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AI革命的开端
视频
音频
原始脚本
破除十年认知误区， AlphaGo 不是 AI 的开端， Transformer 才是时代拐点。 回望过去十年人工智能的发展舆论，大众始终存在一个根深蒂固的认知误区。 很多人将2017年前后， AlphaGo AlphaZero 的横空出世，视作当代人工智能爆发的起点，认为这套自我博弈、从零学习的算法体系。 开启了真正的人工智能时代。 但经过对算法原理、训练逻辑与技术范式的深度拆解后，我彻底推翻了这个延续十年的固有认知。 AlphaGo 与 AlphaZero 从来不是人工智能的革命性突破，它们只是传统博弈算法的工程极致升级。 真正重塑 AI 行业，开启通用智能时代的核心拐点。 是 Transformer 架构的诞生。 二者看似同属 AI 技术，实则是完全不同的技术赛道。 当代 AI 爆发的根源与 DeepMind 的博弈 AI 体系几乎没有本质关联。 从技术底层溯源， AlphaGo AlphaZero 的内核从未跳出传统机器博弈的框架，本质上是深蓝计算机的高阶迭代产物。 从早年国际象棋 ai 深蓝，到围棋领域的 Alpha 系列模型，其核心逻辑一脉相承，都是基于博弈树搜索加减值优化的封闭场景决策算法。 传统深蓝依靠 minimax 极大极小算法与人工定义的估值函数，通过确定性剪纸完成棋局推演。 而 Alpha 系列模型只是更新了搜索范式，用 mcts 蒙特卡洛树搜索替代了老旧的固定剪纸逻辑，搭配神经网络的概率评估能力，完成了算法的升级。 却从未突破底层技术逻辑。 围棋作为典型的 NP 难问题，拥有361个落子选择，指数级增长的分支空间是暴力穷举算力永远无法破解的难题。 这也是围棋长期被视作 AI 禁区的核心原因。 而 AlphaZero 能够攻克围棋，并非实现了智能层面的突破，而是依靠精妙的工程优化。 将无法计算的 NP 难问题通过概率筛选、软剪枝的方式压缩为多项式时间可解的 P 级问题。 围棋盘面看似选择极多。 但绝大多数落子都是无战略价值、低胜率的无效分支。 即便是人类顶尖棋手，也会本能舍弃这些无效走法。 AlphaZero 的核心优势。 就是用神经网络输出全局条件概率分布，对所有落子动作进行概率排序，仅保留 Top K 高潜力分支做深度推演，直接舍弃海量低概率无效路径。 从根源上规避了搜索树的指数爆炸。 这套逻辑和人类下棋思维高度一致。 普通人看三五步，职业棋手能预判十几步。 本质都是择优推演，取舍分支。 Alphazero 战胜人类九段棋手，依靠的从来不是自主认知与智慧，而是无限堆叠的算力、并行计算能力，以及远超人类的搜索深度和广度。 人类受限于记忆、运算、顺势思考的生理极限。 无法完成数十步的全局推演，而机器可以依靠物理算力优势无限延伸搜索边界。 这是纯粹的物理碾压，而非智能的超越。 大众曾经神话的 AlphaZero 从零自学，无人类棋谱，无人工知识，纯粹强化学习进化。 本质上也是一场认知迷雾。 存在极大的夸大成分。 我们传统认知中的纯强化学习，核心痛点是稀疏延迟奖励，模型只有终局输赢这一个全局反馈。 数百步对弈过程中没有任何纠错信号，无法定位哪一步决策导致胜负。 在低通过率的复杂场景中，有效学习信息量趋近于零。 几乎无法完成有效进化。 而 AlphaZero 看似是自我博弈的强化学习，实则内核完全是逐步骤的监督学习。 它所谓的自我进化。 并非依靠终局输赢的模糊反馈迭代，每一步决策都有 MCTS 搜索生成的精准标准答案作为校准依据。 在每一次落子前， MCTS 都会基于当前局面实时推演。 回溯估值，生成最优的动作概率分布，以此指导神经网络更新权重。 换言之， AlphaZero 的整个训练过程，每一步都有明确的纠错信号、精准的学习目标。 完全规避了纯强化学习信号稀疏、无法归因的致命缺陷。 它舍弃的只是人类棋谱的外部知识库，却通过自我推演搭建了一套实时自我指导体系。 看似无师自通，实则步步有凭有据，根本不是大众理解的盲目试错、自主悟道。 基于这套完整的逻辑，我们可以清晰界定两大 AI 技术体系的本质差异。 AlphaGo AlphaZero 属于封闭场景的专用搜索优化工程，仅适用于围棋、象棋这类有固定规则、固定动作空间、封闭边界的博弈场景。 它无法理解语义、无法提炼抽象逻辑、无法迁移跨领域知识、没有任何通用认知能力，只是一套极致优化的专用决策计算器。 而 Transformer 架构才是当代人工智能真正的范式革命与发展源头。 它彻底跳出了固定规则、封闭场景的束缚，能够处理自然语言、海量人类知识。 逻辑推理、内容创作等开放世界的复杂时序问题。 大语言模型依托 Transformer 的注意力机制，实现了上下文关联知识萃取。 逻辑归纳与通用交互，真正搭建起了机器理解人类世界、适配开放场景的智能体系。 这也是近十年 AI 全民爆发。 落地千行百业的核心根基。 历经十年认知偏差，终于厘清 AI 发展的核心脉络。 Alpha 系列博弈 AI 只是传统机器博弈技术的终点，而非通用人工智能的起点。 它的惊艳表现让大众误判了 AI 的发展方向，掩盖了 Transformer 架构真正的革命性价值。 所谓的 AI 时代爆发，从来不是源于搜索算法的极致优化，而是源于通用认知范式的诞生。 封闭场景的极致搜索，终究只是工程技术的迭代。 开放世界的通用理解，才是人工智能真正的开端。 这也是这场深度拆解带给我最核心、最深刻的认知革新。
修正脚本
破除十年认知误区， AlphaGo 不是 AI 的开端， Transformer 才是时代拐点。 回望过去十年人工智能的发展舆论，大众始终存在一个根深蒂固的认知误区。 很多人将2017年前后， AlphaGo AlphaZero 的横空出世，视作当代人工智能爆发的起点，认为这套自我博弈、从零学习的算法体系。 开启了真正的人工智能时代。 但经过对算法原理、训练逻辑与技术范式的深度拆解后，我彻底推翻了这个延续十年的固有认知。 AlphaGo 与 AlphaZero 从来不是人工智能的革命性突破，它们只是传统博弈算法的工程极致升级。 真正重塑 AI 行业，开启通用智能时代的核心拐点。 是 Transformer 架构的诞生。 二者看似同属 AI 技术，实则是完全不同的技术赛道。 当代 AI 爆发的根源与 DeepMind 的博弈 AI 体系几乎没有本质关联。 从技术底层溯源， AlphaGo AlphaZero 的内核从未跳出传统机器博弈的框架，本质上是深蓝计算机的高阶迭代产物。 从早年国际象棋 ai 深蓝，到围棋领域的 Alpha 系列模型，其核心逻辑一脉相承，都是基于博弈树搜索加减值优化的封闭场景决策算法。 传统深蓝依靠 minimax 极大极小算法与人工定义的估值函数，通过确定性剪枝完成棋局推演。 而 Alpha 系列模型只是更新了搜索范式，用 mcts 蒙特卡洛树搜索替代了老旧的固定剪枝逻辑，搭配神经网络的概率评估能力，完成了算法的升级。 却从未突破底层技术逻辑。 围棋作为典型的 NP 难问题，拥有361个落子选择，指数级增长的分支空间是暴力穷举算力永远无法破解的难题。 这也是围棋长期被视作 AI 禁区的核心原因。 而 AlphaZero 能够攻克围棋，并非实现了智能层面的突破，而是依靠精妙的工程优化。 将无法计算的 NP 难问题通过概率筛选、软剪枝的方式压缩为多项式时间可解的 P 级问题。 围棋盘面看似选择极多。 但绝大多数落子都是无战略价值、低胜率的无效分支。 即便是人类顶尖棋手，也会本能舍弃这些无效走法。 AlphaZero 的核心优势。 就是用神经网络输出全局条件概率分布，对所有落子动作进行概率排序，仅保留 Top K 高潜力分支做深度推演，直接舍弃海量低概率无效路径。 从根源上规避了搜索树的指数爆炸。 这套逻辑和人类下棋思维高度一致。 普通人看三五步，职业棋手能预判十几步。 本质都是择优推演，取舍分支。 Alphazero 战胜人类九段棋手，依靠的从来不是自主认知与智慧，而是无限堆叠的算力、并行计算能力，以及远超人类的搜索深度和广度。 人类受限于记忆、运算、瞬时思考的生理极限。 无法完成数十步的全局推演，而机器可以依靠物理算力优势无限延伸搜索边界。 这是纯粹的物理碾压，而非智能的超越。 大众曾经神话的 AlphaZero 从零自学，无人类棋谱，无人工知识，纯粹强化学习进化。 本质上也是一场认知迷雾。 存在极大的夸大成分。 我们传统认知中的纯强化学习，核心痛点是稀疏延迟奖励，模型只有终局输赢这一个全局反馈。 数百步对弈过程中没有任何纠错信号，无法定位哪一步决策导致胜负。 在低回报的复杂场景中，有效学习信息量趋近于零。 几乎无法完成有效进化。 而 AlphaZero 看似是自我博弈的强化学习，实则内核完全是逐步骤的监督学习。 它所谓的自我进化。 并非依靠终局输赢的模糊反馈迭代，每一步决策都有 MCTS 搜索生成的精准标准答案作为校准依据。 在每一次落子前， MCTS 都会基于当前局面实时推演。 回溯估值，生成最优的动作概率分布，以此指导神经网络更新权重。 换言之， AlphaZero 的整个训练过程，每一步都有明确的纠错信号、精准的学习目标。 完全规避了纯强化学习信号稀疏、无法归因的致命缺陷。 它舍弃的只是人类棋谱的外部知识库，却通过自我推演搭建了一套实时自我指导体系。 看似无师自通，实则步步有凭有据，根本不是大众理解的盲目试错、自主悟道。 基于这套完整的逻辑，我们可以清晰界定两大 AI 技术体系的本质差异。 AlphaGo AlphaZero 属于封闭场景的专用搜索优化工程，仅适用于围棋、象棋这类有固定规则、固定动作空间、封闭边界的博弈场景。 它无法理解语义、无法提炼抽象逻辑、无法迁移跨领域知识、没有任何通用认知能力，只是一套极致优化的专用决策计算器。 而 Transformer 架构才是当代人工智能真正的范式革命与发展源头。 它彻底跳出了固定规则、封闭场景的束缚，能够处理自然语言、海量人类知识、 逻辑推理、内容创作等开放世界的复杂时序问题。 大语言模型依托 Transformer 的注意力机制，实现了上下文关联知识萃取、 逻辑归纳与通用交互，真正搭建起了机器理解人类世界、适配开放场景的智能体系。 这也是近十年 AI 全民爆发、 落地千行百业的核心根基。 历经十年认知偏差，终于厘清 AI 发展的核心脉络。 Alpha 系列博弈 AI 只是传统机器博弈技术的终点，而非通用人工智能的起点。 它的惊艳表现让大众误判了 AI 的发展方向，掩盖了 Transformer 架构真正的革命性价值。 所谓的 AI 时代爆发，从来不是源于搜索算法的极致优化，而是源于通用认知范式的诞生。 封闭场景的极致搜索，终究只是工程技术的迭代。 开放世界的通用理解，才是人工智能真正的开端。 这也是这场深度拆解带给我最核心、最深刻的认知革新。
英文翻译
Dispelling a Decade of Misconceptions: AlphaGo Is Not the Beginning of AI; Transformer Is the True Turning Point. Looking back at the public discourse on AI development over the past decade, there has been a deeply ingrained misconception. Many regard the emergence of AlphaGo and AlphaZero around 2017 as the starting point of the contemporary AI boom, believing that this system of self-play and zero-learning algorithms ushered in the true era of artificial intelligence. However, after a thorough deconstruction of the algorithmic principles, training logic, and technical paradigms, I have completely overturned this entrenched view that has persisted for ten years. AlphaGo and AlphaZero were never revolutionary breakthroughs in artificial intelligence—they were merely extreme engineering optimizations of traditional game-playing algorithms. The true inflection point that reshaped the AI industry and ushered in the era of general intelligence was the birth of the Transformer architecture. Although both appear to belong to the same AI technology domain, they are actually entirely different technical tracks. The root of the contemporary AI explosion has almost no essential connection with DeepMind's game-playing AI systems. Tracing back to the technical foundations, the core of AlphaGo and AlphaZero never escaped the framework of traditional machine game-playing. In essence, they are advanced iterations of Deep Blue. From the early chess AI Deep Blue to the Alpha series models for Go, the core logic has been consistent: closed-scenario decision algorithms based on game-tree search combined with value optimization. Traditional Deep Blue relied on the minimax algorithm and manually defined evaluation functions, using deterministic pruning to simulate chess positions. The Alpha series models merely updated the search paradigm, replacing the old fixed pruning logic with Monte Carlo tree search (MCTS), and enhanced it with neural network probability estimation to achieve an algorithmic upgrade—but never broke through the underlying technical logic. Go, as a classic NP-hard problem with 361 possible moves per turn, has a branching space that grows exponentially, making brute-force computation forever impossible. This is the core reason why Go was long regarded as an AI forbidden zone. AlphaZero's ability to conquer Go was not due to a breakthrough in intelligence but rather to clever engineering optimization: it compressed the unsolvable NP-hard problem into a polynomial-time solvable P-level problem through probabilistic screening and soft pruning. A Go board may seem to offer countless choices, but the vast majority of moves are strategically worthless, low-probability losing branches. Even top human players instinctively discard these ineffective moves. AlphaZero's core advantage lies in using a neural network to output a global conditional probability distribution for all moves, ranking them by probability, and retaining only the top-K high-potential branches for deep simulation, directly discarding the vast number of low-probability dead ends. This fundamentally avoids the exponential explosion of the search tree. This logic is highly consistent with human playing strategies. Ordinary players think three to five moves ahead; professional players can anticipate dozens. In essence, both are about selecting the best branches to explore while discarding others. AlphaZero defeated top human Go players not through autonomous cognition or wisdom, but through infinitely stacked computing power, parallel processing capabilities, and a search depth and breadth far beyond human limits. Humans are constrained by the physiological limits of memory, computation, and instantaneous thinking—they cannot perform dozens of steps of global simulation. Machines, however, can leverage physical computing advantages to infinitely extend the search boundary. This is purely physical dominance, not an intellectual transcendence. The once-mythologized claim that AlphaZero learned from scratch—no human game records, no human knowledge, purely through reinforcement learning evolution—is essentially a cognitive fog, heavily exaggerated. The core pain point of traditional pure reinforcement learning is sparse and delayed rewards: the model receives only one global feedback signal at the end of the game (win or lose). Over hundreds of moves, there is no corrective signal to pinpoint which decision led to the outcome. In low-reward complex scenarios, the amount of effective learning information approaches zero, making it nearly impossible to achieve meaningful evolution. AlphaZero, despite appearing to be self-play reinforcement learning, actually has a core that is step-by-step supervised learning. Its so-called self-evolution does not rely on the vague feedback of final wins or losses. Each decision step has a precise standard answer generated by MCTS search as a calibration target. Before each move, MCTS performs real-time simulation from the current position, backpropagates valuation, and generates an optimal action probability distribution to guide the neural network's weight update. In other words, throughout AlphaZero's training process, every single step has a clear corrective signal and a precise learning target. It completely avoids the fatal flaw of pure reinforcement learning—sparse signals and the inability to attribute outcomes. It merely discards the external knowledge base of human game records, but builds a real-time self-guidance system through self-simulation. Seemingly self-taught, every step is actually evidence-based—far from the random trial-and-error and spontaneous insight that the public imagines. Based on this complete logic, we can clearly define the essential differences between the two major AI technical systems. AlphaGo and AlphaZero belong to closed-scenario dedicated search optimization engineering, applicable only to game scenarios with fixed rules, fixed action spaces, and closed boundaries, such as Go and chess. They cannot understand semantics, extract abstract logic, transfer cross-domain knowledge, or possess any general cognitive ability—they are just extremely optimized special-purpose decision calculators. In contrast, the Transformer architecture is the true paradigm shift and origin of modern AI. It completely breaks free from the constraints of fixed rules and closed scenarios, capable of handling complex open-world sequential problems such as natural language, vast human knowledge, logical reasoning, and content creation. Large language models, built on Transformer's attention mechanism, achieve contextual knowledge extraction, logical induction, and general interaction—truly establishing an intelligent system for machines to understand the human world and adapt to open scenarios. This is also the core foundation of the AI explosion over the past decade and its implementation across countless industries. After a decade of cognitive bias, the core trajectory of AI development is finally clarified. The Alpha series game-playing AI is merely the endpoint of traditional machine game-playing technology, not the starting point of general artificial intelligence. Its stunning performance misled the public about the direction of AI development, obscuring the truly revolutionary value of the Transformer architecture. The so-called AI explosion was never driven by extreme optimization of search algorithms, but by the birth of a general-purpose cognitive paradigm. Extreme search in closed scenarios is ultimately just an iteration of engineering technology. General understanding of the open world is the true beginning of artificial intelligence—this is the most core and profound cognitive revolution brought to me by this deep deconstruction.
back to top