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: “从情绪到情感，最是人工智能未来前进的方向.” From emotions to feelings, this is the most important direction for AI’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.

Seven years later, in a 2024 interview with Issues in Science and Technology, she said it again, more directly: “I naturally think about compassion and love. I think this is what defines us as human.” And then: “It’s not clear there is a mathematical path toward that.”

In November 2025, she published “From Words to Worlds” on Substack, calling LLMs “wordsmiths in the dark, eloquent but inexperienced, knowledgeable but ungrounded.” The essay argued that spatial intelligence is AI’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.

Her 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.

Descartes’ 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.

Antonio Damasio’s clinical research says otherwise.

In Descartes’ 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.

The result was not purer reasoning. It was paralysis.

His 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.

Damasio’s Iowa Gambling Task demonstrated this experimentally. Healthy subjects develop measurable skin conductance responses, physical “hunches” 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.

The 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’s evidence suggests this is not a cosmetic gap. It is a structural one. Without emotional tagging, the reasoning system itself degrades.

Rosalind Picard saw this coming. In 1997, she published Affective Computing, arguing that “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.” 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’s attention has gone.

Ghosts, 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.

His metaphor was precise: “We’re not building animals. We’re summoning ghosts.”

In “The Space of Minds” (November 29, 2025), he laid out the optimization pressures that shaped each kind of intelligence side by side.

Animal 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 “huge amounts of compute” to EQ, theory of mind, and coalition dynamics. And for exploration, driven by curiosity and play.

LLM 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 “deeply craves an upvote from average user,” producing the sycophancy that plagues every chatbot).

His conclusion: “LLMs aren’t failed humans; they’re successful alien intelligences shaped by entirely different evolutionary pressures.”

This reframes the “missing half” 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.

This 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.

The 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.

The 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’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.

If Dunbar is right, social cognition is not a nice-to-have module you bolt onto “real” 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.

Combine 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.

Current benchmarks confirm the gap. GPT-4 achieves roughly 90% on simple Theory of Mind questions (“Where will Sally look for her marble?”) 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.

What Emotion AI Actually Measures (And What It Doesn’t)

The emotion AI market (companies like Affectiva, Realeyes, and the now-acqui-hired Hume AI) is built largely on Paul Ekman’s basic emotion theory: six or seven universal emotions, each with a distinct facial expression. Detect the facial action units, classify the emotion.

Lisa Feldman Barrett’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.

If 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 “furrowed brow” as “angry” 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.

Hume 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.

But even Hume’s approach operates at the level of signal detection, not comprehension. There are three levels to emotional intelligence: detecting what someone’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.

If 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.

Three Paths Forward

Not everyone agrees on what to build, but several thinkers have proposed frameworks that take the missing half seriously.

Karl Friston’s active inference is the most theoretically complete account of social cognition in AI. In “Designing Ecosystems of Intelligence from First Principles” (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 “Narrative as Active Inference” (2024), he showed how shared cultural narratives can be modeled as agents jointly minimizing free energy. His key insight: “Perhaps the most important determinants of our behaviour are beliefs about the intentions and behaviour of others.” 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.

Zhu Songchun’s cognitive architecture at PKU’s Beijing Institute for General Artificial Intelligence takes a different approach. His “small data, big task” 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 “climbing Mt. Everest when the real goal is to reach the moon”: impressive but in the wrong direction. BIGAI’s “Tongtong” 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.

Cognitive 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.

But none of these hybrid systems have incorporated emotional processing. The module that Damasio’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.

All 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.

Where 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.

Karpathy 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.

And 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.

The 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 “correctly” is itself contested, culturally situated, and dependent on context that no benchmark captures.

But 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’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.

The 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.

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：“从情绪到情感，最是人工智能未来前进的方向。“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。

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：“I naturally think about compassion and love. I think this is what defines us as human.” 然后她补了一句：“It’s not clear there is a mathematical path toward that.” 目前还不清楚是否存在一条通往那里的数学路径。

2025 年 11 月，她在 Substack 上发表了“From Words to Worlds”，称 LLM 是"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。“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。

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。

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。

安东尼奥·达马西奥的临床研究给出了不同的答案。

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。

结果不是更纯粹的推理，而是瘫痪。

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"这个选项比那个更重要"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。

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"预感”，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。

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。

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"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。“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。

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。

他的比喻很精确：“We’re not building animals. We’re summoning ghosts.” 我们不是在造动物，而是在召唤鬼魂。

在“The Space of Minds”（2025 年 11 月 29 日）中，他并排列出了塑造每种智能的优化压力。

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"大量算力"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。

LLM 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"深深渴望普通用户点赞"的实体，由此产生了困扰每个聊天机器人的谄媚问题）。

他的结论是：“LLMs aren’t failed humans; they’re successful alien intelligences shaped by entirely different evolutionary pressures.” 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。

这重新定义了"缺失的另一半"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。

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。

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。

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。

如果邓巴是对的，那么社会认知不是你装在"真正的"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。

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。

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（“Sally 会去哪里找她的弹珠？"）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。

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。

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。

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"皱眉"解读为"愤怒"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。

Hume 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。

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。

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。

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。

**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"从第一性原理设计智能生态系统”（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“叙事作为主动推理”（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：“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。“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。

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"小数据、大任务"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"攀登珠穆朗玛峰，而真正的目标是登月”：令人印象深刻但方向不对。BIGAI 的"通通” 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。

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。

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。

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。

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。

Karpathy 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。

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。

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"正确"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。

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。

AI 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。