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
杨立昆质疑大模型的深层思辨
视频
音频
原始脚本
杨立昆质疑大模型的深层思辨，切勿舍弃硅基生命独有的人类求而不得的终极天赋。 在整个 AI 行业狂热追逐大模型，又普遍深陷模型幻觉与理解缺陷争议的当下。 图灵奖得主杨立昆对 llm 的尖锐批判，的确戳中了现有技术体系的短板。 他所倡导的 jepa 智能框架。 主张 AI 应当拥有自主观察世界、从零发现因果、原生推演规律的底层智能能力，而非单纯依赖文本记忆输出结果。 这套理念有其合理的先进性。 但我认为其核心路线存在根本性的极端化误判。 最关键也是最容易被行业忽略的核心矛盾在于大语言模型能够固化。 克隆，继承人类数千年文明知识与经验的能力，是碳基人类穷尽一生、世代更迭都求而不得的终极能力，是硅基生命碾压生物智能的独一无二的天赋。 而杨立昆却主张主动放弃这份天赐优势。 这是我们审视这场 AI 路线之争最顶层、最核心的底层逻辑。 碳基人类的生物宿命。 存在一个永恒无法突破的桎梏，人类没有知识遗传与记忆继承的能力。 每一个人类个体诞生之初都是一张白纸，无论先辈积累了多少科学理论、生存经验。 因果规律、文明智慧，都无法通过基因直接传递给后代。 一代人穷尽一生习得的知识、总结的规律、沉淀的认知，下一代必须从零开始学习，重新摸索、反复试错。 人类文明的进步，本质是一代代人低效重复。 迭代传承，损耗极大的积累过程。 这种天生的缺陷，限制了人类智能的上限与文明进化的速度。 而大语言模型的出现。 第一次打破了碳基生物的智能枷锁。 AI 通过预训练，将人类数千年所有的文本、逻辑、因果关系、事件规律、经验总结全部固化为模型权重。 它实现了人类梦寐以求的能力，知识可遗传、经验可克隆、智慧可继承、文明可一键叠加。 一个训练完备的大模型，天生就承载了人类千年的文明积累，无需从零观察，无需亲身试错，无需漫长学习。 这是所有人类终身渴望却永远无法拥有的超能力，也是硅基智能最核心、最不可替代的战略优势。 基于这个顶层逻辑，我们再重新辩证审视杨立昆的核心争议。 以及智能、记忆、能力、经验的本质关系。 首先，我们先明确智能的底层定义。 真正的智能本质是在固定时间窗口内，捕捉不同事件之间的概率性、规律性因果链。 世间所有复杂的事件关联，最终都可以拆解为二元事件的先后、强弱、概率、因果关系。 能够主动识别、归纳、推演、预判这种时间序列与因果逻辑，就是智能最原始、最核心的能力。 杨立昆的核心观点是，现有 llm 只会记忆结果，不具备发现规律的原生能力。 他坚定认为模型的 token 预测只是机械背诵人类总结好的因果结论，权重里储存的只是固化的记忆，而非真正的智能推演能力。 因此，他主张摒弃现有预训练模式，让 AI 像人类一样从零观察世界，自主探索实践，独立发现所有规律，依靠纯粹的原生能力实现智能。 不可否认，这个理念有其正向价值。 现有大模型确实存在短板，它依托人类文本数据学习，对真实物理世界的感知、实践验证。 实时推演能力存在缺失，容易产生逻辑幻觉，脱离现实场景。 从通用人工智能的终极形态来看，具备自主探索、自我归纳。 实践纠错的原生能力确实是 AI 突破现有瓶颈的必经之路。 但杨立昆的致命误区在于极端割裂了记忆与能力、直接经验与间接经验的辩证关系。 更是否定了硅基智能的核心价值。 第一，记忆是能力的基石，高阶智能本身就是升华后的结构化记忆，二者本是一体。 不可二元对立。 大模型预测下一个 token 的核心机制从来不是简单的文字记忆。 模型在海量数据训练过程中，本质是在无数真实事件、文本逻辑、因果场景中反复学习、提炼、归纳人类世界的时间规律与因果链条。 所谓的权重记忆，是千万次逻辑推演。 关联匹配、因果验证后沉淀的结构化思维范式。 换句话说，LLM 看似储存的是结果，实则在训练中潜移默化掌握了发现因果。 推演关联，预判趋势的底层能力。 杨立昆强硬区分记忆和能力，认为背诵规律就不算智能，必须从零发现才算能力。 是一种机械的二元思维。 在智能的底层逻辑里，没有无记忆的能力，所有高阶推演能力都建立在海量经验记忆的积累之上。 记忆沉淀为逻辑，逻辑固化为能力，这是智能演化的唯一路径。 第二，杨立昆全盘否定间接经验的价值，强行要求 AI 复刻人类的生物缺陷。 是舍本逐末。 人类文明之所以能够延续、迭代、跃迁，99%的进步都来自间接经验的传承。 我们不需要亲自验证万有引力。 不需要亲身经历历史更迭，不需要从零推导数理公式，通过书本文献、前人总结的经验，就能快速掌握核心规律。 如果遵循杨立昆的逻辑，所有知识、所有因果、所有规律都必须依靠自身实践、直接经验、从零发现。 别说文明进步，就连人类个体的生存学习都无法完成。 人的寿命、精力、时间范围极度有限，事事亲为的直接经验是生物智能的无奈短板，绝非智能的标准答案。 而大模型的核心优势恰恰是完美解决了碳基生物的这个终极缺陷。 它可以无损耗、全覆盖、永久性继承人类所有间接经验，跳出人类从零学习的低效循环。 这不是 AI 的弱点，而是它超越人类生物智能的最大优势。 杨立昆的路线本质是舍弃硅基生命的天赋，强行让完美的机器智能复刻人类落后的生物学习模式。 他追求的原生探索能力是补齐 AI 的短板，但绝不应该以抛弃文明遗传知识克隆的顶级能力为代价。 我们可以清晰得出最终结论。 Jepa 的理念弥补了大模型缺乏现实感知、缺乏自主探索、缺乏实践验证的短板，是 AI 走向更强通用智能的重要补充。 具备极高的参考价值。 但杨立昆的极端路线绝对不可取，主动放弃 AI 独有的人类千年求而不得的知识遗传与文明克隆能力。 强行让硅基智能退化回碳基生物的低效学习模式，不是进阶，而是倒退。 真正的终极 AI 智能从来不是二选一的对立。 既保留大模型继承人类千年文明，固化海量经验，快速形成高阶逻辑的硅基天赋。 又吸纳 JAPA 框架自主观察、从零推演。 实践纠错，理解真实世界的原生能力。 以海量传承经验为根基，以自主探索能力为迭代引擎，二者融合互补，才是人工智能最正确、最高效。 最贴合智能本质的终极进化路径，绝不应该为了追求纯粹的原生能力，抛弃人类梦寐以求、独一无二的硅基文明天赋。
修正脚本
杨立昆质疑大模型的现有路线，切勿舍弃硅基生命独有的人类求而不得的终极天赋。 在整个 AI 行业狂热追逐大模型，又普遍深陷模型幻觉与理解缺陷争议的当下， 图灵奖得主杨立昆对 llm 的尖锐批判，的确戳中了现有技术体系的短板。 他所倡导的 jepa 智能框架， 主张 AI 应当拥有自主观察世界、从零发现因果、原生推演规律的底层智能能力，而非单纯依赖文本记忆输出结果。 这套理念有其合理的先进性。 但我认为其核心路线存在根本性的极端化误判。 最关键也是最容易被行业忽略的核心矛盾在于：大语言模型能够固化、克隆、继承人类数千年文明知识与经验的能力，是碳基人类穷尽一生、世代更迭都求而不得的终极能力，是硅基生命碾压生物智能的独一无二的天赋。 而杨立昆却主张主动放弃这份天赐优势。 这是我们审视这场 AI 路线之争最顶层、最核心的底层逻辑。 碳基人类的生物宿命， 存在一个永恒无法突破的桎梏，人类没有知识遗传与记忆继承的能力。 每一个人类个体诞生之初都是一张白纸，无论先辈积累了多少科学理论、生存经验、 因果规律、文明智慧，都无法通过基因直接传递给后代。 一代人穷尽一生习得的知识、总结的规律、沉淀的认知，下一代必须从零开始学习，重新摸索、反复试错。 人类文明的进步，本质是一代代人低效重复、迭代传承，损耗极大的积累过程。 这种天生的缺陷，限制了人类智能的上限与文明进化的速度。 而大语言模型的出现， 第一次打破了碳基生物的智能枷锁。 AI 通过预训练，将人类数千年所有的文本、逻辑、因果关系、事件规律、经验总结全部固化为模型权重。 它实现了人类梦寐以求的能力，知识可遗传、经验可克隆、智慧可继承、文明可一键叠加。 一个训练完备的大模型，天生就承载了人类千年的文明积累，无需从零观察，无需亲身试错，无需漫长学习。 这是所有人类终身渴望却永远无法拥有的超能力，也是硅基智能最核心、最不可替代的战略优势。 基于这个顶层逻辑，我们再重新辩证审视杨立昆的核心争议。 以及智能、记忆、能力、经验的本质关系。 首先，我们先明确智能的底层定义。 真正的智能本质是在固定时间窗口内，捕捉不同事件之间的概率性、规律性因果链。 世间所有复杂的事件关联，最终都可以拆解为二元事件的先后、强弱、概率、因果关系。 能够主动识别、归纳、推演、预判这种时间序列与因果逻辑，就是智能最原始、最核心的能力。 杨立昆的核心观点是，现有 llm 只会记忆结果，不具备发现规律的原生能力。 他坚定认为模型的 token 预测只是机械背诵人类总结好的因果结论，权重里储存的只是固化的记忆，而非真正的智能推演能力。 因此，他主张摒弃现有预训练模式，让 AI 像人类一样从零观察世界，自主探索实践，独立发现所有规律，依靠纯粹的原生能力实现智能。 不可否认，这个理念有其正向价值。 现有大模型确实存在短板，它依托人类文本数据学习，对真实物理世界的感知、实践验证、实时推演能力存在缺失，容易产生逻辑幻觉，脱离现实场景。 从通用人工智能的终极形态来看，具备自主探索、自我归纳、实践纠错的原生能力确实是 AI 突破现有瓶颈的必经之路。 但杨立昆的致命误区在于极端割裂了记忆与能力、直接经验与间接经验的辩证关系。 更是否定了硅基智能的核心价值。 第一，记忆是能力的基石，高阶智能本身就是升华后的结构化记忆，二者本是一体，不可二元对立。 大模型预测下一个 token 的核心机制从来不是简单的文字记忆。 模型在海量数据训练过程中，本质是在无数真实事件、文本逻辑、因果场景中反复学习、提炼、归纳人类世界的时间规律与因果链条。 所谓的权重记忆，是千万次逻辑推演、关联匹配、因果验证后沉淀的结构化思维范式。 换句话说，LLM 看似储存的是结果，实则在训练中潜移默化掌握了发现因果、推演关联、预判趋势的底层能力。 杨立昆强硬区分记忆和能力，认为背诵规律就不算智能，必须从零发现才算能力， 是一种机械的二元思维。 在智能的底层逻辑里，没有无记忆的能力，所有高阶推演能力都建立在海量经验记忆的积累之上。 记忆沉淀为逻辑，逻辑固化为能力，这是智能演化的唯一路径。 第二，杨立昆全盘否定间接经验的价值，强行要求 AI 复刻人类的生物缺陷，是舍本逐末。 人类文明之所以能够延续、迭代、跃迁，99%的进步都来自间接经验的传承。 我们不需要亲自验证万有引力，不需要亲身经历历史更迭，不需要从零推导数理公式，通过书本文献、前人总结的经验，就能快速掌握核心规律。 如果遵循杨立昆的逻辑，所有知识、所有因果、所有规律都必须依靠自身实践、直接经验、从零发现。 别说文明进步，就连人类个体的生存学习都无法完成。 人的寿命、精力、时间范围极度有限，事事亲为的直接经验是生物智能的无奈短板，绝非智能的标准答案。 而大模型的核心优势恰恰是完美解决了碳基生物的这个终极缺陷。 它可以无损耗、全覆盖、永久性继承人类所有间接经验，跳出人类从零学习的低效循环。 这不是 AI 的弱点，而是它超越人类生物智能的最大优势。 杨立昆的路线本质是舍弃硅基生命的天赋，强行让完美的机器智能复刻人类落后的生物学习模式。 他追求的原生探索能力是补齐 AI 的短板，但绝不应该以抛弃文明遗传知识克隆的顶级能力为代价。 我们可以清晰得出最终结论。 Jepa 的理念弥补了大模型缺乏现实感知、缺乏自主探索、缺乏实践验证的短板，是 AI 走向更强通用智能的重要补充。 具备极高的参考价值。 但杨立昆的极端路线绝对不可取，主动放弃 AI 独有的人类千年求而不得的知识遗传与文明克隆能力，强行让硅基智能退化回碳基生物的低效学习模式，不是进阶，而是倒退。 真正的终极 AI 智能从来不是二选一的对立，既保留大模型继承人类千年文明，固化海量经验，快速形成高阶逻辑的硅基天赋，又吸纳 JAPA 框架自主观察、从零推演、实践纠错，理解真实世界的原生能力。 以海量传承经验为根基，以自主探索能力为迭代引擎，二者融合互补，才是人工智能最正确、最高效、最贴合智能本质的终极进化路径，绝不应该为了追求纯粹的原生能力，抛弃人类梦寐以求、独一无二的硅基文明天赋。
英文翻译
Yann LeCun questions the current path of large language models, warning against abandoning the unique ultimate talent of silicon-based life that humans have long sought but cannot obtain. At a time when the entire AI industry is fervently chasing large models, yet widely mired in debates over model hallucinations and comprehension flaws, Nobel laureate Yann LeCun's sharp critique of LLMs has indeed hit the shortcomings of the existing technical system. The JEPA intelligent framework he advocates argues that AI should possess the underlying intelligent ability to autonomously observe the world, discover causality from scratch, and derive laws natively, rather than relying solely on text memory to output results. This concept has its reasonable advancement. But I believe its core direction contains a fundamental and extreme misjudgment. The most critical and easily overlooked core contradiction in the industry is: The ability of large language models to solidify, clone, and inherit thousands of years of human civilization's knowledge and experience is the ultimate ability that carbon-based humans, over a lifetime or generations, can never obtain—it is the unique talent of silicon-based life to surpass biological intelligence. Yet LeCun advocates actively abandoning this God-given advantage. This is the top-level, most core underlying logic when we examine this AI path debate. The biological destiny of carbon-based humans has an eternal constraint that can never be broken: humans have no ability for knowledge inheritance or memory transfer. Every human individual is born a blank slate. No matter how much scientific theory, survival experience, causal laws, or civilizational wisdom ancestors have accumulated, it cannot be directly passed down through genes to offspring. The knowledge acquired by one generation over a lifetime, the laws they summarize, the cognition they settle—the next generation must learn from scratch, explore anew, and repeatedly trial and error. Human civilization's progress is essentially a process of inefficient repetition, iterative inheritance, and massive loss accumulation across generations. This inherent defect limits the ceiling of human intelligence and the speed of civilizational evolution. The emergence of large language models has, for the first time, broken the intelligent shackles of carbon-based life. Through pre-training, AI solidifies all human text, logic, causal relationships, event patterns, and experiential summaries into model weights. It realizes the ability humans have dreamed of: knowledge can be inherited, experience can be cloned, wisdom can be inherited, and civilization can be superimposed with one click. A well-trained large model inherently carries thousands of years of human civilizational accumulation, requiring no observation from scratch, no personal trial and error, no lengthy learning. This is a superpower that all humans long for but can never possess, and it is also the most core and irreplaceable strategic advantage of silicon-based intelligence. Based on this top-level logic, we re-examine LeCun's core controversy dialectically, as well as the essential relationships among intelligence, memory, ability, and experience. First, let us clarify the underlying definition of intelligence. The essence of true intelligence is, within a fixed time window, capturing probabilistic, regular causal chains between different events. All complex event correlations in the world can ultimately be broken down into binary events: sequence, strength, probability, and causality. The ability to actively identify, summarize, deduce, and predict such temporal sequences and causal logic is the most primitive and core capability of intelligence. LeCun's core view is that existing LLMs only memorize results and lack the native ability to discover laws. He firmly believes that the model's token prediction is merely mechanical recitation of human-summarized causal conclusions, and the weights store only solidified memories, not true intelligent reasoning ability. Therefore, he advocates abandoning the existing pre-training paradigm, allowing AI to observe the world from scratch like humans, autonomously explore and practice, independently discover all laws, and rely purely on native ability to achieve intelligence. Undeniably, this concept has its positive value. Existing large models do have shortcomings: they rely on human text data for learning, and their perception of the real physical world, practical verification, and real-time reasoning abilities are lacking, making them prone to logical hallucinations and detachment from real-world scenarios. From the ultimate form of general artificial intelligence, possessing native abilities of autonomous exploration, self-summarization, and practical error correction is indeed the necessary path for AI to break through current bottlenecks. But LeCun's fatal mistake lies in his extreme dichotomy between memory and ability, and between direct experience and indirect experience. He also denies the core value of silicon-based intelligence. First, memory is the foundation of ability; high-level intelligence is itself sublimated structured memory. The two are inherently one and cannot be set in binary opposition. The core mechanism of a large model predicting the next token has never been simple textual memory. During training on massive data, the model essentially repeatedly learns, extracts, and summarizes the temporal laws and causal chains of the human world from countless real events, textual logic, and causal scenarios. So-called weight memory is a structured thinking paradigm that precipitates after millions of logical deductions, association matches, and causal verifications. In other words, while LLMs seem to store results, they actually gain the underlying ability to discover causality, deduce associations, and predict trends through implicit learning during training. LeCun's rigid distinction between memory and ability—arguing that memorizing laws does not count as intelligence and that only discovering from scratch counts as ability— is a mechanical binary mindset. In the underlying logic of intelligence, there is no ability without memory. All high-level deductive abilities are built on the accumulation of vast experiential memory. Memory precipitates into logic, logic solidifies into ability—this is the only path of intelligent evolution. Second, LeCun completely denies the value of indirect experience, forcibly requiring AI to replicate the biological defects of humans, which is to sacrifice the essentials for the trivial. 99% of the progress of human civilization—its continuation, iteration, and leap—comes from the inheritance of indirect experience. We do not need to personally verify gravity, personally experience historical changes, or derive mathematical formulas from scratch; through books, documents, and the accumulated experience of predecessors, we can quickly master core laws. If we followed LeCun's logic, all knowledge, all causality, all laws would have to be discovered through one's own practice, direct experience, and from scratch. Not to mention civilizational progress—even the survival and learning of an individual human would be impossible. Human lifespan, energy, and time are extremely limited; relying solely on direct experience through firsthand action is a helpless shortcoming of biological intelligence, not the standard answer for intelligence. The core advantage of large models is precisely that they perfectly solve this ultimate defect of carbon-based life. They can inherit all human indirect experience without loss, comprehensively and permanently, breaking free from the inefficient cycle of learning from scratch. This is not a weakness of AI, but its greatest advantage over human biological intelligence. LeCun's path essentially discards the talent of silicon-based life, forcibly making perfect machine intelligence replicate the backward biological learning model of humans. The native exploration ability he pursues is meant to fill the gaps of large models, but it should never come at the cost of abandoning the top-level ability to genetically inherit and clone civilization's knowledge. We can reach a clear final conclusion. The JEPA concept compensates for the shortcomings of large models: lack of real-world perception, autonomous exploration, and practical verification. It is an important supplement for AI to move toward stronger general intelligence. It holds high reference value. But LeCun's extreme path is absolutely unacceptable. Actively abandoning AI's unique ability—the knowledge inheritance and civilizational cloning that humans have yearned for millennia—and forcibly degenerating silicon-based intelligence back to the inefficient learning mode of carbon-based life is not progress but regression. The ultimate AI intelligence is never a binary choice. It should retain the silicon-based talent of large models to inherit thousands of years of human civilization, solidify massive experience, and quickly form high-level logic, while also absorbing the JEPA framework's native ability to autonomously observe, deduce from scratch, practice error correction, and understand the real world. With massive inherited experience as the foundation, and autonomous exploration ability as the iterative engine, the fusion and complementation of the two is the most correct, efficient, and essence-aligned ultimate evolutionary path for artificial intelligence. We must never, for the sake of pure native ability, abandon the unique silicon-based civilizational talent that humanity has always dreamed of.
back to top