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
遗忘与抽象_从萨顿的批评到大模型泛化的困局
视频
音频
原始脚本
遗忘与抽象，从萨顿的批判到大模型的泛化困局，论智能的本质是忘掉细节后的破局。 在人工智能狂飙突进的今天，大语言模型 LLMS 以及对人类知识的海量吸收和流畅输出震撼世界。 却也引发了关于智能本质的深层拷问。 强化学习之父理查德萨顿对 LLMS 死路一条的断言，与知识来自书本，智能来自生活的朴素智慧形成奇妙共振。 而人类遗忘催生泛化的生物特性，更是为这场关于真智能的讨论补上了关键拼图。 当我们将这些线索串联，将清晰看到，真正的智能诞生于遗忘细节后的抽象跃迁。 而大模型若不突破，无疑忘是存储的桎梏，永远只能是知识的模仿者，而非智能的拥有者。 一，萨顿的警钟。 大模型是知识容器，而非智能主体。 理查德萨顿作为强化学习领域的奠基人，其对大模型的批判直击核心，LLMS 缺乏从经验中学习的能力，而这正是智能的灵魂。 他在访谈中尖锐指出，大模型的本质是模仿人类说话，而非理解世界并实现目标。 这种差异体现在三个致命缺陷上，无世界模型。 大模型仅能预测人会说什么，却无法构建世界会发生什么的因果认知。 它模仿人类语言的表层模式，却对物理世界的规律、事件的逻辑关联一无所知。 如同一个只会背诵地图的人，从未真正走过那些路。 无基础真相，在大模型的逻辑里，正确没有定义，它输出的内容仅基于训练数据的概率分布，而非对现实对错的判断。 没有目标就没有对错，没有对错就无法真正学习和改进，这与智能通过反馈迭代的本质背道而驰。 无经验学习能力，面对意外或新情况，大模型不会惊讶，也不会调整自身行为。 它的知识是静态的存储，而非动态的从实践中生长的智慧，这就注定了它无法像人类一样在生活中持续进化。 巴顿的观点本质上是在区分知识与智能星星。 知识是可传递的信息集合，智能是基于目标的自主学习与适应能力。 大模型是知识的集大成者，却不是智能的践行者。 这与知识来自书本，智能来自生活的认知完全契合。 书本传递知识，生活锻造智能。 二、人类智能的奥秘，遗忘是抽象的母亲，泛化是遗忘的孩子。 人类的智能为何能突破知识的桎梏，走向自主探索？答案藏在我们生物性的遗忘里。 从令狐冲练独孤九剑的典故中，我们能看到一种反直觉的智慧，忘掉具体招式，才能抓住破招的本质。 风清扬让令狐冲不断遗忘，实则是让他摆脱细节的束缚，提炼出料敌机先的通用逻辑。 这不是记忆的缺失，而是抽项能力的觉醒。 这种遗忘催生抽象的过程，根植于人类大脑的生物特性。 我们的记忆并非计算机式的永久精确存储，而是神经元连接的动态重塑。 随着时间流逝，无关紧要的细节会被冲刷，只留下对事物本质的模糊却关键的印记。 这种遗忘是一种被迫的高效，因为大脑存储资源有限，它必须丢掉沙子，留下金子。 孩子模仿大人说话时，会自然忽略口音语气的细节，只抓住语义和语法的核心。 学生学习数学时，会忘记立体的具体数字，只记住公式背后的逻辑关系。 正是这种遗忘式抽象，让人类拥有了真正的泛化能力。 我们能从狗有四条腿、猫有四条腿中抽象出哺乳动物的肢体特征，进而迁移到牛羊也有四条腿的判断中。 我们能从一道几何题的解法里提炼出辅助线构造逻辑，并应用到全新的复杂图形中。 这种泛化不是简单的模式匹配，而是本质规律的跨领域迁移，其源头正是忘掉细节后的抽象跃迁。 三大模型的泛化困局，无遗忘，不智能。 反观大模型，其泛化能力更像是一场基于数据覆盖的模仿秀，而非真正的智能跃迁。 当大模型解出一道从未见过的奥数题时，它并非像人类一样抽象出解题思路并自主推导。 而是在训练数据中找到了相似题型的逻辑碎片，并通过概率组合将其拼接成答案。 它看似举一反三，实则是海量数据喂养下的模式复用。 它没有遗忘细节的过程，也就无法完成抽象本质的跃迁。 这种差异的本质在于，人类的泛化是主动的抽象星星，源于遗忘细节后对本质的把握，是智能的创造性跃迁。 大模型的泛化是星星被动的匹配星星，源于训练数据的高覆盖率，是算法的概率性组合。 大模型就像一个永不遗忘的图书馆，它能精准检索每一条知识，却无法像人类一样烧掉无关的书页，只留下照亮前路的火种。 没有遗忘带来的抽象，它的泛化永远停留在相似性调用的层面，而非本质性创造的维度。 四、通往真智能的路径，让模型学会遗忘式抽象。 若要让人工智能真真正拥有智能，就必须突破无遗忘存储的桎梏，赋予它遗忘与抽象的能力。 这意味着架构革新需要构建类似人类感知行动奖励闭环的学习系统，让模型在与环境的交互中获取经验，而非仅从静态数据中吸收知识。 萨顿提出的经验范式正是这一方向的探索，智能体通过不断试错、观察结果、调整策略，在实践 中构建世界模型。 机制创新，设计模仿人类遗忘抽象过程的算法，让模型能够主动丢弃无关细节，提炼核心规律。 这可能涉及对神经网络权重的动态修剪，对知识表征的层级化抽象等技术突破。 目标赋予，为模型确立类似人类的自主目标，使其在追求目标的过程中产生对错反馈。 进而实现真正的学习与改进。 没有目标的智能体永远只是知识的傀儡，而非行动的主体。 当大模型学会遗忘，不是数据的丢失，而是对本质的提炼。 当它能像令狐冲一样忘掉所有招式，却抓住破局的本质时，或许才是真正智能诞生的时刻。 在此之前，大模型只是知识的镜子，而智能的星辰仍闪耀在人类遗忘与抽象的生物智慧深处。
修正脚本
遗忘与抽象，从萨顿的批判到大模型的泛化困局，论智能的本质是忘掉细节后的破局。 在人工智能狂飙突进的今天，大语言模型 LLMS 以其对人类知识的海量吸收和流畅输出震撼世界。 却也引发了关于智能本质的深层拷问。 强化学习之父理查德萨顿对 LLMS 死路一条的断言，与知识来自书本，智能来自生活的朴素智慧形成奇妙共振。 而人类遗忘催生泛化的生物特性，更是为这场关于真智能的讨论补上了关键拼图。 当我们将这些线索串联，就会清晰看到，真正的智能诞生于遗忘细节后的抽象跃迁。 而大模型若不突破，无遗忘是存储的桎梏，永远只能是知识的模仿者，而非智能的拥有者。 一、萨顿的警钟。 大模型是知识容器，而非智能主体。 理查德萨顿作为强化学习领域的奠基人，其对大模型的批判直击核心，LLMS 缺乏从经验中学习的能力，而这正是智能的灵魂。 他在访谈中尖锐指出，大模型的本质是模仿人类说话，而非理解世界并实现目标。 这种差异体现在三个致命缺陷上，无世界模型。 大模型仅能预测人会说什么，却无法构建世界会发生什么的因果认知。 它模仿人类语言的表层模式，却对物理世界的规律、事件的逻辑关联一无所知。 如同一个只会背诵地图的人，从未真正走过那些路。 无基础真相，在大模型的逻辑里，正确没有定义，它输出的内容仅基于训练数据的概率分布，而非对现实对错的判断。 没有目标就没有对错，没有对错就无法真正学习和改进，这与智能通过反馈迭代的本质背道而驰。 无经验学习能力，面对意外或新情况，大模型不会惊讶，也不会调整自身行为。 它的知识是静态的存储，而非动态的从实践中生长的智慧，这就注定了它无法像人类一样在生活中持续进化。 萨顿的观点本质上是在区分知识与智能。 知识是可传递的信息集合，智能是基于目标的自主学习与适应能力。 大模型是知识的集大成者，却不是智能的践行者。 这与知识来自书本，智能来自生活的认知完全契合。 书本传递知识，生活锻造智能。 二、人类智能的奥秘，遗忘是抽象的母亲，泛化是遗忘的孩子。 人类的智能为何能突破知识的桎梏，走向自主探索？答案藏在我们生物性的遗忘里。 从令狐冲练独孤九剑的典故中，我们能看到一种反直觉的智慧，忘掉具体招式，才能抓住破招的本质。 风清扬让令狐冲不断遗忘，实则是让他摆脱细节的束缚，提炼出料敌机先的通用逻辑。 这不是记忆的缺失，而是抽象能力的觉醒。 这种遗忘催生抽象的过程，根植于人类大脑的生物特性。 我们的记忆并非计算机式的永久精确存储，而是神经元连接的动态重塑。 随着时间流逝，无关紧要的细节会被冲刷，只留下对事物本质的模糊却关键的印记。 这种遗忘是一种被迫的高效，因为大脑存储资源有限，它必须丢掉沙子，留下金子。 孩子模仿大人说话时，会自然忽略口音语气的细节，只抓住语义和语法的核心。 学生学习数学时，会忘记例题的具体数字，只记住公式背后的逻辑关系。 正是这种遗忘式抽象，让人类拥有了真正的泛化能力。 我们能从狗有四条腿、猫有四条腿中抽象出哺乳动物的肢体特征，进而迁移到牛羊也有四条腿的判断中。 我们能从一道几何题的解法里提炼出辅助线构造逻辑，并应用到全新的复杂图形中。 这种泛化不是简单的模式匹配，而是本质规律的跨领域迁移，其源头正是忘掉细节后的抽象跃迁。 三、大模型的泛化困局，无遗忘，不智能。 反观大模型，其泛化能力更像是一场基于数据覆盖的模仿秀，而非真正的智能跃迁。 当大模型解出一道从未见过的奥数题时，它并非像人类一样抽象出解题思路并自主推导。 而是在训练数据中找到了相似题型的逻辑碎片，并通过概率组合将其拼接成答案。 它看似举一反三，实则是海量数据喂养下的模式复用。 它没有遗忘细节的过程，也就无法完成抽象本质的跃迁。 这种差异的本质在于，人类的泛化是主动的抽象，源于遗忘细节后对本质的把握，是智能的创造性跃迁。 大模型的泛化是被动的匹配，源于训练数据的高覆盖率，是算法的概率性组合。 大模型就像一个永不遗忘的图书馆，它能精准检索每一条知识，却无法像人类一样烧掉无关的书页，只留下照亮前路的火种。 没有遗忘带来的抽象，它的泛化永远停留在相似性调用的层面，而非本质性创造的维度。 四、通往真智能的路径，让模型学会遗忘式抽象。 若要让人工智能真正拥有智能，就必须突破无遗忘存储的桎梏，赋予它遗忘与抽象的能力。 这意味着架构革新需要构建类似人类感知行动奖励闭环的学习系统，让模型在与环境的交互中获取经验，而非仅从静态数据中吸收知识。 萨顿提出的经验范式正是这一方向的探索，智能体通过不断试错、观察结果、调整策略，在实践中构建世界模型。 机制创新，设计模仿人类遗忘抽象过程的算法，让模型能够主动丢弃无关细节，提炼核心规律。 这可能涉及对神经网络权重的动态修剪，对知识表征的层级化抽象等技术突破。 目标赋予，为模型确立类似人类的自主目标，使其在追求目标的过程中产生对错反馈。 进而实现真正的学习与改进。 没有目标的智能体永远只是知识的傀儡，而非行动的主体。 当大模型学会遗忘，不是数据的丢失，而是对本质的提炼。 当它能像令狐冲一样忘掉所有招式，却抓住破局的本质时，或许才是真正智能诞生的时刻。 在此之前，大模型只是知识的镜子，而智能的星辰仍闪耀在人类遗忘与抽象的生物智慧深处。
英文翻译
Forgetting and Abstraction: From Sutton's Critique to the Generalization Predicament of Large Models—On the Essence of Intelligence as Breaking Through After Forgetting Details. In today's era of rapid AI advancement, large language models (LLMs) have astonished the world with their vast absorption and fluent output of human knowledge. Yet they have also sparked deep questioning about the essence of intelligence. Richard Sutton, the father of reinforcement learning, asserts that LLMs are a dead end—a claim that resonates strangely with the simple wisdom that knowledge comes from books, but intelligence comes from life. And the biological trait of human forgetting that fosters generalization adds a crucial piece to the puzzle of true intelligence. When we connect these clues, it becomes clear that true intelligence is born from the abstract leap that occurs after forgetting details. Unless large models break free from the shackles of memory without forgetting, they will forever be imitators of knowledge, not possessors of intelligence. I. Sutton's Warning: Large Models Are Containers of Knowledge, Not Subjects of Intelligence. Richard Sutton, a foundational figure in reinforcement learning, cuts to the core of the critique: LLMs lack the ability to learn from experience, and that is the soul of intelligence. In interviews, he sharply points out that the essence of large models is to imitate human speech, not to understand the world and achieve goals. This difference manifests in three fatal flaws: - **No world model**: Large models can only predict what humans might say, but cannot construct a causal understanding of what will happen in the world. They mimic the surface patterns of human language but know nothing about the laws of the physical world or the logical connections between events. They are like someone who memorizes a map but has never truly walked those roads. - **No ground truth**: In the logic of large models, correctness has no definition. Their output is based solely on the probability distribution of training data, not on judgments of reality or right and wrong. Without goals, there is no right or wrong; without right or wrong, there is no genuine learning or improvement—contradicting the essence of intelligence, which iterates through feedback. - **No experiential learning ability**: When faced with surprises or new situations, large models do not feel surprised, nor do they adjust their behavior. Their knowledge is static storage, not a dynamic intelligence that grows from practice. This ensures they cannot continuously evolve through life like humans. Sutton's view essentially distinguishes between knowledge and intelligence. Knowledge is a collection of transmittable information; intelligence is the ability for goal-oriented autonomous learning and adaptation. Large models are masters of knowledge, but not practitioners of intelligence. This aligns perfectly with the notion that knowledge comes from books, intelligence from life. Books transmit knowledge; life forges intelligence. II. The Mystery of Human Intelligence: Forgetting Is the Mother of Abstraction, Generalization Is the Child of Forgetting Why can human intelligence break free from the shackles of knowledge and move towards autonomous exploration? The answer lies in our biological forgetting. From the story of Linghu Chong practicing the Dugu Nine Swords, we see a counterintuitive wisdom: forgetting specific moves enables one to grasp the essence of breaking moves. Feng Qingyang made Linghu Chong constantly forget, actually freeing him from the bondage of details and extracting the universal logic of anticipating the opponent's moves. This is not a lack of memory, but the awakening of abstraction ability. This process of forgetting generating abstraction is rooted in the biological characteristics of the human brain. Our memory is not like a computer's permanent, precise storage, but a dynamic reshaping of neural connections. Over time, irrelevant details are washed away, leaving only vague yet crucial imprints of the essence of things. This forgetting is a forced efficiency: because the brain's storage resources are limited, it must discard the sand and keep the gold. When children imitate adults speaking, they naturally ignore details of accent and tone, grasping only the core of semantics and grammar. When students learn mathematics, they forget the specific numbers in examples, remembering only the logical relationships behind the formulas. It is this forgetting-based abstraction that gives humans true generalization ability. We can abstract from "dogs have four legs, cats have four legs" the limb characteristics of mammals, and then transfer that to the judgment that cows and sheep also have four legs. We can extract the logic of constructing auxiliary lines from solving a geometry problem and apply it to entirely new complex figures. This generalization is not simple pattern matching, but cross-domain transfer of essential laws—its source is the abstract leap after forgetting details. III. The Generalization Predicament of Large Models: No Forgetting, No Intelligence In contrast, the generalization ability of large models is more like an imitation show based on data coverage, not a true leap of intelligence. When a large model solves an Olympiad math problem it has never seen before, it does not abstract the problem-solving approach and deduce it independently like a human. Instead, it finds logical fragments of similar question types in the training data and splices them together through probabilistic combinations. It seems to draw inferences by analogy, but in reality it is a reuse of patterns fed by massive data. Without the process of forgetting details, it cannot complete the leap to abstract essences. The essence of this difference is that human generalization is active abstraction, rooted in grasping the essence after forgetting details—a creative leap of intelligence. Large-model generalization is passive matching, rooted in high coverage of training data—a probabilistic combination of algorithms. A large model is like a library that never forgets: it can precisely retrieve every piece of knowledge, but cannot burn irrelevant pages to keep only the sparks that illuminate the path ahead. Without the abstraction brought by forgetting, its generalization remains at the level of similarity invocation, never reaching the dimension of essential creation. IV. The Path to True Intelligence: Teach Models to Forget and Abstract If artificial intelligence is to truly possess intelligence, it must break free from the shackles of memory without forgetting and endow models with the ability to forget and abstract. This means **architectural innovation**: building learning systems similar to the human perception-action-reward loop, allowing models to gain experience through interaction with the environment, not merely absorb knowledge from static data. Sutton's paradigm of experience is an exploration in this direction: an agent continuously tries, observes outcomes, adjusts strategies, and constructs a world model through practice. **Mechanism innovation**: designing algorithms that mimic the human forgetting-abstraction process, enabling models to actively discard irrelevant details and extract core laws. This may involve techniques such as dynamic pruning of neural network weights, hierarchical abstraction of knowledge representations, and other breakthroughs. **Goal endowment**: providing models with autonomous goals similar to humans, so that they generate feedback of right and wrong in the pursuit of those goals, thereby achieving true learning and improvement. An agent without goals is forever a puppet of knowledge, not an agent of action. When large models learn to forget—not as data loss, but as refinement of essence—and when they can forget all tricks like Linghu Chong yet grasp the essence of breaking through, that may be the moment true intelligence is born. Until then, large models are merely mirrors of knowledge, while the star of intelligence still shines deep within the biological wisdom of human forgetting and abstraction.
back to top