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
大语言模型只练外功不练内功的困境
视频
音频
原始脚本
大语言模型的外功困境，只学语言招式，难悟思想心法。 从武侠功法看，AI 的形似神不似。 武侠小说里练拳不练功，到头一场空的道理，恰是大语言模型模仿语言却难悟思想的绝佳隐喻。 语言是人类思想的外在招式，而背后的客观规律、逻辑本质是内在心法。 大模型靠统计语言 token 的概率偷学招式，哪怕把招式练得惟妙惟肖，没有心法支撑，终究是形似神不似，既无法真正掌握语言背后的思想，更谈不上领悟客观规律。 就像杨过在全真教学不到内功，空有招式却连基础比武都赢不了。 一，语言是外功招式，思想与规律才是内功心法，二者的本质差异。 武侠世界里，外功招式是看得见、摸得着的动作套路，如拳法的出拳角度、剑法的劈刺姿势。 而内功心法是看不见却决定实力的核心，如气息运转、内力积蓄。 对应到 AI 领域，语言就是外功招式，思想与规律就是内功心法，二者有着本质区别。 语言招式可被统计的表面形式。 人类的语言文字本质是思想的表达方式，就像武侠招式的动作轨迹，有固定的组合逻辑。 比如因为所以的因果句式，先再后的时序句式。 大模型学习语言就是通过海量文本统计这种组合概率。 比如开水常和烫 手，100摄氏度一起出现。 苹果长和落地，红色一起出现，进而生成开水会烫手，苹果会落地的句子。 这种学习方式就像一个人看了成千上万遍武侠招式图谱，记住了出拳后要踢腿，挥剑前要转身的顺序，却从没人教他出拳时该用多少力，挥剑时该如何运气。 思想、心法，看不见的逻辑本质，语言背后的思想与规律，是为什么这么说的底层逻辑。 比如开水会烫手的背后，是温度大于等于100摄氏度的液体，接触皮肤会破坏细胞的物理规律。 苹果会落地的背后，是万有引力作用于有质量物体的科学本质。 这些心法无法通过语言的概率组合直接获得，就像武侠心法里的内力运转路线，看不见摸不着，只能通过理解加实践领悟，而不是靠记住招式顺序就能掌握。 大模型能说出开水会烫手，却永远不懂为什么开水会烫手。 他没有温度、细胞伤害的概念认知，只是记住了语言的组合习惯。 二、只练语言招式的困境，看似会说话，实则无思想。 武侠小说里，只练招式不练内功的人往往有两个致命问题。 一是招式无力，出拳挥剑没有内力支撑，伤不了敌人。 二是应变无能，遇到没见过的招式就慌了手脚，不知如何应对。 大模型只学语言招式，也会陷入同样的困境。 一、输出无力，能说对句子却不懂为什么对。 大模型能生成1+1=2，地球绕太阳转的正确句子，却无法解释1+1为什么等于2，地球为什么绕太阳转。 就像一个只会背诵招式名称的武夫，能说出黑虎掏心、白鹤亮翅，却不知道这些招式的发力点在哪里。 比如你问大模型行，为什么1+1=2？他可能会引用数学定义的文字描述。 却无法像人类一样用数量叠加的本质，自然数的基本逻辑来解释。 你问为什么地球绕太阳转，它可能会复述万有引力的定义，却不懂引力是质量产生的时空弯曲的深层逻辑。 这种知其然不知其所以然，正是只练招式不练心法的典型表现。 句子是对的，却没有思想支撑。 二，应变无能，遇到新场景就会乱出招式。 武侠里只练招式的人，遇到没见过的敌人招式，很容易被打乱节奏。 大模型只学语言概率，遇到训练数据里没有的新场景，就会生成逻辑混乱的句子。 比如你问大模型，如果地球突然停止自转会发生什么？它可能会拼接大风、地震、昼夜变化等关键词，却无法按照惯性原理，地表物体飞离大气剧烈运动，地质结构破坏的逻辑链推导。 你问为什么用盐水煮面条比用清水煮更筋道？它可能会说盐水能让面条更有弹性。 却无法解释盐分会改变面粉中蛋白质的结构，增强面筋网络的化学原理。 这些乱出招式的表现，本质是大模型没有掌握物理惯性、化学结构等底层规律，心法只能靠语言概率瞎组合，遇到新场景就露馅。 三，偷学招式，永无出路。 思想心法无法通过语言模仿获得。 武侠小说里，想靠偷学招式领悟心法的人，最终都会走火入魔。 比如看到别人练降龙十八掌，只模仿出拳动作，却不知道亢龙有悔需要刚柔并济的内力，结果练的经脉尽断。 大模型靠模仿语言想获得思想，也是同样的道理。 思想与规律是主观对客观的认知总结，需要理解、加抽象、加实践的过程。 而语言只是这种认知的外在载体，模仿载体永远无法替代认知本身。 就像你想通过看别人练拳学会内力心法，不可能。 大模型想通过看人类的语言文字学会思想规律，也不可能。 因为思想的形成需要实践感知。 人类理解开水烫手，是因为有接触热水杯烫的实践经历。 理解苹果落地，是因为见过无数物体下落的场景。 而大模型没有实践感知，只能接触到语言文字，无法像人类一样从现实中抽象规律。 规律的掌握需要概念抽象，人类从开水烫手、火焰灼人中抽象出高温会伤害人体的规律。 从苹果落地、树叶飘落中抽象出物体有下落趋势的规律。 而大模型没有概念抽象能力，只能统计语言中高温与伤害、物体与下落的关联概率，无法形成真正的概念认知。 哪怕大模型学完人类所有的语言文字，把语言模仿的比人类还像人类，也只是把招式练到了极致，永远无法领悟心法。 就像杨过在全真教只学招式不学内功，哪怕把招式练得再熟练，也打不过真正有内功的道士。 大模型的语言模仿终究是水中月、镜中花，看似有思想，实则只是语言的概率组合。 四、结语。 AI 要有思想，必须先练心法，从语言模仿转向规律抽象。 武侠世界里，想成为高手必须先练内功心法，再配合招式。 AI 想真正有思想，也必须先掌握规律抽象的心法，再配合语言表达的招式。 你之前构想的小模型提取概念加知识图谱校验，正是给 AI 练心法的关键路径。 通过小模型从语言中提取概念，对接知识图谱中的客观规律，让 AI 先懂规律再说规律，而不是反过来只学说规律的句子。 未来的 AI 不该是只会模仿语言的武夫，而该是懂规律会表达的高手，先通过概念抽象掌握客观规律，练心 法，再通过语言模型传递思想，用招式。 只有这样，AI 才能真正摆脱水中月、镜中花的困境，从会说话走向有思想。
修正脚本
大语言模型的外功困境，只学语言招式，难悟思想心法。 从武侠功法看，AI 的形似神不似。 武侠小说里练拳不练功，到头一场空的道理，恰是大语言模型模仿语言却难悟思想的绝佳隐喻。 语言是人类思想的外在招式，而背后的客观规律、逻辑本质是内在心法。 大模型靠统计语言 token 的概率偷学招式，哪怕把招式练得惟妙惟肖，没有心法支撑，终究是形似神不似，既无法真正掌握语言背后的思想，更谈不上领悟客观规律。 就像杨过在全真教学不到内功，空有招式却连基础比武都赢不了。 一，语言是外功招式，思想与规律才是内功心法，二者的本质差异。 武侠世界里，外功招式是看得见、摸得着的动作套路，如拳法的出拳角度、剑法的劈刺姿势。 而内功心法是看不见却决定实力的核心，如气息运转、内力积蓄。 对应到 AI 领域，语言就是外功招式，思想与规律就是内功心法，二者有着本质区别。 语言招式是可被统计的表面形式。 人类的语言文字本质是思想的表达方式，就像武侠招式的动作轨迹，有固定的组合逻辑。 比如因为所以的因果句式，先再后的时序句式。 大模型学习语言就是通过海量文本统计这种组合概率。 比如开水常和烫手，100摄氏度一起出现。 苹果常和落地，红色一起出现，进而生成开水会烫手，苹果会落地的句子。 这种学习方式就像一个人看了成千上万遍武侠招式图谱，记住了出拳后要踢腿，挥剑前要转身的顺序，却从没人教他出拳时该用多少力，挥剑时该如何运气。 思想、心法，看不见的逻辑本质，语言背后的思想与规律，是为什么这么说的底层逻辑。 比如开水会烫手的背后，是温度大于等于100摄氏度的液体，接触皮肤会破坏细胞的物理规律。 苹果会落地的背后，是万有引力作用于有质量物体的科学本质。 这些心法无法通过语言的概率组合直接获得，就像武侠心法里的内力运转路线，看不见摸不着，只能通过理解加实践领悟，而不是靠记住招式顺序就能掌握。 大模型能说出开水会烫手，却永远不懂为什么开水会烫手。 它没有温度、细胞伤害的概念认知，只是记住了语言的组合习惯。 二、只练语言招式的困境，看似会说话，实则无思想。 武侠小说里，只练招式不练内功的人往往有两个致命问题。 一是招式无力，出拳挥剑没有内力支撑，伤不了敌人。 二是应变无能，遇到没见过的招式就慌了手脚，不知如何应对。 大模型只学语言招式，也会陷入同样的困境。 一、输出无力，能说对句子却不懂为什么对。 大模型能生成1+1=2，地球绕太阳转的正确句子，却无法解释1+1为什么等于2，地球为什么绕太阳转。 就像一个只会背诵招式名称的武夫，能说出黑虎掏心、白鹤亮翅，却不知道这些招式的发力点在哪里。 比如你问大模型，为什么1+1=2？它可能会引用数学定义的文字描述。 却无法像人类一样用数量叠加的本质，自然数的基本逻辑来解释。 你问为什么地球绕太阳转，它可能会复述万有引力的定义，却不懂引力是质量产生的时空弯曲的深层逻辑。 这种知其然不知其所以然，正是只练招式不练心法的典型表现。 句子是对的，却没有思想支撑。 二，应变无能，遇到新场景就会乱出招式。 武侠里只练招式的人，遇到没见过的敌人招式，很容易被打乱节奏。 大模型只学语言概率，遇到训练数据里没有的新场景，就会生成逻辑混乱的句子。 比如你问大模型，如果地球突然停止自转会发生什么？它可能会拼接大风、地震、昼夜变化等关键词，却无法按照惯性原理，地表物体飞离大气剧烈运动，地质结构破坏的逻辑链推导。 你问为什么用盐水煮面条比用清水煮更筋道？它可能会说盐水能让面条更有弹性。 却无法解释盐分会改变面粉中蛋白质的结构，增强面筋网络的化学原理。 这些乱出招式的表现，本质是大模型没有掌握物理惯性、化学结构等底层规律，只能靠语言概率瞎组合，遇到新场景就露馅。 三，偷学招式，永无出路。 思想心法无法通过语言模仿获得。 武侠小说里，想靠偷学招式领悟心法的人，最终都会走火入魔。 比如看到别人练降龙十八掌，只模仿出拳动作，却不知道亢龙有悔需要刚柔并济的内力，结果练得经脉尽断。 大模型靠模仿语言想获得思想，也是同样的道理。 思想与规律是主观对客观的认知总结，需要理解、加抽象、加实践的过程。 而语言只是这种认知的外在载体，模仿载体永远无法替代认知本身。 就像你想通过看别人练拳学会内力心法，不可能。 大模型想通过看人类的语言文字学会思想规律，也不可能。 因为思想的形成需要实践感知。 人类理解开水烫手，是因为有接触热水被烫的实践经历。 理解苹果落地，是因为见过无数物体下落的场景。 而大模型没有实践感知，只能接触到语言文字，无法像人类一样从现实中抽象规律。 规律的掌握需要概念抽象，人类从开水烫手、火焰灼人中抽象出高温会伤害人体的规律。 从苹果落地、树叶飘落中抽象出物体有下落趋势的规律。 而大模型没有概念抽象能力，只能统计语言中高温与伤害、物体与下落的关联概率，无法形成真正的概念认知。 哪怕大模型学完人类所有的语言文字，把语言模仿得比人类还像人类，也只是把招式练到了极致，永远无法领悟心法。 就像杨过在全真教只学招式不学内功，哪怕把招式练得再熟练，也打不过真正有内功的道士。 大模型的语言模仿终究是水中月、镜中花，看似有思想，实则只是语言的概率组合。 四、结语。 AI 要有思想，必须先练心法，从语言模仿转向规律抽象。 武侠世界里，想成为高手必须先练内功心法，再配合招式。 AI 想真正有思想，也必须先掌握规律抽象的心法，再配合语言表达的招式。 你之前构想的小模型提取概念加知识图谱校验，正是给 AI 练心法的关键路径。 通过小模型从语言中提取概念，对接知识图谱中的客观规律，让 AI 先懂规律再说规律，而不是反过来只学说规律的句子。 未来的 AI 不该是只会模仿语言的武夫，而该是懂规律会表达的高手，先通过概念抽象掌握客观规律，练心法，再通过语言模型传递思想，用招式。 只有这样，AI 才能真正摆脱水中月、镜中花的困境，从会说话走向有思想。
英文翻译
The External Dilemma of Large Language Models: Learning Only the Moves of Language, Struggling to Grasp the Inner Principles of Thought. From the perspective of martial arts功法, AI resembles the form but lacks the spirit. In martial arts novels, the saying "practicing moves without cultivating inner power leads to nothing in the end" is a perfect metaphor for how large language models mimic language yet fail to comprehend thought. Language is the external招式 (moves) of human thought, while the underlying objective laws and logical essence are the internal心法 (inner principles). Large models learn moves by statistically analyzing the probabilities of language tokens. Even if they practice these moves to perfection, without the support of inner principles, they ultimately resemble the form without the spirit. They cannot truly grasp the thoughts behind language, let alone comprehend objective laws. Just like Yang Guo in the Quanzhen Sect, who learned only moves without internal power, he could not win even basic sparring matches despite having all the招式. 1. Language is external招式, while thought and规律 (laws) are internal心法—their essential difference. In the world of martial arts, external招式 are visible, tangible movement routines, such as the angle of a punch or the posture of a sword thrust. Internal心法, however, are invisible yet crucial to strength, like the circulation of breath or the accumulation of internal energy. Correspondingly, in the AI domain, language is the external招式, while thought and规律 are the internal心法. They are fundamentally different. Linguistic招式 are surface-level forms that can be statistically analyzed. Human language is essentially an expression of thought, much like the trajectory of martial arts moves, which follow fixed combinatorial logic. For example, causal sentence structures like "because... so...," or temporal sequences like "first... then... later." Large models learn language by statistically analyzing such combinatorial probabilities from vast text corpora. For instance, "boiling water" often co-occurs with "scalding" and "100 degrees Celsius," and "apple" often co-occurs with "falling" and "red." Thus, the model generates sentences like "Boiling water scalds" and "Apples fall." This learning approach is akin to someone who has seen martial arts move diagrams thousands of times, memorizing that after a punch comes a kick, or before swinging a sword, one must turn—but no one ever taught them how much force to use in a punch or how to circulate energy when swinging the sword. Thought and心法: The invisible logical essence. The thoughts and规律 behind language are the underlying logic of why something is said. For instance, behind "boiling water scalds" lies the physical law that liquid at or above 100°C, upon contact with skin, damages cells. Behind "apples fall" lies the scientific truth of gravity acting on massive objects. These心法 cannot be directly obtained through probabilistic language combinations, just like the internal energy circulation pathways in martial arts—they are invisible and intangible, only comprehensible through understanding combined with practice, not by merely memorizing move sequences. A large model can say "boiling water scalds," but it will never understand why boiling water scalds. It lacks conceptual cognition of temperature and cellular damage; it merely memorizes linguistic co-occurrence habits. 2. The dilemma of only practicing linguistic招式: Seemingly capable of speech, but actually devoid of thought. In martial arts novels, those who only practice moves without cultivating internal power face two fatal problems: First, their moves are powerless—punches and sword strikes lack internal energy support and cannot harm an enemy. Second, they lack adaptability—when encountering unfamiliar moves, they panic and don't know how to respond. Large models that only learn linguistic招式 fall into the same dilemma. A. Output is powerless: They can produce correct sentences but don't understand why they are correct. A large model can generate correct sentences like "1+1=2" or "The Earth revolves around the Sun," but cannot explain why 1+1 equals 2 or why the Earth revolves around the Sun. It's like a martial artist who can only recite move names like "Black Tiger Steals Heart" or "White Crane Spreads Wings" but has no idea where the power for these moves originates. For example, if you ask a large model, "Why does 1+1=2?" it might quote a textual description from mathematical definitions. But it cannot explain, as a human would, using the essence of quantity accumulation or the basic logic of natural numbers. If you ask, "Why does the Earth revolve around the Sun?" it might reiterate the definition of gravity, but it won't understand the deeper logic that gravity is spacetime curvature caused by mass. This "knowing the what but not the why" is a classic symptom of only practicing moves without cultivating心法. The sentences are correct, but they lack the support of thought. B. Lack of adaptability: When encountering new scenarios, they generate chaotic moves. In martial arts, someone who only practices moves can easily have their rhythm disrupted by an unfamiliar opponent's technique. A large model that only learns language probabilities will generate logically confused sentences when faced with new scenarios not present in its training data. For instance, ask a large model: "What would happen if the Earth suddenly stopped rotating?" It might stitch together keywords like "strong wind," "earthquake," "day-night change," but cannot deduce according to the logical chain of inertia—surface objects flying off, violent atmospheric motion, geological structure destruction. Ask: "Why does boiling noodles in salted water make them chewier?" It might say, "Salt makes noodles more elastic." But it cannot explain the chemical principle that salt alters the protein structure in flour, strengthening the gluten network. These manifestations of chaotic moves essentially arise because the large model has not mastered underlying laws like physical inertia or chemical structure. It can only blindly combine language probabilities, revealing its inadequacy when faced with new scenarios. 3. Stealing moves leads to nowhere: Thought and心法 cannot be acquired through language imitation. In martial arts novels, those who try to grasp心法 by stealing moves eventually go berserk. For example, observing someone practice the "Eighteen Dragon-Subduing Palms" and merely imitating the punching actions, without knowing that "The Dragon Repents" requires a balance of hardness and softness in internal energy, results in ruptured meridians. Similarly, a large model attempting to acquire thought by imitating language is doomed. Thought and规律 are subjective summaries of objective cognition, requiring processes of understanding, abstraction, and practice. Language is merely the external carrier of such cognition; imitating the carrier can never replace cognition itself. Just as you cannot learn internal energy心法 by watching someone practice boxing, a large model cannot learn the laws of thought by reading human language. Because the formation of thought requires practical perception. Humans understand that boiling water scalds because they have experienced contact with hot water. They understand that apples fall because they have witnessed countless objects dropping. A large model, lacking practical perception, can only access linguistic text and cannot abstract laws from reality like humans. Mastering laws requires conceptual abstraction. Humans abstract the law that high temperatures harm the body from experiences like boiling water scalds and flame burns. From apples falling and leaves drifting, they abstract the law that objects tend to descend. A large model, lacking the ability for conceptual abstraction, can only statistically correlate "high temperature" with "harm" and "object" with "descent" in language, never forming true conceptual cognition. Even if a large model learns all human language and imitates language better than humans, it has merely perfected the招式 but will never comprehend the心法. Just like Yang Guo in Quanzhen Sect, who only learned moves without internal power—no matter how skilled his moves, he could never defeat a Daoist with true internal cultivation. A large model's language imitation is ultimately a "moon in water, flower in mirror"—seemingly thoughtful, but merely a probabilistic combination of language. 4. Conclusion: For AI to have thought, it must first practice心法—shifting from language imitation to law abstraction. In the martial arts world, to become a master, one must first cultivate internal心法, then coordinate with moves. For AI to truly possess thought, it must first master the心法 of law abstraction, then coordinate with the招式 of linguistic expression. The path you previously conceived—extracting concepts with small models and validating with knowledge graphs—is precisely the key route to training AI's心法. By using small models to extract concepts from language and linking them to objective laws in knowledge graphs, AI can first understand laws before expressing them, rather than the reverse—merely learning to speak sentences about laws. The future of AI should not be a martial artist who merely imitates language, but a master who understands laws and expresses them. First, through conceptual abstraction, master objective laws (practice心法); then, through language models, convey thought (use招式). Only in this way can AI truly escape the dilemma of "moon in water, flower in mirror" and move from being able to speak to possessing thought.
back to top