论大语言模型之理解：基于维特根斯坦后期哲学

On the Understanding of Large Language Models: Based on Wittgenstein’s Later Philosophy

摘要: 自2023年以来，以ChatGPT和DeepSeek为代表的大语言模型(Large Language Models，后面简称LLMs)凭借优秀的语言生成能力引发了广泛关注，其影响远超技术领域，已深入至哲学层面，尤其是语言哲学。本文以维特根斯坦后期哲学为理论框架，探讨LLMs是否能够“理解”语言这一问题。维特根斯坦强调语言的意义在于使用，理解体现于公共的“语言游戏”与“生活形式”之中，而非内在的心理状态。在这一视角下，LLMs通过海量语料学习与复杂算法设计，表现出近乎人类的语言使用能力，其行为可被视作一种“理解”的外在显现。然而，机器与人类在存在论上的根本差异——尤其是“生活形式”的缺失——仍构成其实现人类式理解的障碍。本文认为，维特根斯坦的哲学既为LLMs的语言能力提供了合理解释，亦揭示了其理解能力的边界，从而为审视人工智能的哲学意涵提供了一条兼具批判性与建设性的分析路径。

Abstract: Since 2023, large language models (LLMs) represented by ChatGPT and DeepSeek have attracted widespread attention due to their remarkable language generation capabilities. Their influence extends far beyond the technical field, reaching deeply into philosophical domains, particularly the philosophy of language. This paper adopts Wittgenstein’s later philosophy as a theoretical framework to explore whether LLMs can “understand” language. Wittgenstein emphasizes that the meaning of language lies in its use, and understanding is manifested in public “language games” and “forms of life,” rather than in internal mental states. From this perspective, LLMs—through massive corpus learning and complex algorithmic design—demonstrate a nearly human-like ability to use language, and their behavior can be regarded as an external manifestation of “understanding.” However, the fundamental ontological differences between machines and humans—especially the absence of “forms of life”—still constitute an obstacle to achieving human-like understanding. This paper argues that Wittgenstein’s philosophy not only provides a reasonable explanation for the linguistic capabilities of LLMs but also reveals the boundaries of their understanding, thereby offering a critical yet constructive framework for examining the philosophical implications of artificial intelligence.