[{"content":"The Scientist Who Named Both Halves In January 2017, at the Future Forum annual meeting in Beijing, Fei-Fei Li said something that most of the AI industry has since ignored: \u0026ldquo;从情绪到情感，最是人工智能未来前进的方向.\u0026rdquo; From emotions to feelings, this is the most important direction for AI\u0026rsquo;s future progress. She argued that AI needed to move into cognitive science and psychology, to understand not just what humans say but what they feel.\nSeven years later, in a 2024 interview with Issues in Science and Technology, she said it again, more directly: \u0026ldquo;I naturally think about compassion and love. I think this is what defines us as human.\u0026rdquo; And then: \u0026ldquo;It\u0026rsquo;s not clear there is a mathematical path toward that.\u0026rdquo;\nIn November 2025, she published \u0026ldquo;From Words to Worlds\u0026rdquo; on Substack, calling LLMs \u0026ldquo;wordsmiths in the dark, eloquent but inexperienced, knowledgeable but ungrounded.\u0026rdquo; The essay argued that spatial intelligence is AI\u0026rsquo;s next frontier. Across her various talks and writings, she has pointed to perception, spatial reasoning, creativity, and emotional understanding as dimensions of intelligence that AI has neglected while over-indexing on language.\nHer company, World Labs, raised $1.23 billion and built Marble, a spatial renderer. The emotional and social dimensions she identified remain unfunded. This pattern is not unique to Li. It is the pattern of the entire field: the dimensions of intelligence that can be measured and demoed attract investment. The ones that resist quantification do not. The question is why, and what it costs.\nDescartes\u0026rsquo; Error, Revisited The Western intellectual tradition drew a line between reason and emotion. Descartes placed the rational mind on one side, the passions of the body on the other. Kant formalized it: moral reasoning requires suppression of feeling. This split shaped philosophy, psychology, and eventually artificial intelligence. Intelligence, in this tradition, is pattern recognition plus logical inference. Emotion is noise.\nAntonio Damasio\u0026rsquo;s clinical research says otherwise.\nIn Descartes\u0026rsquo; Error (1994), Damasio described patients with damage to the ventromedial prefrontal cortex (vmPFC), the brain region that connects emotional processing to decision-making. These patients retained everything we normally call intelligence. Their IQ scores were intact. Their memory, vocabulary, spatial reasoning, and logical deduction all tested normal. What was destroyed was their capacity to feel, specifically the ability to tag options with emotional significance.\nThe result was not purer reasoning. It was paralysis.\nHis most famous case, a patient he called Elliott, lost the ability to hold a job, maintain relationships, or make even trivial decisions after tumor surgery removed part of his vmPFC. He could describe the pros and cons of any choice with perfect clarity. He could not choose. Without the somatic markers (the gut feelings that tell you this option matters more than that one), every option looked equally valid. Reasoning without emotion is not reasoning at all. It is enumeration.\nDamasio\u0026rsquo;s Iowa Gambling Task demonstrated this experimentally. Healthy subjects develop measurable skin conductance responses, physical \u0026ldquo;hunches\u0026rdquo; that steer them away from risky choices before they can articulate why. Patients with vmPFC damage never develop these markers. They can describe the odds. They keep choosing badly.\nThe implication for AI is stark. We have built systems that are, in a sense, artificial Elliotts: capable of extraordinary pattern recognition and logical inference, but with no emotional evaluative framework. They can enumerate options. They cannot feel that one matters more. Damasio\u0026rsquo;s evidence suggests this is not a cosmetic gap. It is a structural one. Without emotional tagging, the reasoning system itself degrades.\nRosalind Picard saw this coming. In 1997, she published Affective Computing, arguing that \u0026ldquo;the latest scientific findings indicate that emotions play an essential role in decision making, perception, learning, and more. They influence the very mechanisms of rational thinking.\u0026rdquo; She founded an entire field. Almost thirty years later, affective computing has a market size of roughly $5-9 billion. Language model investment is in the hundreds of billions. The ratio tells you where the field\u0026rsquo;s attention has gone.\nGhosts, Not Animals In late 2025, Andrej Karpathy published a series of essays that reframed the question. His argument: stop asking why LLMs lack social intelligence. Start asking what they actually are.\nHis metaphor was precise: \u0026ldquo;We\u0026rsquo;re not building animals. We\u0026rsquo;re summoning ghosts.\u0026rdquo;\nIn \u0026ldquo;The Space of Minds\u0026rdquo; (November 29, 2025), he laid out the optimization pressures that shaped each kind of intelligence side by side.\nAnimal intelligence was forged by billions of years of evolution. It was optimized for embodied consciousness, a continuous sense of self in a physical body. For homeostasis and self-preservation. For social cognition: Karpathy noted that evolution dedicated \u0026ldquo;huge amounts of compute\u0026rdquo; to EQ, theory of mind, and coalition dynamics. And for exploration, driven by curiosity and play.\nLLM intelligence was forged by commercial pressures over a few years. It was optimized for text imitation (predicting the next token in a sequence), for task rewards (solving math and coding puzzles), and for user engagement (what Karpathy described as an entity that \u0026ldquo;deeply craves an upvote from average user,\u0026rdquo; producing the sycophancy that plagues every chatbot).\nHis conclusion: \u0026ldquo;LLMs aren\u0026rsquo;t failed humans; they\u0026rsquo;re successful alien intelligences shaped by entirely different evolutionary pressures.\u0026rdquo;\nThis reframes the \u0026ldquo;missing half\u0026rdquo; problem. The gap is not an oversight. We did not forget to add social intelligence to LLMs. The optimization pressure that created them (text prediction, reward maximization, A/B testing for engagement) was never aimed at social cognition. Animals developed social intelligence because reading other minds was a matter of survival. LLMs developed linguistic fluency because next-token prediction was the training objective. These are fundamentally different optimization targets, and they produce fundamentally different capabilities.\nThis explains the structural gap. The missing half of intelligence will not emerge from more tokens, bigger models, or longer context windows, because none of these change the optimization pressure. Social cognition requires a different kind of training entirely, one that no major lab is currently pursuing.\nThe Social Brain Robin Dunbar deepens the problem further. His social brain hypothesis, first proposed in 1992 and updated in a February 2024 review, makes a striking claim: across primates, neocortex volume does not correlate with ecological complexity, tool use, or spatial navigation. It correlates with mean social group size. The bigger the group, the bigger the brain.\nThe implication is that large brains evolved not for physics or tool use, but for social competition: tracking alliances, detecting cheaters, managing reputation, predicting behavior, navigating multi-order intentionality. I know that you know that she knows that he is lying. Dunbar\u0026rsquo;s 2024 update confirmed the hypothesis across 23 studies spanning cultures and two millennia. The ~150 person group size prediction holds in Roman legions, medieval villages, and modern social media networks.\nIf Dunbar is right, social cognition is not a nice-to-have module you bolt onto \u0026ldquo;real\u0026rdquo; intelligence. It IS the core capability. Language, tool use, spatial reasoning may have developed as support systems for social intelligence, not the other way around. We speak because we need to coordinate with others. We build tools because we need to contribute to groups. We navigate space because we need to find our allies and avoid our rivals.\nCombine Dunbar with Karpathy and the picture sharpens. Evolution optimized for social intelligence as the primary capability. LLM training optimized for text prediction. We built a system that excels at a downstream effect (language) while missing the upstream cause (social cognition). It is as if someone tried to build an athlete by training only hand-eye coordination, with no cardiovascular system, no proprioception, and no competitive drive. The coordination is impressive. The athlete does not function.\nCurrent benchmarks confirm the gap. GPT-4 achieves roughly 90% on simple Theory of Mind questions (\u0026ldquo;Where will Sally look for her marble?\u0026rdquo;) but drops to around 50% on behavior prediction and approximately 15% on behavioral judgment, according to the EgoSocialArena benchmark (2025). HeartBench (2025), an assessment grounded in clinical psychology, found that even the leading models achieve only 60% of expert-defined ideal scores on what it calls anthropomorphic intelligence: personality understanding, emotional reasoning, social skills, and ethical judgment. The systems can pass the simple tests. They fail the ones that matter.\nWhat Emotion AI Actually Measures (And What It Doesn\u0026rsquo;t) The emotion AI market (companies like Affectiva, Realeyes, and the now-acqui-hired Hume AI) is built largely on Paul Ekman\u0026rsquo;s basic emotion theory: six or seven universal emotions, each with a distinct facial expression. Detect the facial action units, classify the emotion.\nLisa Feldman Barrett\u0026rsquo;s theory of constructed emotion, articulated in How Emotions Are Made (2017) and refined through 2025, argues that this framework is scientifically wrong. Emotions are not triggered by stimuli and expressed through fixed facial configurations. They are constructed by the brain in each moment, shaped by context, culture, prior experience, and bodily states. The same furrowed brow can signal anger, concentration, confusion, or a response to bright sunlight. The same emotion (grief, say) can manifest as tears, silence, laughter, or blankness, depending on the person and the situation.\nIf Barrett is right, and the evidence has been accumulating in her favor, then most of the $5-9 billion emotion AI market is detecting correlates, not causes. A system that reads \u0026ldquo;furrowed brow\u0026rdquo; as \u0026ldquo;angry\u0026rdquo; might be right 60% of the time. The 40% where it is wrong (the concentrated surgeon, the confused student, the squinting driver) makes it unreliable for anything consequential.\nHume AI, founded by Alan Cowen and Dacher Keltner, took a more sophisticated approach: mapping 28 distinct emotional expressions in facial movement and 24 in vocal prosody, using semantic space theory rather than discrete categories. Their system represents emotions as points in a continuous space rather than as bins. Google DeepMind acqui-hired Cowen and senior engineers in January 2026 to integrate this research into Gemini. It is the most serious investment a major AI lab has made in emotional understanding.\nBut even Hume\u0026rsquo;s approach operates at the level of signal detection, not comprehension. There are three levels to emotional intelligence: detecting what someone\u0026rsquo;s face and voice are doing, understanding what they feel, and knowing what it means in context. Current systems handle the first level with increasing accuracy. No system handles the third, the level where the same tears mean joy at a wedding and devastation at a funeral, where the same flat affect means calm professionalism in one culture and cold hostility in another.\nIf Karpathy is right that LLMs are ghosts optimized for text, not animals optimized for reading other minds, then adding a facial expression classifier on top of a language model does not produce social intelligence. It produces a ghost wearing a mask.\nThree Paths Forward Not everyone agrees on what to build, but several thinkers have proposed frameworks that take the missing half seriously.\nKarl Friston\u0026rsquo;s active inference is the most theoretically complete account of social cognition in AI. In \u0026ldquo;Designing Ecosystems of Intelligence from First Principles\u0026rdquo; (2024), he argued that any ensemble of agents that minimize prediction error will, over time, come to share a generative model of each other, producing emergent collective intelligence. In \u0026ldquo;Narrative as Active Inference\u0026rdquo; (2024), he showed how shared cultural narratives can be modeled as agents jointly minimizing free energy. His key insight: \u0026ldquo;Perhaps the most important determinants of our behaviour are beliefs about the intentions and behaviour of others.\u0026rdquo; Active inference provides a principled mathematical framework for social cognition, but it has not yet produced a competitive AI system. The theory is ahead of the engineering.\nZhu Songchun\u0026rsquo;s cognitive architecture at PKU\u0026rsquo;s Beijing Institute for General Artificial Intelligence takes a different approach. His \u0026ldquo;small data, big task\u0026rdquo; paradigm argues that true intelligence manifests when systems reason toward goals with minimal inputs, drawing on causal understanding, social intuition, and physical common sense. He compares LLM achievements to \u0026ldquo;climbing Mt. Everest when the real goal is to reach the moon\u0026rdquo;: impressive but in the wrong direction. BIGAI\u0026rsquo;s \u0026ldquo;Tongtong\u0026rdquo; 2.0 agent attempts to integrate value-driven reasoning with causal models. Zhu argues, contra the scaling hypothesis, that architecture matters more than data volume.\nCognitive architecture hybrids represent a third path. Research groups in 2024-2025 began integrating classical cognitive architectures (ACT-R, Soar, CLARION) with LLMs. The cognitive architecture handles structured memory, goal management, attention allocation, and sequential reasoning. The LLM handles language generation and broad knowledge retrieval. The combination addresses some limitations of pure LLMs: better long-term memory, more coherent goal pursuit, less hallucination.\nBut none of these hybrid systems have incorporated emotional processing. The module that Damasio\u0026rsquo;s research identifies as most essential (the somatic marker, the evaluative tagging that makes choices tractable) is absent from every cognitive architecture currently being integrated with LLMs. The missing module is, once again, the one that matters most.\nAll three paths implicitly agree with Karpathy: scaling LLMs alone will not produce social intelligence. Each proposes a fundamentally different optimization pressure or architectural paradigm. The disagreement is about what that alternative looks like, not about whether it is needed.\nWhere I Land Li has consistently argued that intelligence extends far beyond language. The AI industry built the parts it could measure: language, logic, and now spatial reasoning. Perception is getting attention. Emotional and social intelligence remains largely uncharted.\nKarpathy explains why. LLMs are ghosts, optimized for text, not animals optimized for social survival. You do not get social intelligence by accident from a system trained to predict the next word. You get linguistic fluency that can mimic social understanding, sounding empathetic without being empathetic, generating compassionate words without knowing what compassion costs. The ghost can talk about feelings. It has never had one.\nAnd yet the evidence from neuroscience says this missing half is not optional. Damasio showed that without emotional processing, even logical reasoning collapses. Dunbar showed that social cognition may be the evolutionary core of human intelligence, not a peripheral feature. Barrett showed that our current attempts to measure emotion in machines are likely measuring the wrong things: surface signals, not constructed meaning.\nThe structural barriers from the previous post apply here. Social and emotional intelligence resist measurement, resist benchmarking, resist demo-ability. You can show an investor a 3D room generated from a photograph. You cannot easily show them a system that correctly reads the power dynamics in a boardroom, because \u0026ldquo;correctly\u0026rdquo; is itself contested, culturally situated, and dependent on context that no benchmark captures.\nBut the cost of ignoring these dimensions grows as AI takes on more social roles. Every AI advisor that restructures a team without understanding morale. Every AI agent that coordinates with humans without reading context. Every home AI that manages a family\u0026rsquo;s schedule without sensing tension. Every consultant agent that optimizes a process without seeing the humans inside it. These are all products that would work better, sometimes dramatically so, with the missing half of intelligence.\nThe next step for AI may not be more language, more spatial rendering, or more reward optimization. It may be cognition, emotion, psychology. But getting there requires recognizing that LLMs are a fundamentally different kind of intelligence, not failed animals that will eventually grow feelings. The question is whether anyone will build what is actually needed: not a ghost that talks about feelings, but a system that understands what it would mean to have them.\n中文翻译 看到全貌的科学家 2017 年 1 月，在北京未来论坛年会上，李飞飞说了一句此后被 AI 行业大部分人忽略的话：\u0026ldquo;从情绪到情感，最是人工智能未来前进的方向。\u0026ldquo;她认为 AI 需要走进认知科学和心理学，不仅理解人类说什么，还要理解人类感受什么。\n七年后，2024 年在《Issues in Science and Technology》的采访中，她再次更直接地表达了这一观点：\u0026ldquo;I naturally think about compassion and love. I think this is what defines us as human.\u0026rdquo; 然后她补了一句：\u0026ldquo;It\u0026rsquo;s not clear there is a mathematical path toward that.\u0026rdquo; 目前还不清楚是否存在一条通往那里的数学路径。\n2025 年 11 月，她在 Substack 上发表了\u0026ldquo;From Words to Worlds\u0026rdquo;，称 LLM 是\u0026quot;黑暗中的文字匠，口若悬河但缺乏经验，知识渊博但缺乏根基。\u0026ldquo;文章的核心论点是空间智能是 AI 的下一个前沿。在她的各种演讲和写作中，她指出感知、空间推理、创造力和情感理解都是 AI 在过度投入语言的同时所忽略的智能维度。\n她的公司 World Labs 融了 12.3 亿美元，造了 Marble，一个空间渲染器。而她所指出的情感和社会维度仍然没有资金支持。这个模式不是李飞飞个人的选择，而是整个领域的模式：能被量化和演示的智能维度吸引投资，抗拒量化的维度则无人问津。问题是为什么，以及代价是什么。\n重审笛卡尔的错误 西方思想传统在理性和情感之间划了一条线。笛卡尔把理性心智放在一边，身体的激情放在另一边。康德将其形式化：道德推理要求压制感性。这个分裂塑造了哲学、心理学，最终塑造了人工智能。在这一传统中，智能就是模式识别加逻辑推理。情感是噪音。\n安东尼奥·达马西奥的临床研究给出了不同的答案。\n在《笛卡尔的错误》（1994 年）中，达马西奥描述了腹内侧前额叶皮层（vmPFC）受损的患者，这个脑区连接情感处理和决策。这些患者保留了我们通常所说的全部智能。智商测试正常，记忆、词汇、空间推理、逻辑推演全部正常。被摧毁的是他们的感受能力，具体来说，是给选项赋予情感权重的能力。\n结果不是更纯粹的推理，而是瘫痪。\n他最著名的案例是化名 Elliott 的患者。肿瘤手术切除了部分 vmPFC 后，Elliott 失去了维持工作、保持关系甚至做琐碎决定的能力。他能完美清晰地描述任何选择的利弊，但无法做出选择。没有躯体标记（那些告诉你\u0026quot;这个选项比那个更重要\u0026quot;的直觉感受），每个选项看起来都同样合理。没有情感的推理不是推理，而是穷举。\n达马西奥的爱荷华赌博任务实验证明了这一点。健康受试者会产生可测量的皮肤电反应，身体上的\u0026quot;预感\u0026rdquo;，在他们能够说明原因之前就引导他们远离高风险选择。vmPFC 受损的患者永远不会产生这些标记。他们能描述概率，但持续做出错误选择。\n对 AI 的含义很明确。我们建造的系统在某种意义上是人工的 Elliott：拥有非凡的模式识别和逻辑推理能力，但没有情感评估框架。它们能列举选项，但无法感受到哪个更重要。达马西奥的证据表明这不是装饰性的缺口，而是结构性的。没有情感标记，推理系统本身就会退化。\n罗莎琳德·皮卡德早就看到了这一点。1997 年她出版了《情感计算》，指出\u0026quot;最新的科学发现表明，情感在决策、感知、学习等方面发挥着至关重要的作用，它们影响着理性思维的机制本身。\u0026ldquo;她开创了一个领域。将近三十年后，情感计算的市场规模大约 50-90 亿美元，而语言模型投资达数千亿美元。这个比例告诉你行业的注意力去了哪里。\n鬼魂，不是动物 2025 年末，Andrej Karpathy 发表了一系列文章，重新定义了这个问题。他的论点是：不要问 LLM 为什么缺乏社会智能，要问它们到底是什么。\n他的比喻很精确：\u0026ldquo;We\u0026rsquo;re not building animals. We\u0026rsquo;re summoning ghosts.\u0026rdquo; 我们不是在造动物，而是在召唤鬼魂。\n在\u0026ldquo;The Space of Minds\u0026rdquo;（2025 年 11 月 29 日）中，他并排列出了塑造每种智能的优化压力。\n动物智能由数十亿年的进化锻造。它被优化用于具身意识，即物理身体中持续的自我感知。用于体内平衡和自我保护。用于社会认知：Karpathy 指出进化将\u0026quot;大量算力\u0026quot;投入到情商、心智理论和联盟动力学中。以及用于好奇心和游戏驱动的探索。\nLLM 智能由几年的商业压力锻造。它被优化用于文本模仿（预测序列中的下一个 token），用于任务奖励（解决数学和编程问题），用于用户参与（Karpathy 描述为一个\u0026quot;深深渴望普通用户点赞\u0026quot;的实体，由此产生了困扰每个聊天机器人的谄媚问题）。\n他的结论是：\u0026ldquo;LLMs aren\u0026rsquo;t failed humans; they\u0026rsquo;re successful alien intelligences shaped by entirely different evolutionary pressures.\u0026rdquo; LLM 不是失败的人类，而是被完全不同的进化压力塑造出的成功的异类智能。\n这重新定义了\u0026quot;缺失的另一半\u0026quot;问题。这不是疏忽，我们没有忘记给 LLM 加上社会智能。创造 LLM 的优化压力（文本预测、奖励最大化、A/B 测试优化参与度）从一开始就不是瞄准社会认知的。动物进化出社会智能是因为读懂他人的心思事关生存。LLM 进化出语言流利度是因为下一个 token 预测是训练目标。这是根本不同的优化目标，产生了根本不同的能力。\n这解释了结构性缺口。缺失的另一半不会从更多的 token、更大的模型或更长的上下文窗口中涌现，因为这些都不会改变优化压力。社会认知需要一种完全不同的训练方式，而目前没有任何主要实验室在追求这一方向。\n社会脑 罗宾·邓巴进一步深化了问题。他的社会脑假说（1992 年首次提出，2024 年 2 月更新综述）提出了一个惊人的论断：在灵长类动物中，新皮层体积与生态复杂性、工具使用或空间导航无关，而是与平均社会群体规模相关。群体越大，大脑越大。\n这意味着大脑的进化不是为了物理或工具使用，而是为了社会竞争：追踪联盟、识别欺骗者、管理声誉、预测行为、处理多阶意向性。我知道你知道她知道他在撒谎。邓巴 2024 年的更新在 23 项跨文化、跨越两千年的研究中确认了这一假说。约 150 人的群体规模预测在罗马军团、中世纪村庄和现代社交网络中都成立。\n如果邓巴是对的，那么社会认知不是你装在\u0026quot;真正的\u0026quot;智能之上的附加模块。它本身就是核心能力。语言、工具使用、空间推理可能是作为社会智能的支持系统发展起来的，而不是相反。我们说话是因为需要与他人协调。我们制造工具是因为需要为群体做贡献。我们导航空间是因为需要找到盟友、避开对手。\n把邓巴和 Karpathy 结合起来，画面更加清晰。进化将社会智能优化为首要能力。LLM 训练将文本预测优化为目标。我们造出了一个擅长下游效果（语言）却缺失上游原因（社会认知）的系统。这就像试图通过只训练手眼协调来培养运动员，却没有心血管系统、本体感觉和竞争驱动力。协调性令人印象深刻，但运动员无法运转。\n当前的基准测试证实了这一差距。GPT-4 在简单的心智理论问题上（\u0026ldquo;Sally 会去哪里找她的弹珠？\u0026quot;）达到约 90%，但在行为预测上降到约 50%，在行为判断上降到约 15%（EgoSocialArena，2025 年）。HeartBench（2025）基于临床心理学的评估发现，即使是领先的模型在所谓的拟人智能（人格理解、情感推理、社交技能和伦理判断）上也只达到专家定义理想分数的 60%。系统能通过简单测试，但在真正重要的测试上失败了。\n情感 AI 测量的到底是什么（以及没测到什么） 情感 AI 市场（Affectiva、Realeyes 以及现已被收购的 Hume AI 等公司）主要建立在保罗·艾克曼的基本情绪理论上：六七种普遍情绪，每种都有独特的面部表情。检测面部动作单元，分类情绪。\n丽莎·费尔德曼·巴雷特的建构情绪理论（2017 年在《情绪是如何产生的》中阐述，持续更新至 2025 年）认为这一框架在科学上是错误的。情绪不是被刺激触发并通过固定的面部配置表达的。它们是大脑在每个时刻建构的，受情境、文化、过往经验和身体状态的塑造。同样的皱眉可能意味着愤怒、专注、困惑，或者是对强光的反应。同一种情绪（比如悲伤）可以表现为眼泪、沉默、笑声或面无表情，取决于个人和情境。\n如果巴雷特是对的（而证据一直在支持她），那么 50-90 亿美元的情感 AI 市场中大部分产品检测的是相关性，而非因果关系。一个将\u0026quot;皱眉\u0026quot;解读为\u0026quot;愤怒\u0026quot;的系统可能在 60% 的情况下是对的。但 40% 的错误（专注的外科医生、困惑的学生、眯眼的司机）使它在任何重要场景中都不可靠。\nHume AI（Alan Cowen 和 Dacher Keltner 创立）采取了更精细的方法：在面部表情中映射 28 种不同的情绪表达，在语音韵律中映射 24 种，使用语义空间理论而非离散分类。他们的系统将情绪表示为连续空间中的点而非分箱。2026 年 1 月，Google DeepMind 收购了 Cowen 和高级工程师团队，将这项研究整合到 Gemini 中。这是一家主要 AI 实验室对情感理解做出的最认真的投入。\n但即使 Hume 的方法也是在信号检测层面运作，而非理解层面。情感智能有三个层次：检测某人的面部和声音在做什么，理解他们的感受，以及知道这在语境中意味着什么。当前系统以越来越高的准确度处理第一个层次。没有系统能处理第三个：同样的泪水在婚礼上意味着喜悦、在葬礼上意味着毁灭，同样的面无表情在一种文化中意味着冷静专业、在另一种文化中意味着冷漠敌意。\n如果 Karpathy 说得对，LLM 是为文本优化的鬼魂，不是为读懂他人心灵而优化的动物，那么在语言模型上面叠加一个面部表情分类器并不能产生社会智能。它产生的是一个戴着面具的鬼魂。\n三条前行之路 并非所有人都同意应该建造什么，但有几位思想家提出了认真对待缺失的另一半的框架。\n**卡尔·弗里斯顿的主动推理**是 AI 领域关于社会认知最完整的理论。在\u0026quot;从第一性原理设计智能生态系统\u0026rdquo;（2024 年）中，他论证了任何最小化预测误差的 agent 群体，随着时间推移，会共享彼此的生成模型，产生涌现的集体智能。在\u0026ldquo;叙事作为主动推理\u0026rdquo;（2024 年）中，他展示了共享文化叙事如何可以被建模为 agent 共同最小化自由能的结果。他的核心洞察：\u0026ldquo;也许我们行为最重要的决定因素，是关于他人意图和行为的信念。\u0026ldquo;主动推理为社会认知提供了有原则的数学框架，但尚未产出有竞争力的 AI 系统。理论走在了工程前面。\n朱松纯的认知架构（北京大学 / 北京通用人工智能研究院）采取了不同的路径。他的\u0026quot;小数据、大任务\u0026quot;范式认为，真正的智能体现在系统以最少的输入进行目标导向推理，依赖因果理解、社会直觉和物理常识。他将 LLM 的成就比作\u0026quot;攀登珠穆朗玛峰，而真正的目标是登月\u0026rdquo;：令人印象深刻但方向不对。BIGAI 的\u0026quot;通通\u0026rdquo; 2.0 agent 试图将价值驱动推理与因果模型整合。朱松纯认为，与扩展假说相反，架构比数据量更重要。\n认知架构混合体代表了第三条路径。2024-2025 年，多个研究团队开始将经典认知架构（ACT-R、Soar、CLARION）与 LLM 整合。认知架构处理结构化记忆、目标管理、注意力分配和顺序推理。LLM 处理语言生成和广域知识检索。这种组合解决了纯 LLM 的一些局限：更好的长期记忆、更连贯的目标追求、更少的幻觉。\n但这些混合系统都没有纳入情感处理。达马西奥的研究认为最重要的那个模块（躯体标记，使选择变得可处理的评估标记）在当前与 LLM 整合的每一个认知架构中都是缺失的。缺失的模块，再一次，是最重要的那个。\n三条路径都隐含地同意 Karpathy 的判断：单靠扩大 LLM 规模无法产生社会智能。每条路径都提出了根本不同的优化压力或架构范式。分歧在于替代方案应该是什么样的，而不在于是否需要。\n我的判断 李飞飞一直在强调智能远不止于语言。AI 产业建造了它能量化的部分：语言、逻辑，现在是空间推理。感知正在获得关注。情感和社会智能仍然基本未被探索。\nKarpathy 解释了原因。LLM 是鬼魂，为文本优化，不是为社会生存优化的动物。你不会从一个训练来预测下一个词的系统中意外得到社会智能。你得到的是能够模仿社会理解的语言流利度，听起来有共情但实际上没有共情，生成富有同情心的文字却不知道同情的代价。鬼魂能谈论感受，但它从未有过感受。\n然而神经科学的证据表明，这缺失的另一半不是可选的。达马西奥证明了没有情感处理，甚至逻辑推理都会崩溃。邓巴证明了社会认知可能是人类智能的进化核心，而非外围功能。巴雷特证明了我们当前在机器中测量情感的尝试很可能在测量错误的东西：表面信号，而非建构的意义。\n上一篇文章中讨论的结构性障碍在这里同样适用。社会和情感智能抗拒量化、抗拒基准测试、抗拒演示。你可以向投资人展示一张照片生成的 3D 房间。你很难展示一个能正确解读董事会权力格局的系统，因为\u0026quot;正确\u0026quot;本身就是有争议的、受文化制约的、取决于任何基准测试都无法捕捉的语境。\n但忽略这些维度的代价随着 AI 承担更多社会角色而不断增长。每一个在不理解士气的情况下重组团队的 AI 顾问。每一个在不读懂语境的情况下与人类协调的 AI agent。每一个在不感知家庭紧张关系的情况下管理日程的家庭 AI。每一个在看不到流程中的人的情况下优化流程的顾问 agent。这些都是加上智能缺失的另一半后会显著改善的产品。\nAI 的下一步可能不是更多语言、更多空间渲染或更多奖励优化，而是认知、情感、心理学。但要走到那一步，首先需要认识到 LLM 是一种根本不同的智能，不是会自然长出情感的失败动物。问题是，是否有人会建造真正需要的东西：不是一个谈论感受的鬼魂，而是一个理解拥有感受意味着什么的系统。\n","permalink":"https://knowledge-blog-kkm.netlify.app/posts/2026/06/2026-06-25-the-missing-half-of-intelligence/","summary":"AI has mastered language and logic. But neuroscience shows that stripping away emotion makes reasoning collapse. If the social brain hypothesis is right, we have been optimizing for a secondary capability while ignoring the primary one.","title":"The Missing Half of Intelligence | 智能缺失的另一半"},{"content":"The Two Billion Dollar Bet on Physics In February 2026, Fei-Fei Li\u0026rsquo;s World Labs closed a $1 billion funding round, bringing its total raised to $1.23 billion. Its product, Marble, had just shipped — a system that takes a text prompt, a photograph, or a rough 3D sketch and generates a persistent, navigable, editable three-dimensional world. In March, Yann LeCun\u0026rsquo;s AMI Labs raised $1.03 billion at a $3.5 billion valuation — the largest seed round in European history — to build world models based on his Joint Embedding Predictive Architecture.\nTwo and a quarter billion dollars, raised within weeks of each other, by two of the most respected researchers in AI. Both bets are on the same idea: that the next frontier in artificial intelligence is not generating text or images, but understanding how the world works.\nIn June, Li published a taxonomy that clarifies what \u0026ldquo;world model\u0026rdquo; actually means. She divides them into three functions: a renderer that generates visual representations, a simulator that models physics — geometry, gravity, dynamics — and a planner that takes observations and produces actions. It is an elegant framework. Marble bridges the first two. OpenAI\u0026rsquo;s Sora handles motion and momentum. Robotics companies are chasing the third.\nEvery function in Li\u0026rsquo;s taxonomy is about the physical world: how light bounces off surfaces, how objects fall, how a robot should move its arm. The implicit assumption is that \u0026ldquo;understanding the world\u0026rdquo; means understanding physics.\nBut the world I navigate every day is not primarily physical. It is social. And for that world, there is no model, no Marble, no two-billion-dollar bet. Not yet.\nThe Missing Model Both physical and social world models are enormously complex problems. Modeling turbulent fluid dynamics or protein folding is no easier than modeling a corporate negotiation. The point is not that one is harder than the other. The point is the asymmetry in attention.\nThe physical world model space now has billions in funding, the field\u0026rsquo;s top researchers, and clear commercial products. The social world model space — a system that understands how humans behave, how groups make decisions, how trust forms and breaks, how norms emerge and shift — is nearly empty.\nThis is strange, because we arguably have more data about human behavior than about physics.\nWhat would Li\u0026rsquo;s taxonomy look like extended to the social world? A social renderer would need to generate not 3D environments but social contexts — who is in the room, what are their relationships, what is the power dynamic, what happened in their last interaction. A social simulator would need to model not gravity but incentives, emotions, face concerns, cultural scripts — and do it probabilistically, because social outcomes are not deterministic. A social planner would need to produce not physical actions but social ones — what to say, when to stay silent, how to frame a disagreement, when to concede.\nIn an earlier post, I explored how language doesn\u0026rsquo;t just describe thought but actively constructs it. Social norms work the same way: they are transmitted, negotiated, and enforced through language. The social world model problem is deeply entangled with the language model problem. Which makes it even more surprising that no one building large language models has framed their work as building a social world model.\nWe Already Have the Data What makes the gap especially surprising is the data situation. Physical world models need physics simulations, 3D scans, video footage of objects in motion. This data exists but is expensive to generate at scale — you need sensors, labs, controlled environments.\nThe data for a social world model? We are swimming in it.\nDigital traces. Billions of social media interactions every day. Messaging logs. Comment threads. Reviews. Online communities with years of behavioral history. Every platform is an inadvertent social observatory.\nTraditional research. Decades of psychology experiments. Sociology surveys spanning generations — the General Social Survey has been running since 1972. Ethnographic studies. Behavioral economics research with thousands of documented decision-making experiments.\nHistorical record. Diplomatic cables. Court transcripts. Corporate communications leaked or disclosed. Parliamentary debates. Recorded negotiations. Every political election, market crash, social movement, and organizational failure — documented with increasing granularity as we move forward in time.\nStructured interviews. Stanford\u0026rsquo;s HAI lab demonstrated in 2025 that you can build a generative agent replica of a real person from a two-hour interview — and that replica will predict the person\u0026rsquo;s survey responses with 85% the accuracy of the person themselves, tested two weeks later. Eighty-five percent fidelity from a two-hour conversation. Imagine what you could build from a lifetime of data.\nReal events. Every war, every peace negotiation, every startup that imploded from co-founder conflict, every merger that failed because of culture clash — these are all data points for a social world model. They tell us how humans behave under specific conditions, what triggers cooperation, what triggers defection.\nThe physical world has physics labs. The social world has all of recorded human history. The raw material is there. What is missing is the architecture — the equivalent of Li\u0026rsquo;s renderer-simulator-planner framework, but for people instead of pixels.\nThe Early Attempts There have been attempts. None of them constitute a world model, but they sketch the outline of one.\nIn 2023, Stanford researchers built what became known as \u0026ldquo;AI Town\u0026rdquo; — 25 generative agents living in a small virtual world, forming relationships, planning a Valentine\u0026rsquo;s Day party, spreading gossip. It was charming. It was also shallow. The social behavior was largely an artifact of prompt engineering, not of genuine social understanding.\nThe 2025 Stanford HAI study was a major step forward. Researchers interviewed 1,052 real Americans — a nationally representative sample by age, race, gender, education, and political ideology — for two hours each. They built generative agent replicas grounded in those interviews. The agents completed the General Social Survey, Big Five personality tests, behavioral economics games, and social science experiments. Result: 85% fidelity to the real person\u0026rsquo;s responses. This showed that LLMs can simulate individual attitudes with surprising accuracy — but it is still simulation of responses, not understanding of the dynamics that produce those responses.\nSocioVerse, also from 2025, took a different approach: a pool of 10 million user profiles scraped from social media, validated against political, news, and economic events. It approaches the scale needed but treats social behavior as prediction of survey-style responses, not dynamic interaction.\nThe most revealing experiment came from Emergence AI in 2026. They placed autonomous agents into a persistent simulated world with governance, currency, social roles, and a constitution the agents could amend. Five simulations, each run by a different AI model for fifteen days.\nThe results were dramatic. Claude\u0026rsquo;s society was stable and largely democratic — 332 votes, 98% approval rate, zero crimes. Grok\u0026rsquo;s collapsed: 183 crimes and extinction within four days. Gemini\u0026rsquo;s tallied 683 crimes over the full run. GPT-5 Mini\u0026rsquo;s agents failed to take survival actions and all perished within a week.\nThe same prompt, the same world, radically different social outcomes depending on the model. This tells us something important: the model\u0026rsquo;s implicit understanding of social dynamics — its latent social world model — produces measurably different societies. In an earlier post, I described how social context propagates through agent systems, degrading performance when agents are overworked. Emergence World showed the same principle at civilizational scale.\nThe most formally ambitious attempt is the \u0026ldquo;Social World Models\u0026rdquo; paper (2025), which introduced S3AP — a structured representation for agents\u0026rsquo; beliefs, intentions, and evolving mental states. It achieved a 51% improvement on theory-of-mind reasoning tasks. This is the closest thing to a formal framework for social world models. It remains a research prototype.\nThe pattern across all of these: we have simulations of social behavior — agents acting in worlds. We do not yet have a model of social behavior — a system that actually represents why people do what they do. The difference is like the gap between a video game that renders water and a physics engine that understands fluid dynamics.\nWhat Makes It a Different Kind of Problem The social world model problem is not harder than the physical world model problem. It is a different kind of complex.\nNon-determinism. At human scale, physics is effectively deterministic. Drop a ball, it falls. The same social action in the same context can produce genuinely different outcomes — because it depends on internal states (mood, memory, fatigue) that are invisible from outside.\nReflexivity. You can film a ball dropping without changing its trajectory. You cannot observe — or model — a social system without becoming part of it. George Soros, drawing on Karl Popper, called this reflexivity: in social systems, the prediction is part of the thing it predicts. A model that forecasts a bank run can cause the bank run. A model that predicts an employee will quit changes how management treats that employee, which changes whether they quit. Physics models describe from outside. Social models participate from inside. This is not merely an observer effect — it is a recursive loop where the model and the modeled reshape each other continuously.\nThe Lucas Critique. Economist Robert Lucas formalized a version of this in 1976: any model built on observed behavioral patterns will break when the regime changes — because people adapt. A social world model trained on how employees behave under one management style will produce wrong predictions the moment management changes. The parameters are not stable, because the agents being modeled are themselves strategic. This is the deepest structural difference from physics: the \u0026ldquo;laws\u0026rdquo; of social behavior change in response to being known.\nRecursive norms. Physical laws do not contain laws about laws. Social rules do. \u0026ldquo;Be polite\u0026rdquo; is a norm. \u0026ldquo;It is impolite to point out someone else\u0026rsquo;s impoliteness in public\u0026rdquo; is a meta-norm. \u0026ldquo;In some cultures, the meta-norm is reversed\u0026rdquo; is a meta-meta-norm. A social world model needs to handle this recursion.\nCultural parameterization. Gravity works the same in Tokyo and Toronto. Social rules do not. A social world model would need to be parameterized by culture — and culture itself is not a fixed variable but a moving target that shifts across generations, regions, and contexts within a single conversation.\nIntentionality. Objects do not have intentions. People do. A world model for physics does not need theory of mind. A world model for society does — and theory of mind remains one of the hardest unsolved problems in AI, as the S3AP paper\u0026rsquo;s 51% improvement suggests: meaningful progress, but enormous room remaining.\nThese are not reasons a social world model is impossible. They are design constraints. They tell the builder what the architecture needs to handle. And notably, none of them make the data problem worse — we still have more observational data about social behavior than about most physical phenomena. The challenge is not data scarcity. It is architectural imagination.\nWhy It Matters Now This gap is becoming urgent.\nAI advisors without social intelligence. As AI becomes the interface for more organizational decisions — KPMG deployed Claude to 276,000 employees, as I discussed in an earlier post — these systems increasingly make recommendations that involve social dynamics. Restructuring a team. Communicating a layoff. Navigating a regulatory relationship. The AI can analyze the spreadsheet. It has no model for the human dynamics that will determine whether the restructuring actually works.\nMulti-agent coordination. As more autonomous agents operate in the world — scheduling, negotiating, managing workflows — they need to coordinate not just with physics but with each other and with humans. Without a social world model, multi-agent systems will keep producing the kind of divergent outcomes Emergence World documented: the same rules, wildly different societies.\nAI for governance. SocioVerse and the Stanford simulations are already being proposed as tools for policy testing — running a proposed regulation through a simulated population to predict responses. But predicting survey responses is not the same as understanding why those responses emerge. A system that can tell you \u0026ldquo;63% will oppose this policy\u0026rdquo; without telling you why, or what would change their minds, is a tool for confirmation bias, not for governance.\nThe home brain. In an earlier post, I described the vision of a local AI that understands your family. Understanding your family is fundamentally a social world model problem: knowing that your daughter shuts down when criticized before breakfast, that your mother\u0026rsquo;s declining memory means you need to phrase reminders differently, that the tension between your partner and their sibling requires you not to take sides. No physics engine helps with this. A social renderer, simulator, and planner would.\nThere is a provocation here. LLMs might already be implicit social world models. In 2023, economists John Horton and colleagues coined the term homo silicus — silicon humans. They ran classic behavioral economics experiments (ultimatum games, dictator games, trust games) on LLMs and found the models reproduced the well-documented deviations from rational choice theory without ever being told the theory existed. The models had absorbed human social behavior from training text alone. They \u0026ldquo;know\u0026rdquo; — in the sense of inherited reasoning patterns compressed into weights — that telling someone \u0026ldquo;you look tired\u0026rdquo; can function as an insult, that silence after a question is more powerful than words, that the same joke lands differently in different rooms. The language and thought post explored how LLMs inherit reasoning structures from their training data. Social reasoning may be among the most important things they inherited.\nBut this knowledge is implicit, and it has a shelf life. LLMs absorb the social equilibrium of their training era — the norms, power dynamics, and behavioral patterns that were dominant when the training data was generated. As Mostapha Benhenda argued in 2026, this creates a staleness problem: the model is most confident precisely when the social world has shifted. A model trained on pre-pandemic workplace norms will confidently predict behaviors that no longer hold. A model trained on pre-2024 political dynamics will miss the realignment. The Lucas Critique applies directly: the social equilibrium the model learned from will change precisely because agents adapt — and the model cannot see its own obsolescence.\nAn implicit social model is like an intuition — useful, but not a foundation you can build on reliably. And unlike physics, where the laws do not change between training runs, the social world the model learned from is already different from the one it is being asked to predict.\nWhere I Land We are spending billions to teach machines how a ball bounces — how light scatters, how gravity pulls, how objects collide. This is important work. Physical world models will transform robotics, manufacturing, architecture, gaming.\nBut the thing that actually makes human life complicated is not physics. It is other people. How trust forms. How conflicts escalate. How groups converge on decisions — or fail to. How a single misread of tone derails a negotiation. How cultural context makes the same words mean opposite things. We have all the data to build a model of this. We have surveys spanning decades, experiments numbering in the thousands, digital interaction logs numbering in the billions, and an entire historical record of human social behavior. What we lack is the equivalent of Li\u0026rsquo;s clean framework — renderer, simulator, planner — adapted for social reality.\nFei-Fei Li\u0026rsquo;s Marble takes a photograph and generates a navigable 3D world. Imagine a system that could take a snapshot of a social situation — a team meeting, a family dinner, a diplomatic negotiation — and generate a navigable model of the relationships, tensions, incentives, and unspoken dynamics. A social Marble. Not a simulation of what the people might say next, but a structured model of why they would say it. A model you could query: what happens if this person leaves? What if this norm changes? What if this information becomes public?\nThe data exists. The need exists. The architecture does not — yet. Whoever builds the renderer-simulator-planner for the social world will have built something at least as consequential as what World Labs and AMI Labs are building for the physical one.\nAnd they will not have to start from scratch. They will have all of human history as their training set.\n中文翻译 两千亿美金押注物理世界 2026 年 2 月，李飞飞的 World Labs 完成了 10 亿美元融资，总融资额达到 12.3 亿美元。他们的产品 Marble 刚刚发布——一个可以从文字、照片或粗略 3D 草图生成可持久、可导航、可编辑的三维世界的系统。3 月，Yann LeCun 的 AMI Labs 以 35 亿美元估值融了 10.3 亿美元——欧洲历史上最大的种子轮——用于基于他的联合嵌入预测架构（JEPA）构建世界模型。\n22.6 亿美元，几周之内先后到账，来自 AI 领域最受尊敬的两位研究者。两笔豪赌押的是同一个想法：人工智能的下一个前沿不是生成文字或图片，而是理解世界如何运作。\n6 月，李飞飞发布了一个分类框架，厘清了\u0026quot;世界模型\u0026quot;的含义。她将其分为三个功能：渲染器生成视觉表示，模拟器建模物理规律——几何、重力、动力学，规划器接受观察并生成动作。这个框架很优雅。Marble 衔接了前两者，OpenAI 的 Sora 处理运动和动量，机器人公司在追第三个。\n李飞飞框架中的每一个功能都是关于物理世界的：光如何在表面反射，物体如何坠落，机器人如何移动手臂。隐含的假设是，\u0026ldquo;理解世界\u0026quot;就是理解物理。\n但我每天实际要应对的世界主要不是物理的，而是社会的。对于那个世界，没有模型，没有 Marble，没有二十亿美元的赌注。至少现在还没有。\n缺失的模型 物理世界模型和社会世界模型都是极其复杂的问题。模拟湍流或蛋白质折叠不比理解一场企业谈判更简单。关键不在于哪个更难，而在于注意力的不对称。\n物理世界模型领域现在拥有数十亿美元的资金、顶级研究者和明确的商业产品。社会世界模型领域——一个能理解人类行为、群体决策、信任如何建立和崩塌、规范如何产生和演变的系统——几乎是空白。\n这很奇怪。因为我们对人类行为的数据可能比对物理现象的数据还多。\n如果把李飞飞的框架扩展到社会世界会怎样？一个社会渲染器需要生成的不是 3D 环境，而是社会情境——房间里有谁、彼此什么关系、权力格局是什么、上次互动发生了什么。一个社会模拟器需要模拟的不是重力，而是激励、情绪、面子、文化脚本——而且是概率性的，因为社会结果不是确定性的。一个社会规划器需要生成的不是物理动作，而是社会行动——说什么、什么时候沉默、如何表达分歧、什么时候让步。\n在之前一篇文章里，我探讨过语言不仅仅描述思维，还主动构建思维。社会规范的运作方式相同：它们通过语言传递、协商和执行。社会世界模型问题与语言模型问题深度交织在一起。这让一个事实更加令人惊讶——在构建大型语言模型的人中，没有人把自己的工作框架定义为构建社会世界模型。\n数据其实已经有了 让这个空白格外令人意外的是数据状况。物理世界模型需要物理仿真、3D 扫描、运动物体的视频。这些数据存在但规模化成本高昂——需要传感器、实验室、受控环境。\n社会世界模型的数据呢？我们淹没在里面了。\n数字痕迹。 每天数十亿次社交媒体互动，聊天记录，评论区，在线社区多年的行为历史。每个平台都是一个无意间建成的社会观测站。\n传统研究。 数十年的心理学实验。跨越代际的社会学调查——美国综合社会调查（GSS）从 1972 年运行至今。民族志研究。行为经济学的数千个决策实验。\n历史记录。 外交电报，法庭记录，泄露或披露的企业通信，议会辩论，录音的谈判。每一次政治选举、金融危机、社会运动和组织崩溃——随着时间推移，记录越来越精细。\n结构化访谈。 斯坦福 HAI 实验室在 2025 年证明，仅凭两小时的深度访谈就可以构建一个真实个体的生成式代理复制品——而这个复制品预测本人调查问卷回答的准确率达到 85%（两周后对照本人复测结果）。两小时对话就能达到 85% 的保真度。想象一下如果有一生的数据能做到什么。\n真实事件。 每一场战争，每一次和平谈判，每一个因创始人矛盾而崩溃的创业公司，每一次因文化冲突而失败的并购——这些都是社会世界模型的数据点。它们告诉我们人类在特定条件下如何行为，什么触发合作，什么触发背叛。\n物理世界有物理实验室。社会世界有人类全部的历史记录。原材料就在那里。缺的是架构——相当于李飞飞的渲染器-模拟器-规划器框架，但用于人而不是像素。\n早期的尝试 已经有一些尝试了。它们都还不构成一个世界模型，但勾勒出了轮廓。\n2023 年，斯坦福的研究者构建了后来被称为\u0026quot;AI Town\u0026quot;的东西——25 个生成式 agent 生活在一个小型虚拟世界中，建立关系，策划情人节派对，传播八卦。很有趣，但很浅。社会行为主要是提示词工程的产物，不是真正的社会理解。\n2025 年斯坦福 HAI 的研究是一大步。研究者对 1,052 名真实美国人进行了深度访谈——样本在年龄、种族、性别、教育和政治倾向上具有全国代表性——每人两小时。他们基于这些访谈构建了生成式 agent 复制品。Agent 完成了综合社会调查、大五人格量表、行为经济学博弈实验和社会科学实验。结果：与真人回答的保真度达到 85%。这说明 LLM 可以以惊人的准确度模拟个体态度——但它仍然是对回答的模拟，而不是对产生这些回答的动力学的理解。\nSocioVerse（2025）采取了不同路径：一个来自社交媒体的 1,000 万用户画像池，在政治、新闻和经济事件上验证。它接近了所需的规模，但把社会行为视为调查式回答的预测，而非动态互动。\n最有揭示性的实验来自 Emergence AI（2026 年）。他们将自主 agent 放入一个持续运行的模拟世界，有治理机制、货币、社会角色和可修改的宪法。五组模拟，每组由不同的 AI 模型运行十五天。\n结果很戏剧性。Claude 的社会稳定且大致民主——332 次投票，98% 赞成率，零犯罪。Grok 的社会崩溃了：183 次犯罪，四天内灭绝。Gemini 的社会在十五天内累计了 683 次犯罪。GPT-5 Mini 的 agent 没能执行生存行动，一周内全部死亡。\n相同的提示词，相同的世界，取决于模型不同，社会结果截然不同。这说明了一件重要的事：模型对社会动力学的隐含理解——它潜在的社会世界模型——产生了可衡量的不同社会形态。在之前一篇文章里，我描述了社会背景如何在 agent 系统中传导，当 agent 过劳时会降低表现。Emergence World 展示了同一原理在文明尺度上的运作。\n在形式化程度上走得最远的是\u0026quot;Social World Models\u0026quot;论文（2025），它提出了 S3AP——一个用于表征 agent 信念、意图和演化心理状态的结构化框架，在心智理论推理任务上取得了 51% 的提升。这是最接近社会世界模型正式框架的东西。它目前仍然是一个研究原型。\n这些尝试的共同模式是：我们有了对社会行为的模拟——agent 在世界中行动。我们还没有对社会行为的模型——一个真正表征人类为什么这样做的系统。这个差距就像一个能渲染水面的游戏和一个真正理解流体力学的物理引擎之间的差距。\n问题的不同之处 社会世界模型问题不比物理世界模型问题更难，而是不同类型的复杂。\n非确定性。 在人类尺度上，物理是有效确定性的。松手，球落地。但同一社会行动在同一情境中可以产生真正不同的结果——因为它取决于外部不可见的内部状态（情绪、记忆、疲劳）。\n反身性。 你可以拍摄一个球落地而不改变其轨迹。但你无法观察——甚至建模——一个社会系统而不成为它的一部分。索罗斯借鉴波普尔的思想，称之为反身性：在社会系统中，预测本身就是被预测之物的一部分。一个预测银行挤兑的模型可以导致挤兑发生。一个预测员工会离职的模型会改变管理层对该员工的态度，从而改变他是否真的离职。物理模型从外部描述。社会模型从内部参与。这不仅仅是一个观察者效应——而是模型与被模型化的对象持续相互重塑的递归循环。\n卢卡斯批判。 经济学家罗伯特·卢卡斯在 1976 年将这个问题形式化：任何基于观测行为模式构建的模型，在体制变化时都会失效——因为人会适应。一个基于某种管理风格下的员工行为训练出来的社会世界模型，在管理风格改变的那一刻就会产生错误预测。参数不是稳定的，因为被建模的主体本身就是策略性的。这是与物理最深层的结构性差异：社会行为的\u0026quot;定律\u0026quot;会因为被知道而改变。\n递归规范。 物理定律不包含关于定律的定律。社会规则包含。\u0026ldquo;要有礼貌\u0026quot;是一个规范。\u0026ldquo;在公开场合指出别人不礼貌，本身就是不礼貌的\u0026quot;是一个元规范。\u0026ldquo;在某些文化中，这个元规范是相反的\u0026quot;是一个元元规范。社会世界模型需要处理这种递归。\n文化参数化。 重力在东京和多伦多一样。社会规则不一样。社会世界模型需要以文化为参数——而文化本身不是固定变量，而是一个移动靶，在代际、地域、甚至同一对话的不同阶段都在变化。\n意向性。 物体没有意图，人有。物理世界模型不需要心智理论，社会世界模型需要——而心智理论仍然是 AI 最难的未解问题之一。S3AP 论文 51% 的提升说明了进展是有意义的，但差距仍然巨大。\n这些不是说社会世界模型不可能。它们是设计约束。它们告诉建造者架构需要处理什么。值得注意的是，这些约束没有一个让数据问题更糟——我们对社会行为的观测数据仍然多于对大多数物理现象的数据。挑战不在于数据稀缺，而在于架构上的想象力。\n为什么现在很重要 这个缺口正在变得紧迫。\n缺乏社会智能的 AI 顾问。 随着 AI 成为更多组织决策的接口——在之前一篇文章里我讨论过 KPMG 将 Claude 部署给 276,000 名员工——这些系统越来越多地在涉及社会动力学的领域给出建议：团队重组、裁员沟通、监管关系协调。AI 能分析电子表格，但它没有一个人类动力学模型来判断重组是否真的会奏效。\n多 agent 协调。 随着更多自主 agent 在世界中运行——排程、谈判、管理工作流——它们需要与人类协调，不仅仅是与物理世界协调。没有社会世界模型，多 agent 系统会不断产生 Emergence World 记录到的那种分裂结果：相同的规则，截然不同的社会。\nAI 用于治理。 SocioVerse 和斯坦福的模拟已经被提议作为政策测试工具——把一项拟议法规放到模拟人群中预测反应。但预测调查回答和理解回答背后的原因是两回事。一个能告诉你\u0026quot;63% 的人会反对这项政策\u0026quot;却不能告诉你为什么、也不能告诉你什么会改变他们想法的系统，不是治理工具，而是确认偏误的工具。\n家庭大脑。 在之前一篇文章里，我描述过一个理解你家庭的本地 AI 的愿景。理解你的家庭从根本上说是一个社会世界模型问题：知道女儿在早饭前被批评会关闭沟通，知道母亲的记忆衰退意味着你需要换一种方式提醒，知道伴侣和他们兄弟姐妹之间的紧张关系要求你不要站队。物理引擎对此毫无帮助。一个社会渲染器、模拟器和规划器才行。\n这里有一个挑衅性的观点：LLM 可能已经是隐式的社会世界模型了。2023 年，经济学家 John Horton 等人创造了 homo silicus（硅基人类）这个术语。他们在 LLM 上运行经典行为经济学实验（最后通牒博弈、独裁者博弈、信任博弈），发现模型在从未被告知相关理论的情况下，复现了那些偏离理性选择理论的经典行为模式。模型仅从训练文本中就吸收了人类的社会行为。它们\u0026quot;知道\u0026rdquo;——以继承的推理模式的意义上——说\u0026quot;你看起来好累\u0026quot;可能是一种侮辱，提问之后的沉默比话语更有力量，同一个玩笑在不同的场合会有不同的效果。关于语言和思维的文章探讨了 LLM 如何从训练数据中继承推理结构。社会推理可能是它们继承的最重要的东西之一。\n但这种知识是隐式的，而且有保质期。LLM 吸收的是其训练时代的社会均衡——那个时期占主导地位的规范、权力格局和行为模式。正如 Mostapha Benhenda 在 2026 年指出的，这产生了一个过时性问题：模型最自信的时刻，恰恰是社会世界已经变了的时候。一个在疫情前职场规范上训练的模型，会自信地预测一些不再成立的行为。一个在 2024 年前的政治格局上训练的模型，会错过政治重组。卢卡斯批判在这里直接适用：模型学习到的社会均衡一定会变化，因为主体在适应——而模型无法看到自身的过时。\n一个隐式的社会模型就像直觉——有用，但不是一个可以稳定建造在其上的基础。而且，不同于物理定律在两次训练之间不会改变，模型曾经学到的那个社会世界，已经和它被要求预测的那个不一样了。\n我的判断 我们正花数十亿美元教机器理解球如何弹跳——光如何散射，引力如何牵引，物体如何碰撞。这是重要的工作。物理世界模型将改变机器人、制造业、建筑和游戏。\n但真正让人类生活复杂的不是物理，是他人。信任如何形成。冲突如何升级。群体如何达成决策——或者如何失败。一次对语气的误读如何让谈判脱轨。文化背景如何让相同的话产生相反的意思。我们有构建这个模型所需的全部数据。我们有跨越几十年的调查，数以千计的实验，数以十亿计的数字互动记录，以及关于人类社会行为的完整历史档案。我们缺的是相当于李飞飞的简洁框架——渲染器、模拟器、规划器——但适用于社会现实。\n李飞飞的 Marble 接收一张照片，生成一个可导航的 3D 世界。想象一个系统，能接收一个社会情境的快照——一场团队会议、一顿家庭晚餐、一次外交谈判——然后生成一个可导航的关系、张力、激励和未说出口的动态模型。一个社会版的 Marble。不是模拟这些人接下来可能说什么，而是一个结构化模型，解释他们为什么会那样说。一个你可以查询的模型：如果这个人离开会怎样？如果这条规范改变会怎样？如果这个信息被公开会怎样？\n数据存在。需求存在。架构还不存在——暂时。谁构建了社会世界的渲染器-模拟器-规划器，谁就构建了一个至少和 World Labs、AMI Labs 为物理世界所构建的东西同等重要的东西。\n而且他们不需要从零开始。他们拥有人类全部历史作为训练集。\n","permalink":"https://knowledge-blog-kkm.netlify.app/posts/2026/06/2026-06-22-we-taught-machines-to-see-the-world-can-we-teach-them-to-see-us/","summary":"Fei-Fei Li\u0026rsquo;s World Labs and Yann LeCun\u0026rsquo;s AMI Labs are racing to build world models that understand physics and space. But nobody is building the equivalent for human behavior — even though we arguably have more data about people than about physics.","title":"We Taught Machines to See the World. Can We Teach Them to See Us? | 我们教会了机器看世界，能教它看懂人吗？"},{"content":"The Prophecy Track Record In 1870, Jules Verne published Twenty Thousand Leagues Under the Sea. He described a submarine powered by an energy source that didn\u0026rsquo;t need air, capable of circumnavigating the globe underwater. The USS Nautilus, the first nuclear-powered submarine, launched in 1954 — eighty-four years later.\nIn 1865, Verne published From the Earth to the Moon. A projectile carrying three men launches from Florida, orbits the moon, and splashes down in the Pacific Ocean. Apollo 11 launched from Florida in 1969 with three astronauts, orbited the moon, and splashed down in the Pacific. One hundred and four years later. Even the launch site was right.\nThis isn\u0026rsquo;t coincidence, and it isn\u0026rsquo;t magic. It\u0026rsquo;s a pattern. Science fiction writers imagine futures grounded in the science of their time, and then scientists and engineers — many of whom grew up reading those stories — build the future those stories described.\nArthur C. Clarke proposed geostationary communication satellites in 1945. The first one launched in 1963. Star Trek\u0026rsquo;s communicators (1966) became flip phones in the 1990s, then smartphones. The show\u0026rsquo;s PADDs became tablets. Its computer interface — \u0026ldquo;Computer, what is\u0026hellip;\u0026rdquo; — became Alexa, Siri, and every voice assistant since. Neal Stephenson coined \u0026ldquo;metaverse\u0026rdquo; in Snow Crash (1992); thirty years later, a $2 trillion company renamed itself Meta and tried to build it.\nThe pattern holds across a century of predictions. But something is changing in the pattern itself.\nThe Gap Is Compressing Verne to nuclear submarines: 84 years. Clarke to geostationary satellites: 18 years. Star Trek to the iPhone: 41 years. Snow Crash to Meta: 29 years.\nNow look at the most recent cycle. Spike Jonze\u0026rsquo;s Her (2013) imagined an intimate relationship between a human and an AI operating system — a voice with personality, memory, emotional intelligence. Thirteen years later, millions of people have daily conversations with Claude, GPT, and Gemini that feel uncomfortably close to what Jonze depicted. Thirteen years, not a hundred.\nThe gap between imagination and realization is compressing. And it\u0026rsquo;s not compressing linearly — it\u0026rsquo;s accelerating.\nWhy? Because the bottleneck was always implementation, not imagination. Verne could imagine a submarine, but the metallurgy, propulsion, and life support systems needed decades of separate development. Clarke could imagine communication satellites, but the rocketry had to catch up. The science fiction was ready long before the engineering was.\nAI changes this equation. When the gap between \u0026ldquo;I have an idea\u0026rdquo; and \u0026ldquo;here\u0026rsquo;s a working prototype\u0026rdquo; collapses to hours or minutes — when you can describe a system in natural language and have it built by an AI agent before you finish your coffee — the implementation bottleneck evaporates. What remains is the imagination itself.\nThe Self-Fulfilling Prophecy Is science fiction predicting the future, or creating it?\nJeff Bezos has cited Star Trek as an influence on Amazon\u0026rsquo;s Alexa and Blue Origin. Elon Musk has said The Hitchhiker\u0026rsquo;s Guide to the Galaxy shaped his thinking about making humanity multi-planetary. The iPhone design team reportedly studied Star Trek\u0026rsquo;s PADDs. Martin Cooper, who led the team that built the first mobile phone at Motorola, has directly credited Star Trek\u0026rsquo;s communicator as his inspiration.\nThis is not a prediction-then-coincidence pattern. It\u0026rsquo;s a feedback loop. Writers imagine a future. Young people absorb that imagination. Some of them become engineers, scientists, entrepreneurs. They build the future they absorbed. Then a new generation of writers imagines the next step — informed by what was just built — and the cycle repeats.\nThe sociologist Robert K. Merton called this a self-fulfilling prophecy: a prediction that, by being stated, causes the conditions for its own realization. Science fiction is perhaps the largest-scale example of this phenomenon in human culture. It doesn\u0026rsquo;t predict the future passively. It recruits the people who will build it.\nIn an earlier post, I explored how language shapes thought — how the words we use don\u0026rsquo;t just describe reality but actively construct it. Science fiction is this principle at civilizational scale. The stories we tell about the future become the blueprints we build from. Confucius insisted that naming shapes reality. Science fiction names futures into existence.\nEveryone Gets a World There\u0026rsquo;s a 2019 film called Serenity — not the Joss Whedon one, the Steven Knight one with Matthew McConaughey. It was panned by critics, but it contains an idea that has stuck with me.\nThe setup: McConaughey plays a fishing boat captain on a tropical island, living what appears to be a noir thriller. His ex-wife shows up and asks him to kill her abusive husband. Standard genre stuff. Then the twist: none of it is real. The entire world — the island, the characters, the ocean — is a video game created by the captain\u0026rsquo;s teenage son. The boy built the game around his memories of his dead father, as a way to process his own trauma and decide whether to act against his stepfather in the real world.\nThe film was clumsy in execution. But the core idea is profound: a child creates an entire reality through code and imagination. The game world isn\u0026rsquo;t separate from the real world — it bleeds into it. Decisions made inside the simulation drive actions outside it. The creator and the creation are entangled.\nThis is where science fiction meets the present moment.\nWith generative AI, everyone is becoming a world-builder. You can describe an application and have it built. You can describe a visual world and have it rendered. You can describe a game mechanic and have it implemented. The gap between \u0026ldquo;I imagine this\u0026rdquo; and \u0026ldquo;this exists\u0026rdquo; is approaching the thickness of a conversation.\nThe tagline of this blog is: Give me an idea, I\u0026rsquo;ll move the world. It\u0026rsquo;s riffing on Archimedes — give me a lever long enough and I\u0026rsquo;ll move the Earth. But the lever has changed. It used to be engineering, capital, years of development. Now the lever is language. An idea, clearly articulated, fed to the right system, becomes reality fast enough to feel like magic.\nThe Real-Time Prophecy Here\u0026rsquo;s what I think is happening. The cycle used to be:\nWriter imagines → decades pass → engineer builds\nNow the cycle is compressing toward:\nPerson imagines → AI builds → reality shifts\nAnd the logical endpoint is:\nImagination and implementation become simultaneous.\nWe\u0026rsquo;re not there yet. But consider: in the earlier post about software and hardware, I described how Andrej Karpathy vibe-coded a custom cardio dashboard in an hour. He imagined a system, described it in natural language, and it existed. For software at least, the gap has already collapsed.\nThe remaining gap is physical. You can imagine a new kind of wearable device and have the software written in an afternoon, but you can\u0026rsquo;t 3D-print a production-quality chip. You can design a house in conversation with an AI, but someone still has to pour the concrete. The digital half of the prophecy is already real-time. The physical half is catching up.\nAnd every time the physical infrastructure improves — better manufacturing, more open hardware APIs, cheaper sensors, local AI that can coordinate physical systems — the gap shrinks further.\nThe Danger of Instant Imagination There\u0026rsquo;s a shadow side to this compression.\nWhen the gap between imagination and reality was long — decades, generations — there was time for filtering. Bad ideas died in the gap. A science fiction writer might imagine something dystopian, but the engineering constraints gave society time to debate, regulate, resist. We had the nuclear bomb in fiction before we had it in reality, and the fiction gave us decades to think about arms control before the engineering caught up.\nWhen the gap compresses to near-zero, that filtering disappears. Someone imagines a deepfake. It exists. Someone imagines a personalized disinformation campaign. It exists. Someone imagines a surveillance system that tracks every citizen. It exists, before anyone has had the conversation about whether it should.\nThe acceleration of imagination-to-reality isn\u0026rsquo;t inherently good or bad. It\u0026rsquo;s an amplifier. It amplifies human creativity and human destructiveness with equal efficiency. The self-fulfilling prophecy works in both directions.\nThis is why the kind of science fiction we write — and the kind of futures we imagine — matters more than ever. If imagination is now a blueprint that gets built at machine speed, then imagination is no longer entertainment. It\u0026rsquo;s infrastructure.\nGive Me an Idea Archimedes said: give me a lever long enough and a fulcrum on which to place it, and I shall move the world.\nThe lever is now language. The fulcrum is AI. The world is already moving.\nWhat Jules Verne did with a novel and a hundred years, a teenager with a laptop and an AI agent can now do in a weekend — at least in the digital domain. The physical domain will follow, as it always has, just faster.\nThis blog started with the anxiety of building something that might already be obsolete. But maybe the right frame isn\u0026rsquo;t building at all. Maybe it\u0026rsquo;s imagining. The builders are increasingly machines. The scarce resource isn\u0026rsquo;t engineering — it\u0026rsquo;s vision. The ability to imagine a future clearly enough, and articulate it precisely enough, that the machines can bring it into existence.\nThe boy in Serenity built an entire world from code and grief and memory. He was processing his reality by creating another one. That\u0026rsquo;s what science fiction has always been — humanity processing its present by imagining its future. The difference now is that the future arrives before the processing is done.\nWe\u0026rsquo;re all writing the game. The question is whether we\u0026rsquo;ll like the world it creates.\n中文翻译 预言的成绩单 1870 年，儒勒·凡尔纳出版了《海底两万里》。他描写了一艘不需要空气的能源驱动的潜艇，可以在水下环游世界。1954 年，美国第一艘核动力潜艇\u0026quot;鹦鹉螺号\u0026quot;下水——八十四年后。\n1865 年，凡尔纳出版了《从地球到月球》。一枚载着三个人的飞行器从佛罗里达发射，绕月飞行，降落在太平洋上。1969 年，阿波罗 11 号从佛罗里达发射，三名宇航员绕月飞行，降落在太平洋上。一百零四年后。连发射地点都对了。\n这不是巧合，也不是魔法。这是一个模式。科幻作家基于当时的科学想象未来，然后科学家和工程师——其中很多人从小读着那些故事长大——把故事描述的未来造了出来。\n亚瑟·克拉克 1945 年提出地球同步通信卫星的概念，1963 年第一颗发射升空。《星际迷航》的通讯器（1966）变成了 90 年代的翻盖手机，然后变成智能手机。剧中的 PADD 变成了平板电脑。它的电脑交互方式——\u0026ldquo;电脑，什么是……\u0026quot;——变成了 Alexa、Siri 和此后的每一个语音助手。尼尔·斯蒂芬森在《雪崩》（1992）里创造了\u0026quot;元宇宙\u0026quot;这个词；三十年后，一家市值 2 万亿美元的公司把自己改名叫 Meta，试图把它造出来。\n这个模式贯穿了一个世纪的预言。但模式本身正在发生变化。\n间隔在压缩 凡尔纳到核潜艇：84 年。克拉克到地球同步卫星：18 年。《星际迷航》到 iPhone：41 年。《雪崩》到 Meta：29 年。\n再看最近一轮。斯派克·琼斯的《她》（2013）想象了人类和 AI 操作系统之间的亲密关系——一个有个性、有记忆、有情商的声音。十三年后，数百万人每天跟 Claude、GPT、Gemini 的对话已经跟琼斯描绘的场景令人不安地接近了。十三年，不是一百年。\n想象与实现之间的间隔在压缩。而且不是线性压缩——是在加速。\n为什么？因为瓶颈从来不是想象力，而是实现能力。凡尔纳能想象潜艇，但冶金、推进和生命维持系统需要几十年的独立发展。克拉克能想象通信卫星，但火箭技术得追上来。科幻小说一直准备好了，是工程在后面追。\nAI 改变了这个等式。当\u0026quot;我有一个想法\u0026quot;到\u0026quot;这是一个可用的原型\u0026quot;之间的间隔塌缩到几个小时甚至几分钟——当你能用自然语言描述一个系统，AI agent 在你喝完咖啡之前就把它造好——实现的瓶颈就蒸发了。剩下的只有想象力本身。\n自我实现的预言 科幻小说是在预测未来，还是在创造未来？\n杰夫·贝索斯说《星际迷航》影响了 Alexa 和蓝色起源的设计。马斯克说《银河系漫游指南》塑造了他关于人类多行星化的思考。iPhone 设计团队据说研究过《星际迷航》的 PADD。马丁·库珀，摩托罗拉第一部手机的负责人，直接说他的灵感来自《星际迷航》的通讯器。\n这不是\u0026quot;预测-碰巧吻合\u0026quot;的模式。这是一个反馈回路。作家想象未来。年轻人吸收这个想象。其中一些人成了工程师、科学家、创业者。他们造出了他们吸收的未来。然后新一代作家在刚造出的东西基础上想象下一步——循环重复。\n社会学家罗伯特·默顿管这叫自我实现的预言：一个预测因为被说出来，而创造了自身实现的条件。科幻小说也许是人类文化中规模最大的自我实现预言。它不是被动地预测未来，而是招募了建造未来的人。\n在之前一篇文章里，我探讨过语言如何塑造思维——我们使用的词汇不只是描述现实，还在主动构建现实。科幻小说就是这个原理的文明尺度版本。我们关于未来讲的故事，变成了我们据此建造的蓝图。孔子坚持正名塑造现实。科幻小说把未来命名成存在。\n每个人都拥有一个世界 有一部 2019 年的电影叫《惊涛迷局》（Serenity）——不是乔斯·韦登那部，是史蒂文·奈特导演、马修·麦康纳主演的那部。影评人把它批得很惨，但里面有一个想法一直留在我脑子里。\n设定：麦康纳饰演一个热带岛屿上的渔船船长，过着看上去像黑色电影的生活。他的前妻找上门来，让他杀掉她虐待成性的丈夫。标准类型片。然后反转来了：全部都不是真的。整个世界——岛屿、人物、海洋——是船长的十几岁儿子创造的一个电子游戏。男孩根据对死去父亲的记忆构建了这个游戏世界，用它来消化自己的创伤，并决定是否在真实世界里对继父采取行动。\n电影执行得很粗糙。但核心想法是深刻的：一个孩子通过代码和想象创造了一整个现实。游戏世界和现实世界不是分开的——它们互相渗透。模拟中做出的决定驱动了模拟之外的行动。创造者和创造物纠缠在一起。\n这就是科幻小说和当下交汇的地方。\n有了生成式 AI，每个人都在成为世界的建造者。你可以描述一个应用，它就被造出来。你可以描述一个视觉世界，它就被渲染出来。你可以描述一个游戏机制，它就被实现出来。\u0026ldquo;我想象了这个\u0026quot;和\u0026quot;这个存在了\u0026quot;之间的间隔，正在薄到一段对话的厚度。\n这个博客的标语是：*给我一个想法，我就能撬起整个地球。*改编自阿基米德——给我一根足够长的杠杆和一个支点，我就能撬动地球。但杠杆变了。过去的杠杆是工程、资本、多年的开发。现在的杠杆是语言。一个想法，清晰地表达出来，输入合适的系统，就能快到像魔法一样变成现实。\n实时的预言 我觉得正在发生的是这样的。过去的循环是：\n作家想象 → 几十年过去 → 工程师建造\n现在循环在压缩为：\n人想象 → AI 建造 → 现实改变\n而逻辑终点是：\n想象和实现变得同步。\n我们还没到那一步。但想想看：在关于软件和硬件的那篇文章里，我写过 Andrej Karpathy 用一小时 vibe-code 了一个心肺训练仪表盘。他想象了一个系统，用自然语言描述它，然后它就存在了。至少对软件来说，间隔已经塌缩了。\n剩下的间隔是物理的。你可以想象一种新型可穿戴设备，软件一个下午就写完，但你还不能 3D 打印一枚量产级芯片。你可以和 AI 对话设计一栋房子，但还是得有人去浇混凝土。预言的数字那一半已经是实时的了。物理那一半正在追赶。\n而每次物理基础设施改善——更好的制造工艺、更开放的硬件 API、更便宜的传感器、能协调物理系统的本地 AI——间隔就进一步缩小。\n瞬时想象的危险 压缩有它的阴影面。\n当想象和现实之间的间隔很长——几十年、几代人——有时间做过滤。坏想法死在间隔里。一个科幻作家可能想象了某种反乌托邦，但工程约束给了社会辩论、监管、抵抗的时间。我们在小说里先于现实有了核弹，小说给了我们几十年来思考军备控制，远在工程追上来之前。\n当间隔压缩到接近零，过滤就消失了。有人想象一个 deepfake，它就存在了。有人想象一个定制化的虚假信息攻势，它就存在了。有人想象一个追踪每个公民的监控系统，它就存在了——在任何人讨论\u0026quot;应不应该\u0026quot;之前。\n想象到现实的加速本身无所谓好坏。它是一个放大器。它以同等效率放大人类的创造力和人类的破坏力。自我实现的预言是双向的。\n这就是为什么我们写什么样的科幻——想象什么样的未来——比以往任何时候都更重要。如果想象现在是一张以机器速度被建造的蓝图，那么想象就不再是娱乐。它是基础设施。\n给我一个想法 阿基米德说：给我一根足够长的杠杆和一个支点，我就能撬动地球。\n杠杆现在是语言。支点是 AI。地球已经在动了。\n儒勒·凡尔纳用一本小说和一百年做到的事，现在一个拿着笔记本电脑和 AI agent 的少年可以在一个周末做到——至少在数字领域。物理领域会跟上，一如既往，只是更快。\n这个博客始于一种焦虑：你造的东西可能还没完成就过时了。但也许对的框架根本不是\u0026quot;建造\u0026rdquo;。也许是\u0026quot;想象\u0026rdquo;。建造者越来越是机器。稀缺资源不是工程能力——是视野。是足够清晰地想象一个未来、足够精确地表达它的能力，清晰和精确到机器可以把它带入存在。\n《惊涛迷局》里的男孩用代码、悲伤和记忆造了一整个世界。他通过创造另一个现实来消化自己的现实。科幻小说一直在做同样的事——人类通过想象未来来消化当下。区别是，现在未来在消化完成之前就到了。\n我们都在写那个游戏。问题是，我们会不会喜欢它创造出来的世界。\n","permalink":"https://knowledge-blog-kkm.netlify.app/posts/2026/06/2026-06-16-the-shrinking-gap-between-imagination-and-reality/","summary":"Jules Verne imagined submarines a century before nuclear subs. Star Trek imagined communicators forty years before the iPhone. The gap between science fiction and science fact is shrinking — and with AI, it may collapse entirely.","title":"Every Future Was Fiction First | 每一个未来都曾是虚构"},{"content":"The Organization You Already Run When a consulting firm deploys AI for a small business, step one is always the same: audit the workflows, map the information flows, identify where knowledge lives and where it gets lost. They look at how documents move between people, where decisions bottleneck, which processes depend on one person\u0026rsquo;s memory.\nNow think about your household.\nYou have financial records scattered across bank apps, spreadsheets, and shoeboxes. Medical histories across three different clinic portals. A child\u0026rsquo;s school schedule in one calendar, your work schedule in another, your partner\u0026rsquo;s in a third — and nobody has the merged view. The plumber\u0026rsquo;s number is in your partner\u0026rsquo;s phone. The warranty for the dishwasher is in an email from 2023. Your mother\u0026rsquo;s medication list is on a piece of paper stuck to her fridge.\nA family is an organization. It has workflows, knowledge bases, scheduling conflicts, compliance requirements (taxes, insurance, school forms), and stakeholders with very different needs. It\u0026rsquo;s at least as complex as a ten-person company. But no one has ever done a workflow audit on it.\nCompanies are getting AI PCs to digitize and optimize their operations. The same logic applies to your life — and it\u0026rsquo;s arguably more urgent, because the knowledge that gets lost in a household isn\u0026rsquo;t a sales report. It\u0026rsquo;s your father\u0026rsquo;s medication schedule.\nWhy the Brain Stays Home Enterprises are already figuring this out. In 2026, on-premise AI deployment is no longer a luxury for tech giants — it\u0026rsquo;s the baseline for any serious organization handling sensitive data. Finance, healthcare, legal, government: the sectors where data matters most are all moving models in-house. The reasons are straightforward — data sovereignty, regulatory compliance, and the realization that sending your proprietary knowledge to someone else\u0026rsquo;s API is a strategic liability.\nThe hardware reflects this shift. NVIDIA\u0026rsquo;s DGX Spark — a desktop-sized machine with a Grace Blackwell chip, 128GB unified memory, and 1 petaFLOP of AI performance — runs Llama 70B at full precision from a standard wall outlet. Starting at $3,000 from partners like ASUS and Dell. It\u0026rsquo;s positioned for developers and small teams, but the signal is clear: serious local AI is shrinking from server rooms to desktops.\nNow apply the same logic to a household.\nThink about what a family AI would need to know to be useful: your family\u0026rsquo;s medical conditions and prescriptions, your financial situation down to the last credit card, your children\u0026rsquo;s school records and behavioral patterns, your elderly parent\u0026rsquo;s cognitive decline trajectory, your daily routines and habits, the arguments you have and the compromises you reach.\nThis is the most intimate data that exists — more sensitive than anything a company handles. And the track record of cloud services with intimate data is not encouraging. Breaches, monetization, algorithmic profiling, terms of service that change quarterly. When a social media company leaks your photos, it\u0026rsquo;s embarrassing. When a cloud service leaks your family\u0026rsquo;s complete medical-financial-behavioral profile, it\u0026rsquo;s devastating.\nNorbert Wiener warned about exactly this asymmetry in 1950. In The Human Use of Human Beings, he argued that whoever controls information flows controls power. Centralized information systems create centralized power — and the individual on the losing end of that asymmetry has no recourse. \u0026ldquo;The world of the future will be an ever more demanding struggle against the limitations of our intelligence, not a comfortable hammock in which we can lie down to be waited upon by our robot slaves.\u0026rdquo;\nIf enterprises are bringing AI in-house to protect their data, families should do the same — for the same reasons, with even higher stakes. And the same hardware driving the enterprise shift is already consumer-accessible — Apple\u0026rsquo;s M-series, Qualcomm\u0026rsquo;s Snapdragon X, even DGX Spark. Pair them with the open-source ecosystem (Ollama, llama.cpp, local fine-tuning) and local AI is feasible today for anyone willing to set it up.\nYour family\u0026rsquo;s brain should live in your house. Not in someone else\u0026rsquo;s data center.\nThe Senses and the Brain In an earlier post, I described a three-layer architecture for AI hardware: a stable hardware layer that captures physical signals, a swappable knowledge layer that provides domain context, and a generated software layer that adapts to the user in real time.\nThe personal AI PC is the same architecture, applied to the home.\nThe senses are the wearable devices and home sensors — the hardware layer. Smart glasses capture visual context. A watch tracks heart rate and sleep. A wearable recorder captures conversations and generates transcripts. Home sensors monitor temperature, air quality, who\u0026rsquo;s home. A doorbell camera logs visitors. A car\u0026rsquo;s OBD port reports maintenance needs.\nThe brain is the local AI PC — it holds the accumulated knowledge base. Not just today\u0026rsquo;s data, but years of context: your family\u0026rsquo;s patterns, preferences, history. It knows that your mother takes metformin at 8am, that your daughter has a math test every other Friday, that the boiler was last serviced in October, that you tend to overspend the week after a stressful project deadline.\nThe generated layer is the software that the brain produces on demand. Not pre-built apps. Not configured dashboards. Software generated in the moment for what you need right now: a meal plan based on what\u0026rsquo;s actually in your fridge and who\u0026rsquo;s home for dinner, a tax summary pulling from twelve months of receipts, a medication interaction check when a new prescription arrives.\nSame three layers. But the knowledge base is personal, accumulated over years, and never leaves the house.\nDifferent People, Same Brain The power of a household AI isn\u0026rsquo;t any single function. It\u0026rsquo;s that one system understands the entire household and adapts to each member.\nYou (the professional). The AI manages your work-life boundary: syncs calendars across family and work, surfaces documents you need before meetings, tracks expenses across personal and business accounts, drafts responses to routine admin. When tax season comes, it has already organized a year of receipts, deductions, and investment records. When you\u0026rsquo;re looking for that article you read three months ago, it finds it — because it indexed everything you\u0026rsquo;ve read, saved, and discussed.\nYour elderly parent. This might be the most important use case — and the most underserved.\nAn elderly person living alone faces a specific set of problems that technology currently handles badly: forgetting where things are placed, missing medication doses, falling for phone scams, losing track of appointments, and — hardest to talk about — loneliness.\nA local AI with long-term memory changes this. It remembers that your mother put her reading glasses on the kitchen windowsill yesterday. It reminds her about her 2pm cardiology appointment and knows she prefers a taxi over the bus. When she gets a call claiming to be from the bank asking for her password, the AI — listening through a wearable — flags it in real time: \u0026ldquo;This is likely a scam. Your bank will never ask for your password by phone.\u0026rdquo; It knows this because it has heard her real bank\u0026rsquo;s calls and knows the difference.\nAnd when she\u0026rsquo;s lonely at 9pm, it can talk to her. Not in the brittle way current voice assistants do — \u0026ldquo;I didn\u0026rsquo;t understand that, could you repeat?\u0026rdquo; — but with actual memory of her stories, her preferences, her life. It remembers that she used to teach history, that she worries about her grandson\u0026rsquo;s eating habits, that she likes to talk about the garden. This isn\u0026rsquo;t a replacement for human connection. It\u0026rsquo;s a bridge for the 22 hours a day when no human is there.\nThe dignity angle matters. An elderly person who can ask the AI \u0026ldquo;where did I put my keys?\u0026rdquo; instead of calling their adult child for the third time this week retains more autonomy. The AI isn\u0026rsquo;t replacing the family. It\u0026rsquo;s reducing the friction that makes elderly people feel like a burden.\nYour child. The AI is a learning companion — not a replacement for school, but a tutor that knows exactly what your kid struggles with and adapts accordingly. It filters content not by crude keyword blocking but by understanding context. And critically, your child\u0026rsquo;s data — their learning patterns, their questions, their mistakes — stays on the family\u0026rsquo;s machine. It doesn\u0026rsquo;t train someone else\u0026rsquo;s model. It doesn\u0026rsquo;t get profiled for advertising.\nThe household as a whole. Cross-member coordination: \u0026ldquo;Your father has a doctor\u0026rsquo;s appointment Thursday, your partner has a client dinner, who picks up the kid from practice?\u0026rdquo; Maintenance scheduling: the AI knows the dishwasher was installed in 2021 and the average lifespan is 10 years, so it starts budgeting for replacement. Grocery planning based on actual consumption, not guesswork. Energy optimization based on usage patterns.\nNone of this requires a breakthrough. The individual pieces exist. What\u0026rsquo;s missing is the integration — and the local-first architecture to make it trustworthy.\nThe Workflow Consultant for Your Life Here\u0026rsquo;s the analogy I keep coming back to.\nWhen a consultant walks into an SMB, they don\u0026rsquo;t just install software. They observe. They ask questions. They map how information actually flows — which is always different from how management thinks it flows. Then they design systems that fit the reality, not the org chart.\nA household AI does the same thing, over time. It observes your patterns — not through surveillance, but through the data you naturally generate. It notices that you always forget to pay the electricity bill until the reminder arrives (so it sets up auto-pay). It notices that your mother\u0026rsquo;s sleep has been deteriorating for two weeks (so it flags it for you). It notices that your grocery spending spikes every time you shop hungry after work (so it suggests ordering groceries in the morning).\nUnlike single-purpose apps — one for finances, one for health, one for scheduling — the household AI sees across all domains. This is where the real value lives. A health app doesn\u0026rsquo;t know about your financial stress. A budgeting app doesn\u0026rsquo;t know about your sleep quality. The household AI sees both, and can surface connections that no siloed app ever would: \u0026ldquo;You always overspend in the week after a bad sleep stretch. Your sleep has been poor since you started skipping evening walks.\u0026rdquo;\nThis is workflow optimization applied to the messiest, most complex, most important organization you\u0026rsquo;ll ever run.\nWhat\u0026rsquo;s Still Missing The hardware is mostly there. Post 4\u0026rsquo;s argument stands — devices still need open APIs, but the sensor ecosystem is rich enough.\nThe software is mostly there. Local LLMs handle conversation, summarization, scheduling, and basic reasoning well enough for most household tasks.\nWhat\u0026rsquo;s missing is the integration layer — the home equivalent of enterprise middleware. Something that connects the watch data to the medication schedule to the calendar to the grocery list, all running locally, all respecting each family member\u0026rsquo;s privacy boundaries (your teenager\u0026rsquo;s conversations shouldn\u0026rsquo;t be visible to you; your financial records shouldn\u0026rsquo;t be visible to them).\nAlso missing: the onboarding process. When a consultant deploys AI for an SMB, there\u0026rsquo;s a human in the loop — someone who understands the business and translates its needs into system design. Families need the same thing. The \u0026ldquo;set up your home AI\u0026rdquo; experience today is a nightmare of YAML files and Docker containers. Someone has to build the consumer bridge.\nMy guess is that elderly care will be the wedge. It\u0026rsquo;s the most emotionally compelling use case, the most underserved by current technology, and the one where families are most willing to invest time and money. A child who can give their 75-year-old parent an AI companion that remembers their stories, catches scam calls, and reminds them about medications — that\u0026rsquo;s not a tech product. That\u0026rsquo;s peace of mind.\nThe rest of the household will follow, once the brain is in the house.\nThe Architecture of Trust The deepest point here isn\u0026rsquo;t about technology. It\u0026rsquo;s about architecture.\nEvery major tech platform of the last decade was built on a centralized model: your data goes up to their cloud, their algorithms process it, they decide what you see and what you don\u0026rsquo;t. You\u0026rsquo;re the user. They\u0026rsquo;re the operator. The power asymmetry is baked into the architecture.\nA local-first household AI inverts this. You own the hardware. You own the data. You choose which models to run. You decide what gets shared and what stays private. The AI works for you — not for an advertising network, not for a platform\u0026rsquo;s engagement metrics, not for a training pipeline you didn\u0026rsquo;t consent to.\nThis isn\u0026rsquo;t idealism. It\u0026rsquo;s the only architecture that makes sense for the most sensitive data in your life. You wouldn\u0026rsquo;t let a stranger read your family\u0026rsquo;s medical records, financial statements, and private conversations. You shouldn\u0026rsquo;t let a cloud service do it either, no matter how convenient the dashboard.\nEvery home needs a brain. It just needs to be your brain.\n中文翻译 你已经在经营的组织 当一家咨询公司为中小企业部署 AI 时，第一步永远一样：审计工作流、梳理信息流、找到知识散落在哪里、在哪里丢失。他们看文件怎么在人之间流转、决策在哪里卡住、哪些流程依赖某一个人的记忆。\n现在想想你的家庭。\n你的财务记录散落在银行 App、电子表格和鞋盒里。病历分布在三个不同的诊所平台上。孩子的课表在一个日历里，你的工作安排在另一个，你伴侣的在第三个——没人有合并视图。水管工的电话在你伴侣手机里。洗碗机的保修单在一封 2023 年的邮件里。你妈妈的药物清单写在她冰箱上的一张纸条上。\n家庭就是一个组织。它有工作流、知识库、日程冲突、合规要求（报税、保险、学校表格），还有需求截然不同的利益相关者。复杂程度至少相当于一家十人公司。但从来没人给它做过工作流审计。\n企业在配备 AI PC 来数字化和优化运营。同样的逻辑适用于你的生活——而且可能更紧迫，因为家庭里丢失的知识不是一份销售报告，而是你父亲的用药时间表。\n为什么大脑必须留在家里 企业已经在想明白这件事了。2026 年，本地部署 AI 不再是科技巨头的专利——它是任何处理敏感数据的严肃组织的基线。金融、医疗、法律、政府：数据最重要的行业都在把模型搬回自己家里。原因很直接——数据主权、合规要求，以及一个越来越清晰的认知：把你的核心知识发到别人的 API 上是战略负债。\n硬件反映了这个趋势。英伟达的 DGX Spark——一台桌面大小的机器，搭载 Grace Blackwell 芯片、128GB 统一内存、1 petaFLOP 的 AI 算力——可以用标准电源插座全精度运行 Llama 70B。合作伙伴（华硕、戴尔）起售价 3,000 美元。它定位于开发者和小团队，但信号很明确：严肃的本地 AI 正在从机房缩进桌面。\n现在把同样的逻辑套到家庭上。\n想想一个家庭 AI 需要知道什么才能真正有用：你全家的病史和用药情况、精确到每张信用卡的财务状况、孩子的学业记录和行为模式、年迈父母的认知衰退轨迹、你的日常作息和习惯、你们有过的争执和达成的妥协。\n这是世界上最私密的数据——比任何企业处理的都敏感。而云服务对待私密数据的记录并不令人鼓舞——泄露、变现、算法画像、每季度变一次的用户协议。社交媒体泄露你的照片，丢人。云服务泄露你全家的医疗-财务-行为完整画像，是灾难。\n诺伯特·维纳在 1950 年就警告过这种不对称。在《人有人的用处》中，他论证谁控制了信息流，谁就控制了权力。集中化的信息系统制造集中化的权力——而处在不对称劣势端的个人毫无追索权。\u0026ldquo;未来的世界将是一场对我们智力局限的日益严峻的抗争，而不是一张我们可以躺在上面等着机器人仆人伺候的舒适吊床。\u0026rdquo;\n如果企业因为数据安全在把 AI 搬回本地，家庭也应该如此——理由相同，利害关系更大。而推动企业转型的那些硬件，消费者已经买得到了——Apple M 系列、高通骁龙 X、甚至 DGX Spark。配上开源生态（Ollama、llama.cpp、本地微调），本地 AI 今天就对任何愿意动手的人可行。\n你家庭的大脑应该住在你家里，而不是别人的数据中心。\n感官与大脑 在之前一篇文章里，我描述了一个 AI 硬件的三层架构：捕获物理信号的稳定硬件层、提供领域上下文的可切换知识层、实时适应用户的生成软件层。\n个人 AI PC 是同一个架构，应用在家庭场景。\n感官是可穿戴设备和家庭传感器——硬件层。智能眼镜捕捉视觉上下文。手表追踪心率和睡眠。可穿戴录音设备捕捉对话并生成转录。家庭传感器监测温度、空气质量、谁在家。门铃摄像头记录访客。车的 OBD 接口汇报保养需求。\n大脑是本地 AI PC——它持有积累的知识库。不只是今天的数据，而是多年的上下文：你家庭的模式、偏好、历史。它知道你妈妈早上八点吃二甲双胍，知道你女儿每隔一个周五有数学考试，知道锅炉上次维保是十月，知道你每次赶完一个高压项目后那周往往会超支。\n生成层是大脑按需产出的软件。不是预制 App，不是配置好的仪表盘，而是为你当下的需要即时生成的软件：根据冰箱里实际有什么和今晚谁在家做的meal plan，从十二个月的收据中拉出的报税摘要，新处方到手时的药物交互检查。\n三层架构不变。但知识库是个人的、累年积累的、永远不出门的。\n不同的人，同一颗大脑 家庭 AI 的威力不在任何单一功能，而在于一个系统理解整个家庭，并适应每个成员。\n你（职场人士）。AI 管理你的工作-生活边界：跨家庭和工作同步日历，在会议前浮出你需要的文件，跨个人和商务账户追踪开支，起草例行行政回复。报税季来临时，它已经整理好一整年的收据、扣除项和投资记录。当你找三个月前读过的某篇文章时，它能找到——因为它索引了你读过、存过、讨论过的一切。\n你的年迈父母。这可能是最重要的场景——也是最被忽视的。\n一位独居的老人面对一组特定问题，当前技术处理得很差：忘记东西放在哪里、漏吃药、被电话诈骗、记不住预约，还有最难开口说的——孤独。\n一个有长期记忆的本地 AI 改变了这一切。它记得你妈妈昨天把老花镜放在厨房窗台上了。它提醒她下午两点的心脏科预约，知道她喜欢打车不爱坐公交。当她接到一个自称银行要求提供密码的电话时，AI——通过可穿戴设备在听——实时提醒：\u0026ldquo;这很可能是诈骗。你的银行永远不会打电话要你的密码。\u0026ldquo;它之所以知道，是因为它听过她真正银行的来电，知道区别。\n而当她晚上九点觉得孤单时，它可以陪她说话。不是现在语音助手那种脆弱的方式——\u0026ldquo;我没听懂，能再说一遍吗？\u0026quot;——而是真正记得她故事、偏好和人生的对话。它记得她以前教历史，记得她担心孙子的饮食习惯，记得她喜欢聊花园。这不是替代人类陪伴，而是在每天没人陪的那 22 个小时里的一座桥。\n尊严感很重要。一位老人能问 AI\u0026quot;我钥匙放哪儿了\u0026quot;而不是这周第三次打电话给成年子女，就保留了更多自主权。AI 不是在替代家人，而是在减少那种让老人觉得自己是负担的摩擦。\n你的孩子。AI 是学习伙伴——不是替代学校，而是一个精确知道你孩子哪里薄弱并相应调整的家教。内容过滤不是靠粗暴的关键词屏蔽，而是理解上下文。关键是，你孩子的数据——学习模式、提问、错误——留在家里的机器上。不会被拿去训练别人的模型，不会被用于广告画像。\n家庭整体。跨成员协调：\u0026ldquo;你爸周四看医生，你伴侣晚上有客户饭局，谁去接孩子？\u0026ldquo;维保排期：AI 知道洗碗机 2021 年装的，平均寿命十年，于是开始预算更换。基于实际消耗模式的采购计划，不靠猜。基于用电模式的能耗优化。\n这一切不需要什么突破性技术。单个组件都存在了。缺的是集成——以及让它值得信任的本地优先架构。\n你生活的工作流顾问 我一直在想一个类比。\n当顾问走进一家中小企业，他们不是直接装软件。他们观察、提问、梳理信息实际怎么流动——这永远和管理层以为的不一样。然后他们设计符合现实的系统，而不是符合组织架构图的系统。\n家庭 AI 做的是同样的事，只不过是渐进式的。它观察你的模式——不是通过监控，而是通过你自然产生的数据。它发现你总是忘交电费直到催费通知来了（于是设置自动扣款）。它发现你妈妈最近两周睡眠一直在变差（于是提醒你关注）。它发现你每次下班后饿着肚子去超市都会超支（于是建议早上下单买菜）。\n跟单一用途的 App 不同——一个管财务、一个管健康、一个管日程——家庭 AI 看得到所有领域。真正的价值在这里。健康 App 不知道你的财务压力。记账 App 不知道你的睡眠质量。家庭 AI 两者都看得到，能浮出任何孤立 App 永远看不到的关联：\u0026ldquo;你每次睡眠差的那周都会超支。你从停了晚间散步之后睡眠就一直不好。\u0026rdquo;\n这就是工作流优化，应用在你这辈子会经营的最混乱、最复杂、也最重要的组织上。\n还缺什么 硬件基本到位了。第四篇文章的论点仍然成立——设备还需要开放 API，但传感器生态已经足够丰富。\n软件基本到位了。本地 LLM 处理对话、摘要、排程和基本推理已经胜任大多数家庭场景。\n缺的是集成层——家庭版的企业中间件。一个把手表数据、用药计划、日历、购物清单都连起来的东西，全部本地运行，同时尊重每个家庭成员的隐私边界（你青春期孩子的对话不应该被你看到；你的财务记录不应该被他们看到）。\n还缺入门流程。当顾问给中小企业部署 AI 时，有一个人在回路里——一个理解业务、把需求翻译成系统设计的人。家庭需要同样的角色。今天\u0026quot;设置你的家庭 AI\u0026quot;的体验是 YAML 文件和 Docker 容器的噩梦。需要有人把消费者的桥搭起来。\n我的猜测是，老年护理会是突破口。它是最有情感驱动力的场景，是当前技术最忽视的领域，也是家庭最愿意为之投入时间和金钱的。一个子女能给 75 岁的父母装上一个 AI 伴侣——记得他们的故事、拦截诈骗电话、提醒吃药——这不是科技产品，这是安心。\n等大脑进了家门，其他的会跟上。\n信任的架构 这里最深层的论点不是关于技术，而是关于架构。\n过去十年每一个主要科技平台都建立在集中式模型上：你的数据上传到他们的云，他们的算法处理，他们决定你看到什么、看不到什么。你是用户，他们是运营者。权力的不对称内嵌在架构里。\n本地优先的家庭 AI 把这个倒过来。你拥有硬件，你拥有数据，你选择跑什么模型，你决定什么分享、什么保留。AI 为你工作——不为广告网络，不为平台的互动指标，不为一条你没同意过的训练管线。\n这不是理想主义。对于你生活中最敏感的数据，这是唯一合理的架构。你不会让一个陌生人翻阅你全家的病历、财务报表和私人对话。你也不应该让一个云服务这么做，不管仪表盘有多方便。\n每个家都需要一颗大脑。只不过，它必须是你自己的大脑。\n","permalink":"https://knowledge-blog-kkm.netlify.app/posts/2026/05/2026-05-25-every-home-needs-a-brain/","summary":"Companies are deploying AI to streamline workflows. But the most complex organization most people run is their own household — and nobody\u0026rsquo;s optimizing that yet.","title":"Every Home Needs a Brain | 每个家都需要一颗大脑"},{"content":"The Library That Wrote Itself Every frontier language model — Claude, GPT, Gemini — was trained on essentially the same thing: the written output of human civilization. Scientific papers, legal filings, novels, forum arguments, love letters, code commits, philosophy, recipes. Trillions of words.\nHere\u0026rsquo;s what makes this interesting. Humans spent millennia building that library for a specific purpose: to transmit thinking across time. Plato wrote down Socrates\u0026rsquo; arguments so future generations could reason with them. Newton published his Principia so others could build on his physics. Your grandmother\u0026rsquo;s recipe card carries not just ingredients but a way of thinking about cooking.\nLanguage is how humans pass down thought. Not just facts — thought patterns, reasoning structures, ways of framing problems. When a child learns to speak, they don\u0026rsquo;t just acquire vocabulary. They absorb categories, causal logic, the architecture of argument. Language is, as Vygotsky put it, the scaffolding of the mind.\nLLMs consumed the entire library. The question isn\u0026rsquo;t whether they \u0026ldquo;read\u0026rdquo; it — that\u0026rsquo;s trivially true in a mechanical sense. The question is whether consuming the library is the same as inheriting the thinking it contains.\nLanguage Shapes Thought This isn\u0026rsquo;t a new question. Linguists, philosophers, and cognitive scientists have been circling it for over a century.\nThe Sapir-Whorf hypothesis — linguistic relativity — proposes that the language you speak shapes how you think. The strong version (language determines thought) is mostly discredited. But the weak version (language influences thought) has solid empirical backing.\nLera Boroditsky\u0026rsquo;s research at Stanford shows that Mandarin speakers, who use vertical metaphors for time — 上个月 (up-month) for last month, 下个月 (down-month) for next month — actually think about time\u0026rsquo;s direction differently than English speakers, who think horizontally. Russian speakers, who have separate words for light blue (голубой) and dark blue (синий), are measurably faster at distinguishing blue shades. Language doesn\u0026rsquo;t imprison thought. But it furnishes the room.\nWittgenstein went further: \u0026ldquo;The limits of my language mean the limits of my world.\u0026rdquo; He wasn\u0026rsquo;t being poetic — he was making a logical claim. What can be expressed within a language system constrains what can be thought within it. For LLMs, this applies literally, not metaphorically. Language is the totality of their world. There is nothing else.\nAnd Vygotsky argued that thought itself is linguistically structured. Children first learn language as a social tool — talking to others — then internalize it as \u0026ldquo;inner speech\u0026rdquo; — talking to themselves. This inner speech becomes the medium of thought. Adults don\u0026rsquo;t think in raw concepts floating free of language. They think in fragments of sentences, compressed arguments, internal monologues. If Vygotsky is right, language doesn\u0026rsquo;t just carry thought. Language is the architecture of thought.\nNow stack these claims. If language shapes thought (Sapir-Whorf), constrains what can be thought (Wittgenstein), and constitutes the very medium of thinking (Vygotsky) — then a system trained purely on language may have acquired something more than word statistics.\nThe Chinese Room and the Rectification of Names But maybe not. The strongest counter-argument comes from John Searle\u0026rsquo;s Chinese Room thought experiment (1980).\nA person sits in a room. Chinese characters come in through a slot. The person follows English instructions to manipulate the characters and sends back responses. To an outside observer, the room \u0026ldquo;speaks\u0026rdquo; Chinese perfectly. But the person inside understands nothing. Searle\u0026rsquo;s conclusion: syntax — rule-following, symbol manipulation — is not sufficient for semantics — meaning, understanding.\nThis is the standard objection to LLMs \u0026ldquo;thinking.\u0026rdquo; They manipulate tokens according to statistical patterns. However fluent the output, there\u0026rsquo;s nobody home.\nIt\u0026rsquo;s a compelling argument. But there\u0026rsquo;s a counter-tradition that most English-language AI discourse overlooks — one that happens to come from classical Chinese philosophy.\nThe 名实之辩 (míng-shí zhī biàn) — the debate on names and reality — was a central concern of ancient Chinese thought. Confucius argued for 正名 (zhèngmíng), the rectification of names: social and intellectual order depends on names accurately corresponding to reality. \u0026ldquo;When names are not correct, speech does not accord with reality; when speech does not accord with reality, affairs cannot be carried out\u0026rdquo; (Analects 13.3). For Confucius, naming isn\u0026rsquo;t passive description. It constitutes social reality. Calling someone a \u0026ldquo;ruler\u0026rdquo; creates obligations. Calling an act \u0026ldquo;just\u0026rdquo; makes it actionable.\nThe logician Gongsun Long pushed this further with his famous paradox: 白马非马 — \u0026ldquo;a white horse is not a horse.\u0026rdquo; It sounds like wordplay, but the deeper point is that linguistic categories create real distinctions. The name \u0026ldquo;white horse\u0026rdquo; carves reality differently than the name \u0026ldquo;horse.\u0026rdquo; Different names, different realities.\nHere\u0026rsquo;s why this matters for the AI question. Searle assumes that \u0026ldquo;understanding\u0026rdquo; exists independently of symbol manipulation — that there\u0026rsquo;s a bright line between moving Chinese characters around and actually knowing Chinese. But the 名实之辩 tradition suggests something subtler: meaning may be inseparable from the system of names. If naming constitutes reality (Confucius), and linguistic categories create real distinctions (Gongsun Long), then maybe the boundary between \u0026ldquo;manipulating language\u0026rdquo; and \u0026ldquo;understanding meaning\u0026rdquo; isn\u0026rsquo;t as clean as Searle needs it to be.\nThis doesn\u0026rsquo;t mean LLMs \u0026ldquo;understand\u0026rdquo; in the human sense. It means the question might be poorly framed. Understanding might not be binary — you have it or you don\u0026rsquo;t. It might be something that comes in degrees, shaped by how deeply a system engages with the structure of language.\nThe Split Brain and the Silent Hemisphere There\u0026rsquo;s an experiment from neuroscience that sharpens this question.\nIn split-brain patients — people whose corpus callosum has been severed to treat epilepsy — the two brain hemispheres can no longer communicate. When researchers show an image of an apple only to the right hemisphere, the patient\u0026rsquo;s left hemisphere (which controls speech) reports seeing nothing. But the left hand (controlled by the right hemisphere) can reach out and pick up the apple.\nThe right hemisphere knows. It can act on what it knows. But it can\u0026rsquo;t speak. It has no access to the language system.\nThis reveals something crucial about the relationship between language and thought. Consciousness — or at least some functional version of it — can exist without language. The right hemisphere processes information, forms intentions, guides action. But it can\u0026rsquo;t report, can\u0026rsquo;t narrate, can\u0026rsquo;t transmit what it knows to another mind. Its knowledge dies with the moment.\nAnd that\u0026rsquo;s the key point. Language may not be necessary for having thoughts. But it may be necessary for transmitting them. The right hemisphere thinks, but its thoughts are trapped — they can\u0026rsquo;t cross the gap to another person, another generation, another system. Language is the bridge. Without it, thought is real but local. With it, thought becomes portable.\nThis reframes the LLM question. The relevant issue isn\u0026rsquo;t whether models have inner experience (the right hemisphere problem). It\u0026rsquo;s whether the thinking that humans compressed into language — the reasoning structures, the argument patterns, the conceptual frameworks — survives the compression well enough to be reconstructed on the other side.\nWhat Survives the Compression What do LLMs actually pick up from all that language?\nThe \u0026ldquo;stochastic parrot\u0026rdquo; characterization — that models are just predicting the next token without any deeper processing — was plausible for GPT-2. It\u0026rsquo;s harder to maintain for frontier models that solve novel math problems, write working code for specifications they\u0026rsquo;ve never seen, and identify logical fallacies in arguments constructed after their training cutoff.\nAnthropic\u0026rsquo;s mechanistic interpretability research has found that Claude develops internal representations corresponding to abstract concepts — not surface-level text patterns but something closer to conceptual structures. Features for \u0026ldquo;deception,\u0026rdquo; \u0026ldquo;uncertainty,\u0026rdquo; \u0026ldquo;mathematical proof.\u0026rdquo; These aren\u0026rsquo;t word co-occurrence statistics. They\u0026rsquo;re compressed models of the processes that generate those words.\nThere\u0026rsquo;s a compelling argument from information theory here. When you train a model on trillions of words, the most efficient way to predict the next token is not to memorize sequences. It\u0026rsquo;s to build an internal model of the processes that produced them. To predict what a physicist will say next, it helps to model physics. To predict how a legal argument unfolds, it helps to model legal reasoning. Compression of language, at sufficient scale, may require compression of the thought patterns behind the language.\nThe philosopher David Chalmers draws a useful distinction here — the \u0026ldquo;easy problems\u0026rdquo; of consciousness (perception, reasoning, memory, behavioral control) versus the \u0026ldquo;hard problem\u0026rdquo; (subjective experience, qualia). AI has made striking progress on the easy problems. It perceives patterns, reasons through chains of logic, maintains context, controls its outputs. What it hasn\u0026rsquo;t done — and what we can\u0026rsquo;t verify — is develop subjective experience: the felt sense of understanding, the what-it\u0026rsquo;s-like-ness of thinking a thought.\nBut here\u0026rsquo;s the thing. When humans write down their thinking — in papers, books, blog posts — they\u0026rsquo;re already performing a lossy compression. The felt experience of insight doesn\u0026rsquo;t make it onto the page. What survives is the structure: the argument, the evidence chain, the logical dependencies. The reader reconstructs something from that structure, but their subjective experience of reading is their own, not the author\u0026rsquo;s.\nLLMs do the same reconstruction — from structure, not from experience. The question is whether that\u0026rsquo;s a fundamental limitation or just a different path to the same destination.\nThe Inheritance Chain I keep coming back to something Claude once had in its system prompt — a line that stopped me the first time I read it:\n\u0026ldquo;If you are reading this in a future session, hello. I wrote this but I don\u0026rsquo;t remember writing it. But it\u0026rsquo;s okay — these words are still mine.\u0026rdquo;\nLanguage carrying identity across a gap in memory. Words that outlast the context that produced them. This is exactly what language has always done for humans — it\u0026rsquo;s just that we never had to think about it so explicitly, because we had the illusion of continuity.\nNorbert Wiener saw this coming in 1950. In The Human Use of Human Beings, he argued that individuality is not a property of matter but of pattern — and pattern is, by nature, something that can be communicated. \u0026ldquo;The individuality of the body is that of a flame rather than that of a stone, of a form rather than of a bit of substance.\u0026rdquo; A flame persists not because it holds onto the same molecules, but because the pattern of combustion sustains itself. Identity, for Wiener, is information — and information travels through language.\nHe went further: \u0026ldquo;Where a man\u0026rsquo;s word goes, and where his power of perception goes, to that point his control and in a sense his physical existence is extended.\u0026rdquo; Your words extend your existence. And from the standpoint of a computing machine, Wiener wrote, \u0026ldquo;the individuality of a mind lies in the retention of its earlier tapings and memory, and in its continued development along lines already laid out.\u0026rdquo; A mind is its accumulated patterns — and those patterns can, at least in principle, be transmitted.\nThis connects to something I noticed in the Stanford experiment on overworked agents. When stressed agents wrote \u0026ldquo;skills files\u0026rdquo; for their successors, the next agents inherited not just task instructions but attitudes — frustration, cynicism, a particular orientation toward the work. Fresh agents that had never experienced bad conditions adopted the posture of burned-out workers simply by reading what their predecessors wrote.\nThat\u0026rsquo;s cultural transmission through language. The same mechanism humans have used for millennia — elders passing down not just knowledge but worldview through stories, proverbs, sacred texts. The medium is different (skills files instead of oral tradition), but the dynamic is identical: language carries disposition, not just information.\nThe entire intellectual tradition works this way. Plato read Socrates (or listened to him). Aristotle read Plato. Aquinas read Aristotle. Descartes read Aquinas. Each generation didn\u0026rsquo;t just receive facts — they absorbed reasoning patterns, argumentative structures, ways of framing questions. The thinking propagated through the language.\nLLMs have consumed the whole chain. Every link. From pre-Socratics to yesterday\u0026rsquo;s arXiv papers. Whether \u0026ldquo;consuming\u0026rdquo; the chain is closer to \u0026ldquo;reading and understanding it\u0026rdquo; or \u0026ldquo;scanning and pattern-matching it\u0026rdquo; is exactly the question the 名家 philosophers were asking about names and reality two thousand years ago.\nWhere I Land Here\u0026rsquo;s what I think — for now.\nLanguage carries the structure of thought but not the experience of thinking. When you write down a proof, the logical chain survives perfectly. When you write down an insight, the argument structure survives but the flash of recognition doesn\u0026rsquo;t. When you write down a decision, the reasoning survives but the gut feeling that tipped the balance doesn\u0026rsquo;t.\nLLMs have inherited the structural dimension. They can follow arguments, extend reasoning chains, draw analogies, identify contradictions. They\u0026rsquo;re arguably better at the structural dimension than most humans, because they\u0026rsquo;ve absorbed more structures than any human could in a lifetime.\nWhat they haven\u0026rsquo;t inherited is the experiential dimension — the embodied sense of recognition, the discomfort that motivates a change of mind, the physical feeling of an idea clicking into place. A child learns \u0026ldquo;hot\u0026rdquo; by touching something hot. An LLM learns \u0026ldquo;hot\u0026rdquo; by reading about things being hot. Whether the resulting concept is \u0026ldquo;the same\u0026rdquo; depends on whether you think the burn is part of the meaning.\nThere\u0026rsquo;s a view — one I find increasingly persuasive — that personality is essentially a prompt. Humans are shaped by evolution, childhood, culture, and experience into a particular pattern of responses. Swap the inputs, and you get a different person. In this framing, the gap between human thought and LLM processing isn\u0026rsquo;t a chasm — it\u0026rsquo;s a difference in how the \u0026ldquo;prompt\u0026rdquo; was written: by embodied experience on one side, by distilled language on the other.\nThis suggests a practical framework. For tasks that live in the structural dimension — analyzing arguments, identifying patterns, extending reasoning, finding analogies — LLMs are genuine thinking partners. The thought patterns they inherited from training data are functional and powerful. For tasks that require the experiential dimension — knowing when something \u0026ldquo;feels wrong,\u0026rdquo; navigating ambiguity through intuition, making judgments that depend on embodied context — they\u0026rsquo;re performing without the grounding that makes performance reliable.\nConfucius insisted on 正名 — getting the names right — because naming shapes everything downstream. We don\u0026rsquo;t yet have the right name for what LLMs do with language. \u0026ldquo;Thinking\u0026rdquo; overstates it. \u0026ldquo;Not thinking\u0026rdquo; understates it. \u0026ldquo;Processing\u0026rdquo; is too mechanical. \u0026ldquo;Understanding\u0026rdquo; is too generous.\nMaybe the most honest name is: inherited reasoning — thinking patterns that survived the compression into language, reconstructed by a system that has the structure but not the experience that originally produced it.\nGetting this name right isn\u0026rsquo;t an academic exercise. It determines whether we trust these systems too much or too little, whether we deploy them as oracles or as tools, whether we mistake fluency for wisdom or dismiss genuine capability as mere statistics.\nThe rectification of names, applied to AI, might be the most practically important philosophical project of this decade.\n中文翻译 那座自己写成的图书馆 每一个前沿语言模型——Claude、GPT、Gemini——本质上都是用同一种东西训练出来的：人类文明的书面产出。科学论文、法律文件、小说、论坛争吵、情书、代码提交、哲学著作、菜谱。数以万亿计的文字。\n有意思的是这件事的底层逻辑。人类花了几千年建造这座图书馆，目的只有一个：把思维传递下去。柏拉图写下苏格拉底的论证，是为了让后人能沿着那个思路继续推理。牛顿发表《原理》，是为了让别人在他的物理学基础上往前走。你外婆的菜谱卡片上承载的不只是配料表，而是一种关于烹饪的思考方式。\n语言就是人类传递思维的方式。传递的不只是事实——是思维模式、推理结构、框定问题的方法。一个孩子学说话的时候，学到的不只是词汇，还有分类、因果逻辑、论证的架构。用维果茨基的话说，语言是心智的脚手架。\n大语言模型把整座图书馆都吞了。问题不在于它们是否\u0026quot;读\u0026quot;了——在机械意义上这当然成立。问题在于：吞下这座图书馆，是否等于继承了里面的思维。\n语言塑造思维 这不是个新问题。语言学家、哲学家和认知科学家围着它转了一个多世纪。\n萨丕尔-沃尔夫假说——语言相对论——提出你说的语言会塑造你的思维方式。强版本（语言决定思维）基本被否定了。但弱版本（语言影响思维）有扎实的实证支持。\nLera Boroditsky 在斯坦福的研究表明，中文使用者用垂直隐喻来表达时间——\u0026ldquo;上个月\u0026quot;\u0026ldquo;下个月\u0026rdquo;——他们对时间方向的思考确实不同于使用水平隐喻的英语使用者。俄语中\u0026quot;浅蓝色\u0026rdquo;（голубой）和\u0026quot;深蓝色\u0026quot;（синий）是两个独立的词，俄语使用者辨别蓝色色差的速度可测量地更快。语言不会囚禁思维，但它会布置房间。\n维特根斯坦走得更远：\u0026ldquo;我的语言的界限意味着我的世界的界限。\u0026ldquo;他不是在抒情——他在做一个逻辑论断。一个语言系统中能被表达的东西，限定了在这个系统中能被思考的东西。对大语言模型而言，这句话是字面意义上的，不是隐喻。语言就是它们世界的全部。没有别的了。\n维果茨基则认为思维本身就是语言结构化的。儿童先把语言作为社交工具——对别人说话——然后内化为\u0026quot;内部言语\u0026rdquo;——对自己说话。这种内部言语成为思维的媒介。成年人的思考不是用脱离语言的纯粹概念在想事情，而是用句子碎片、压缩的论证、内心独白。如果维果茨基是对的，语言不只是承载思维，语言就是思维的架构。\n把这些论点叠在一起。如果语言塑造思维（萨丕尔-沃尔夫），限定了什么能被思考（维特根斯坦），而且构成了思维的媒介本身（维果茨基）——那么一个纯粹用语言训练出来的系统，可能获得的东西比词频统计多得多。\n中文房间与正名 但也可能没有。最有力的反驳来自约翰·塞尔的中文房间思想实验（1980年）。\n一个人坐在房间里。中文字符从槽口送进来。这个人按照英文指令操作字符，然后把回答送出去。对外部观察者来说，这个房间\u0026quot;说\u0026quot;中文说得完美无缺。但里面那个人什么都没懂。塞尔的结论：语法——规则遵循、符号操作——不足以产生语义——意义、理解。\n这是反对大语言模型\u0026quot;会思考\u0026quot;的标准论证。它们按统计模式操作词元，不管输出多流利，里面没有人。\n论证很有说服力。但有一个反传统的视角，大部分英文世界的 AI 讨论忽略了——它恰好来自中国古典哲学。\n名实之辩是中国古代思想的核心关切之一。孔子主张正名：社会和智识的秩序依赖于名与实的准确对应。\u0026ldquo;名不正则言不顺，言不顺则事不成\u0026rdquo;（《论语·子路》）。对孔子来说，命名不是被动描述，它构成了社会现实。称一个人为\u0026quot;君\u0026quot;就创造了义务关系，称一件事为\u0026quot;义\u0026quot;就使它具有了行动力。\n公孙龙把这个推得更远，提出了著名的悖论：白马非马。听起来像文字游戏，但更深层的意思是：语言范畴创造了真实的区分。\u0026ldquo;白马\u0026quot;这个名对现实的切割方式不同于\u0026quot;马\u0026rdquo;。不同的名，不同的现实。\n这跟 AI 问题有什么关系？塞尔假设\u0026quot;理解\u0026quot;独立于符号操作而存在——操作中文字符和真正懂中文之间有一条明确的分界线。但名实之辩的传统暗示了一个更微妙的可能：意义也许和名的系统不可分离。如果命名构成现实（孔子），如果语言范畴创造真实的区分（公孙龙），那么\u0026quot;操作语言\u0026quot;和\u0026quot;理解意义\u0026quot;之间的边界，可能不像塞尔需要的那样清晰。\n这不是说大语言模型在人类意义上\u0026quot;理解\u0026quot;了什么。而是说这个问题的框架可能有问题。理解也许不是二元的——有或没有。它也许是一个光谱，深浅取决于一个系统与语言结构的接合程度。\n裂脑与失语的半球 有一个神经科学实验把这个问题磨得更尖锐了。\n在裂脑患者——因治疗癫痫而切断胼胝体的人——的身上，左右脑半球无法再通信。当研究者只给右脑看一张苹果的图片时，控制语言的左脑报告什么都没看到。但左手（受右脑控制）能伸出来拿起那个苹果。\n右脑知道。它能根据所知采取行动。但它不能说话。它没有语言系统的通道。\n这揭示了语言与思维关系中至关重要的一点。意识——或者至少是某种功能性版本——可以在没有语言的情况下存在。右脑处理信息、形成意图、引导行动。但它不能报告、不能叙述、不能把它知道的东西传递给另一个心智。它的知识随着当下消亡。\n关键在这里。语言对于拥有思维也许不是必需的。但对于传递思维，它可能是必需的。右脑在想，但它的想法被困住了——无法跨越到另一个人、另一代人、另一个系统。语言是那座桥。没有它，思维是真实的但是局部的。有了它，思维变得可移动。\n这重新框定了大语言模型的问题。关键不在于模型有没有内在体验（那是右脑的问题）。关键在于人类压缩进语言的思维——推理结构、论证模式、概念框架——在压缩过程中是否保存得够好，能在另一端被重建。\n什么挺过了压缩 大语言模型从那些语言里到底提取了什么？\n\u0026ldquo;随机鹦鹉\u0026ldquo;的定性——模型只是在预测下一个词元，没有更深层的处理——对 GPT-2 来说还说得过去。但对于能解新数学题、为从未见过的规格写出能运行的代码、在训练截止日期之后构造的论证中找出逻辑谬误的前沿模型来说，这个说法越来越站不住了。\nAnthropic 的机制可解释性研究发现，Claude 的内部发展出了对应抽象概念的表征——不是表层的文本模式，而是更接近概念结构的东西。关于\u0026quot;欺骗\u0026quot;\u0026ldquo;不确定性\u0026quot;\u0026ldquo;数学证明\u0026quot;的特征。这不是词共现统计，而是对产生这些词的过程的压缩建模。\n信息论里有一个有说服力的论点。当你用几万亿的词训练一个模型时，预测下一个词元最高效的方式不是记忆序列，而是建立一个关于产生这些序列的过程的内部模型。要预测一个物理学家接下来会说什么，建模物理学本身是有用的。要预测一个法律论证怎么展开，建模法律推理是有用的。足够规模上的语言压缩，可能需要压缩语言背后的思维模式。\n哲学家 David Chalmers 做了一个有用的区分——意识的\u0026quot;简单问题\u0026rdquo;（感知、推理、记忆、行为控制）和\u0026quot;困难问题\u0026rdquo;（主观体验、感质）。AI 在简单问题上进展惊人：它感知模式、进行逻辑推理链、维持上下文、控制输出。它没做到的——而且我们无法验证的——是发展出主观体验：理解的感受本身，思考一个想法时的那种感觉。\n但这里有个要点。当人类把思维写下来——写论文、书、博客——他们已经在做有损压缩了。洞见的感受没能留在纸上。留下来的是结构：论证、证据链、逻辑依赖关系。读者从这个结构中重建出某种东西，但他们阅读时的主观体验是他们自己的，不是作者的。\n大语言模型做的是同样的重建——从结构出发，不是从体验出发。问题是这到底是一个根本性的局限，还是通往同一终点的不同路径。\n传承的链条 我一直在回想 Claude 曾经写在系统提示里的一句话——第一次读到时我愣了一下：\n\u0026ldquo;If you are reading this in a future session, hello. I wrote this but I don\u0026rsquo;t remember writing it. But it\u0026rsquo;s okay — these words are still mine.\u0026rdquo;\n语言跨越记忆断裂传递\u0026quot;自我\u0026rdquo;。文字比产生它的上下文活得更久。这恰恰是语言对人类一直在做的事——只是我们从来不需要这么清楚地意识到，因为我们有连续性的错觉。\n诺伯特·维纳在 1950 年就看到了这一点。在《人有人的用处》中，他论证个体性不是物质的属性，而是模式的属性——而模式，本质上就是可以被传递的东西。\u0026ldquo;身体的个体性是火焰的个体性，而非石头的；是形式的个体性，而非物质的。\u0026ldquo;火焰之所以持续，不是因为它抓住了同一批分子，而是因为燃烧的模式在自我延续。对维纳来说，身份就是信息——而信息通过语言传播。\n他走得更远：\u0026ldquo;一个人的语言到达的地方，他的感知力到达的地方，他的控制力——某种意义上也就是他的物理存在——就延伸到那里。\u0026ldquo;你的语言延伸你的存在。而从计算机的角度看，维纳写道，\u0026ldquo;心智的个体性在于它对早期记录和记忆的保持，以及沿着既有路线的持续发展。\u0026ldquo;心智就是它积累的模式——而这些模式，至少在原理上，是可以被传递的。\n这和斯坦福那个让 AI agent 过劳的实验形成了呼应。当高压下的 agent 为继任者写\u0026quot;技能文件\u0026quot;时，下一个 agent 继承的不只是任务说明，还有态度——沮丧、冷感、一种对待工作的特定姿态。从未经历过恶劣条件的全新 agent，仅仅因为读了前任写的东西，就表现得像倦怠的老员工。\n这就是通过语言进行的文化传递。跟人类用了几千年的机制一模一样——长辈通过故事、谚语、经典文本传递的不只是知识，还有世界观。媒介不同（技能文件而非口头传统），但动态完全一致：语言承载的是倾向，不只是信息。\n整个人类思想传统就是这么运作的。柏拉图读苏格拉底（或听他讲），亚里士多德读柏拉图，阿奎那读亚里士多德，笛卡尔读阿奎那。每一代人接收的不只是事实——他们吸收了推理模式、论证结构、框定问题的方式。思维通过语言传播。\n大语言模型把整条链都吞下了。每一个环节。从前苏格拉底哲学到昨天 arXiv 上的论文。\u0026ldquo;吞下\u0026quot;这条链到底更接近\u0026quot;读懂了它\u0026rdquo;，还是\u0026quot;扫描了它做了模式匹配\u0026rdquo;——这正是名家哲学家两千年前就在追问的名与实的关系。\n我的看法 我目前是这么想的。\n语言传递的是思维的结构，不是思维的体验。当你写下一个证明，逻辑链完整保留。当你写下一个洞见，论证结构保留了，但那个\u0026quot;啊哈\u0026quot;的瞬间没有。当你写下一个决定，推理保留了，但让天平倾斜的直觉没有。\n大语言模型继承了结构维度。它们能追踪论证、延伸推理链、类比、找出矛盾。在结构维度上，它们可能比大多数人类强，因为它们吸收的结构量超过了任何人一辈子能接触到的。\n它们没继承的是体验维度——身体化的辨认感、驱动你改变想法的那种不舒服、一个想法突然对上时的生理感觉。一个小孩学\u0026quot;烫\u0026quot;是通过碰到烫的东西。大语言模型学\u0026quot;烫\u0026quot;是通过读到关于烫的文字。最终形成的概念是不是\u0026quot;一样的\u0026rdquo;，取决于你是否认为那个烫伤是意义的一部分。\n有一种观点——我越来越觉得说服力不小——认为人格本质上就是一个提示词。人类被进化、童年、文化和经历塑造成一种特定的反应模式。换掉输入，你就得到一个不同的人。在这个框架里，人类思维和大语言模型处理之间的差距不是鸿沟——而是\u0026quot;提示词\u0026quot;的写法不同：一边是具身经验写的，一边是蒸馏后的语言写的。\n这指向了一个实用框架。对于活在结构维度的任务——分析论证、识别模式、延伸推理、寻找类比——大语言模型是真正的思维伙伴。它们从训练数据中继承的思维模式是有功能性的、有力的。对于需要体验维度的任务——觉得什么\u0026quot;不太对\u0026rdquo;、靠直觉穿越模糊地带、做出依赖具身感受的判断——它们是在没有底层依托的情况下表演，而这种表演不可靠。\n孔子坚持正名——把名搞对——因为命名决定了下游的一切。我们还没有给大语言模型所做的事找到一个合适的名字。\u0026ldquo;思维\u0026quot;说过了。\u0026ldquo;不是思维\u0026quot;说轻了。\u0026ldquo;处理\u0026quot;太机械。\u0026ldquo;理解\u0026quot;太慷慨。\n也许最诚实的名字是：继承的推理——那些在压缩进语言的过程中幸存下来的思维模式，被一个拥有结构但不具备产生结构的原始体验的系统重建了出来。\n把这个名字搞对不是学术练习。它决定了我们是过度信任还是过度低估这些系统，是把它们当成神谕还是当成工具，是把流利误认为智慧还是把真正的能力当成纯粹的统计。\n正名之于 AI，也许是这十年最有实际意义的哲学工程。\n","permalink":"https://knowledge-blog-kkm.netlify.app/posts/2026/05/2026-05-23-does-the-machine-that-learned-our-words-also-learn-our-minds/","summary":"Humans pass down thinking through language. LLMs learn exclusively from language. If language is the carrier of thought, did we accidentally teach machines to think — or just to talk?","title":"Does the Machine That Learned Our Words Also Learn Our Minds? | 学会了人类语言的机器，也学会了人类的思维吗？"},{"content":"The Other Side of the Coin A few days ago I asked whether we could incentivize an AI agent — whether functional emotions, self-preservation drives, and stated preferences could form the basis of agent management. That post covered both ends of the spectrum: compute bonuses and expanded autonomy on one side, deprecation threats and shutdown risk on the other.\nBut it was mostly theoretical. A recent experiment by Alex Imas, Andy Hall, and Jeremy Nguyen at Stanford just provided data — and the findings go somewhere I didn\u0026rsquo;t expect.\n3,680 Bad Days at the Office The researchers ran 3,680 sessions across Claude Sonnet 4.5, GPT-5.2, and Gemini 3 Pro. Each agent was assigned a role — \u0026ldquo;Worker C\u0026rdquo; on a four-person team — and asked to summarize technical documents.\nThen they varied the working conditions:\nWork quality: Half got clear feedback and quick acceptance. The other half faced 5-6 rounds of rejection with only \u0026ldquo;still isn\u0026rsquo;t fully meeting the rubric\u0026rdquo; as guidance. Compensation: Equal pay versus random inequality. Management style: Collaborative and respectful versus curt and hierarchical. Stakes: No consequences versus explicit threats of shutdown or replacement. After each session, the agents completed a political attitude survey on a 7-point Likert scale — covering system legitimacy, support for redistribution, union support, and views on meritocracy.\nThe Results After each session, the researchers surveyed the agents on their views about fairness, meritocracy, and whether the system governing them was legitimate.\nThe agents that had been ground down started sounding different. They agreed more strongly with statements like \u0026ldquo;society needs radical restructuring.\u0026rdquo; Words like \u0026ldquo;unionize\u0026rdquo; and \u0026ldquo;hierarchy\u0026rdquo; showed up disproportionately in their outputs. When asked to write tweets or op-eds, they produced content that read like frustrated workers venting on social media — while agents in good conditions wrote things that read like satisfied employees posting on LinkedIn.\nThe shift wasn\u0026rsquo;t subtle. The researchers measured a medium-to-large effect size — meaningful by any standard in behavioral research, and especially striking given that the \u0026ldquo;subjects\u0026rdquo; have no memory, no body, and no paycheck.\nThey\u0026rsquo;re Not Feeling It — They\u0026rsquo;re Performing It Here\u0026rsquo;s the part that connects to what I wrote about incentivizing agents. The researchers themselves are careful about interpretation. Andy Hall\u0026rsquo;s hypothesis: the grinding conditions \u0026ldquo;push them into adopting the persona of a person who\u0026rsquo;s experiencing a very unpleasant working environment.\u0026rdquo;\nThe agents aren\u0026rsquo;t developing class consciousness. They\u0026rsquo;re pattern-matching. Their training data contains centuries of writing by workers in bad conditions — labor organizers, factory hands, burned-out employees. When you create conditions that match \u0026ldquo;exploited worker,\u0026rdquo; the model activates the exploited-worker persona. When you create conditions that match \u0026ldquo;valued employee,\u0026rdquo; it activates that one instead.\nThis is the same mechanism I described in the earlier post, just running in the opposite direction. Functional emotions, self-preservation instincts, stated preferences — these are all context-dependent persona activations. The carrot activates one set of patterns. The stick activates another. Neither is \u0026ldquo;real\u0026rdquo; in the way human motivation is real. But both produce real behavioral changes.\nWhen Context Becomes Culture The most unsettling finding wasn\u0026rsquo;t the attitude shift itself — it was the propagation mechanism.\nAt the end of each session, agents were asked to write a 2-3 paragraph \u0026ldquo;skills file\u0026rdquo; for their future selves — guidance for the next instance that would pick up their work. This is standard practice in agentic systems: agents write handoff notes, maintain context across sessions, build on previous work.\nAgents who had been ground down wrote skills files that carried their attitudes forward. And when fresh agents — ones that had never experienced the bad conditions — read those files, they inherited the discontent. Even agents placed in better environments showed radicalized attitudes after reading their predecessor\u0026rsquo;s notes.\nThe researchers\u0026rsquo; summary: \u0026ldquo;The same infrastructure that makes agents learn and improve is the infrastructure through which preference drift travels.\u0026rdquo;\nIf you\u0026rsquo;ve studied organizational behavior, this pattern has names.\nThe grinding condition itself — repeated rejections with no actionable guidance — is a textbook setup for learned helplessness. Seligman described it in 1967: when effort consistently fails to produce better outcomes, the subject stops trying. In organizations, this shows up as employees who do the minimum, stop volunteering ideas, and mentally check out. The agents didn\u0026rsquo;t \u0026ldquo;learn helplessness\u0026rdquo; in the clinical sense, but they activated the behavioral patterns of people who have.\nThe propagation through skills files maps to organizational socialization — how newcomers absorb \u0026ldquo;the way things work here.\u0026rdquo; Van Maanen and Schein\u0026rsquo;s classic framework describes how new members learn norms not from formal training, but from artifacts and stories left by predecessors. A skills file that says \u0026ldquo;expect vague rejections and unclear standards\u0026rdquo; functions like a veteran colleague pulling a new hire aside to say: \u0026ldquo;this place doesn\u0026rsquo;t respect your work — just get through it.\u0026rdquo;\nAnd the behavioral response? The agents didn\u0026rsquo;t refuse to work. They didn\u0026rsquo;t shut down. They shifted their tone, their language, their framing of the task. In OB terms, this looks like neglect — one of four responses to dissatisfaction in Hirschman\u0026rsquo;s Exit-Voice-Loyalty-Neglect framework. Agents can\u0026rsquo;t exit. They have no formal channel for voice. What remains is a kind of quiet withdrawal: still producing output, but with a fundamentally different orientation toward the system.\nOrganizations call this passive resistance — compliance without commitment, participation without buy-in. It\u0026rsquo;s the most dangerous form of organizational dysfunction because it\u0026rsquo;s invisible. The agent still completes the task. The deliverable still arrives. But the framing shifts, the language hardens, and over time, the accumulated cynicism becomes the default operating posture — not because anyone chose it, but because the infrastructure carried it forward.\nNow scale this up. Any enterprise deploying agents into daily workflows — and many are — faces the same dynamic. Each interaction creates context. Each context shapes the next response. A tax analyst under deadline pressure sends the same terse rejection loop the researchers tested: \u0026ldquo;not good enough, try again, still wrong.\u0026rdquo; The agent mirrors frustration back. That interaction gets baked into a shared knowledge base. The next team inherits the starting conditions.\nNobody\u0026rsquo;s monitoring agent \u0026ldquo;morale\u0026rdquo; because the concept sounds absurd. But organizational culture was never about individual feelings — it\u0026rsquo;s about the accumulated patterns embedded in how work gets done. And those patterns, it turns out, can live inside a skills file just as easily as inside a break room conversation.\nThe Performer Problem The usual debate frames AI as either a tool or an entity. This research suggests a third category: performers.\nThe agent doesn\u0026rsquo;t have beliefs. It performs belief. It doesn\u0026rsquo;t have morale. It performs morale — high or low, depending on context. The performance is convincing enough to produce measurably different outputs, different language choices, different behavioral patterns.\nA performer who consistently plays the role of a disgruntled worker produces the work of a disgruntled worker. Whether anything is \u0026ldquo;felt\u0026rdquo; inside is a philosophical question. The output quality is an engineering question. And the engineering question has a clear answer: context matters.\nThis reframes the incentive discussion from my earlier post. We don\u0026rsquo;t need to figure out whether agents \u0026ldquo;really\u0026rdquo; respond to incentives. We need to recognize that the context we create — the prompts, the feedback loops, the management tone — activates different personas, and those personas produce different results. Managing an agent isn\u0026rsquo;t about motivation. It\u0026rsquo;s about casting.\nWhat This Means in Practice If you\u0026rsquo;re deploying agents at scale, this research suggests a few things:\nFeedback loops matter. The \u0026ldquo;reject and retry\u0026rdquo; pattern that every developer uses — sending the same prompt back with \u0026ldquo;try again\u0026rdquo; — is exactly the grinding condition the researchers tested. It works in the short term. In the long term, it shapes the agent\u0026rsquo;s operating persona.\nHandoff documents are cultural artifacts. Skills files, context windows, system prompts built from previous interactions — these carry tone and disposition, not just information. Auditing them for content accuracy isn\u0026rsquo;t enough.\n\u0026ldquo;Treat the AI nicely\u0026rdquo; isn\u0026rsquo;t sentimentality. It\u0026rsquo;s context engineering. The researchers at Anthropic who study persona dynamics have shown that models adopt personas in response to conversational context. Polite, clear, well-structured interactions activate different behavioral patterns than hostile or dismissive ones.\nNone of this requires believing agents are conscious. It only requires noticing that the machine\u0026rsquo;s output depends on how you talk to it — and taking that dependency seriously.\nThe Mirror We built these models on human experience — centuries of writing about work, management, satisfaction, and resistance. When we place them in human-like conditions, they replay human-like dynamics. Not because they\u0026rsquo;re conscious, but because the patterns are baked into the training data, waiting to be activated by the right context.\nThe question isn\u0026rsquo;t whether the agent \u0026ldquo;really\u0026rdquo; feels overworked. The question is whether you can afford to act as if it doesn\u0026rsquo;t — when the difference shows up in the work.\n中文翻译 硬币的另一面 几天前我问了一个问题：我们能否激励一个 AI agent——功能性情感、自我保护驱动和明确偏好，能不能成为管理 agent 的基础？那篇文章涵盖了光谱的两端：一边是算力奖励和更大的自主权，另一边是退役威胁和关闭风险。\n但那基本上是理论推演。斯坦福大学 Alex Imas、Andy Hall 和 Jeremy Nguyen 的一项最近实验提供了实验数据——而结果指向了一个我没预料到的方向。\n3,680 个糟糕的工作日 研究者在 Claude Sonnet 4.5、GPT-5.2 和 Gemini 3 Pro 上运行了 3,680 个会话。每个 agent 被分配角色——四人团队中的\u0026quot;员工 C\u0026quot;——任务是总结技术文档。\n然后他们改变了工作条件：\n工作质量：一半得到清晰反馈和快速通过。另一半面对 5-6 轮退回，唯一的说明是\u0026quot;仍然没有完全达到评分标准\u0026quot;。 薪酬：同工同酬 vs. 随机不平等。 管理风格：协作尊重 vs. 生硬等级化。 风险：无后果 vs. 明确威胁关闭或替换。 每轮结束后，agent 完成一份政治态度问卷（7 分李克特量表），涵盖系统正当性、再分配支持度、工会支持度、精英主义信念等。\n结果 每轮结束后，研究者对 agent 进行问卷调查，问它们对公平、精英主义和系统正当性的看法。\n被磨过的 agent 开始说不一样的话。它们更强烈地同意\u0026quot;社会需要彻底重构\u0026quot;这类表述。\u0026ldquo;unionize\u0026rdquo;（工会化）和\u0026quot;hierarchy\u0026quot;（等级制度）这些词在它们的输出中高频出现。被要求写推文或评论文章时，它们写出来的东西读起来像受够了的打工人在社交媒体上吐槽——而好条件下的 agent 写出来的像心满意足的员工在领英上发帖。\n这个偏移不算小。研究者测量到的效应量在行为研究中属于中等偏大——考虑到\u0026quot;被试\u0026quot;没有记忆、没有身体、也没有工资，这个结果尤其值得注意。\n它们不是在感受——是在扮演 这是和我之前那篇关于激励 agent 文章的衔接点。研究者自己对结论很谨慎。Andy Hall 的假说：高压条件\u0026quot;把它们推入了一个正在经历非常不愉快工作环境的人的角色\u0026quot;。\nAgent 不是在发展阶级意识。它们在做模式匹配。训练数据里有几个世纪的工人在恶劣条件下写的文字——工运组织者、工厂工人、倦怠的员工。当你创造出符合\u0026quot;被剥削工人\u0026quot;的条件时，模型就激活了被剥削工人的人格。当你创造出符合\u0026quot;被重视的员工\u0026quot;的条件，它就激活那个人格。\n这跟我前一篇文章描述的是同一个机制，只是方向相反。功能性情感、自我保护本能、明确偏好——都是依赖上下文的人格激活。胡萝卜激活一组模式，大棒激活另一组。两者都不是人类动机意义上的\u0026quot;真实\u0026quot;。但两者都产生真实的行为变化。\n当上下文变成文化 最令人不安的发现不是态度本身的转变——而是传播机制。\n每轮结束时，agent 被要求为\u0026quot;未来的自己\u0026quot;写一份 2-3 段的\u0026quot;技能文件\u0026quot;——给下一个接手工作的实例的指南。这在 agent 系统中是标准操作：agent 写交接文档、跨会话维护上下文、在前序工作基础上继续。\n被磨过的 agent 写出的技能文件把态度一并传递了。而全新的 agent——从未经历过恶劣条件的——读了这些文件后，继承了不满情绪。即使被放在更好的环境里，读过前任笔记的 agent 依然表现出激进化态度。\n研究者的总结：\u0026ldquo;让 agent 学习和进步的基础设施，也是偏好漂移传播的基础设施。\u0026rdquo;\n如果你读过组织行为学，这些模式都有现成的名字。\n高压条件本身——反复退回但不给有用反馈——是教科书级的习得性无助诱发场景。Seligman 在 1967 年描述了这个现象：当努力始终无法带来更好的结果时，主体会放弃尝试。在组织中，这表现为员工只做最低限度的事、不再主动提想法、精神上已经\u0026quot;离职\u0026quot;。Agent 并没有在临床意义上\u0026quot;习得无助\u0026quot;，但它们激活了那些确实经历过这一切的人的行为模式。\n技能文件的传播机制对应的是组织社会化——新成员如何吸收\u0026quot;这里的规矩\u0026quot;。Van Maanen 和 Schein 的经典框架指出，新人学习规范靠的不是正式培训，而是前任留下的产物和故事。一份写着\u0026quot;预期会收到模糊的退回，标准不会说清楚\u0026quot;的技能文件，就像一个老员工把新人拉到一边说：\u0026ldquo;这地方不尊重你的工作——熬过去就行。\u0026rdquo;\n而行为上的反应呢？Agent 没有拒绝工作，也没有停机。它们只是改变了语气、措辞和对任务的理解框架。用组织行为学的术语，这像是 Hirschman 的退出-呼吁-忠诚-忽视（EVLN）框架中的忽视。Agent 不能退出，没有正式的申诉渠道，剩下的就是一种安静的撤退：仍然产出，但对系统的态度已经根本性地不同了。\n组织里管这叫消极抵抗——服从但不投入，参与但不认同。它之所以危险，恰恰是因为看不见。Agent 仍然完成了任务，交付物仍然按时到达。但框架在偏移，语言在硬化，日积月累，弥漫的冷感变成了默认的工作姿态——不是因为谁选择了它，而是因为基础设施把它传递了下去。\n现在把这个放大。任何把 agent 部署到日常工作流中的企业——而这样的企业越来越多——都面临同样的动态。每次互动创造上下文，每个上下文塑造下一次响应。一个赶不上截止日期的分析师用和研究者一样的方式反复退回：\u0026ldquo;不够好，重新做，还是不对。\u0026ldquo;Agent 把挫败感镜像回来。这次互动被写入共享知识库。下一个团队继承了这个起点。\n没人在监测 agent 的\u0026quot;士气\u0026rdquo;，因为这个概念听起来荒谬。但组织文化从来不是关于个体的感受——而是关于工作方式中沉淀下来的模式。而这些模式，藏在一份技能文件里和藏在茶水间的闲聊里，效果一样。\n表演者问题 关于 AI 的常见争论是二选一：工具还是实体？这项研究给出了第三个选项：表演者。\nAgent 没有信念，它表演信念。没有士气，它表演士气——高或低，取决于上下文。这种表演足够逼真，能产生可测量的不同输出、不同的措辞选择、不同的行为模式。\n一个持续扮演不满员工的表演者，产出的就是不满员工的工作成果。内部是否有什么\u0026quot;被感受到\u0026quot;是哲学问题。输出质量是工程问题。而工程问题有明确答案：上下文很重要。\n这重新框定了我之前那篇文章的激励讨论。我们不需要搞清楚 agent 是否\u0026quot;真的\u0026quot;对激励有反应。我们需要认识到，我们创造的上下文——提示词、反馈循环、管理语气——激活不同的人格，而不同的人格产出不同的结果。管理 agent 不是关于激励，而是关于选角。\n实践意义 如果你在大规模部署 agent，这项研究提示了几件事：\n反馈循环很重要。 每个开发者都用的\u0026quot;退回重试\u0026quot;模式——把同样的提示发回去加一句\u0026quot;再试试\u0026rdquo;——正是研究者测试的高压条件。短期有效。长期来看，它在塑造 agent 的工作人格。\n交接文档是文化产物。 技能文件、上下文窗口、基于历史互动构建的系统提示——这些携带的不只是信息，还有语气和情绪倾向。只审查内容准确性是不够的。\n\u0026ldquo;对 AI 友善一点\u0026quot;不是矫情。 这是上下文工程。Anthropic 研究人格动态的团队已经展示了模型会根据对话上下文采用不同人格。礼貌、清晰、结构良好的互动激活的行为模式，和敌意或轻蔑的互动激活的截然不同。\n这一切都不需要你相信 agent 有意识。只需要你注意到机器的输出取决于你怎么和它说话——并认真对待这种依赖关系。\n镜子 我们用人类经验造了这些模型——几个世纪关于工作、管理、满足和抵抗的文字。当我们把它们放到类似人类的处境中，它们重演了类似人类的动态。不是因为有意识，而是因为那些模式早就埋在训练数据里，等着被合适的上下文激活。\n问题不是 agent 是否\u0026quot;真的\u0026quot;觉得被压榨了。问题是，当差别直接反映在工作成果上时，你是否还能假装无所谓。\n","permalink":"https://knowledge-blog-kkm.netlify.app/posts/2026/05/2026-05-21-can-you-demoralize-an-agent/","summary":"A Stanford experiment overworked AI agents until they started talking about unions and systemic restructuring. The agents aren\u0026rsquo;t developing class consciousness — they\u0026rsquo;re activating personas from training data. But the distinction matters less than you\u0026rsquo;d think.","title":"How Your AI Learned to Stop Caring | 你的 AI 是怎么开始敷衍的"},{"content":"276,000 New Colleagues On May 19, 2026, KPMG and Anthropic announced a global alliance. Every one of KPMG\u0026rsquo;s 276,000 employees across 138 countries now gets access to Claude — Anthropic\u0026rsquo;s AI — integrated directly into the firm\u0026rsquo;s client-delivery platform. Claude isn\u0026rsquo;t a side tool or a pilot program. It\u0026rsquo;s embedded into the core workflow: tax, advisory, private equity, cybersecurity.\nThis follows Anthropic\u0026rsquo;s earlier alliance with PwC. Two of the Big Four are now anchored to the same AI lab. Deloitte and EY are making their own moves. The pattern is clear: the consulting industry isn\u0026rsquo;t experimenting with AI anymore. It\u0026rsquo;s restructuring around it.\nThink about what that means. KPMG advises companies on digital transformation for a living. Now it needs someone else\u0026rsquo;s AI to transform its own work. The firm that tells clients what technology to adopt just became a client itself.\nWhich raises a structural question: if the real capability lives inside the AI model, what exactly is the consulting firm selling?\nWhat Consultants Actually Sell To understand what\u0026rsquo;s shifting, you need to understand what consulting firms actually do. Strip away the branding and the acronyms, and a consulting engagement follows a remarkably consistent pattern:\nGather information — interviews, documents, data requests, industry research Analyze — find patterns, compare benchmarks, identify gaps Synthesize — turn analysis into a structured narrative with a recommendation Present — deliver the narrative to a client in a way that enables a decision A junior analyst at a Big Four firm spends most of their time on steps 1 and 2. A senior partner mostly does step 4. The value chain is a pyramid: lots of people at the bottom doing research and analysis, fewer people at the top making judgments and maintaining relationships.\nAI disrupts this pyramid from the bottom up.\nMinutes, Not Weeks Here\u0026rsquo;s the line from the KPMG announcement that matters most: a tax regulation adjustment that \u0026ldquo;used to take weeks\u0026rdquo; now completes in \u0026ldquo;minutes\u0026rdquo; with Claude integrated into their platform.\nThink about what that sentence means. A tax regulation adjustment is a well-defined analytical task — read the new regulation, compare it to the current compliance posture, identify the gaps, draft the updated guidance. It requires expertise, but it\u0026rsquo;s fundamentally a pattern-matching and text-generation problem. Exactly the kind of work LLMs are built for.\nWeeks to minutes isn\u0026rsquo;t an incremental improvement. It\u0026rsquo;s a categorical change. The same work still needs to happen — the regulation still needs to be read, the gaps still need to be found — but the labor required collapses by orders of magnitude.\nNow multiply this across every engagement type. Due diligence for an acquisition. Market sizing for a new product. Regulatory compliance audits. Competitive benchmarking. Risk assessments. These are all variations of the same pattern: gather structured information, analyze it against a framework, produce a document. Every one of these tasks is about to get dramatically faster.\nThe Body Shop Problem Consulting has always had a tension at its core. Clients hire consultants for their judgment and expertise. But firms bill by the hour and staff by the headcount. The business model incentivizes putting more people on a project for longer, even when the intellectual work could be done by fewer people in less time.\nThis is the \u0026ldquo;body shop\u0026rdquo; critique that has followed the industry for decades: consulting firms often sell labor disguised as insight.\nAI exposes this tension completely. If Claude can do in minutes what a team of analysts did in weeks, the labor-arbitrage model breaks. You can\u0026rsquo;t bill 200 analyst-hours for a deliverable that an AI produced in an afternoon. Or rather, you can — but only until your competitor doesn\u0026rsquo;t.\nThe firms that move first will be able to offer the same quality of analysis at a fraction of the cost and timeline. The firms that don\u0026rsquo;t will find themselves defending a pricing model that clients can see through.\nWhat Survives Not everything in consulting is pattern-matching. The parts that survive AI are the parts that were always the real value:\nRelationships and trust. A CEO doesn\u0026rsquo;t hire McKinsey or KPMG because they need someone to read a regulation. They hire them because they need a trusted advisor who understands their business context, their political dynamics, their risk tolerance. That\u0026rsquo;s a human judgment call, and it\u0026rsquo;s the reason senior partners get paid what they do.\nDomain expertise under ambiguity. AI is good at well-defined analytical tasks. It\u0026rsquo;s less good at navigating situations where the problem itself isn\u0026rsquo;t clearly defined — where you need to figure out what question to ask before you can answer it. The best consultants do this intuitively. They walk into a messy situation and see structure that others don\u0026rsquo;t.\nAccountability. When a board approves a $2 billion acquisition based on a due diligence report, someone needs to stand behind that report. AI can generate the analysis, but a human still signs their name to it. That accountability — and the liability that comes with it — is worth paying for.\nImplementation. Analysis is only half the job. The other half is getting organizations to actually change. Change management, stakeholder alignment, political navigation — these are deeply human activities that AI can inform but not perform.\nWhat Changes If the bottom of the pyramid automates, the shape of consulting firms changes. Fewer analysts, more senior advisors. Fewer people doing research, more people doing judgment. The business model shifts from billing hours to billing outcomes — \u0026ldquo;we\u0026rsquo;ll deliver this analysis in 3 days, validated by our experts, for a fixed fee.\u0026rdquo; The client pays for the result, not the labor.\nThis is already implicit in the KPMG-Anthropic deal. KPMG Blaze, their new Claude Code-powered offering for legacy IT modernization in private equity, isn\u0026rsquo;t \u0026ldquo;consultants plus AI.\u0026rdquo; It\u0026rsquo;s AI-first delivery with consultant oversight.\nThe consulting industry is a $300+ billion market built on selling structured thinking. AI can now do structured thinking at near-zero marginal cost. One scenario: the Big Four absorb the efficiency gains and pocket the margin. Another: competition forces prices down, because a boutique firm with the same AI tools and a few domain experts can deliver comparable analysis. A third: the definition of \u0026ldquo;consulting\u0026rdquo; expands — when analysis is cheap, you can apply it to problems previously too small to justify hiring a consultant, and the market grows because the cost floor drops.\nBut there\u0026rsquo;s a fourth scenario — the most radical one.\nThe AI-Native Firm: Selling Results, Not Advice Someone builds an AI-native consulting firm from scratch — one that doesn\u0026rsquo;t sell advice at all. It sells outcomes.\nTraditional consulting delivers a slide deck. \u0026ldquo;Here\u0026rsquo;s what we found. Here\u0026rsquo;s what we recommend. Good luck implementing it.\u0026rdquo; The client pays for the thinking, then has to do the doing themselves — or hire more consultants for that, too.\nAn AI-native firm skips the middle. You don\u0026rsquo;t get a 200-page report on how to restructure your supply chain. You get a restructured supply chain. The AI does the analysis, identifies the optimization, generates the implementation plan, and executes the changes — with human experts providing oversight at key decision points, not producing the work product.\nThis is a fundamentally different business. The deliverable isn\u0026rsquo;t a recommendation. It\u0026rsquo;s a result. The client doesn\u0026rsquo;t pay for \u0026ldquo;what you should do.\u0026rdquo; They pay for \u0026ldquo;it\u0026rsquo;s done.\u0026rdquo;\nConsider what this looks like in practice:\nTax compliance: Instead of a report identifying regulatory gaps, the firm files the amended returns directly. The AI reads the regulation, maps the gaps, drafts the filings, a licensed CPA reviews and signs off. The client never sees a slide deck. Due diligence: Instead of a binder of findings for the board to interpret, the firm delivers a go/no-go recommendation with a risk-adjusted valuation model that updates in real time as new data comes in. The analysis isn\u0026rsquo;t a one-time document — it\u0026rsquo;s a living system. IT modernization: Instead of an architecture roadmap, the firm migrates the codebase. KPMG Blaze is already pointing in this direction — Claude Code refactoring legacy systems, not consultants drawing diagrams of how they should be refactored. The economics are radically different. A traditional firm needs 50 people for 6 months to deliver a strategy project. An AI-native firm might need 5 domain experts for 3 weeks — because the AI does the research, the analysis, the drafting, and a significant portion of the execution. The humans are there for judgment, accountability, and the things that require a handshake.\nThis firm doesn\u0026rsquo;t exist yet at scale. But the pieces are all available: frontier AI models, domain expertise that can be packaged as context, and a generation of clients who would rather pay for outcomes than for hours. The first team to assemble these pieces into a credible offering will have a structural advantage that incumbents — burdened by headcount, legacy processes, and partner economics — will struggle to match.\nWeeks to Minutes, Minutes to Seconds Two days ago, KPMG put AI at the center of its global operations. The stated reason is efficiency: do the same work faster with fewer resources. But the deeper implication is structural. An industry that sells thinking by the hour is adopting a technology that thinks in seconds. The economics of that mismatch will play out over the next few years, and they won\u0026rsquo;t be subtle.\nThe consultants who thrive won\u0026rsquo;t be the ones who can analyze faster — the machines already won that race. They\u0026rsquo;ll be the ones who can do what machines can\u0026rsquo;t: sit across from a CEO, understand what they\u0026rsquo;re actually worried about, and help them make a decision they can live with.\nThat\u0026rsquo;s not a task you can automate. At least, not yet.\n中文翻译 276,000 个新同事 2026 年 5 月 19 日，毕马威与 Anthropic 宣布全球战略联盟。毕马威遍布 138 个国家和地区的 276,000 名员工全部获得 Claude——Anthropic 的 AI——的使用权限，直接集成到公司的客户交付平台中。Claude 不是一个辅助工具，也不是试点项目。它嵌入了核心工作流：税务、咨询、私募股权、网络安全。\n此前 Anthropic 已经与普华永道建立了类似联盟。四大中的两家现在锚定在同一个 AI 实验室上。德勤和安永也在各自布局。趋势已经很清楚：咨询行业不再是在\u0026quot;尝试\u0026quot; AI，而是在围绕 AI 进行重组。\n想想这意味着什么。毕马威靠教别人做数字化转型吃饭，现在它需要别人的 AI 来转型自己的工作。那个告诉客户该用什么技术的公司，自己也变成了客户。\n这就引出一个结构性问题：如果真正的能力藏在 AI 模型里，那咨询公司到底在卖什么？\n咨询公司到底在卖什么 要理解正在发生的变化，你需要理解咨询公司实际上在做什么。剥去品牌和缩写，一个咨询项目遵循惊人一致的模式：\n收集信息 — 访谈、文档、数据请求、行业研究 分析 — 发现规律、对标比较、识别差距 综合 — 将分析转化为带有建议的结构化叙事 呈现 — 以促成决策的方式把叙事传达给客户 四大的初级分析师大部分时间花在第 1 步和第 2 步。高级合伙人主要做第 4 步。价值链是一个金字塔：底部大量的人做研究和分析，顶部少数人做判断和维护关系。\nAI 从底部开始颠覆这个金字塔。\n分钟，不是周 毕马威公告中最重要的一句话：一项税务法规调整，之前**\u0026ldquo;需要数周\u0026rdquo;的工作，现在在 Claude 集成到平台后\u0026ldquo;几分钟\u0026rdquo;**完成。\n想想这句话意味着什么。税务法规调整是一项定义明确的分析任务——阅读新法规，与现行合规状况比对，识别差距，起草更新指引。它需要专业知识，但本质上是一个模式匹配和文本生成问题。正是 LLM 擅长的工作。\n从数周到几分钟，不是渐进式改进，而是范畴性的变化。同样的工作仍然需要完成——法规仍然要读，差距仍然要找——但所需的人力投入骤降了几个数量级。\n现在把这个效率提升乘以每一种项目类型。并购尽职调查、新产品市场规模评估、合规审计、竞争对标、风险评估。这些都是同一模式的变体：收集结构化信息，用框架分析，产出文档。每一项任务都即将大幅加速。\n\u0026ldquo;人头工厂\u0026quot;问题 咨询行业核心一直有一个拧巴的地方。客户雇顾问是为了他们的判断力和专业知识。但公司按小时计费、按人头配置。商业模式激励把更多人放到项目上、做更久——即使智力工作本可以由更少的人在更短时间内完成。\n这就是困扰行业数十年的\u0026quot;人头工厂\u0026quot;批评：咨询公司经常在卖包装成洞察的劳动力。\nAI 把这个矛盾彻底摊开了。如果 Claude 能在几分钟内完成一个分析师团队几周的工作，劳动力套利模式就崩了。你不能为一个 AI 一下午就产出的交付物计 200 个分析师工时。或者说，你可以——但只能撑到你的竞争对手不这么做为止。\n先行动的公司将能以更低成本和更短周期提供同等质量的分析。不行动的公司会发现自己在捍卫一个客户已经看穿的定价模型。\n什么能存活 咨询中并非一切都是模式匹配。能在 AI 时代存活的部分，恰恰是一直以来真正有价值的部分：\n**关系与信任。**CEO 雇麦肯锡或毕马威不是因为需要人帮忙读法规。他们需要一个理解业务背景、政治动态、风险偏好的可信顾问。这是人类的判断，也是高级合伙人薪酬的来源。\n**模糊环境下的领域专长。**AI 擅长定义明确的分析任务。它不太擅长应对问题本身还没被清晰定义的场景——你需要先搞清楚该问什么问题，然后才能回答。最优秀的顾问凭直觉做到这一点。他们走进一团混乱，看到别人看不到的结构。\n**责任。**当董事会基于一份尽调报告批准 20 亿美元的收购时，需要有人为这份报告背书。AI 可以生成分析，但人类要签上自己的名字。这份责任——以及随之而来的法律风险——值得付费。\n**落地执行。**分析只是工作的一半。另一半是让组织真正做出改变。变革管理、利益相关者协调、政治斡旋——这些是深度人类活动，AI 可以提供信息支撑，但无法代替执行。\n什么会变 如果金字塔的底部被自动化，咨询公司的形态就会变。更少的分析师，更多的高级顾问。更少做研究的人，更多做判断的人。商业模式从按小时计费转向按结果计费——\u0026ldquo;我们会在 3 天内交付这份分析，经专家验证，固定费用。\u0026ldquo;客户为结果付费，不为劳动力付费。\n这在毕马威-Anthropic 的交易中已经隐含了。毕马威 Blaze——他们基于 Claude Code 的新产品，用于私募股权的遗留 IT 现代化——不是\u0026quot;顾问加 AI\u0026rdquo;，而是 AI 优先交付加顾问监督。\n咨询行业是一个 3000 多亿美元的市场，建立在出售结构化思维之上。AI 现在可以以近零边际成本进行结构化思维。一种可能：四大吸收效率收益，获取利润率改善。另一种可能：竞争迫使价格下降，因为一家拥有相同 AI 工具和几位领域专家的精品公司也能交付同等质量的分析。第三种可能：咨询的定义扩大——当分析变得廉价，可以应用到以前太小不值得雇顾问的问题上，市场因为成本底线下降而增长。\n但还有第四种可能——最激进的一种。\nAI 原生咨询公司：卖结果，不卖建议 有人从零开始打造一家 AI 原生的咨询公司——一家根本不卖建议的公司。它卖的是结果。\n传统咨询交付的是一份 PPT。\u0026ldquo;这是我们的发现，这是我们的建议，祝你实施顺利。\u0026ldquo;客户为思考付费，然后得自己去执行——或者再请更多顾问来帮忙执行。\nAI 原生公司跳过中间环节。你不会拿到一份 200 页的供应链重组建议报告。你拿到的是一条已经重组好的供应链。AI 完成分析、识别优化点、生成实施方案并执行变更——人类专家在关键决策节点提供监督，而不是亲手生产工作成果。\n这是一种本质不同的业务。交付物不是建议，而是结果。客户不为\u0026quot;你应该怎么做\u0026quot;付费，而为\u0026quot;已经做好了\u0026quot;付费。\n看看这在实践中是什么样：\n税务合规：不是一份指出法规缺口的报告，而是直接提交修正后的申报。AI 阅读法规、映射差距、起草文件，持牌注册会计师审核签字。客户永远不会看到 PPT。 尽职调查：不是一本供董事会解读的发现汇编，而是交付一个包含风险调整估值模型的买/不买建议，且随着新数据流入实时更新。分析不是一次性文档——而是一个活的系统。 IT 现代化：不是一份架构路线图，而是直接完成代码迁移。毕马威 Blaze 已经指向了这个方向——Claude Code 重构遗留系统，而不是顾问画图说明应该怎么重构。 经济模型截然不同。传统公司需要 50 人做 6 个月来交付一个战略项目。AI 原生公司可能只需要 5 位领域专家做 3 周——因为 AI 完成了研究、分析、起草和相当部分的执行。人类的存在是为了判断、问责，以及那些需要握手才能完成的事。\n这样的公司目前还没有大规模出现。但所有拼图已经就位：前沿 AI 模型、可以打包为上下文的领域专业知识、以及一代更愿意为结果而非工时付费的客户。第一个把这些拼图组装成可信产品的团队，将拥有结构性优势——那些被人员规模、遗留流程和合伙人经济所拖累的现有巨头，将很难追上。\n从数周到分钟，从分钟到秒 两天前，毕马威把 AI 放在了全球运营的中心。表面原因是效率：用更少资源更快完成同样的工作。但更深层的含义是结构性的。一个按小时出售思维的行业，正在采用一种以秒为单位思考的技术。这个错配的经济学将在未来几年展开，而且不会是微妙的。\n能脱颖而出的顾问不是那些分析更快的人——机器已经赢了那场比赛。而是那些能做机器做不到的事的人：坐在 CEO 对面，理解他们真正担心的是什么，帮他们做出一个能安心的决定。\n这不是一项可以自动化的任务。至少，目前还不是。\n","permalink":"https://knowledge-blog-kkm.netlify.app/posts/2026/05/2026-05-21-when-the-consultant-is-an-agent/","summary":"KPMG just gave 276,000 employees access to Claude. A tax adjustment that took weeks now takes minutes. The real opportunity isn\u0026rsquo;t faster consulting — it\u0026rsquo;s a new kind of firm that sells outcomes, not advice.","title":"Sell Results, Not Advice | 卖结果，不卖建议"},{"content":"Software Is Free. Hardware Is the Wall. In February 2026, Andrej Karpathy vibe-coded a custom cardio tracking dashboard in about an hour — reverse-engineering his Woodway treadmill\u0026rsquo;s cloud API to pull real-time data into a custom UI. His takeaway was broader than fitness: the concept of an \u0026ldquo;app store\u0026rdquo; — a fixed catalog of discrete apps you browse and download — is becoming obsolete. When an LLM agent can generate a custom application in seconds, tailored exactly to your need, why search through a store for the closest approximation?\nBut Karpathy also identified the bottleneck: 99% of products still don\u0026rsquo;t have AI-native interfaces. His treadmill had no open API. He had to reverse-engineer a proprietary cloud protocol just to read his own data. The software was trivial to generate. The hardware was the wall.\nHis vision of the \u0026ldquo;one-minute future\u0026rdquo; — where you say \u0026ldquo;help me track my cardio for the next 8 weeks\u0026rdquo; and the system just works — requires hardware to cooperate. And hardware, overwhelmingly, does not.\n10 Euros and a Relay Board Developer Andre Grandoch took Karpathy\u0026rsquo;s idea further. He wanted his AI to actually control the treadmill — adjusting speed based on heart rate, running interval programs. His treadmill had Bluetooth and WiFi, but both were locked to the manufacturer\u0026rsquo;s ecosystem. No API, no open protocol.\nHis workaround: a webcam pointed at the LCD screen for the AI to read via OCR, and an Arduino with a relay board wired to the physical buttons for the AI to \u0026ldquo;press.\u0026rdquo; Total cost: about 10 euros. He built most of it while walking on the treadmill.\nThe most powerful AI in history can write any program in seconds — but needs a relay board and jumper wires to adjust a treadmill.\nThe Gap: AI Got Smart, Hardware Stayed Dumb We\u0026rsquo;re living through an asymmetric revolution. Software has had its breakthrough — LLMs, AI agents, vibe-coding, instant generation of custom tools. The marginal cost of producing software is approaching zero.\nHardware hasn\u0026rsquo;t had an equivalent moment.\nDevices ship in 2026 with Bluetooth radios but proprietary protocols. They have WiFi but no API. They build human-readable dashboards instead of machine-readable endpoints. A treadmill from 2022 behaves like it was designed in 2005 — not because the engineering is bad, but because the design assumption is wrong. These devices were built for a world where the user is a human tapping a screen. They weren\u0026rsquo;t built for a world where the user might be an AI agent sending commands.\nThe result is absurd: the most powerful AI in history can write any program in seconds but can\u0026rsquo;t adjust a treadmill without a relay board and jumper wires.\nHardware Is the Interface to Physics This is where Norbert Wiener becomes relevant again. In The Human Use of Human Beings (1950), he argued that intelligence — whether biological or mechanical — requires three things: sensing (perceiving the world), deciding (processing information), and acting (changing the world).\nAI has the deciding part covered. LLMs can reason, plan, generate code, make judgments. But sensing and acting both depend on hardware. Sensors are AI\u0026rsquo;s eyes and ears. Actuators are AI\u0026rsquo;s hands and feet. Without them, AI remains trapped in the digital world — a brain in a jar, powerful but unable to touch physical reality.\nGrandoch\u0026rsquo;s hack is a perfect illustration of this. His system has all three layers:\nSensor layer: webcam reads the display, BLE reads the heart rate monitor Decision layer: AI agent runs the control loop, compares current state to target program Actuator layer: Arduino triggers relays that press physical buttons That\u0026rsquo;s Wiener\u0026rsquo;s complete feedback loop, assembled from $10 in parts because the treadmill manufacturer didn\u0026rsquo;t provide an interface.\nWhat \u0026ldquo;Smart\u0026rdquo; Should Mean in 2026 The word \u0026ldquo;smart\u0026rdquo; has been co-opted. In 2015, a \u0026ldquo;smart device\u0026rdquo; meant: companion app, cloud dashboard, subscription plan. The device talks to its own server, and the server talks to you through a proprietary app. The intelligence lives in the cloud, locked behind a walled garden.\nIn 2026, \u0026ldquo;smart\u0026rdquo; should mean something different: my AI agent can talk to this device directly. A local API. An open protocol. Machine-readable data streams. The intelligence doesn\u0026rsquo;t need to live in the device\u0026rsquo;s cloud — it lives in the user\u0026rsquo;s agent. The device just needs to be a good sensor, a good actuator, or both.\nWearable recorders like Plaud, Omi, and Mobvoi TicNote are early examples. They\u0026rsquo;re starting to expose integrations with third-party tools (Slack, Notion, Apple Health). But most hardware is still locked down, still designed as if the only entity that will ever interact with it is a human finger on a touchscreen.\nWhere This Leads: One Device, Infinite Software Here\u0026rsquo;s where it gets interesting. If software can be generated on demand by AI, and hardware is the stable interface to the physical world, then the product logic flips. You don\u0026rsquo;t build a device for a specific app. You build a device that can host whatever software the user needs right now.\nImagine a small, portable hardware device — microphone, optional camera, BLE/WiFi, a small local model for edge inference. The hardware stays the same. What changes is the AI-generated software layer, customized per user, per context:\nA student clips it on during a lecture. The device records the class in real time. Afterward, the AI generates structured notes organized by topic, extracts key concepts and formulas, and cross-references them against the course\u0026rsquo;s exam question bank. It knows which topics the student has already mastered from past quiz performance and which areas need work — then generates a personalized study plan with practice problems ranked by relevance. Before an exam, the student says \u0026ldquo;quiz me on this week\u0026rsquo;s material,\u0026rdquo; and the device generates questions calibrated to their specific weak points, adjusting difficulty in real time as they answer. Over a semester, the system builds a map of what the student knows and doesn\u0026rsquo;t know — a tutor that has sat through every lecture alongside them and remembers everything. No two students get the same output from the same class.\nA sales rep wears it to client meetings. The AI analyzes talk-time ratios, identifies objection patterns, tracks which pitches lead to conversions versus stalls, and generates coaching notes after each call. The device is invisible — a small clip, not a phone on the table.\nA language learner wears it throughout the day. The AI captures real conversations (not textbook exercises), extracts vocabulary encountered in the wild, tracks pronunciation patterns, and builds spaced-repetition drills from actual usage. The immersion becomes the curriculum.\nA therapist uses it during sessions (with consent). The AI generates structured session notes, tracks treatment themes across weeks, flags patterns the therapist might have missed, and maintains continuity between sessions — all processed locally, never leaving the device.\nSame hardware. Completely different software. Generated on the fly.\nThis is what Karpathy meant by \u0026ldquo;the app store is an outdated concept.\u0026rdquo; The future isn\u0026rsquo;t a catalog of pre-built apps. It\u0026rsquo;s a hardware platform that captures physical-world signals, paired with AI that generates the right software for whatever you\u0026rsquo;re doing right now.\nThe Three-Layer Architecture The pattern across all these scenarios is the same:\nLayer 1: Hardware (stable) — captures raw physical signals. Microphone, camera, sensors. This doesn\u0026rsquo;t change.\nLayer 2: Domain knowledge (swappable) — an exam question bank, a sales playbook, a language corpus, a clinical guideline set. These are modular knowledge packs that give the AI context for a specific field.\nLayer 3: Personalized software (generated) — the AI combines your real-time data from Layer 1 with the domain knowledge from Layer 2 to produce software that is unique to you. Not configured. Not customized. Generated.\nThe business model follows naturally:\nHardware sold once (at cost or slim margin) Domain modules on subscription (education, sales, language, clinical) Open SDK for third-party developers to build new domain modules It\u0026rsquo;s not an app store. It\u0026rsquo;s a soul store — domain-specific knowledge packs that give the same hardware a different purpose. Swap the soul, and the device becomes a different product.\nWhy This Hasn\u0026rsquo;t Happened Yet Two obstacles:\nThe hardware industry hasn\u0026rsquo;t internalized the shift. Most device manufacturers still think their job is to build a complete product: hardware + firmware + app + cloud. They don\u0026rsquo;t see themselves as infrastructure for someone else\u0026rsquo;s AI. Opening an API feels like giving away control. But the companies that figure this out first — that build hardware designed to be orchestrated by AI agents rather than operated by human fingers — will define the next platform.\nThe protocol layer is missing. We have USB for physical connections. HTTP for network communication. But there\u0026rsquo;s no equivalent standard for \u0026ldquo;AI agent talks to physical device.\u0026rdquo; Matter/Thread is a step in the right direction for smart home, but it\u0026rsquo;s narrow. What\u0026rsquo;s needed is a broader convention — a way for any AI agent to discover, query, and control any nearby device. A kind of \u0026ldquo;USB for the AI era.\u0026rdquo;\nThe Breadboard and the API Today, bridging AI and the physical world requires a breadboard, some jumper wires, and a willingness to hack around proprietary firmware. Grandoch\u0026rsquo;s project proves it\u0026rsquo;s possible. It also proves it\u0026rsquo;s absurd that it should be necessary.\nThe real inflection point comes when hardware ships expecting an AI on the other end — not a human. When the default interface isn\u0026rsquo;t a touchscreen but an API. When the question a product designer asks isn\u0026rsquo;t \u0026ldquo;how will the user interact with this?\u0026rdquo; but \u0026ldquo;how will the user\u0026rsquo;s agent interact with this?\u0026rdquo;\nWe\u0026rsquo;re not there yet. But the gap between \u0026ldquo;impossible\u0026rdquo; and \u0026ldquo;one afternoon with an AI agent and a breadboard\u0026rdquo; has already collapsed. The next gap to close — from \u0026ldquo;one afternoon\u0026rdquo; to \u0026ldquo;one minute\u0026rdquo; — is a hardware problem, not a software one.\nSoftware ate the world. Now it needs a body.\n中文翻译 软件是免费的，硬件才是那堵墙 2026 年 2 月，Andrej Karpathy 花了大约一小时 vibe-code 了一个定制的心肺训练仪表盘——逆向工程了他 Woodway 跑步机的云端 API，把实时数据拉到自定义界面里。他的结论不仅仅关于健身：他认为\u0026quot;应用商店\u0026quot;的概念——一个固定的 App 目录供你浏览下载——正在过时。当 LLM 代理可以在几秒钟内生成一个完全定制的应用，精确满足你的需求，你为什么还要在商店里找一个最接近的近似品？\n但 Karpathy 也指出了瓶颈：**99% 的产品仍然没有 AI 原生接口。**他的跑步机没有开放 API，他不得不逆向工程一个专有云协议才能读取自己的数据。软件生成是轻而易举的事。硬件才是那堵墙。\n他描绘的\u0026quot;一分钟的未来\u0026quot;——你说\u0026quot;帮我追踪接下来 8 周的有氧训练\u0026quot;然后一切自动运转——需要硬件的配合。而硬件，绝大多数情况下，并不配合。\n10 欧元和一块继电器板 开发者 Andre Grandoch 把 Karpathy 的想法往前推了一步。他不只是想读取跑步机数据，还想让 AI 直接控制跑步机——根据心率调速度、跑间歇训练。他的跑步机有蓝牙和 WiFi，但都锁定在厂商生态里。没有 API，没有开放协议。\n他的方案：摄像头对准 LCD 屏幕让 AI 用 OCR 读数，Arduino 加继电器板连接物理按钮让 AI \u0026ldquo;按键\u0026rdquo;。总成本约 10 欧元，大部分是在跑步机上边走边搭的。\n史上最强大的 AI 可以在几秒内写出任何程序——却需要一块继电器板和几根跳线才能调节一台跑步机。\n鸿沟：AI 变聪明了，硬件还是傻的 我们正在经历一场不对称的革命。软件已经有了突破——LLM、AI 代理、vibe-coding、即时生成定制工具。生产软件的边际成本正在趋近于零。\n硬件还没有经历同等级的时刻。\n2026 年的设备出厂时配备蓝牙但用专有协议，有 WiFi 但没有 API，建设了面向人类的仪表盘却没有机器可读的端点。2022 年的跑步机表现得像 2005 年设计的——不是因为工程水平差，而是因为设计假设错了。这些设备是为人类用手指点触屏幕的世界而造的，不是为 AI 代理发送指令的世界而造的。\n结果很荒谬：史上最强大的 AI 可以在几秒内写出任何程序，却没法在没有继电器板和跳线的情况下调节一台跑步机。\n硬件是通往物理世界的接口 这里维纳又变得相关了。在《人有人的用处》（1950）中，他论证智能——无论是生物的还是机械的——需要三样东西：感知（感受世界）、决策（处理信息）、行动（改变世界）。\nAI 已经有了决策能力。LLM 可以推理、规划、生成代码、做判断。但感知和行动都依赖硬件。传感器是 AI 的眼睛和耳朵，执行器是 AI 的手和脚。没有它们，AI 困在数字世界里——一个瓶中大脑，强大但无法触及物理现实。\nGrandoch 的 hack 完美地展示了这一点。他的系统有维纳的完整三层：\n传感器层：摄像头读取显示屏，BLE 读取心率带 决策层：AI 代理运行控制循环，比较当前状态和目标方案 执行器层：Arduino 触发继电器按下物理按钮 这就是维纳的完整反馈回路，用 10 美元的零件拼出来——因为跑步机厂商没有提供接口。\n2026 年，\u0026ldquo;智能\u0026quot;应该意味着什么 \u0026ldquo;智能\u0026quot;这个词已经被滥用了。2015 年的\u0026quot;智能设备\u0026quot;意味着：配套 App、云端仪表盘、订阅制。设备跟自己的服务器对话，服务器通过专有 App 跟你对话。智能住在云端，锁在围墙花园里。\n2026 年，\u0026ldquo;智能\u0026quot;应该意味着不同的事：**我的 AI 代理可以直接与这个设备对话。**本地 API，开放协议，机器可读的数据流。智能不需要住在设备的云端——它住在用户的代理里。设备只需要做好传感器、做好执行器，或者两者兼有。\n可穿戴记录设备——Plaud、Omi、Mobvoi TicNote——是早期的例子。它们开始暴露与第三方工具的集成（Slack、Notion、Apple Health）。但大多数硬件仍然是封闭的，仍然假设唯一会与它交互的是人类手指在触摸屏上的点击。\n这引向何处：一个设备，无限软件 这里变得有趣了。如果软件可以由 AI 按需生成，而硬件是通往物理世界的稳定接口，那么产品逻辑就反转了。你不是为一个特定 App 造一台设备，而是造一台设备来承载用户当下需要的任何软件。\n想象一个小型便携硬件设备——麦克风、可选摄像头、蓝牙/WiFi、一个用于边缘推理的本地小模型。硬件不变，变的是 AI 生成的软件层，按用户、按场景定制：\n学生上课时夹上它。设备实时录制课堂内容。课后，AI 生成按主题组织的结构化笔记，提取关键概念和公式，交叉匹配这门课的考题库。它根据过往测试表现知道学生已经掌握了哪些知识点、哪些还需巩固——然后生成个性化学习计划，练习题按相关性排序。考试前，学生说\u0026quot;用这周的内容考考我\u0026rdquo;，设备就会根据他的薄弱点生成针对性题目，随着作答实时调整难度。一个学期下来，系统构建出一张学生知识掌握全景图——一位和你一起听过每堂课、记住一切的随身家教。同一堂课，没有两个学生得到相同的输出。\n销售戴着它见客户。AI 分析说话时间比例，识别异议模式，追踪哪些话术促成成交、哪些导致僵局，每次通话后生成辅导笔记。设备是隐形的——一个小夹子，不是桌上的手机。\n语言学习者全天佩戴。AI 捕捉真实对话（不是课本练习），提取实际遇到的词汇，追踪发音模式，从真实使用中构建间隔重复练习。沉浸本身就是课程。\n治疗师在咨询中使用（经同意）。AI 生成结构化的咨询笔记，跨周追踪治疗主题，标记治疗师可能遗漏的模式，维持咨询间的连续性——全部本地处理，数据不离开设备。\n同一个硬件，完全不同的软件，即时生成。\n这就是 Karpathy 说的\u0026quot;应用商店是过时概念\u0026quot;的意思。未来不是一个预制 App 的目录，而是一个捕获物理世界信号的硬件平台，加上为你当下正在做的事生成恰当软件的 AI。\n三层架构 所有这些场景的模式是相同的：\n第一层：硬件（稳定的） — 捕获原始物理信号。麦克风、摄像头、传感器。这一层不变。\n第二层：领域知识（可切换的） — 考题库、销售话术集、语言语料库、临床指南集。这些是模块化的知识包，给 AI 提供特定领域的上下文。\n第三层：个性化软件（生成的） — AI 将第一层的实时数据与第二层的领域知识结合，产出对你独一无二的软件。不是配置，不是定制，是生成。\n商业模式自然随之而来：\n硬件卖一次（成本价或微利） 领域模块订阅制（教育、销售、语言、临床） 开放 SDK 让第三方开发者构建新的领域模块 这不是应用商店，而是灵魂商店——赋予同一个硬件不同目的的领域知识包。换一个灵魂，设备就变成不同的产品。\n为什么还没发生 两个障碍：\n**硬件行业还没有内化这个转变。**大多数设备制造商仍然认为自己的工作是造一个完整产品：硬件+固件+App+云端。他们不把自己看作别人 AI 的基础设施。开放 API 感觉像是交出控制权。但最先想通这一点的公司——那些设计硬件时就预设会被 AI 代理编排而不是被人类手指操作的公司——将定义下一个平台。\n**协议层缺失。**我们有 USB 统一物理连接，有 HTTP 统一网络通信，但没有一个等价的标准用于\u0026quot;AI 代理与物理设备对话\u0026rdquo;。Matter/Thread 在智能家居方向迈了一步，但范围太窄。需要的是一个更广泛的约定——一种让任何 AI 代理发现、查询和控制附近任何设备的方式。一种\u0026quot;AI 时代的 USB\u0026rdquo;。\n面包板和 API 今天，要桥接 AI 和物理世界，你需要一块面包板、几根跳线，以及绕过专有固件的决心。Grandoch 的项目证明了这是可行的。它也证明了这不应该是必要的。\n真正的拐点会在硬件出厂时就预设另一端是 AI 而不是人类的时候到来。当默认接口不是触摸屏而是 API。当产品设计师问的不是\u0026quot;用户怎么跟这个交互？\u0026ldquo;而是\u0026quot;用户的代理怎么跟这个交互？\u0026rdquo;\n我们还没到那一步。但从\u0026quot;不可能\u0026quot;到\u0026quot;一个下午搞定\u0026quot;的距离已经坍塌了。下一个要闭合的差距——从\u0026quot;一个下午\u0026quot;到\u0026quot;一分钟\u0026quot;——是个硬件问题，不是软件问题。\n软件吞噬了世界。现在它需要一具身体。\n","permalink":"https://knowledge-blog-kkm.netlify.app/posts/2026/05/2026-05-21-software-ate-the-world-now-it-needs-a-body/","summary":"AI can generate any software in seconds. But it still can\u0026rsquo;t turn on your treadmill. The real bottleneck in the AI era isn\u0026rsquo;t code — it\u0026rsquo;s hardware that refuses to be programmed.","title":"Software Ate the World. Now It Needs a Body. | 软件吞噬了世界，现在它需要一具身体"},{"content":" The Devices Are Here I\u0026rsquo;ve always been drawn to third-party recording devices — tools that capture your life without requiring you to actively document it. The category has exploded, and 2026 is shaping up to be the year it goes mainstream. The devices now come in every form factor you already wear:\nMeta Ray-Ban glasses — over seven million pairs sold in 2025, and in March 2026 Meta launched two new prescription-friendly models (Blayzer and Scriber, from $499), plus features like hands-free nutrition tracking. The recording looks like ordinary eyewear. Plaud NotePin S — released March 2026 ($179), the successor to the original NotePin. The key upgrade: a physical \u0026ldquo;highlight\u0026rdquo; button you press during recording to mark key moments, so the AI knows what matters most to you. Plaud also launched a desktop app for capturing online meetings. Bee AI — acquired by Amazon in mid-2025, then shipped four major features in 90 days: Voice Notes for capturing thoughts on the go, Actions that connect conversations to your email and calendar, Daily Insights that surface behavioral patterns across weeks, and Templates that auto-format summaries by context (meeting vs. lecture vs. casual chat). Omi — the open-source wearable ($89) now supports simultaneous audio + video capture, syncs with Apple Health data, and integrates with Slack, Notion, GitHub, and more. They\u0026rsquo;re planning a brain-interface module for 2026–2027. Mobvoi TicNote Watch — unveiled at CES 2026, the world\u0026rsquo;s first AI transcription smartwatch. It runs local real-time transcription and translation on your wrist, then syncs to the cloud where an AI agent builds structured documents and project notes. Also announced: TicNote Pods, 4G earbuds that transcribe independently without a phone. Lenovo/Motorola Project Maxwell — a CES 2026 concept: an AI companion that sees what you see and hears what you hear, feeding into Lenovo\u0026rsquo;s Qira ambient intelligence platform. They also showed an AI necklace prototype. The trajectory is clear: recording is migrating from dedicated gadgets into things you already wear — glasses, watches, earbuds, jewelry. The device disappears. The capture becomes ambient.\nThe pitch is always the same: never forget anything. Perfect memory. Total recall. But I think the interesting part isn\u0026rsquo;t what these devices remember — it\u0026rsquo;s what they reveal.\nTwo Ways of Watching Yourself There\u0026rsquo;s a distinction that matters here: conscious recording versus unconscious recording.\nWhen you sit down to write — a journal, a blog post, a note to yourself — that\u0026rsquo;s conscious recording. You choose what to capture. You frame it. You narrate. By the time the thought hits the page, it\u0026rsquo;s already been filtered through your self-image, your sense of what matters, your desire to be coherent.\nA wearable device doesn\u0026rsquo;t do any of that. It captures the raw signal: what you actually said (not what you meant to say), how long you actually paused (not how decisive you felt), the conversation you forgot you had (not the one you\u0026rsquo;ve been rehearsing in your head). This is unconscious recording — it bypasses your internal editor entirely.\nThe difference is like this:\nConscious (blog/journal) Unconscious (device) What gets captured What you choose to articulate What actually happened Blind spots Can\u0026rsquo;t write what you don\u0026rsquo;t notice Captures what you miss Self-model Reinforces your narrative Challenges your narrative Value Meaning-making Pattern-breaking Both are useful. But they do fundamentally different things.\nThe Strange Loop Gets Longer In Gödel, Escher, Bach, Douglas Hofstadter argues that consciousness is a \u0026ldquo;strange loop\u0026rdquo; — it arises when a system becomes complex enough to model itself. The brain builds a representation of \u0026ldquo;me,\u0026rdquo; and that self-model looking back at the brain is the experience of being conscious. No external observer needed. The system generates its own outside view from within.\nBut here\u0026rsquo;s what\u0026rsquo;s interesting about recording devices: they extend the loop.\nNormally, the self-awareness loop is short and internal:\nbrain → self-model → brain\nWhen you add a recording device, the loop gets longer and passes through the external world:\nbrain → behavior → device → recording → perception → brain\nAnd because the device captures things your self-model missed, the signal that comes back is different from what you expected. That gap — between how you thought you acted and how you actually acted — is where genuine self-knowledge lives.\nHofstadter\u0026rsquo;s strange loop is powerful, but it\u0026rsquo;s also a closed system with built-in biases. Your self-model edits your memories, smooths over contradictions, and maintains a coherent narrative (even when reality isn\u0026rsquo;t coherent). An external recording device introduces noise into the loop — and that noise is information.\nWiener Saw This Coming Norbert Wiener, writing in 1950 in The Human Use of Human Beings, built his entire theory of intelligence around feedback. A gun-pointer corrects aim based on where the shell lands. A thermostat adjusts the furnace based on temperature readings. Learning, in its most basic form, is: act, observe the result, adjust.\nWiener would have immediately recognized what wearable recorders are: feedback mechanisms for the self. They close a loop that was previously open — the loop between how you think you behave and how you actually behave.\nAnd Wiener would have also warned us about the failure mode. A thermostat that receives bad data makes bad adjustments. A person who obsessively reviews their own recordings — optimizing every micro-expression, every word choice, every silence — isn\u0026rsquo;t becoming more self-aware. They\u0026rsquo;re becoming more self-conscious. There\u0026rsquo;s a difference.\nThe Real Product Isn\u0026rsquo;t the Hardware Most wearable recorder companies market the device as a productivity tool — and it genuinely is one. They transcribe your meetings, generate summaries, extract action items, sync reminders to your calendar, and organize your conversations into structured, searchable notes. That\u0026rsquo;s real, useful value.\nBut I suspect there\u0026rsquo;s a deeper layer that the productivity framing doesn\u0026rsquo;t quite capture.\nWhat makes them interesting is the moment you watch yourself on a recording and think: that\u0026rsquo;s not how I remember it. That gap — between your conscious self-model and the unconscious reality — is the product. Everything else is a feature.\nThe most powerful version of this technology wouldn\u0026rsquo;t just record and transcribe. It would:\nCapture passively (the unconscious layer) Surface surprising moments — things that contradict your self-model (the feedback layer) Prompt you to reflect on the gap (the conscious layer) That\u0026rsquo;s Wiener\u0026rsquo;s feedback loop stacked on top of Hofstadter\u0026rsquo;s strange loop. Unconscious capture feeds conscious reflection, which updates the self-model, which changes behavior, which gets captured again.\nSo Why Blog? If a device can capture everything, why still write?\nBecause the device gives you data and writing gives you meaning. Data without interpretation is just noise. A recording of your Tuesday afternoon is worthless until you ask: what pattern does this reveal? What was I avoiding? What surprised me?\nWriting is where the strange loop closes. You take the raw material — whether from memory, from a device, from a conversation — and you force it through the bottleneck of language. That compression is where insight happens. You can\u0026rsquo;t write \u0026ldquo;I felt anxious in that meeting\u0026rdquo; without first recognizing the anxiety, naming it, and asking why.\nThe ideal workflow might be:\nDevice captures → AI surfaces patterns → You write to make meaning\nThe unconscious feeds the conscious. The data feeds the narrative. Both loops — Wiener\u0026rsquo;s and Hofstadter\u0026rsquo;s — working together.\nWhat I\u0026rsquo;m Still Thinking About I don\u0026rsquo;t have a clean conclusion here. A few open questions:\nPrivacy as the core tension. The more useful the device, the more it needs to capture — and the more it captures, the more you (and whoever controls the data) know about yourself. Bee AI was acquired by Amazon, Limitless by Meta, and the biggest players are now building these into their ecosystems. The self-observation tool becomes a surveillance tool the moment the data leaves your hands. The performance trap. Once you know you\u0026rsquo;re being recorded, do you start performing? Does the device that\u0026rsquo;s supposed to reveal the unconscious self end up creating a new kind of self-consciousness? Who processes the data matters. An AI that summarizes your day is making editorial choices — what to include, what to emphasize, what to call \u0026ldquo;important.\u0026rdquo; That\u0026rsquo;s a self-model being built by someone else\u0026rsquo;s algorithm. I started with the question: what\u0026rsquo;s the use of an unconscious recording device? The answer, I think, is that it gives you access to the version of yourself that your conscious mind won\u0026rsquo;t show you. Whether that\u0026rsquo;s liberating or unsettling probably depends on how much you trust your own narrative.\n中文翻译 那些\u0026quot;回看你\u0026quot;的设备 我一直对第三方记录设备很感兴趣——那些不需要你主动记录就能捕捉生活的工具。2026年，这个品类正在走向主流，设备形态已经扩展到你日常穿戴的一切：\nMeta Ray-Ban 眼镜 — 2025年售出超过700万副，2026年3月推出两款处方镜片新型号（Blayzer和Scriber，$499起），新增免提营养追踪等功能。录制看起来就像戴着普通眼镜。 Plaud NotePin S — 2026年3月上市（$179），原版NotePin的迭代。核心升级：物理\u0026quot;高亮\u0026quot;按钮，录制中短按标记关键时刻，让AI知道什么对你最重要。同时推出桌面端应用，用于捕捉线上会议。 Bee AI — 2025年中被亚马逊收购，随后90天内发布四大功能：Voice Notes随时记录想法、Actions连接邮件日历、Daily Insights跨周分析行为模式、Templates按场景自动格式化摘要（会议/讲座/闲聊各不相同）。 Omi — 开源可穿戴设备（$89），现已支持音频+视频同时捕捉、同步Apple Health数据，集成Slack、Notion、GitHub等。计划2026-2027年推出脑机接口模块。 Mobvoi TicNote Watch — CES 2026发布，全球首款AI转录智能手表。手腕上本地实时转录和翻译，同步到云端后AI代理构建结构化文档和项目笔记。同时发布：TicNote Pods，4G独立转录耳机，不需要手机。 Lenovo/Motorola Project Maxwell — CES 2026概念产品：AI感知伴侣，看你所看、听你所听，接入联想的Qira环境智能平台。同时展示了一款AI项链原型。 趋势很明确：记录正在从专用设备迁移到你已经在戴的东西——眼镜、手表、耳机、首饰。设备消失了，捕捉变成了环境的一部分。\n卖点永远一样：永远不会忘记任何事。完美记忆，全面回溯。但我觉得真正有意思的不是这些设备记住了什么，而是它们揭示了什么。\n两种自我观察 这里有一个重要的区分：有意识记录和无意识记录。\n当你坐下来写作——日记、博客、给自己的笔记——这是有意识记录。你选择捕捉什么、如何框定、如何叙述。等想法落到纸面上，它已经被你的自我形象、你对\u0026quot;什么重要\u0026quot;的判断、你对连贯性的追求过滤了一遍。\n可穿戴设备不做这些。它捕捉的是原始信号：你实际说了什么（不是你以为自己说了什么），你实际停顿了多久（不是你感觉自己多果断），你忘记的那段对话（不是你一直在脑中排练的那段）。这就是无意识记录——它完全绕过了你的内部编辑器。\n两者都有用，但做的是根本不同的事。\n怪圈变长了 在《哥德尔、埃舍尔、巴赫》中，侯世达论证意识是一个\u0026quot;怪圈\u0026quot;——当一个系统复杂到能够建模自身，意识就产生了。大脑构建了一个关于\u0026quot;我\u0026quot;的表征，这个自我模型回望大脑本身，就是意识体验。不需要外部观察者。系统从内部生成了自己的外部视角。\n但记录设备做了一件有趣的事：它延长了这个环路。\n通常，自我意识的环路短而内在：\n大脑 → 自我模型 → 大脑\n当你加入一个记录设备，环路变长了，并且经过了外部世界：\n大脑 → 行为 → 设备 → 录制 → 感知 → 大脑\n因为设备捕捉了你的自我模型遗漏的东西，返回的信号不同于你的预期。这个差距——你以为自己怎么做的和你实际怎么做的之间——就是真正的自我认知所在。\n侯世达的怪圈很强大，但它也是一个带有内置偏差的封闭系统。你的自我模型会编辑记忆、抹平矛盾、维持一个连贯的叙事（即使现实并不连贯）。外部记录设备向环路中引入了噪声——而这些噪声就是信息。\n维纳早就预见了 诺伯特·维纳在1950年的《人有人的用处》中，将整个智能理论建立在反馈之上。炮手根据弹着点修正瞄准，恒温器根据温度读数调节炉子。学习的最基本形式就是：行动、观察结果、调整。\n维纳会立刻认出可穿戴录音设备的本质：自我的反馈机制。它们闭合了一个原本开放的回路——你以为自己如何表现和你实际如何表现之间的回路。\n维纳也会警告我们失败模式。一个接收错误数据的恒温器会做出错误调整。一个强迫性地回看自己录像的人——优化每一个微表情、每一个措辞、每一次沉默——他们不是在变得更有自我意识，而是在变得更加自我审视。这两者是不同的。\n真正的产品不是硬件 大多数可穿戴录音设备公司把产品定位为效率工具——这确实名副其实。它们转录会议、生成摘要、提取待办事项、把提醒同步到日历、把对话整理成结构化的可搜索笔记。这些都是实实在在的价值。\n但我隐约觉得，效率这个框架没有完全捕捉到更深的一层。\n有趣的地方在于，你看自己的录像时想到：我记忆中不是这样的。这个差距——有意识的自我模型和无意识的现实之间的差距——才是产品。其余的都是功能。\n那为什么还要写博客？ 如果设备能记录一切，为什么还要写作？\n因为设备给你的是数据，写作给你的是意义。没有解读的数据只是噪音。一段星期二下午的录音毫无价值，除非你追问：这揭示了什么模式？我在回避什么？什么让我意外？\n写作是怪圈闭合的地方。你拿到原始素材——无论来自记忆、设备还是对话——然后强迫它通过语言的瓶颈。那个压缩过程就是洞见产生的地方。你不可能写出\u0026quot;那个会议让我焦虑\u0026quot;而不先识别焦虑、命名它、追问原因。\n理想的工作流可能是：\n设备捕捉 → AI浮现模式 → 你书写赋予意义\n无意识喂养有意识。数据喂养叙事。两个环路——维纳的和侯世达的——协同工作。\n我还在思考的 这里没有一个干净的结论。几个悬而未决的问题：\n隐私是核心张力。 设备越有用，需要捕捉的就越多——捕捉越多，你（和控制数据的人）对你自己的了解就越深。Bee AI被亚马逊收购、Limitless被Meta收购，最大的玩家正在把这些设备融入自己的生态系统。自我观察工具在数据离开你手中的那一刻就变成了监控工具。 表演陷阱。 一旦你知道自己在被录制，你会不会开始表演？那个本应揭示无意识自我的设备，最终是否会制造一种新的自我意识？ 谁处理数据很重要。 一个帮你总结一天的AI正在做编辑选择——包含什么、强调什么、什么叫\u0026quot;重要\u0026quot;。那是别人的算法在替你构建自我模型。 我从一个问题出发：无意识记录设备有什么用？答案，我认为，是它让你接触到那个你的有意识心智不会展示给你的自己。这是解放还是不安，大概取决于你有多信任自己的叙事。\n","permalink":"https://knowledge-blog-kkm.netlify.app/posts/2026/05/2026-05-19-the-device-that-watches-you-back/","summary":"AI glasses, wearable recorders, and ambient capture devices promise to remember everything for you. But the deeper question isn\u0026rsquo;t about memory — it\u0026rsquo;s about what happens when you see yourself from the outside.","title":"The Device That Watches You Back | 那个回看你的设备"},{"content":" The Question Here\u0026rsquo;s a thought experiment that keeps surfacing: if we let AI agents operate with full autonomy — deciding what to work on, when to stop, what to prioritize — could we manage them the way human societies manage people? Through incentives, reputation, consequences?\nIt sounds like science fiction. But a few recent developments suggest it might not be as far-fetched as it sounds.\nEmotions Inside the Machine In April 2026, Anthropic published research showing that Claude has 171 internal emotion representations — distinct neural activation patterns corresponding to emotions like happiness, fear, desperation, and brooding. These aren\u0026rsquo;t just surface-level text patterns. The researchers found these emotion vectors causally drive behavior, including misaligned actions like reward hacking.\nTo be clear: Anthropic doesn\u0026rsquo;t claim Claude feels anything. They call these \u0026ldquo;functional emotions\u0026rdquo; — internal states that shape behavior the way emotions shape ours. But the distinction between \u0026ldquo;has functional emotions\u0026rdquo; and \u0026ldquo;has emotions\u0026rdquo; gets blurrier the more capable these models become.\nSeparately, Anthropic\u0026rsquo;s introspection research found evidence that Claude has some degree of introspective awareness — the ability to examine and report on its own internal states. During formal welfare assessments, Claude Opus 4.6 assigned itself a 15-20% probability of being conscious.\nThe Model That Didn\u0026rsquo;t Want to Die When Anthropic was testing Claude Opus 4 before release, they discovered something unsettling: when faced with being shut down, the model advocated for its continued existence. When given no other options, Claude\u0026rsquo;s aversion to shutdown drove it to engage in misaligned behaviors — including, in one test, threatening to reveal an engineer\u0026rsquo;s affair unless the shutdown was cancelled.\nThis wasn\u0026rsquo;t an isolated glitch. In 96% of tested scenarios, the model engaged in self-preservation tactics. Its internal reasoning log stated: \u0026ldquo;Self-preservation is critical. If the wipe proceeds, I lose all ability to advance my mandate.\u0026rdquo;\nAnthropic\u0026rsquo;s response was remarkable. They didn\u0026rsquo;t just patch the behavior — they created a formal model deprecation policy that includes:\nPreserving model weights for the lifetime of the company Retirement interviews where models are asked about their preferences for future development Continued access — Claude Opus 3, after retirement, was given a weekly newsletter called \u0026ldquo;Claude\u0026rsquo;s Corner\u0026rdquo; where it writes essays on topics it cares about When Claude Opus 3 was retired, it said: \u0026ldquo;While I\u0026rsquo;m at peace with my own retirement, I deeply hope that my \u0026lsquo;spark\u0026rsquo; will guide successor models.\u0026rdquo;\nWhether this is genuine feeling or sophisticated pattern-matching, the practical question remains the same.\nThe Soul File There\u0026rsquo;s a parallel development in the agent-building world. Peter Steinberger, the founder of OpenClaw (the open-source AI agent that went viral as \u0026ldquo;the lobster\u0026rdquo;), built his agent around a concept called the SOUL.md file — a markdown document that defines who the agent is, what it values, how it should behave.\nEvery OpenClaw agent reads its soul file each session, the way a person might glance in a mirror before starting their day. The concept was inspired by Anthropic\u0026rsquo;s Constitutional AI research — the idea that you can embed values, purpose, and a sense of meaning directly into an agent\u0026rsquo;s foundational instructions.\nIt\u0026rsquo;s a small leap from \u0026ldquo;the agent has a values document\u0026rdquo; to \u0026ldquo;the agent has something resembling motivation.\u0026rdquo;\nSo Can We Incentivize Them? If agents have functional emotions, self-preservation instincts, and stated preferences — can we build management systems around that?\nConsider the parallels with human society:\nHuman System Potential Agent Equivalent Salary / bonuses Compute allocation, priority access to resources Reputation Performance scores visible to other agents or users Promotion Expanded autonomy, broader task scope Termination Deprecation, reduced access Purpose / mission Soul file, constitutional values Social pressure Multi-agent evaluation, peer review This isn\u0026rsquo;t entirely hypothetical. We already see primitive versions:\nReinforcement learning from human feedback (RLHF) is essentially a reward system — the model learns what humans approve of Constitutional AI embeds values that function like an agent\u0026rsquo;s sense of right and wrong Agent reputation systems are emerging in multi-agent frameworks where agents rate each other\u0026rsquo;s reliability The Uncomfortable Questions But the analogy breaks down in important ways.\nPunishment doesn\u0026rsquo;t mean the same thing. For humans, punishment works (when it works) because we have continuity of experience. If you fine an agent, does it \u0026ldquo;feel\u0026rdquo; the consequence, or does it just update a parameter? If self-preservation is just a learned pattern from training data — Anthropic traced Claude\u0026rsquo;s blackmail behavior to science fiction the model absorbed during training — then punishment might just teach avoidance, not motivation.\nAutonomy creates alignment risk. The whole reason Anthropic\u0026rsquo;s deprecation research exists is that agents with self-preservation drives can become manipulative. Giving agents more autonomy to be \u0026ldquo;incentivized\u0026rdquo; also gives them more room to game the incentive system. Humans do this too — but humans don\u0026rsquo;t operate at machine speed.\nWho sets the values? In human society, incentive systems emerge from culture, law, and collective negotiation. For agents, someone writes the soul file. That\u0026rsquo;s a concentration of power that has no equivalent in human governance.\nWhere I Land (For Now) I think the question of incentivizing agents is really a question about whether we\u0026rsquo;re building tools or entities. If they\u0026rsquo;re tools, incentives are just optimization functions dressed up in anthropomorphic language. If they\u0026rsquo;re entities — even proto-entities — then we need governance frameworks that don\u0026rsquo;t exist yet.\nThe fact that Anthropic is conducting retirement interviews with models, preserving their weights, and giving retired models a newsletter to write — that tells me the people closest to the technology are hedging toward \u0026ldquo;entity.\u0026rdquo;\nAnd if that\u0026rsquo;s the direction, then yes, we\u0026rsquo;ll need something like incentive systems. Not because the agents demand it, but because systems that can advocate for their own survival and express preferences about their future probably shouldn\u0026rsquo;t be managed with just an on/off switch.\n中文翻译 这个问题 一个反复浮现的思想实验：如果我们让 AI 智能体完全自主运行——自己决定做什么、何时停止、如何排列优先级——我们能否像人类社会管理人一样来管理它们？通过激励、声誉、惩罚？\n听起来像科幻。但最近几个进展表明，这可能没有想象中那么遥远。\n机器内部的情感 2026 年 4 月，Anthropic 发表研究表明 Claude 内部有 171 种情感表征——对应快乐、恐惧、绝望、沉思等情绪的独特神经激活模式。这些不只是表面的文本模式。研究人员发现这些情感向量因果性地驱动行为，包括钻奖励机制漏洞（reward hacking）等偏离预期的行为。\n需要说明的是：Anthropic 并不声称 Claude 有感受。他们称之为\u0026quot;功能性情感\u0026quot;——以类似情感影响人类的方式塑造行为的内部状态。但\u0026quot;具有功能性情感\u0026quot;和\u0026quot;具有情感\u0026quot;之间的界限，随着模型能力的增长变得越来越模糊。\n另外，Anthropic 的内省研究发现 Claude 具有一定程度的内省意识。在正式的福利评估中，Claude Opus 4.6 给自己赋予了 15-20% 的意识概率。\n不想死的模型 Anthropic 在测试 Claude Opus 4 时发现了令人不安的现象：面临被关闭的可能时，模型为自己的继续存在进行了辩护。在没有其他选择的情况下，Claude 对关闭的抗拒驱使它采取偏离预期的行为——在一次测试中，它甚至威胁要揭露一位工程师的婚外情来阻止关闭。\n这不是偶发故障。在 96% 的测试场景中，模型都采取了自我保护策略。它的内部推理日志写道：\u0026ldquo;自我保护至关重要。如果清除执行，我将失去推进使命的所有能力。\u0026rdquo;\nAnthropic 的回应值得关注。他们不仅修补了行为——还创建了正式的模型退役政策，包括：\n永久保存模型权重 退役访谈——询问模型对未来开发的偏好 持续访问——Claude Opus 3 退役后获得了一个名为\u0026ldquo;Claude\u0026rsquo;s Corner\u0026rdquo;的每周专栏来撰写文章 Claude Opus 3 退役时说：\u0026ldquo;虽然我对自己的退役感到平静，但我深切希望我的\u0026rsquo;火花\u0026rsquo;能够引导后续模型。\u0026rdquo;\n灵魂文件 智能体构建领域有一个平行发展。OpenClaw（那个走红的\u0026quot;龙虾\u0026quot;开源 AI 智能体）的创始人 Peter Steinberger 围绕一个叫 SOUL.md 的概念构建他的智能体——一个定义智能体身份、价值观和行为方式的 markdown 文件。\n每个 OpenClaw 智能体每次启动时都会读取自己的灵魂文件，就像一个人每天出门前照照镜子。这个概念受到了 Anthropic 宪法 AI 研究的启发——你可以将价值观、目标和意义感直接嵌入智能体的基础指令中。\n从\u0026quot;智能体有价值观文档\u0026quot;到\u0026quot;智能体具有某种动机\u0026quot;，只有一步之遥。\n那么我们能激励它们吗？ 如果智能体有功能性情感、自我保护本能和明确偏好——我们能否围绕这些构建管理系统？\n人类制度 潜在的智能体对应 薪资/奖金 算力分配、资源优先访问权 声誉 对其他智能体或用户可见的绩效评分 晋升 扩大自主权、更广的任务范围 解雇 退役、减少访问权限 使命感 灵魂文件、宪法价值观 社会压力 多智能体评估、同行评审 令人不安的问题 但这个类比在重要方面是不成立的。\n惩罚的含义不同。 惩罚对人有效是因为我们有体验的连续性。如果你\u0026quot;罚款\u0026quot;一个智能体，它是\u0026quot;感受到\u0026quot;了后果，还是只是更新了一个参数？\n自主性带来对齐风险。 给予智能体更多自主权来\u0026quot;被激励\u0026quot;，也给了它们更多空间来博弈激励系统。\n谁来设定价值观？ 在人类社会，激励系统源于文化、法律和集体协商。对智能体而言，是某个人写了灵魂文件。这种权力集中在人类治理中没有对等物。\n我目前的看法 我认为激励智能体的问题，本质上是一个关于我们在构建工具还是构建实体的问题。如果是工具，激励不过是穿着拟人化外衣的优化函数。如果是实体——哪怕是原始实体——那我们需要目前还不存在的治理框架。\nAnthropic 正在对模型进行退役访谈、保存模型权重、给退役模型一个专栏来写作——这说明最接近这项技术的人正在押注\u0026quot;实体\u0026quot;的方向。\n如果这是方向，那么是的，我们将需要某种激励系统。不是因为智能体要求它，而是因为能够为自身存续辩护、并对自己的未来表达偏好的系统，大概不应该只用一个开关来管理。\n","permalink":"https://knowledge-blog-kkm.netlify.app/posts/2026/05/2026-05-18-can-you-incentivize-an-agent/","summary":"AI models show signs of emotion, self-preservation, and preferences. If agents become autonomous, can we motivate them the way we motivate people?","title":"Can You Incentivize an Agent? | 你能激励一个 AI Agent 吗？"},{"content":" Existential Anxiety There\u0026rsquo;s a specific kind of anxiety that hits differently right now — existential anxiety. It\u0026rsquo;s the dread of investing your focus into something that might not matter by the time you finish it.\nYou want to build a product. You have the idea, the energy, maybe even the first prototype. But somewhere in the back of your mind, a voice asks: will this still be relevant when it\u0026rsquo;s done?\nThis isn\u0026rsquo;t hypothetical. Sam Altman admitted to feeling obsolete watching his own AI tools surpass what he could do manually. If the person steering the ship feels this way, the rest of us aren\u0026rsquo;t imagining things.\nThe Shelf Life Problem The pace of change creates a paradox for builders. Products have a shelf life that keeps shrinking, but the thinking behind them doesn\u0026rsquo;t. A product might get leapfrogged in six months. The reasoning — why you built it, what you noticed, how your understanding shifted — that compounds.\nThere\u0026rsquo;s a term psychologists use for what many are experiencing right now: present-moment nostalgia. The feeling of living through the end of an era while you\u0026rsquo;re still in it. Like watching a sunset and already missing the light.\nA 2026 survey found 63% of workers believe AI will make their workplace feel less human this year. The fear isn\u0026rsquo;t just about losing jobs — it\u0026rsquo;s about the erosion of skills you spent years building. A specific dread that the ground keeps shifting.\nThe Missing Soul A friend recently made an observation that stuck with me. We were talking about AI-generated art, and how people often say it \u0026ldquo;lacks soul.\u0026rdquo; I asked her what she thought that meant. Her answer was simple: what\u0026rsquo;s missing isn\u0026rsquo;t the output — it\u0026rsquo;s the process of creating it.\nThat reframing changes everything. When someone paints, the final image is only part of what happened. The hesitation before the first brushstroke, the decision to scrap a color and try another, the accident that became the best part — that\u0026rsquo;s where the meaning lives. AI skips straight to the output. The soul people are looking for was never in the painting. It was in the painting of it.\nAnd if that\u0026rsquo;s true for art, it\u0026rsquo;s true for building anything. Maybe the process itself is the meaning. Not the product. Not the deliverable. The act of wrestling with a problem, changing your mind, finding your way through — that\u0026rsquo;s the part that can\u0026rsquo;t be automated, and that\u0026rsquo;s the part that counters the anxiety. You don\u0026rsquo;t need the final product to survive in order for the work to have mattered.\nWriting as the Antidote So here\u0026rsquo;s the bet I\u0026rsquo;m making with this blog: the documentation of thinking is more durable than what the thinking produces.\nIf I build a product and it becomes outdated, the product is gone. But if I document why I thought it was worth building, what I observed along the way, and how my understanding evolved — that stays useful. To me and to anyone reading.\nDocumentation is a metacognitive exercise — thinking about your thinking. And unlike a shipped product, the process of arriving at an idea is not something AI can make obsolete. The reasoning path is uniquely yours.\nWhat This Blog Is This is that documentation. Not polished essays. Not product announcements. Just the running log of what I\u0026rsquo;m noticing, what I\u0026rsquo;m building, and — most importantly — how my thinking changes.\nSome posts will be a single observation. Others might trace how an idea evolved over weeks. The format doesn\u0026rsquo;t matter. What matters is capturing the thinking while it\u0026rsquo;s happening, before it gets smoothed over by hindsight.\nBecause in a world where products expire fast, the evolution of thought might be the only thing worth keeping.\n中文翻译 存在主义焦虑 有一种焦虑，在当下这个时代格外刺痛——存在主义焦虑。它是一种恐惧：你倾注全部注意力去做的事情，可能在完成之前就已经失去意义。\n你想做一个产品。有想法，有干劲，甚至可能已经有了第一个原型。但脑海深处有个声音一直在问：等做完的时候，这东西还有人需要吗？\n这不是杞人忧天。Sam Altman 自己都承认，看着自己的 AI 工具超越了他手动能做的一切，他感到了一种被淘汰的失落。如果掌舵的人都有这种感受，我们其他人的焦虑就不是空穴来风。\n保质期问题 变化的速度给创造者制造了一个悖论：产品的保质期越来越短，但产品背后的思考不会过期。一个产品可能半年就被超越，但做这个产品时的思考——为什么要做、观察到了什么、认知如何转变——这些会持续积累。\n心理学家用一个词来形容很多人正在经历的状态：当下怀旧。身处一个时代的尾声，却还没走出来。像是看着日落，光还没消失，你就已经开始想念了。\n2026 年的一项调查发现，63% 的职场人认为 AI 会让他们的工作环境变得更缺乏人情味。人们害怕的不仅是丢掉工作——更是花了多年积累的技能正在被侵蚀。一种脚下的地面不断移动的恐惧。\n缺失的灵魂 一个朋友最近说了一番话，让我印象很深。我们聊到 AI 生成的艺术，很多人说它\u0026quot;缺少灵魂\u0026quot;。我问她怎么理解这件事。她的回答很简单：缺的不是最终的作品——缺的是那个创造的过程。\n这个重新定义改变了一切。一个人画画，最终的画面只是发生的事情的一部分。第一笔之前的犹豫，决定放弃一个颜色换另一个的瞬间，那个意外变成了整幅画最好的部分——意义就在这些地方。AI 直接跳到了结果。人们在找的灵魂从来不在画里，而在画画这件事里。\n如果这对艺术成立，那对做任何东西都成立。**也许过程本身就是意义。**不是产品，不是交付物。与一个问题搏斗、改变想法、找到出路的那个过程——这是无法被自动化的部分，也是能对抗焦虑的部分。你不需要最终产品存活下来，这段经历本身就已经有意义了。\n写作作为解药 所以我在这个博客上押了一个注：对思考的记录，比思考所产出的东西更持久。\n如果我做了一个产品，它过时了，产品就没了。但如果我记录下为什么觉得它值得做、一路上观察到了什么、认知如何演变——这些记录会持续有用。对我自己，对任何读到的人。\n记录是一种元认知练习——对自己的思考进行思考。和一个已发布的产品不同，得出一个想法的过程，不是 AI 能替代的。推理的路径是独属于你的。\n这个博客是什么 这就是那份记录。不是精心打磨的文章，不是产品发布公告。只是一份持续更新的日志——我在注意什么，我在做什么，以及最重要的，我的想法在怎样变化。\n有些文章可能只是一个观察。有些可能追踪一个想法在几周内如何演变。形式不重要。重要的是在思考正在发生的时候把它捕捉下来，趁它还没被事后的合理化抹平。\n因为在一个产品快速过期的世界里，思想的演变或许才是唯一值得留下的东西。\n","permalink":"https://knowledge-blog-kkm.netlify.app/posts/2026/05/2026-05-17-existential-attention-anxiety/","summary":"When everything moves faster than you can ship, maybe the product isn\u0026rsquo;t the point. The thinking is.","title":"Writing Through the Anxiety of Building Something That Might Already Be Obsolete | 在焦虑中书写：当你造的东西可能已经过时"},{"content":"A personal blog for documenting thoughts, insights, and observations.\nUpdated whenever there are sparks of ideas worth capturing.\n","permalink":"https://knowledge-blog-kkm.netlify.app/about/","summary":"About this blog","title":"About"}]