AI 与机器学习2020nature.com
为什么通用人工智能无法实现
Why general artificial intelligence will not be realized
进阶 · 约 30 分钟读完前置知识：图灵测试、默会知识、具身认知
导读
这篇论文主张通用人工智能（强 AI）在原则上不可能实现，核心论据复兴了 Hubert Dreyfus 的观点：计算机没有身体、没有童年、没有文化实践，也不“在世界之中”，因此无法获得人类式的理解，不能通过完整图灵测试。文章主要是哲学论证，而非技术路线分析。
HN 关注点：Nature 为何刊发这样一篇被批评为定义含糊、二元论甚至“神学”的反 AGI 文章，以及具身性和默会知识是否真能否定通用人工智能。
HN 评论精选
这几条评论里：质疑 4
质疑作者的核心论证依赖默会知识的不可编码性，本质上是一种二元论/神学立场。
对我来说，关键句是：“Hubert Dreyfus 认为，计算机没有身体、没有童年、没有文化实践，因此根本不能获得智能。Dreyfus 的主要论据之一是，人类知识部分是默会的，因此无法被清晰表达并写入计算机程序。”文章进一步论证这在原则上不可能，并复兴 Dreyfus 的“计算机不在世界中”的论点。Wiktionary 把 tacit 定义为“并非由形式推理原则导出”。所以主要论点是：人类拥有无法通过理性或编码表达出来的智能。换言之，人类有一个字面意义上的灵魂，与物理世界分离，因此无法在物理世界中表达，从而任何创造人工智能的努力都不可能。这是二元论推理，在我看来不过是披着哲学外衣的神学。我更希望作者干脆说自己是二元论者，或拒绝 Church-Turing 论题。
abetusk查看英文原文
Here is, to me, the relevant line: "Hubert Dreyfus, who argued that computers, who have no body, no childhood and no cultural practice, could not acquire intelligence at all. One of Dreyfus’ main arguments was that human knowledge is partly tacit, and therefore cannot be articulated and incorporated in a computer program. " I have not read the rest of the article but in the introduction it's stated: "The article further argues that this is in principle impossible, and it revives Hubert Dreyfus’ argument that computers are not in the world." Wiktionary defines tacit as "Not derived from formal principles of reasoning" [1]. So the main argument is that humans have intelligence that is impossible to express through reason or codification. In other words, humans have a literal soul, divorced from physical world, that cannot be expressed in our physical world thus making any endeavour to create artificial intelligence impossible. This is a dualist line of reasoning and, in my opinion, is nothing more than theology dressed up in philosophy. I would much rather the author just flat out say they are a dualist or that they reject the Church-Turing thesis.
质疑严厉批评文章论证空洞、关键概念未定义，并回避 Dennett 等主要反驳。
这可能是我在 Nature 见过的最差文章。它勉强达到本科生关于 AI（或作者偏好的 AGI/ANI）的论文水平。关于“计算机不在世界中”的论证只是挥手，靠不定义“计算机”“在……之中”和“世界”来蒙混。它几乎方便地回避了所有更细致的反 Searle/反 Dreyfus 批评者，尤其是 Dennett。作者论证的核心涉及现实本质以及因果性与常规参数科学的关系，但这些 broader 问题被近乎故意地窄化了。Nature 选择发表它令人惊讶。
PaulDavisThe1st查看英文原文
This is possibly the worst article I have ever seen in Nature. It is barely at the level of an undergraduate essay on AI (or AGI or ANI, as the author prefers). A hand-waving argument about "computers not being in the world" that relies on not bothering to define "computers", "being in" or "the world". Convenient avoidance of almost every more subtle anti-Searle/anti-Dreyfus critic (notably Dennett). Almost wilfull narrowing of a much broader argument about the nature of reality and the connection between causality and conventional parametric science, when these things lie at the heart of the author's argument (such as it is). It is amazing that Nature chose to publish this.
质疑所谓默会知识可能只是人类学习与沟通的局限，而非任务本身不可编码的本质。
这是经典的“思维投射谬误”：把思考方式的属性当成现实的属性。对人类来说，确实不可能听人讲完怎么骑自行车就立刻会骑；无论解释得多完整，第一次上车仍会挣扎。错误在于假设这种“必须亲自做才行”的效应是骑车知识本身的内在限制，而不是人类学习和沟通机制的限制，并假设它会推广到所有骑车系统。在计算机系统中，对人类来说是默会知识的东西不再默会。如果你写程序成功控制一个骑车机器人，该程序可以复制到同型号的新机器人上；新机器人第一次被放上车就能成功骑，没有人类那种挣扎阶段。
Strilanc查看英文原文
> In the book [Dreyfus] argued that an important part of human knowledge is tacit. Therefore, it cannot be articulated and implemented in a computer program. [...] [...] [...] These skills cannot just be learned from textbooks. They are acquired by instruction from someone who knows the trade. This is a classic example of the Mind Projection Fallacy [1], where a property of how you think is assumed to be a property of reality. It's true, for humans, that it is simply not possible to be told how to ride a bike and then be good at riding a bike. No matter how carefully and completely you explain to a human what they will have to do, when you put them on a bike for the first time they will struggle. The mistake is assuming that this "have-to-really-do-it" effect is a limitation intrinsic to bike-riding-knowledge instead of a limitation in human learning and communication mechanisms. The mistake is assuming this property will generalize to all bike riding systems. In a computer system, what would be tacit knowledge for a human is no longer tacit. If you create a computer program that can successfully control one bike riding robot, that program can be copied to a freshly built bike riding robot of the same make. The new robot will then successfully ride a bike the first time it is placed on it, without any hint of the human "have-to-really-do-it" struggling phase. It can be intuitively useful to imagine computers as having the ability to "super communicate" in a way that humans simply can't. That has its advantages, and its disadvantages. If you had super communication you could super-explain to a blind person what it was like to see and if they ever did gain their sight there would be no "Oh so that's what you meant" moment. On the other hand, a heroin addict could super-explain being addicted to heroin to you. 1: https://en.wikipedia.org/wiki/Mind_projection_fallacy
质疑具身性不是根本障碍，现代机器翻译已显示处理抽象概念关系的雏形。
我发这篇文章是因为它引人思考，尽管我不同意它。例如文中说：“真正的问题是计算机不在世界中，因为它们没有具身。”又说：“我要论证它们不能通过完整图灵测试，因为它们不在世界中，因此没有理解。”首先，我认为是否具身并不重要。我们的感官最终很可能可以近似为输入计算机的数字数组，所以缺少身体不是大问题。关于机器理解，这确实是当前 AI 的问题，但至少据我所知，现代机器翻译已经在处理概念/抽象词及其关系，在我看来像是抽象推理的开端。
sega_sai查看英文原文
I posted the article because it is thought provoking, although I disagree with it. The reason is for example statements like this "The real problem is that computers are not in the world, because they are not embodied." "And I shall argue that they cannot pass the full Turing test because they are not in the world, and, therefore, they have no understanding." First I think that whether something is embodied or not doesn't matter. Our senses in the end could be likely approximated by arrays of numbers fed to a computer, so I don't think lack of body is such an issue. Regarding understanding by machines, that is clearly the issue for current AI, but at least based on what I know about modern machine translation, there is already something that works with concepts/abstract terms and their relations, which looks to me like a beginning for abstract reasoning...
相关资源
从这篇论文所有 HN 讨论帖的全部评论里整理出来的链接。
历次讨论
- 2020Hacker News 讨论帖181 分 · 386 评
- 2022Hacker News 讨论帖25 分 · 76 评
同领域还值得读
DeepSeek-R1：通过强化学习激励 LLM 的推理能力
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL
研究如何用强化学习激励大语言模型的推理能力，提出 DeepSeek-R1，并探讨 RL 训练对模型推理表现的影响。
AI 与机器学习2025 · 1,351 分 · 1,056 评 · 引用 30进阶 · 约 80 分钟编辑精选1-bit LLM 时代：面向低成本计算的三值参数
The Era of 1-bit LLMs: ternary parameters for cost-effective computing
提出用三值参数 {-1, 0, 1} 替代 LLM 中的浮点权重，即所谓“1-bit LLM”。
AI 与机器学习2024 · 1,040 分 · 447 评 · 引用 19进阶 · 约 45 分钟编辑精选DeepSeek-V3.2：推进开放大语言模型的前沿
DeepSeek-v3.2: Pushing the frontier of open large language models
这是 DeepSeek-V3.2 的技术报告，延续了 DeepSeek 开源权重模型的路线，目标是让开放模型继续逼近甚至追平前沿闭源模型。
AI 与机器学习2025 · 982 分 · 465 评进阶 · 约 90 分钟编辑精选通过自我对弈与通用强化学习掌握国际象棋和将棋（AlphaZero）
Mastering Chess and Shogi by Self-Play with General Reinforcement Learning
DeepMind 的 AlphaZero 论文：用通用强化学习算法（自我博弈 + MCTS + 神经网络），从零开始仅数小时训练便在国际象棋和将棋上击败了世界顶级引擎 Stock…
AI 与机器学习2017 · 539 分 · 270 评 · 引用 13进阶 · 约 90 分钟编辑精选用强化学习发现更快的矩阵乘法算法
Discovering faster matrix multiplication algorithms with reinforcement learning
论文提出 AlphaTensor，用强化学习搜索矩阵乘法算法。
AI 与机器学习2022 · 518 分 · 111 评专业 · 约 90 分钟编辑精选