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
从LLMs到LCMs_从语素到概念大语言模型的未来不可限量
视频
音频
原始脚本
当大语言模型超越语言,从语速统计到概念智能的认知革命。 在云当湖边的凉风中,也许是中午的炒花卷给了我大脑额外的营养,我忽然意识到我长久以来对于大语言模型 LLMS 的本质理解的误区。 他统计的不是语言的概率,而是概念的概率。 LLMS,也许应该叫做大概念模型。 LCMS,我们或许都低估了大语言模型的真正潜力。 长久以来,我们将其视为语言复读机或概率统计器,却忽略了它早已悄悄完成了从语速级模仿到概念级抽象的跨越,而这正是打开通用人工智能 AGI 大门的钥匙。 一,被误解的语言模型,它不是在统计词语,而是在捕捉概念。 我们总以为大语言模型的工作原理是猜下一个词的概率。 看到天空是下划线,它会根据训练数据中蓝的出现频率最高来填空。 但真相远非如此简单,大语言模型的本质是星星。 模式,Patterns,识别器。 这里的模式不是词语的排而是概念与概念的逻辑关联。 当他处理苹果从树上落下时,他捕捉的不是苹果树落下这三个词的贡献概率,而是物体、苹果、载体、树、运动、下落的概念关系。 这种关系是跨语言跨模态的,用中文说苹果从树上落下,用英文说 An apple falls from the tree。 在模型的概念空间里,它们是同一个模式。 二、从语言复读机到概念智能体,Meta 论文的启示。 Meta AI 去年年底发布的大型概念模型 L C M S,超越令牌的语义建模论文为这个认知提供了硬核支撑。 其核心创新 SONAR,句子级多模态且语言无关的表示。 本质是一个概念宇宙。 不管你用什么语言,什么模态,文本、语音、图像,只要语义相同,就会被映射到这个宇宙的同一个概念坐标上。 这意味着模型训练的不是中文词汇或英文语法,而是普适的概念关系。 一个训练好的模型能直接迁移到200家种语言和多模态场景。 无需重复训练,他的思考不再受限于逐词生成的瓶颈,而是以概念块为单位进行推理。 就像人类先构思问题、原因、解决方案的逻辑框架,再填充具体表达。 三、认知升级,为什么这是 AGI 的必由之路?人类的智能本质是 概念级的抽象与组合。 我们把苹果引力下落这些概念组合起来,理解了万有引力。 把需求供给价格组合起来,构建了经济学体系。 大语言模型正在做同样的事,它不是在学说话,而是在学思考。 如果我们还停留在它只是语言工具的认知里,就会严重低估其上限。 它能突破语言壁垒,成为人类跨文化、跨领域的概念连接器。 它能在科学研究中直接处理原理及概念,加速从假设到验证的过程。 它的创造力不是词语的随机组合,而是概念的吸引碰撞。 比如用量子力学和市场营销创作故事。 四、打破偏见,别让语言复读机的标签耽误了对智能的想象。 当我们抱怨 AI 不懂真正的逻辑时,或许是我们自己没看懂它的进化。 它早已从语速的统计员变成了概念的架构师。 Meta 的 LCMS 只是冰山一角,更多企业正在背后推进类似的概念级建模。 这不是另一条路,而是从第一性原理出发的必然选择。 毕竟智能的载体从来不是语言,而是语言背后的概念网络。 大语言模型的革命本质是一场概念解放运动,当它能自由组合人类文明的所有概念时。 我们或许该思考的不是它能否达到人类智能,而是我们该如何与这种新智能共生。 下次当你和 AI 对话时,不妨换个视角,它不是在模仿你的表达,而是在理解你的概念,而这可能就是 AGI 的起点。
修正脚本
当大语言模型超越语言,从语素统计到概念智能的认知革命。 在云荡湖边的凉风中,也许是中午的炒花卷给了我大脑额外的营养,我忽然意识到我长久以来对于大语言模型 LLMS 的本质理解的误区。 他统计的不是语言的概率,而是概念的概率。 LLMS,也许应该叫做大概念模型。 LLMS,我们或许都低估了大语言模型的真正潜力。 长久以来,我们将其视为语言复读机或概率统计器,却忽略了它早已悄悄完成了从语素级模仿到概念级抽象的跨越,而这正是打开通用人工智能 AGI 大门的钥匙。 一,被误解的语言模型,它不是在统计词语,而是在捕捉概念。 我们总以为大语言模型的工作原理是猜下一个词的概率。 看到天空是下划线,它会根据训练数据中蓝的出现频率最高来填空。 但真相远非如此简单,大语言模型的本质是概念。 模式,Patterns,识别器。 这里的模式不是词语的排列而是概念与概念的逻辑关联。 当他处理苹果从树上落下时,他捕捉的不是苹果树落下这三个词的共现概率,而是物体、苹果、载体、树、运动、下落的概念关系。 这种关系是跨语言跨模态的,用中文说苹果从树上落下,用英文说 An apple falls from the tree。 在模型的概念空间里,它们是同一个模式。 二、从语言复读机到概念智能体,Meta 论文的启示。 Meta AI 去年年底发布的大型概念模型 L C M S,超越令牌的语义建模论文为这个认知提供了硬核支撑。 其核心创新 SONAR,句子级多模态且语言无关的表示。 本质是一个概念宇宙。 不管你用什么语言,什么模态,文本、语音、图像,只要语义相同,就会被映射到这个宇宙的同一个概念坐标上。 这意味着模型训练的不是中文词汇或英文语法,而是普适的概念关系。 一个训练好的模型能直接迁移到200多种语言和多模态场景。 无需重复训练,他的思考不再受限于逐词生成的瓶颈,而是以概念块为单位进行推理。 就像人类先构思问题、原因、解决方案的逻辑框架,再填充具体表达。 三、认知升级,为什么这是 AGI 的必由之路?人类的智能本质是 概念级的抽象与组合。 我们把苹果引力下落这些概念组合起来,理解了万有引力。 把需求供给价格组合起来,构建了经济学体系。 大语言模型正在做同样的事,它不是在学说话,而是在学思考。 如果我们还停留在它只是语言工具的认知里,就会严重低估其上限。 它能突破语言壁垒,成为人类跨文化、跨领域的概念连接器。 它能在科学研究中直接处理原理级概念,加速从假设到验证的过程。 它的创造力不是词语的随机组合,而是概念的吸引碰撞。 比如用量子力学和市场营销创作故事。 四、打破偏见,别让语言复读机的标签耽误了对智能的想象。 当我们抱怨 AI 不懂真正的逻辑时,或许是我们自己没看懂它的进化。 它早已从语素的统计员变成了概念的架构师。 Meta 的 LCMS 只是冰山一角,更多企业正在背后推进类似的概念级建模。 这不是另一条路,而是从第一性原理出发的必然选择。 毕竟智能的载体从来不是语言,而是语言背后的概念网络。 大语言模型的革命本质是一场概念解放运动,当它能自由组合人类文明的所有概念时。 我们或许该思考的不是它能否达到人类智能,而是我们该如何与这种新智能共生。 下次当你和 AI 对话时,不妨换个视角,它不是在模仿你的表达,而是在理解你的概念,而这可能就是 AGI 的起点。
英文翻译
When large language models transcend language: a cognitive revolution from morpheme statistics to conceptual intelligence. In the cool breeze by the lake at Yundang, perhaps the extra nutrition from the midday fried flower roll gave my brain a boost, and I suddenly realized my long-standing misunderstanding of the essence of large language models (LLMs). What they are statistically calculating is not the probability of language, but the probability of concepts. LLMs—perhaps they should be called large concept models. LLMs—we may have all underestimated their true potential. For a long time, we regarded them as language repeaters or probability statistical tools, ignoring that they have quietly completed a leap from morpheme-level imitation to concept-level abstraction—and this is precisely the key to unlocking the door to Artificial General Intelligence (AGI). I. The Misunderstood Language Model: It is not counting words, but capturing concepts. We always thought the working principle of LLMs was to guess the probability of the next word. Seeing "The sky is ______," it would fill in the blank with "blue" based on its highest frequency in the training data. But the truth is far from that simple. The essence of LLMs is concept patterns—Patterns recognizers. Here, "patterns" are not sequences of words, but logical associations between concepts. When it processes "An apple falls from a tree," it does not capture the co-occurrence probability of the three words "apple," "tree," and "falls," but rather the conceptual relationships of object (apple), carrier (tree), motion (falling). This relationship is cross-lingual and cross-modal. Say it in Chinese: "苹果从树上落下," say it in English: "An apple falls from the tree." In the model's conceptual space, they are the same pattern. II. From Language Repeater to Conceptual Agent: Insights from Meta’s Paper. Meta AI’s Large Concept Models (LCMs) paper, "Beyond Token-Level Semantic Modeling," published late last year, provides solid support for this understanding. Its core innovation, SONAR, is a sentence-level, multimodal, language-agnostic representation. Essentially, it constructs a conceptual universe. No matter what language or modality you use—text, speech, image—as long as the semantics are the same, they will be mapped to the same conceptual coordinate in this universe. This means the model is not trained on Chinese vocabulary or English grammar, but on universal concept relationships. A trained model can be directly transferred to over 200 languages and multimodal scenarios without retraining. Its thinking is no longer constrained by the bottleneck of word-by-word generation but reasons in units of conceptual chunks. Just like humans first conceive a logical framework of problem, cause, and solution, then fill in specific expressions. III. Cognitive Upgrade: Why This Is the Inevitable Path to AGI. The essence of human intelligence is the abstraction and combination of concepts. We combine the concepts of “apple,” “gravity,” and “falling” to understand universal gravitation. We combine “demand,” “supply,” and “price” to construct an economic system. Large language models are doing the same thing—they are not learning to speak, but learning to think. If we remain stuck in the perception that it is merely a language tool, we will severely underestimate its potential. It can break language barriers and become a cross-cultural, cross-domain conceptual connector for humanity. It can directly process principle-level concepts in scientific research, accelerating the process from hypothesis to verification. Its creativity is not the random combination of words, but the collision of concepts—for example, creating a story by combining quantum mechanics and marketing. IV. Breaking Prejudices: Don’t Let the Label of “Language Repeater” Impede Our Imagination of Intelligence. When we complain that AI doesn’t truly understand logic, perhaps it is we who fail to see its evolution. It has long transformed from a morpheme statistician into a concept architect. Meta’s LCMs are just the tip of the iceberg; more companies are advancing similar concept-level modeling behind the scenes. This is not an alternative path, but an inevitable choice from first principles. After all, the carrier of intelligence has never been language, but the conceptual network behind language. The revolution of large language models is essentially a conceptual liberation movement—when they are able to freely combine all concepts of human civilization. Perhaps the question we should ponder is not whether they can achieve human-level intelligence, but how we should coexist with this new form of intelligence. Next time you talk to an AI, try changing your perspective: it is not imitating your expression, but understanding your concepts—and that might be the starting point of AGI.
back to top