我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎,当追忆的图片变得模糊,我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
认知的世界从来不是认识的世界而是工具的边界
视频
音频
原始脚本
认知的边界从来不是世界的边界,而是我们自己给工具设定的边界。 维特根斯坦的哲学启示,大语言模型的边界与 AGI 的迷途。 当我们将维特根斯坦的思想与大语言模型的发展路径并置,会发现这位哲学家早已用自我颠覆的思考,为当代 AI 技术划定了 理论疆界。 他不仅拆解了语言及世界镜像的幻想,更间接戳破了 LLMs 通向 AGI 的认知误区,甚至从根本上质疑了图灵实验的有效性。 以下从核心逻辑出发,对这种深层关联做一次系统总结。 一、维特根斯坦的思想转向。 LLMs 的天花板早被定义为维特根斯坦前期以逻辑哲学论构建语言镜像论认为语言的逻辑结构能精准对应世界的原子事实。 就像 LLM 的早期构想,用语言的 逻辑训练数据与算法,复现世界的全貌。 但他后期以哲学研究完成自我否定,提出的语言游戏论,恰恰成为 LLM 无法突破的理论边界。 语言的本质是场景适配,而非事实复刻。 维特根斯坦发现语言没有统一的、固定的逻辑,其意义源于在具体生活场景中的使用,如 水在喝水、跳水中用法迥异。 就像 LLMs 的输出只是上下文概率下的合理答案,而非对世界事实的精准描述。 LLMs 能流畅对话,本质是熟练掌握了语言游戏的规则,但从未真正理解规则背后的世界。 就像一个熟记象棋规则的人未必懂 象棋为何存在?更不懂其他棋类的逻辑。 不可说之物与 LLM 的认知盲区。 维特根斯坦始终承认语言有边界,超出边界的伦理、审美、主观体验,如疼痛的感受,无法用语言描述。 LLM 同样如此,它能输出疼痛是身体的不适感,却永远无法体会疼痛本身,能谈论爱,却不懂爱的情感重量。 这种盲区不是技术问题,而是语言工具的先天局限。 用语言构建的 AI 注定无法触达语言之外的世界。 从这个角度看,LLM 的发展就像数学极限,无论算法如何优化,数据如何扩容,它都只能无限逼近人类语言游戏的熟练度,却永远无法突破语言工具的边界。 因为维特根斯坦早已证明,语言不是世界的钥匙,只是我们与世界互动的玩具。 二,LLM 通向 AGI,一条被哲学否定的技术路径。 很多人将 LLM 视为 AGI 的终极道路,核心假设是语言能承载世界的全部信息,掌握语言就能掌握世界。 但维特根斯坦的思想早已否定了这一假设。 AGI 需要认知世界,LLM 只需要玩转语言。 AGI 的核心是拥有类人认,能理解因果,如开水会烫人。 不仅是语言记忆,更是对温度与人体伤害的认知,能判断价值。 如撒谎不对不仅是语言规则,更是对伦理的理解,能自主探索。 如好奇星星为何发光,不仅是语言提问,更是对未知的主动好奇。 但 LLMs 的本质是语言概率生成器,它输出开水会烫人。 只是因为训练数据中,开水与烫人关联度高。 他说撒谎不对,只是因为人类语言中撒谎常与错误绑定。 他从未认知世界,只是模仿人类对世界的语言描述。 沙滩上的城堡,LLM 的可靠性陷阱。 如你所说,语言本身是不精确的、场景化的。 用语言构筑的 LLM 就像建在沙滩上的城堡,看似坚固,能流畅对话、写代码、做分析,实则依赖人类语言的共识。 一旦脱离共识,如面对全新场景、小众文化、抽象概念。 城堡就会崩塌。 如 LLM 生成逻辑矛盾的内容,误解隐喻含义。 而 AGI 需要的是在未知世界中站稳脚跟,不是在已知语言中玩的熟练。 用语言游戏工具去追求认知世界的智能,从起点就偏离了方向。 三、图灵实验的失效,被语言游戏解构的智能标准,图灵实验将语言流畅度作为智能的判定标准,认为能骗过人类的语言交互就是有智能。 但在维特根斯坦 的视角下,这个标准本身就站不住脚。 会玩语言游戏不等于有智能。 维特根斯坦说语言是游戏,而游戏的核心是遵守规则。 一个人能熟练下象棋,不代表他有象棋智能,可能只是熟记了棋谱。 同理,LLM 能流畅对话,不代表它有认知智能,只是熟记了人类语言的对话。 棋谱。 就像会玩游戏的二傻子,即便能赢棋,也不代表他有深度思考。 图灵实验把游戏熟练度错当成了智能本身,而维特根斯坦早已拆穿了这种混淆。 智能的核心是理解规则的意义,而非遵守规则。 人类使用语言不仅是遵守语法规则,更是理解语言背后的意义。 说今天天气好。 可能是想约朋友出门,意义是邀约。 可能是单纯表达心情,意义是情绪。 可能是开启话题,意义是社交。 但 LLMs 只能识别天气好的语言表层含义,无法判断背后的深层意义。 它能输出,那我们去公园吧,却不知道邀约需要考虑对方的时间、意愿。 能输出,你心情不错呀。 却不知道共情需要感知对方的情绪,这种只懂规则不懂意义的交互。 根本不是智能,只是语言的自动化响应。 四,终极启示。 维特根斯坦不是 AI 泡沫的预言家,而是认知的清醒剂。 我们不必将维特根斯坦的思想视为 LLAM 泡沫破裂的预言。 而应将其视为 AI 发展的认知校正。 LLM 不是失败的技术,而是被高估的工具。 它在语言辅助领域写文案、查信息、做翻译有巨大价值。 就像语言在日常交流中有巨大价值,但我们不能要求它超越工具属性,就像不能要求语言直接创造世界。 AGI 的探索需要跳出出语言中心主义。 维特根斯坦告诉我们,认知世界需要的是与世界的直接互动,如感官体验、实践探索、情感共鸣,而非对世界的语言描述。 未来 AGI 的突破可能不再让 AI 更会说话,而在让 AI 更会感知、更会实践、更会体验。 就像人类的智能不是源于会说话,而是源于会用身体感受世界,用行动改变世界。 说到底,维特根斯坦的思想给我们的最大启示是,不要把工具当成了目的,不要把语言当成了世界,更不要把玩转工具的 AI 当成了认知世界的智 当我们清醒的看到 LLAM 的边界,才能更理性的探索 AI 的未来,而不是在语言游戏的狂欢中迷失在 AGI 的幻想里。
修正脚本
认知的边界从来不是世界的边界,而是我们自己给工具设定的边界。 维特根斯坦的哲学启示,大语言模型的边界与 AGI 的迷途。 当我们将维特根斯坦的思想与大语言模型的发展路径并置,会发现这位哲学家早已用自我颠覆的思考,为当代 AI 技术划定了理论疆界。 他不仅拆解了语言及世界镜像的幻想,更间接戳破了 LLMs 通向 AGI 的认知误区,甚至从根本上质疑了图灵实验的有效性。 以下从核心逻辑出发,对这种深层关联做一次系统总结。 一、维特根斯坦的思想转向。 LLMs 的天花板早被维特根斯坦定义:他前期以逻辑哲学论构建语言镜像论,认为语言的逻辑结构能精准对应世界的原子事实。 就像 LLM 的早期构想,用语言的逻辑训练数据与算法,复现世界的全貌。 但他后期以哲学研究完成自我否定,提出的语言游戏论,恰恰成为 LLM 无法突破的理论边界。 语言的本质是场景适配,而非事实复刻。 维特根斯坦发现语言没有统一的、固定的逻辑,其意义源于在具体生活场景中的使用,如水在喝水、跳水中用法迥异。 就像 LLMs 的输出只是上下文概率下的合理答案,而非对世界事实的精准描述。 LLMs 能流畅对话,本质是熟练掌握了语言游戏的规则,但从未真正理解规则背后的世界。 就像一个熟记象棋规则的人未必懂象棋为何存在?更不懂其他棋类的逻辑。 不可说之物与 LLM 的认知盲区。 维特根斯坦始终承认语言有边界,超出边界的伦理、审美、主观体验,如疼痛的感受,无法用语言描述。 LLM 同样如此,它能输出疼痛是身体的不适感,却永远无法体会疼痛本身,能谈论爱,却不懂爱的情感重量。 这种盲区不是技术问题,而是语言工具的先天局限。 用语言构建的 AI 注定无法触达语言之外的世界。 从这个角度看,LLM 的发展就像数学极限,无论算法如何优化,数据如何扩容,它都只能无限逼近人类语言游戏的熟练度,却永远无法突破语言工具的边界。 因为维特根斯坦早已证明,语言不是世界的钥匙,只是我们与世界互动的玩具。 二、LLM 通向 AGI,一条被哲学否定的技术路径。 很多人将 LLM 视为 AGI 的终极道路,核心假设是语言能承载世界的全部信息,掌握语言就能掌握世界。 但维特根斯坦的思想早已否定了这一假设。 AGI 需要认知世界,LLM 只需要玩转语言。 AGI 的核心是拥有类人认知,能理解因果,如开水会烫人。 不仅是语言记忆,更是对温度与人体伤害的认知,能判断价值。 如撒谎不对不仅是语言规则,更是对伦理的理解,能自主探索。 如好奇星星为何发光,不仅是语言提问,更是对未知的主动好奇。 但 LLMs 的本质是语言概率生成器,它输出开水会烫人。 只是因为训练数据中,开水与烫人关联度高。 它说撒谎不对,只是因为人类语言中撒谎常与错误绑定。 它从未认知世界,只是模仿人类对世界的语言描述。 沙滩上的城堡,LLM 的可靠性陷阱。 如你所说,语言本身是不精确的、场景化的。 用语言构筑的 LLM 就像建在沙滩上的城堡,看似坚固,能流畅对话、写代码、做分析,实则依赖人类语言的共识。 一旦脱离共识,如面对全新场景、小众文化、抽象概念。 城堡就会崩塌。 如 LLM 生成逻辑矛盾的内容,误解隐喻含义。 而 AGI 需要的是在未知世界中站稳脚跟,不是在已知语言中玩得熟练。 用语言游戏工具去追求认知世界的智能,从起点就偏离了方向。 三、图灵实验的失效,被语言游戏解构的智能标准,图灵实验将语言流畅度作为智能的判定标准,认为能骗过人类的语言交互就是有智能。 但在维特根斯坦的视角下,这个标准本身就站不住脚。 会玩语言游戏不等于有智能。 维特根斯坦说语言是游戏,而游戏的核心是遵守规则。 一个人能熟练下象棋,不代表他有象棋智能,可能只是熟记了棋谱。 同理,LLM 能流畅对话,不代表它有认知智能,只是熟记了人类语言的棋谱。 就像会玩游戏的二傻子,即便能赢棋,也不代表他有深度思考。 图灵实验把游戏熟练度错当成了智能本身,而维特根斯坦早已拆穿了这种混淆。 智能的核心是理解规则的意义,而非遵守规则。 人类使用语言不仅是遵守语法规则,更是理解语言背后的意义。 说今天天气好。 可能是想约朋友出门,意义是邀约。 可能是单纯表达心情,意义是情绪。 可能是开启话题,意义是社交。 但 LLMs 只能识别天气好的语言表层含义,无法判断背后的深层意义。 它能输出那我们去公园吧,却不知道邀约需要考虑对方的时间、意愿。 能输出你心情不错呀。 却不知道共情需要感知对方的情绪,这种只懂规则不懂意义的交互,根本不是智能,只是语言的自动化响应。 四、终极启示。 维特根斯坦不是 AI 泡沫的预言家,而是认知的清醒剂。 我们不必将维特根斯坦的思想视为 LLM 泡沫破裂的预言。 而应将其视为 AI 发展的认知校正。 LLM 不是失败的技术,而是被高估的工具。 它在语言辅助领域写文案、查信息、做翻译有巨大价值。 就像语言在日常交流中有巨大价值,但我们不能要求它超越工具属性,就像不能要求语言直接创造世界。 AGI 的探索需要跳出语言中心主义。 维特根斯坦告诉我们,认知世界需要的是与世界的直接互动,如感官体验、实践探索、情感共鸣,而非对世界的语言描述。 未来 AGI 的突破可能不再让 AI 更会说话,而在让 AI 更会感知、更会实践、更会体验。 就像人类的智能不是源于会说话,而是源于会用身体感受世界,用行动改变世界。 说到底,维特根斯坦的思想给我们的最大启示是,不要把工具当成了目的,不要把语言当成了世界,更不要把玩转工具的 AI 当成了认知世界的智能。当我们清醒地看到 LLM 的边界,才能更理性地探索 AI 的未来,而不是在语言游戏的狂欢中迷失在 AGI 的幻想里。
英文翻译
The boundary of cognition has never been the boundary of the world, but the boundary we set for our tools. Wittgenstein's philosophical insights, the boundaries of large language models, and the misguided path to AGI. When we juxtapose Wittgenstein's ideas with the development trajectory of large language models, we find that this philosopher, through his self-subverting thinking, had already delineated the theoretical boundaries for contemporary AI technology. He not only dismantled the illusion of language as a mirror of the world but also indirectly exposed the cognitive fallacies of LLMs on the path to AGI, and even fundamentally questioned the validity of the Turing test. Below, starting from the core logic, I offer a systematic summary of this deep connection. 1. Wittgenstein's Philosophical Turn The ceiling of LLMs was defined long ago by Wittgenstein: In his early work, the *Tractatus Logico-Philosophicus*, he constructed the picture theory of language, arguing that the logical structure of language can precisely correspond to atomic facts in the world. This is similar to the early conception of LLMs, using the logical training data and algorithms of language to replicate the full picture of the world. But in his later work, *Philosophical Investigations*, he completed a self-negation, proposing the language-game theory, which precisely becomes the theoretical boundary that LLMs cannot break through. The essence of language is contextual adaptation, not factual reproduction. Wittgenstein discovered that language has no unified, fixed logic; its meaning derives from its use in specific life contexts—just as the word "water" has vastly different usages in "drinking water" and "jumping into water." Similarly, the output of LLMs is merely a plausible answer under contextual probability, not an accurate description of world facts. LLMs can engage in fluent conversations, but their essence is mastering the rules of language games, never truly understanding the world behind those rules. Just as someone who knows the rules of chess by heart may not understand why chess exists, let alone the logic of other board games. The Ineffable and the Cognitive Blind Spots of LLMs Wittgenstein always acknowledged that language has boundaries; beyond those boundaries lie ethics, aesthetics, and subjective experiences—such as the sensation of pain—that cannot be described by language. LLMs are no different: they can output "pain is an uncomfortable physical sensation," but they can never experience pain itself; they can talk about love, but they do not understand the emotional weight of love. This blind spot is not a technical problem but an inherent limitation of language as a tool. An AI built on language is destined never to reach the world beyond language. From this perspective, the development of LLMs resembles a mathematical limit: no matter how algorithms are optimized or data expanded, they can only infinitely approach the proficiency of human language games, yet never break through the boundary of language as a tool. Because Wittgenstein long ago proved that language is not the key to the world; it is merely a toy with which we interact with the world. 2. LLMs as a Path to AGI: A Technical Route Negated by Philosophy Many regard LLMs as the ultimate path to AGI, with the core assumption that language can carry all information about the world, and mastering language means mastering the world. But Wittgenstein's thought has already negated this assumption. AGI requires cognition of the world; LLMs only need to play with language. The core of AGI is possessing human-like cognition: understanding causality (e.g., boiling water burns), which is not just linguistic memory but cognition of temperature and harm to the body; making value judgments (e.g., lying is wrong), which is not just a linguistic rule but an understanding of ethics; and autonomously exploring (e.g., curiosity about why stars shine), which is not just linguistic questioning but active curiosity about the unknown. But the essence of LLMs is a language probability generator: it outputs "boiling water burns" only because in the training data, "boiling water" and "burns" are highly correlated. It says "lying is wrong" only because in human language, "lying" is often associated with "wrong." It has never cognized the world; it only imitates human linguistic descriptions of the world. A Castle Built on Sand: The Reliability Trap of LLMs As you said, language itself is imprecise and context-dependent. An LLM built on language is like a castle built on sand: it appears solid—fluent conversation, code writing, analysis—but actually relies on the consensus of human language. Once it deviates from that consensus—facing entirely new scenarios, niche cultures, abstract concepts—the castle collapses. For example, LLMs generate logically contradictory content or misinterpret metaphorical meanings. What AGI requires is the ability to stand firm in an unknown world, not to play skillfully within known language. Pursuing intelligence that cognizes the world using a language-game tool is a deviation from the very start. 3. The Failure of the Turing Test: An Intelligence Standard Deconstructed by Language Games The Turing test uses linguistic fluency as the criterion for intelligence, asserting that an entity capable of deceiving humans in language interaction is intelligent. But from Wittgenstein's perspective, this criterion itself is untenable. Being good at language games does not equal having intelligence. Wittgenstein said language is a game, and the core of a game is following rules. A person who plays chess skillfully may not possess chess intelligence; they may simply have memorized the openings. Similarly, an LLM that converses fluently does not possess cognitive intelligence; it has merely memorized the playbook of human language. Just like a "fool who plays games well"—even if they can win at chess, it does not mean they engage in deep thinking. The Turing test mistakes game proficiency for intelligence itself, and Wittgenstein long ago exposed this confusion. The core of intelligence is understanding the meaning of rules, not just following them. Humans use language not only by obeying grammatical rules but also by understanding the meaning behind language. Saying "It's a nice day today" could be an invitation to go out—the meaning is an invitation; it could be an expression of mood—the meaning is emotion; it could be starting a conversation—the meaning is social interaction. But LLMs can only identify the surface meaning of "nice day" and cannot judge the deeper meaning behind it. It can output "Then let's go to the park," but it doesn't know that an invitation requires considering the other person's time and willingness. It can output "You seem in a good mood," but it doesn't know that empathy requires sensing the other's emotions. Such interaction—knowing only the rules but not the meaning—is not intelligence at all; it is merely automated linguistic response. 4. Ultimate Revelation Wittgenstein is not a prophet of the AI bubble burst, but a sobering reminder for cognition. We should not view Wittgenstein's thought as a prophecy of the LLM bubble bursting, but as a cognitive correction for AI development. LLMs are not a failed technology; they are an overestimated tool. They have immense value in language-assistance fields—writing copy, searching information, translation—just as language has immense value in everyday communication. But we must not demand that they transcend their tool nature, just as we cannot demand that language directly create the world. The exploration of AGI needs to step out of linguacentrism. Wittgenstein tells us that cognizing the world requires direct interaction with the world—sensory experience, practical exploration, emotional resonance—not merely linguistic descriptions of the world. Future breakthroughs in AGI may lie not in making AI speak better, but in making AI perceive more, practice more, and experience more. Just as human intelligence does not originate from speech, but from using the body to feel the world and actions to change the world. Ultimately, the greatest lesson from Wittgenstein's thought is: do not mistake the tool for the goal, do not mistake language for the world, and do not mistake an AI that plays with tools for an intelligence that cognizes the world. When we clearly see the boundaries of LLMs, we can more rationally explore the future of AI, rather than losing ourselves in the revelry of language games and the illusion of AGI.
back to top