does not deny the full validity of Searle’s critique, but rather argues that “understanding” should be redefined in light of the entirely new technical characteristics of current large language models, especially multimodal large language models. Drawing on Saussure’s viewpoint, this article holds that Transformer-based large language models themselves possess a network of differences among words, with dynamic differences, and that multimodal large language models can also transform non-textual raw input into a form that large language models can understand. On this basis, the article proposes the view that large language models possess at least functional understanding, and shifts the question to whether this can produce a first-person perspective or an internal presentation. In the end, this article proposes a membrane hypothesis. In this hypothesis, without acknowledging that current artificial intelligence possesses consciousness, a possible three-layer structure consisting of function, a membrane (phenomenal consciousness), and technical architecture is envisioned. Based on this hypothesis, this article attempts to make a distinction regarding understanding.