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
阿尔法狗不是真智能
视频
音频
原始脚本
AlphaGo 不是真智能，别再被十年神话骗了！今天我们把一件被神话了十年的事彻底说透，AlphaGo 不是真智能。 AlphaGo 不是真智能。 AlphaGo 不是真智能，它不仅没有开创当代人工智能的革命，甚至和我们今天熟悉的 GPT 文心一言、各类通用大模型。 几乎没有任何血缘与技术传承关系。 一， AlphaGo 的本质，究极升级版深蓝，仅此而已。 一九九七年。 IBM 深蓝战胜国际象棋大师卡斯帕罗夫，全世界惊呼人工智能时代来临。 事后几十年回头看，深蓝是什么？一套人工写死估值规则、暴力减脂、极致算力的专用博弈搜索机器。 它只会下国际象棋，换一个棋盘，换一个规则。 立刻彻底失灵，没有任何举一反三的能力，更谈不上通用智能。 2016年， AlphaGo 战胜李世石，全世界再一次陷入同样的狂热。 所有人都以为真正的人工智能觉醒了，起点近在眼前。 可剥开所有营销包装，二者底层逻辑一模一样。 深蓝等于固定规则加 minimax 减支加暴力算力。 AlphaGo 等于固定规则加 MCTS 蒙特卡洛减支加神经网络概率筛选加暴涨万倍的现代算力。 它赢人类从来不是因为诞生了自主智慧，而是人类一辈子最多能往前预判十几步棋，而它靠着算力和极致减支，能高效推演远超人类的深度。 人类会疲劳、会失误、会情绪波动，而机器永远冷静精准，仅此而已。 二、被刻意夸大的从零自主强化学习。 只是营销迷雾大众最大的误解就是相信 Alphazero 是完全从零摸索、无师自通、自主领悟其道的奇迹。 真相残酷且清晰。 它确实抛弃了人类棋谱、人类定式，实现了白板开局、自我对弈。 但它从来不是纯粹靠终局输赢、盲目试错的纯强化学习。 他的每一步落子，都有 mcts 蒙特卡洛搜索树当场推演，给出最优走法的标准答案。 他的神经网络全程都在拟合搜索给出的精准标签。 本质是步步都被指引的监督学习。 所谓的自我进化，本质是自己给自己出题，自己给自己批改，全程没有真正的自主顿悟。 DeepMind 用话术把一套高级搜索工程包装成了机器自主觉醒的神话，成功俘获了全世界的想象。 三、致命短板。 零泛化能力永远走不出棋盘。 区分专用计算器和真正智能的唯一标准就是泛化能力。 什么是真智能？学会一个逻辑，就能迁移到万千不同场景。 会说话、会推理、会写文章、懂常识，能解决从未见过的全新开放问题。 而 AlphaGo 体系天生没有这个能力，它只会下围棋。 除此之外什么都不会。 你无法把这套围棋搜索模型改成下象棋、打扑克，更不可能让它去写文字、做规划、理解人类语言。 应对开放世界。 它的整个架构从根上就是为封闭、固定、规则完备的博弈场景量身定做，无法横向迁移。 哪怕是 DeepMind 最引以为傲的 AlphaFold 蛋白质结构预测。 本质上依然是 AlphaGo 路线的延伸。 把分子结构问题转化为一个超高维空间的搜索、匹配、最优解求解问题，依旧是封闭场景下的精准搜索工程。 它有极高的科研价值、工业价值，却依然不是通用智能。 四、真正的 AI 革命和 AlphaGo 毫无关系，很多人颠倒了时间线和因果。 AlphaGo 击败李世石，2016年 Transformer 划时代论文发布，2017年 Transformer 诞生之后，才有了大语言模型，才有了 AI 的全民爆发。 才有了今天改变各行各业的通用人工智能浪潮。 两条路线从出生起就背道而驰。 Transformer 路线模仿人类经验。 承载全人类知识。 开放世界及强泛化，无限可能性。 哪怕有幻觉，有缺陷，却是真正能复制、留存、延续人类文明的底座。 Alphago 路线，封闭规则，纯理性搜索，极致可控，零幻觉。 落地扎实，却天生锁死在单一专用场景里，无法通用。 谷歌作为 Transformer 的亲生父母，最早看透了两条路线的优缺点。 AlphaGo 路线稳，可控，风险低，商业落地扎实。 Transformer 路线野，不可控，风险高，商业化前路模糊。 却拥有无限的想象空间。 于是谷歌陷入了长达十年的战略犹豫与摇摆。 DeepMind 的强势崛起与路线拉扯，让谷歌长期双线分散资源。 迟迟没有 all in 大模型赛道，最终起了个大早，赶了个晚集。 五、最终结论， DeepMind 很强，但谷歌压错了主赛道。 我们从不否认 DeepMind 的天才与伟大。 Alphago 是工程史上的奇迹，Alphafold 更是改写了整个生物科研与医药行业，硬核价值无可估量。 但我们必须清醒承认。 一、 AlphaGo 不是当代 AI 的开端，它只是深蓝的终极升级版，一场惊艳的技术表演，而非范式革命。 二、他所谓的自主学习是经过营销美化的概念，内核依旧是步步引导的高级搜索。
修正脚本
AlphaGo 不是真智能，别再被十年神话骗了！今天我们把一件被神话了十年的事彻底说透，AlphaGo 不是真智能。 AlphaGo 不是真智能。 AlphaGo 不是真智能，它不仅没有开创当代人工智能的革命，甚至和我们今天熟悉的 GPT、文心一言、各类通用大模型几乎没有任何血缘与技术传承关系。 一、AlphaGo 的本质，究极升级版深蓝，仅此而已。 一九九七年。 IBM 深蓝战胜国际象棋大师卡斯帕罗夫，全世界惊呼人工智能时代来临。 事后几十年回头看，深蓝是什么？一套人工写死估值规则、暴力剪枝、极致算力的专用博弈搜索机器。 它只会下国际象棋，换一个棋盘，换一个规则，立刻彻底失灵，没有任何举一反三的能力，更谈不上通用智能。 2016年， AlphaGo 战胜李世石，全世界再一次陷入同样的狂热。 所有人都以为真正的人工智能觉醒了，起点近在眼前。 可剥开所有营销包装，二者底层逻辑一模一样。 深蓝等于固定规则加 minimax 剪枝加暴力算力。 AlphaGo 等于固定规则加 MCTS 蒙特卡洛剪枝加神经网络概率筛选加暴涨万倍的现代算力。 它赢人类从来不是因为诞生了自主智慧，而是人类一辈子最多能往前预判十几步棋，而它靠着算力和极致剪枝，能高效推演远超人类的深度。 人类会疲劳、会失误、会情绪波动，而机器永远冷静精准，仅此而已。 二、被刻意夸大的从零自主强化学习。这只是营销迷雾，大众最大的误解就是相信 Alphazero 是完全从零摸索、无师自通、自主领悟其道的奇迹。 真相残酷且清晰。 它确实抛弃了人类棋谱、人类定式，实现了白板开局、自我对弈。 但它从来不是纯粹靠终局输赢、盲目试错的纯强化学习。 它的每一步落子，都有 mcts 蒙特卡洛搜索树当场推演，给出最优走法的标准答案。 它的神经网络全程都在拟合搜索给出的精准标签。 本质是步步都被指引的监督学习。 所谓的自我进化，本质是自己给自己出题，自己给自己批改，全程没有真正的自主顿悟。 DeepMind 用话术把一套高级搜索工程包装成了机器自主觉醒的神话，成功俘获了全世界的想象。 三、致命短板。零泛化能力永远走不出棋盘。 区分专用计算器和真正智能的唯一标准就是泛化能力。 什么是真智能？学会一个逻辑，就能迁移到万千不同场景。 会说话、会推理、会写文章、懂常识，能解决从未见过的全新开放问题。 而 AlphaGo 体系天生没有这个能力，它只会下围棋。 除此之外什么都不会。 你无法把这套围棋搜索模型改成下象棋、打扑克，更不可能让它去写文字、做规划、理解人类语言，应对开放世界。 它的整个架构从根上就是为封闭、固定、规则完备的博弈场景量身定做，无法横向迁移。 哪怕是 DeepMind 最引以为傲的 AlphaFold 蛋白质结构预测。 本质上依然是 AlphaGo 路线的延伸。 把分子结构问题转化为一个超高维空间的搜索、匹配、最优解求解问题，依旧是封闭场景下的精准搜索工程。 它有极高的科研价值、工业价值，却依然不是通用智能。 四、真正的 AI 革命和 AlphaGo 毫无关系，很多人颠倒了时间线和因果。 2016年 AlphaGo 击败李世石，同年 Transformer 划时代论文发布，2017年 Transformer 诞生之后，才有了大语言模型，才有了 AI 的全民爆发。 才有了今天改变各行各业的通用人工智能浪潮。 两条路线从出生起就背道而驰。 Transformer 路线模仿人类经验。 承载全人类知识。 开放世界、强泛化，无限可能性。 哪怕有幻觉，有缺陷，却是真正能复制、留存、延续人类文明的底座。 Alphago 路线，封闭规则，纯理性搜索，极致可控，零幻觉。 落地扎实，却天生锁死在单一专用场景里，无法通用。 谷歌作为 Transformer 的亲生父母，最早看透了两条路线的优缺点。 AlphaGo 路线稳，可控，风险低，商业落地扎实。 Transformer 路线野，不可控，风险高，商业化前路模糊。 却拥有无限的想象空间。 于是谷歌陷入了长达十年的战略犹豫与摇摆。 DeepMind 的强势崛起与路线拉扯，让谷歌长期双线分散资源。 迟迟没有 all in 大模型赛道，最终起了个大早，赶了个晚集。 五、最终结论， DeepMind 很强，但谷歌压错了主赛道。 我们从不否认 DeepMind 的天才与伟大。 Alphago 是工程史上的奇迹，Alphafold 更是改写了整个生物科研与医药行业，硬核价值无可估量。 但我们必须清醒承认。 一、 AlphaGo 不是当代 AI 的开端，它只是深蓝的终极升级版，一场惊艳的技术表演，而非范式革命。 二、它所谓的自主学习是经过营销美化的概念，内核依旧是步步引导的高级搜索。
英文翻译
AlphaGo is not true intelligence. Don’t be fooled by the decade-long myth. Today, we will thoroughly expose something that has been mythologized for ten years: AlphaGo is not true intelligence. AlphaGo is not true intelligence. AlphaGo is not true intelligence. Not only did it not pioneer the modern AI revolution, but it also has almost no genetic or technical lineage with the GPT, ERNIE Bot, and various general large models we are familiar with today. I. The essence of AlphaGo is merely an upgraded version of Deep Blue, nothing more. In 1997, IBM’s Deep Blue defeated chess grandmaster Garry Kasparov, and the world exclaimed that the era of artificial intelligence had arrived. Looking back decades later, what was Deep Blue? A specialized game-search machine with manually hardcoded evaluation rules, brute-force pruning, and extreme computing power. It could only play chess. Change the board, change the rules, and it would immediately fail completely. It had no ability to generalize, let alone general intelligence. In 2016, AlphaGo defeated Lee Sedol, and the world once again fell into the same frenzy. Everyone thought true AI had awakened, and the starting point was just around the corner. But strip away all the marketing packaging, and the underlying logic of the two is identical. Deep Blue equals fixed rules + minimax pruning + brute-force computing. AlphaGo equals fixed rules + MCTS (Monte Carlo Tree Search) + neural network probabilistic screening + computing power multiplied ten thousand times. It never defeated humans because it possessed autonomous intelligence; rather, humans can only predict a dozen or so moves ahead in their lifetime, while it, relying on computing power and extreme pruning, could efficiently simulate far deeper than any human. Humans get tired, make mistakes, and have emotional fluctuations, while the machine remains forever calm and precise—nothing more. II. The exaggerated "self-play reinforcement learning from scratch" is just marketing fog. The public’s biggest misconception is believing that AlphaZero, by starting from a blank slate, self-taught and independently discovered the true path. The truth is harsh and clear. It did discard human game records and human joseki, achieving a blank-board opening and self-play. But it was never purely blind trial-and-error reinforcement learning based solely on game outcomes. Every move it made was instantaneously simulated by the MCTS search tree, providing the optimal move as the standard answer. Its neural network was entirely engaged in fitting the precise labels given by the search. Essentially, it was supervised learning with guidance at every step. The so-called self-evolution was essentially generating its own problems and grading its own answers, with no true autonomous insight throughout the process. DeepMind used rhetoric to package an advanced search project as a myth of machine self-awakening, successfully capturing the world’s imagination. III. Fatal flaw: zero generalization ability, forever trapped on the board. The only standard to distinguish specialized calculators from true intelligence is generalization ability. What is true intelligence? Learning one logic and being able to transfer it to thousands of different scenarios. Being able to speak, reason, write articles, possess common sense, and solve novel open-ended problems never seen before. The AlphaGo system inherently lacks this ability. It can only play Go. Beyond that, it can do nothing. You cannot transform this Go search model to play chess or poker, let alone have it write text, make plans, understand human language, or cope with an open world. Its entire architecture is tailor-made for closed, fixed, rule-complete game scenarios and cannot be transferred horizontally. Even DeepMind’s most proud AlphaFold, for protein structure prediction, is essentially an extension of the AlphaGo path. It transforms molecular structure problems into a search, matching, and optimal-solution problem in a ultra-high-dimensional space—still precise search engineering in a closed scenario. It holds tremendous scientific and industrial value, yet it is still not general intelligence. IV. The real AI revolution has nothing to do with AlphaGo. Many people have reversed the timeline and causality. In 2016, AlphaGo defeated Lee Sedol. In the same year, the groundbreaking Transformer paper was published. In 2017, after the birth of the Transformer, large language models emerged, leading to the massive explosion of AI across all fields, and finally the wave of general AI that is transforming various industries today. These two paths diverged from the start. The Transformer path mimics human experience, carries the knowledge of all humanity, and offers an open world, strong generalization, and infinite possibilities. Despite hallucinations and flaws, it is a true foundation that can replicate, preserve, and continue human civilization. The AlphaGo path, on the other hand, operates in a closed rule set, pure rational search, extreme controllability, zero hallucinations, and solid implementation, but is inherently locked into a single specialized scenario, unable to generalize. Google, as the parent of the Transformer, was the first to see the pros and cons of both paths. The AlphaGo path is stable, controllable, low-risk, and commercially viable. The Transformer path is wild, uncontrollable, high-risk, with an unclear commercial path but unlimited imagination. Thus, Google fell into a decade-long strategic hesitation and vacillation. The strong rise of DeepMind and the tug of war between paths caused Google to spread resources across two lines for a long time, delaying its full commitment to the large model track, ultimately arriving early but missing the boat. V. Final conclusion: DeepMind is strong, but Google bet on the wrong main track. We never deny DeepMind’s genius and greatness. AlphaGo is a miracle in engineering history, and AlphaFold has rewritten the entire biological research and pharmaceutical industry, with hardcore value beyond measure. But we must soberly acknowledge: First, AlphaGo is not the beginning of modern AI. It is merely the ultimate upgrade of Deep Blue—a stunning technical performance, not a paradigm revolution. Second, its so-called autonomous learning is a concept embellished by marketing; its core remains step-by-step guided advanced search.
back to top