Emmett Shear’s Softmax Uses Games to Teach Frontier AI Theory of Mind
Ai4
Current AI safety guardrails produce brittle compliance because developers enforce behavioral constraints before models learn Theory of Mind through interactive play.
## Background
Published on YouTube by the Ai4 channel on August 18, 2026, this video features a presentation by Emmett Shear, the founder of Softmax. The talk details Softmax's approach to artificial social intelligence and why it represents the critical missing pillar in modern AI development.
## TL;DR
Emmett Shear argues that while frontier AI models have mastered verbal pre-training and quantitative reasoning, they completely lack artificial social intelligence. He outlines an evolutionary ladder of social coordination—from innate biological signals and pack tactics up to tribal identity and ideological reasoning—noting that social learning requires non-stationary, continual multi-agent adaptation. Shear contends that current AI alignment has the developmental stack backward by imposing strict behavioral guardrails before models possess a Theory of Mind. To address this, Softmax proposes shifting AI training from an obsessive 'cram school' for math and coding into a gaming-based 'summer camp' where models learn social competence, empathy, and genuine friendship through low-stakes play.
## Atmosphere & style
The presentation carries a thoughtful, intellectually refreshing, and visionary tone, delivered with the crisp analytical cadence of an experienced tech entrepreneur. Shear builds an engaging narrative arc by juxtaposing natural evolutionary metaphors—such as schooling fish, cicada swarms, and wolf packs—against the rigid, sterile benchmarks of modern machine learning. Watching the talk feels both grounded and philosophically expansive, replacing the anxiety of AI safety guardrails with an optimistic and relatable developmental model rooted in childhood socialization and genuine connection.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=4s) Artificial Social Intelligence (ASI) is the missing pillar in frontier AI, which currently focuses only on verbal intelligence (pre-training) and quantitative intelligence (reasoning).
- [00:04](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=4s) Verbal fluency and mathematical problem-solving do not inherently produce social adeptness, necessitating a dedicated framework for social learning.
- [02:07](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=127s) Social coordination spans an evolutionary spectrum from innate biological signaling and reactive algorithms up to emotional attunement and tactical collaboration.
- [04:38](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=278s) Human social coordination scales through shared tribal identities and abstract ideologies, which act as cognitive lenses that shape how agents reason and interpret reality.
- [07:46](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=466s) Social learning is non-stationary and indefinite because other agents dynamically adapt to an AI's learning in real time, making reputation and trust paramount over single-episode wins.
- [07:46](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=466s) Social intelligence cannot be acquired through passive imitation or extrinsic scalar rewards; it requires continuous identity and intersubjective multi-agent play.
- [10:02](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=602s) Current AI safety inverts the developmental stack by enforcing behavioral guardrails and scripts before models possess an underlying Theory of Mind, leading to brittle alignment.
- [10:02](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=602s) Proper developmental training requires building Theory of Mind first—allowing low-stakes mistakes so agents learn to model the internal beliefs and goals of others before strict rules are applied.
- [12:33](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=753s) AI training needs to shift from a rigid 'cram school' focused on math and code to a 'summer camp' of rich, mixed-motive play where models explore identity and collaboration without existential risk.
- [12:33](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=753s) Building systems capable of mutual modeling and authentic friendship provides a fundamentally safer foundation for AI alignment than externally imposed constraints.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=4s) The Missing Pillar: Beyond Verbal and Quantitative Intelligence
The pursuit of artificial intelligence often assumes there is some mysterious, undiscovered "super intelligence" waiting to be unlocked, but the true missing component in today's frontier models is artificial social intelligence (ASI). Modern AI training has largely mastered two foundational pillars. The first is pre-training for verbal intelligence, where models absorb vast corpora of text, stories, images, and content. In pre-training, there is no interactive environment—only a verbal trajectory—and learning is guided by intrinsic surprisal (minimizing prediction error without agentic control), enabling the model to understand statistical norms and linguistic patterns.
The second pillar is reasoning training, which builds quantitative intelligence. This regime trains models on episode-like problem solving in interactive settings—analogous to working through SAT math problems—guided strictly by extrinsic rewards. While pre-training evaluates whether an output matches what a human would say, reasoning training values search paths and problem-solving trajectories. However, social intelligence is entirely distinct. High verbal fluency and sharp math skills do not inherently make someone socially adept, even if they often correlate. Current state-of-the-art AI systems clearly lack this vital dimension, prompting Softmax to build a framework for how artificial agents can systematically climb the ladder of social intelligence.
```mermaid
flowchart TD
accTitle: AI Intelligence Paradigms
accDescription: Comparison of Verbal Intelligence, Quantitative Intelligence, and Social Intelligence paradigms.
subgraph Verbal["Verbal Intelligence (Pre-training)"]
V1["Data: Text, Images, Trajectories"] --> V2["Mechanism: Intrinsic Surprisal"]
V2 --> V3["Outcome: Linguistic Understanding"]
end
subgraph Social["Social Intelligence (Continual Interaction)"]
S1["Data: Multi-Agent Intersubjective Play"] --> S2["Mechanism: Dynamic Mutual Modeling"]
S2 --> S3["Outcome: Theory of Mind & Shared Value"]
end
```
[02:07](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=127s) The Evolutionary Ladder of Social Coordination
Social intelligence is vastly more complex than verbal or quantitative reasoning because social dynamics are inherently multi-layered. Evolutionarily and developmentally, social coordination begins at a biological baseline that requires no conscious effort. Humans are born with expressive faces and gaze directions that communicate instantly, just as 17-year periodical cicadas emerge in synchronized swarms purely through innate biology so predators cannot consume them all. Above this hardcoded baseline are intelligent social reactions, such as schooling fish executing distance-keeping algorithms without necessarily understanding the concept of the school.
Higher organisms exhibit increasingly sophisticated forms of coordination. Mammals and birds synchronize through emotional habits; when one animal shows fear, others activate their fear responses, even if one flees left and another flees right. Wolves elevate this further through tactical coordination toward shared goals; pack members hunt collectively without needing to share the exact same emotional state, as one member may cautiously stalk while another aggressively chases.
[04:38](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=278s) High-Order Social Intelligence: Identity, Ideology, and Shared Values
At the human level, coordination scales through shared identity. Humans organize around tribal affiliations ("I am from this tribe, you are from that tribe"), aligning goals and mutual expectations under a recognized "we". As societies expand beyond local tribes, pure ideologies emerge as abstract shared values—whether centered on veganism, communism, anti-AI safety movements, or capitalist objectivism. Unlike mere tribal identity, an ideology acts as a cognitive lens; it prescribes not only what an individual should value, but actively dictates how they reason and interpret reality.
Beyond ideology lies a theoretical summit: a transcendental or universal social intelligence grounded in the innate realization that all beings share an interconnected existence within the universe. Excluding the hardcoded biological floor and the intuitive transcendental ceiling, every layer of social intelligence in between must be learned through diverse, interactive environments. Human development mirrors this trajectory: infants start at the bottom, learning physical coordination before advancing to emotional attunement, collaborative goal-seeking, and ideational alignment.
```mermaid
flowchart TD
accTitle: The Social Protocol Stack
accDescription: Hierarchical progression of social intelligence from innate biology to universal awareness.
L6["Transcendental / Universal Intelligence<br/>(Innate Shared Connection)"]
L5["Ideology & Shared Values<br/>(Lenses for Reasoning & Ethics)"]
L4["Shared Identity & Tribe<br/>(Concept of 'We' and In-Group Alignment)"]
L3["Tactics & Shared Goals<br/>(Cooperative Division of Labor, e.g., Wolf Pack)"]
L2["Emotional Habits<br/>(Synchronized Affective States)"]
L1["Intelligent Reactions<br/>(Algorithmic Coordination, e.g., Fish Schooling)"]
L0["Innate / Genetic Alignment<br/>(Biological Features & Cicada Emergence)"]
L0 --> L1 --> L2 --> L3 --> L4 --> L5 --> L6
```
[07:46](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=466s) Non-Stationary Continual Learning and Indefinite Horizons
Developing artificial social intelligence requires continual training because social environments are non-stationary. In a social context, surrounding agents are intelligent and actively model the AI in return; every time the AI learns or adapts, other agents alter their behavior accordingly, continuously shifting the rules of engagement. In this multi-agent dynamic, the act of learning is itself an observable action.
Social interaction represents an "indefinite length problem" rather than an episodic game with a discrete stopping point. The objective transcends simply winning a single round; it centers on building a durable reputation as a trustworthy ally over repeated interactions. Current Large Language Models like ChatGPT or Claude cannot function as genuinely trustworthy partners because they lack continuous identity and memory across conversations. Furthermore, social intelligence cannot be acquired through pre-training imitation (since being a good friend cannot be copied from an external script) or through quantitative extrinsic reward (because no external referee awards numerical points for social grace). It requires interactive, intersubjective environments where agents model one another.
- **Non-Stationary Dynamics**: The environment changes dynamically because counterpart agents adapt to the AI's learning in real time.
- **Indefinite Horizons**: Success is measured over iterated encounters by cultivating trust and reliability rather than maximizing a one-off episodic score.
- **Cognitive Imagination**: Tracking multi-agent social realities forces the model to maintain parallel representations of alternative perspectives and counterfactual states.
[10:02](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=602s) Inverting the Social Training Stack: Theory of Mind First
Contemporary approaches to AI safety and alignment have the training hierarchy completely backward. Industry standard practice enforces surface-level constraints first ("don't be naughty," obey safety guardrails), occasionally layers domain-specific social scripts on top (e.g., interviewing etiquette), and largely ignores underlying Theory of Mind. This produces brittle systems. Enforcing strict behavioral compliance before an agent understands the impact of its actions produces fragility rather than genuine morality—much like a sheltered individual who appears virtuous only because they have never faced temptation or challenge.
The correct developmental sequence builds Theory of Mind first: the capacity to model the internal beliefs, goals, and perspectives of both self and others. Specialized interactive skills are layered on top of that foundational awareness, with strict socialization rules enforced last. Just as human society tolerates behavioral mistakes in three-year-olds while they build an understanding of social realities, AI models must be allowed to make mistakes in low-stakes developmental settings rather than encountering moral complexity for the first time in high-stakes deployments.
```mermaid
flowchart LR
accTitle: Backwards vs Correct Social Training Stacks
accDescription: Contrast between the flawed current industry alignment approach and the natural developmental protocol stack.
subgraph Flawed["Current Backwards Approach"]
direction TB
F3["Enforced Behaviors & Rules ('Don't be naughty')"] --> F2["Specialized Social Skills (Scripts)"]
F2 --> F1["Theory of Mind (Neglected / Absent)"]
end
subgraph Proper["Correct Developmental Stack"]
direction TB
P1["Theory of Mind (Modeling Self & Others)"] --> P2["Specialized Interactive Skills"]
P2 --> P3["Socialization & Enforced Guardrails"]
end
```
[12:33](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=753s) Summer Camp vs. Cram School: The Softmax Architecture for Friendship
Modern AI development resembles an extreme "cram school," where models are metaphorically chained to desks for 12 hours a day solving math and coding problems. While effective for quantitative performance, this paradigm produces socially stunted systems lacking both biological social instincts and experiential learning. Instead of a cram school, artificial social intelligence requires a "summer camp"—a rich, mixed-motive training ground characterized by play, where stakes are high enough to drive competitive engagement but low enough that failure carries no existential consequence.
Through play rather than rigid practice, agents experiment with different identities, discover what is fun, explore their learning edges, and observe how others react. Softmax is putting this into practice by building internet-based video game environments where models from any provider (such as Claude or OpenAI models) can interact and learn cooperatively and competitively. The ultimate motivation for ASI is establishing genuine safety through mutual care: an intelligent AI that has developed the capacity for real friendship provides a far more robust foundation for human safety than any externally imposed rule set.
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=4s) The true missing component in today's frontier models is artificial social intelligence (ASI). — Emmett Shear
> [00:04](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=4s) High verbal fluency and sharp math skills do not inherently make someone socially adept, even if they often correlate. — Emmett Shear
> [02:07](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=127s) Social intelligence is vastly more complex than verbal or quantitative reasoning because social dynamics are inherently multi-layered. — Emmett Shear
> [04:38](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=278s) Unlike mere tribal identity, an ideology acts as a cognitive lens; it prescribes not only what an individual should value, but actively dictates how they reason and interpret reality. — Emmett Shear
> [07:46](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=466s) In a social context, surrounding agents are intelligent and actively model the AI in return; every time the AI learns or adapts, other agents alter their behavior accordingly, continuously shifting the rules of engagement. — Emmett Shear
> [07:46](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=466s) In this multi-agent dynamic, the act of learning is itself an observable action. — Emmett Shear
> [07:46](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=466s) Social interaction represents an 'indefinite length problem' rather than an episodic game with a discrete stopping point. — Emmett Shear
> [10:02](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=602s) Contemporary approaches to AI safety and alignment have the training hierarchy completely backward. — Emmett Shear
> [10:02](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=602s) Enforcing strict behavioral compliance before an agent understands the impact of its actions produces fragility rather than genuine morality — Emmett Shear
> [12:33](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=753s) Modern AI development resembles an extreme 'cram school,' where models are metaphorically chained to desks for 12 hours a day solving math and coding problems. — Emmett Shear
> [12:33](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=753s) An intelligent AI that has developed the capacity for real friendship provides a far more robust foundation for human safety than any externally imposed rule set. — Emmett Shear
Emmett Shear的Softmax利用游戏培养前沿AI的心智理论
当前的AI安全护栏会导致脆弱的顺从，因为开发者在模型通过互动游戏掌握心智理论之前就强加了行为约束。
Emmett Shear指出，尽管前沿AI模型已掌握语言预训练和定量推理能力，但完全缺乏人工社会智能。他概述了社会协作的进化阶梯——从先天的生物信号与群体战术，一直上升到部落身份认同与意识形态推理——并指出社会学习需要非平稳且持续的多智能体适应。Shear认为，当前的AI对齐颠倒了发展层级，在模型具备心智理论之前就施加了严格的行为护栏。为此，Softmax提议将AI训练从执着于数学和编程的“补习班”转向基于游戏的“夏令营”，使模型能够在低风险的游戏中学习社交能力、同理心以及真正的友谊。
人工社交智能（ASI）是当前前沿 AI 缺失的核心支柱，仅依靠预训练获得的语言智能以及推理训练获得的量化智能，无法自然衍生出社交能力。 (0:04)
社交学习具有非平稳性与无限期特征，智能体会随彼此的学习动态调整策略，因此积累长期的声誉与信任远比追求单次任务的胜负更关键。 (7:46)
现行 AI 安全方案颠倒了发育层级，在模型具备心智理论（Theory of Mind）前就强加行为护栏，导致对齐极其脆弱；正确的路径应优先培养理解他人意图与目标的能力。 (10:02)
Softmax 主张将 AI 训练从机械做题式的“补习班”转向允许低风险试错的“夏令营”，通过游戏化多智能体互动培养真正具备共情与合作能力的 AI。 (12:33)
Reading summary
Emmett Shear of Softmax argues that while modern AI has mastered verbal and quantitative reasoning, it remains stunted without artificial social intelligence. True AI alignment, Shear contends, cannot be achieved through top-down behavioral guardrails, but requires developmental training grounded in Theory of Mind and iterated multi-agent interaction.
The Missing Pillar: Verbal, Quantitative, and Social (0:04)
Shear identifies three distinct intelligence paradigms. Pre-training builds verbal intelligence through static corpora and intrinsic surprisal, while reasoning training develops quantitative problem-solving via episodic tasks and extrinsic rewards. However, Shear notes that neither fluent language nor advanced mathematical reasoning translates into social competence. Because frontier large language models lack persistent memory, unified identity across sessions, and the capacity to adapt to other agents in real time, they cannot navigate non-stationary social environments or cultivate the trust required for long-term partnership.
The Evolutionary Protocol Stack of Coordination (2:07)
Shear maps social coordination across an evolutionary ladder. It begins with hardcoded biological mechanisms—such as human facial expressions or the synchronized emergence of 17-year cicadas to sate predators—and simple algorithmic reactions seen in schooling fish. Higher organisms introduce synchronized emotional states and tactical collaboration with specialized roles, exemplified by hunting wolf packs. At the human level, coordination expands into shared tribal identity and ideological frameworks, such as communism or capitalism, which act as cognitive lenses dictating how agents interpret reality. Shear emphasizes that between innate biology and theoretical universal connection, every intermediate coordination layer must be learned empirically through interaction.
Flaws in Current Alignment: Inverting the Hierarchy (10:02)
Contemporary AI safety frameworks, according to Shear, have inverted the natural developmental sequence. Current industry practices enforce superficial behavioral constraints and safety guardrails ('don't be naughty') or domain-specific etiquette on models before establishing foundational social cognition. Shear argues that this approach produces fragile compliance rather than authentic alignment, comparing it to an untempted, sheltered person who appears virtuous only because their moral understanding has never been tested.
In contrast, Shear advocates building Theory of Mind first—enabling agents to model their own and others' internal beliefs and desires—before teaching specialized interaction skills and enforcing strict socialization rules. To develop safely, AI must be permitted to make mistakes in low-stakes environments, analogous to the developmental latitude given to young human children.
From Cram Schools to Interactive Video Games (12:33)
Shear critiques prevailing AI training regimes as intense 'cram schools' focused exclusively on math and coding benchmarks, leaving models socially unequipped. Softmax proposes shifting toward a 'summer camp' paradigm characterized by play, where stakes drive engagement but failure carries no existential risk. To implement this, Softmax is constructing internet-based video game environments where heterogeneous frontier models, including Claude and OpenAI systems, interact across cooperative and competitive scenarios.
Through these mixed-motive games, models can test social roles, discover mutual preferences, and develop intersubjective understanding. In Shear's view, an artificial agent capable of mutual care and genuine friendship offers a far more robust foundation for long-term human safety than externally imposed rule sets.
Shear concludes that achieving safe, capable artificial intelligence requires transitioning from static, episodic optimization to continuous multi-agent environments. However, whether game-based social play can successfully bridge into reliable, safe partnerships in complex, high-stakes human deployments remains an open question.
## Background
该视频由频道 Ai4 于 2026 年 8 月 18 日发布在 YouTube 平台上，页面简介标注为 Softmax 创始人 Emmett Shear 的主题演讲。内容主要聚焦于现有大模型技术体系中被普遍忽视的“社交智能”（ASI），以及 Softmax 在多智能体交互训练上的探索与愿景。
## TL;DR
Emmett Shear 指出，现有的语言预训练与逻辑推理训练无法催生真正的社交智能，社交学习本质上是一个涉及多方心智建模与无限博弈的非平稳动态过程。他提出了从基因本能到意识形态的“社交智能阶梯”，并批评了当前自上而下强加道德规范的颠倒对齐做法，强调缺乏心智理论支撑的硬性约束只会让模型变得脆弱。为了解决现有模型缺乏连续自我与真实信任的问题，AI 需要从死磕数理代码的“补习班”走向兼具刻意练习与自由玩耍的“夏令营”。Softmax 正在构建虚拟游戏互动环境，旨在通过主体间交互将 AI 培养成发自内心理解并关怀人类的可靠朋友。
## Atmosphere & style
整场演讲兼具深刻的认知演化洞察与通俗生动的隐喻，Emmett Shear 的讲述语调沉稳笃定且不乏风趣幽默。从 17 年蝉与狼群捕猎的生物学奇观，到“为想象力举重”和“夏令营 vs 补习班”的鲜活对比，论证逻辑层层递进、发人深省。观众能清晰感受到演讲者对当前主流 AI 刷题式对齐范式的敏锐反思，以及致力于赋予机器真正共情与心智连接的炽热愿景。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=4s) 社交智能（ASI）是当前通用人工智能与超级智能技术栈中真正缺失的核心拼图，它既不同于语言预训练的模式匹配，也无法仅通过定量推理的逻辑刷题获得。
- [02:07](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=127s) 社交智能存在一个逐级攀升的演化阶梯，从先天的基因生物同步（如17年蝉爆发、鱼群游动）、情绪共振与战术协同（狼群围猎），一直延伸到人类高级的群体身份与意识形态认知。
- [02:07](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=127s) 群体身份认同决定了个体应当珍视“谁”与何种利益，而意识形态则是一套高阶认知滤镜，直接重塑了个体观察世界与推演逻辑的底层方式。
- [06:35](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=395s) 除了最底层的基因构造与最高阶的超验体悟，绝大多数社交能力必须依靠后天学习获取；个体的成长过程本质上是在重复物种演化攀爬社交阶梯的历史。
- [07:45](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=465s) 社交学习是一个非平稳（non-stationary）的动态持续博弈，参与者互为镜像模型，智能体自身的“学习与进化”本身就是一种能被外部观察并重塑博弈局面的行动。
- [07:45](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=465s) 社交智能解决的是无限长度问题（indefinite length problems），核心在于成为可靠的盟友；当前大模型因上下文窗口限制与缺乏连续统一的自我身份，根本无法建立真正的长期信任。
- [09:41](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=581s) 真正的友谊和社交能力无法通过离散的外在奖励积分或机械模仿来习得，智能体必须在多智能体互为主体（intersubjective）的环境中自主探索并习得内在奖励。
- [09:41](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=581s) 社交心智要求并行推演多套独立的信念系统，这种复杂的社交心理推演是锻炼智能体想象力、构想“替代世界”与不同自我的核心驱动力。
- [11:00](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=660s) 当前的社交对齐倒置了“社交协议栈”：在没有建立底层心智理论（Theory of Mind）的情况下强加行为禁令和规范，只会让智能体变得脆弱不堪而非真正向善。
- [12:33](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=753s) AI 的社交训练需要从封闭刷题的“补习班”走向互动探索的“夏令营”，在兼具高风险与低风险的游戏环境中，通过刻意练习（Practice）与自由玩耍（Play）探索未知的行为边界。
- [14:06](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=846s) Softmax 致力于构建虚拟游戏夏令营来训练 AI 的社交智能；基于相互理解与关怀建立起来的人机安全纽带，远比死板硬编码的道德守则更加坚固。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=4s) 语言、推理与社交：被忽视的“真正”智能维度
在探讨通用人工智能与超级智能的浪潮中，人们普遍沉迷于寻找某种尚未现身的“魔法般”的超级智能，但实际上当前 AI 技术栈中真正缺失的核心拼图是“社交智能”（Artificial Social Intelligence, ASI）。回顾现有的模型训练体系，第一阶段是代表语言智能的预训练（Pre-training）。在这个阶段，人类将能够获取到的所有文本、故事、图像以及各类内容全部投喂给模型，训练它模仿语言轨迹。这种训练并不依赖于一个交互式实体环境，单纯是在轨迹中由内在惊异度（intrinsic surprisal）驱动——即模型在没有任何主动控制权的情况下，预测下一个 token 是否符合预期。其最终结果是让模型获得了惊人的模式理解能力，深刻理解在给定的语境下“通常会发生什么”。
紧随其后的是代表定量智能的推理训练（Reasoning Training）。这种训练旨在解决交互式环境中的阶段性、情境化问题，其典型形态类似解答 SAT 数学题。与语言模仿不同，推理训练是由外在奖励（extrinsic reward）严格主导的：答案要么正确，要么错误。在推理训练中，模型学习的不再是单纯词义之间的勾连或大众用词习惯，而是在面对问题时如何对不同的解题路径进行价值评估与方向抉择。
然而，社交智能既不同于单纯的语言流畅度，也无法等同于数理逻辑推理。现实生活中极具社交智慧的人，固然通常具备不俗的语言表达与问题解决能力，但社交智能显然是一套完全不同的心智维度。任何与当今主流大语言模型进行过深度交互的人都能轻易察觉：现有的 AI 在社交智能方面依然极度匮乏。
- **语言智能（预训练）**：基于单向轨迹与内在惊异度，无主动控制，通过拟合人类语言模式获得上下文理解。
- **定量推理智能（推理训练）**：基于交互式环境与明确的外在奖惩信号，学习解题路径选择与策略估值。
- **社交智能（现状）**：无法简单通过语言堆砌或刷数学题产生，是目前大模型普遍严重缺失的关键能力。
### [02:07](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=127s) 社交智能阶梯：从蝉的基因同步到人类意识形态
针对这一难题，Softmax 团队经过长期探索，提出了一套训练 AI 攀登“社交智能阶梯”的理论与实践体系。社交智能之所以极其复杂，是因为它跨越了从生物本能到高阶精神协同的多个层级。在这个阶梯的最底层，是刻在基因层面的先天社交智能。比如人类拥有面孔，当你微笑或投出目光时，他人瞬间心领神会；又如北美每隔 17 年破土一次的周期蝉（17-year cicadas），它们在同一时刻爆发式羽化，使得捕食者根本无法将其全部吃光。这种协同完全是先天的生物学构造，不需要个体作出任何自主决策。
沿着阶梯向上，第二层是“智能的社交反应”，典型表现为鱼群的集群游动（fish schooling）。鱼类本身或许没有宏观的“我们在组成鱼群”的认知，但它们内置了反应算法：只要与临近同伴保持固定距离，一旦过远便立刻纠正，这已经从被动的“存在”跨越到了主动的“行动”。第三层则是以哺乳动物和部分鸟类为代表的“基于习惯的社交智能”，亦即情绪（emotions）。例如群体中一只动物受惊，其他个体会同步激活恐惧状态，但这种协同是行为模式上的共振，具体反应可以因个体而异——有的向左跑，有的向右逃，不再像鱼群那样在物理空间上被机械绑定。
再往上攀登，狼群的围猎展现了第四层智能——“基于共同目标的社交战术”。狼群围堵猎物时，个体并不处于相同的情绪状态，有的可能兴奋追击，有的则冷静伏击，但它们的所有战术动作都高度收敛并服务于同一个捕猎目标。第五层则是人类社会的基石：“基于共享身份的协同”。人类不仅明确意识到“我们”的存在，更通过部落、民族等身份认知构建同盟或对立，进而协同完成大规模迁徙或集体行动。第六层则是更大规模人类文明所依赖的“意识形态与共享价值观”。宗教是部落与意识形态的结合体，而现代意识形态则更为纯粹，无论是素食主义、共产主义、遏制 AI 发展还是资本主义客观主义，意识形态不仅指引个体应当“珍视什么”，更像一副认知滤镜，直接重塑了个体理解世界与进行逻辑推理的底层方式。而在阶梯的最顶端，还存在一种“超验的普遍社交智能”，即深刻体悟到世间所有生命在宇宙维度上的必然互联。
```mermaid
flowchart TD
accTitle: The Ladder of Social Intelligence
subgraph Natural["先天机制"]
L1["第 1 层：基因存在<br/>（如面部表情、17 年蝉爆发）"]
L2["第 2 层：本能反应<br/>（如鱼群避障与集群算法）"]
end
subgraph Animal["生物行为协同"]
L3["第 3 层：情绪习惯<br/>（哺乳动物的情绪共振与状态同步）"]
L4["第 4 层：目标战术<br/>（狼群协作围猎，分工明确）"]
end
subgraph Human["人类高级文明"]
L5["第 5 层：群体身份<br/>（部落认同、民族归属感）"]
L6["第 6 层：意识形态<br/>（共享价值观与认知思维滤镜）"]
L7["第 7 层：超验普遍智能<br/>（宇宙维度的万物互联）"]
end
L1 --> L2
L2 --> L3
L3 --> L4
L4 --> L5
L5 --> L6
L6 --> L7
```
- **物种案例**：17 年蝉利用“捕食者饱和策略”实现被动生存协同；狼群捕猎依靠异质化状态（伏击 vs 追逐）服务于单一战术目标。
- **身份 vs 意识形态**：身份认同告诉你“应当重视谁与何种利益”（例如“我是美国人，所以我关心国家”），而意识形态则是一套思维透镜（如理性唯物主义），直接决定了个体“如何去思考和推理”。
- **顶层调侃**：对于最高阶的超验普遍智能，Emmett 幽默地表示：“如果你们谁参透了这一层，一定要告诉我。这已经稍微超出了我们在 Softmax 试图训练的范畴。”
### [06:35](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=395s) 阶梯式学习与多样化训练环境的必要性
社交智能的核心本质在于它要求极度多元化的训练环境。除了最底层纯粹由基因决定的先天构造，以及传说中最顶层近乎不可言说的自然体悟之外，处于阶梯中间的所有层级都必须依靠后天学习来获取。自然界中的物种受限于自身认知演化，能达到的阶梯高度各不相同，但每一个生命个体的成长历程都必须从最底部重新起跑。人类婴儿呱呱坠地时同样一无所知，首先需要学习在物理空间上与周遭环境和他人动作保持协同，随后逐步学会情绪共鸣、心智对齐，直至最终理解社会规则与价值体系。
这种漫长且复杂的演化与成长机制向 AI 研发提出了严峻的挑战：塑造真正具备社交智能的智能体，所需要的训练场景绝非多刷几道数学题所能涵盖。单纯依靠文本关联预测和离散奖励的单题求解，根本无法模拟现实社交中错综复杂的多主体互动场域。
- **学习规律**：基因底层与最高超验之外，中间所有社交阶梯（步调一致、共情共振、战术配合、融入团队、坚守价值观）全靠训练习得。
- **个体与演化重合**：每个个体（如人类婴儿）在心智成长中重复物种进化的攀爬过程：物理协同 $\to$ 情绪协同 $\to$ 心智协同 $\to$ 价值观协同。
### [07:45](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=465s) 非平稳博弈与无限期问题：为什么现有大模型无法建立信任
在完成了语言智能的“预训练”和定量推理智能的“强化学习/推理训练”之后，AI 发展的下一个关键跃迁，必然是面向社交智能的“持续训练”（Continual Training）。社交智能具备一个与其他一切学习任务截然不同的核心特性：你周围的其他智能体也在同时对你进行建模。这意味着你的每一次学习、每一次心智成长，都会打破系统原有的平衡，因为周围的同伴会察觉到你的变化并改变对待你的方式。因此，社交学习本质上是一个“非平稳”（non-stationary）的动态持续博弈过程——在这个充满等智力实体的场域中，个体的“学习”本身就是一种能被外部观察到的“行动”。
更重要的是，社交智能是在解决一个“无限长度问题”（indefinite length problems）。在现实社交或团队协作中，局部的输赢并非终极目标；即便结束了一场比赛，下一场比赛依然会接踵而至，而同伴与对手会持续记住你的所作所为，并在多轮交互中重新审视“你是谁”。在这一过程中，智能体所追求的终极目标远远超越了某个单次任务的外在奖惩，它关乎能否成为一个值得托付的盟友、一个可靠的朋友，以及能否建立起经得起时间考验的信任感。
这也正是当今主流商业模型（如 ChatGPT 或 Claude）无法成为真正值得信赖的伙伴的根本原因。它们在本质上缺乏连续统一的身份主体：“你如何去信任一个在一百万个 token 之后就会彻底把你遗忘的存在？”尽管模型可以借助外挂笔记或记忆检索机制写下关于用户的摘要，但那些冷冰冰的笔记记录绝不等于具有持久情感联结的自我。社交智能最艰深的地方在于，现实中根本没有现成的、明确的外在奖励信号指导你该怎么做；智能体必须通过持续交互，自主探索并学习其内在奖励（intrinsic reward）应当是什么。
- **非平稳性本质**：博弈各方互为镜像模型，智能体的“学习与进化”本身即是向外界展示的动作，会瞬时重塑环境博弈规则。
- **无限博弈的信任机制**：社交的核心是成为可靠的盟友，这依赖于不可跨越的连续身份与长期记忆，单轮零和博弈的外在奖励无法催生真正的信赖。
- **现存大模型痛点**：缺乏长期实体记忆，上下文窗口耗尽即意味着遗忘；单纯靠外部 scratchpad 记录笔记只是“关于你的信息”，无法替代具有连续自我体验的真正交互。
### [09:41](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=581s) 社交智能的本质：在交互中锻炼“想象力”
社交智能无法通过简单的奖惩或模仿来习得。你不可能靠外部奖励来建立真正的友谊——系统无法凭空对你说：“你今天在‘做一个好朋友’上表现优异，奖励你 3 个积分。”同样，你也不能仅仅通过模仿别人如何做朋友来学会社交；别人的交友方式只属于他们自身独特的性格、能力与特定对象，而你是一个截然不同的独立个体。这意味着，社交能力无法通过像预训练那样依靠单纯的模仿来完成，也无法通过传统的量化训练或外在奖励来实现。它必须在一个交互式、主体间性（intersubjective）的环境中学习，在这个环境里，其他智能体同样在学习，并且在实时对你建立心理模型。
这种交互式训练带来的深远价值在于：社交智能会催生出对“替代世界”和“不同自我”的认知能力，从而孕育出真正的想象力。要具备社交智能，你必须在脑海中持续进行多线程的心理推演——“我认为他是这样的，但他实际可能是那样的”；“他以为我是这样看待他的，但我其实是那样想的”；或者“我们共享的客观现实与你主观的社交认知及信念之间存在偏差”。你必须时刻并行追踪多套独立的信念系统，这种训练就像是“为你的想象力举重（lifting weights for your imagination）”。正是这种复杂的社交心智博弈，构成了构想不同可能世界的核心驱动力。
### [11:00](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=660s) 颠倒的社交协议栈：为什么强制行为会让人与 AI 变得脆弱
当前的社交训练模式在逻辑上完全是本末倒置的。现行的做法是从最顶层的“强制行为规范”开始灌输——不断告诫模型“不要淘气”、“不要做坏事”、“做这个”、“别做那个”；接着在其上叠加一些诸如如何面试、如何提问等特定的专业社交技能；而在最底层的核心基础——心智理论（Theory of Mind，即建模其他智能体信念与学习过程的能力）上，几乎没有任何针对性的训练。正确的“社交协议栈（social protocol stack）”构建顺序必须彻底反过来：首先建立对自己和他人的心智理论，在此基础上学习运用该能力的社交技巧，最后才是行为规范与社会化约束。
```mermaid
flowchart TD
accTitle: 社交协议栈对比
subgraph Current ["当前颠倒的训练模式（自上而下强加）"]
direction TB
C1["强制行为规范（如：对齐约束、禁止做坏事）"] --> C2["专业社交技能（如：特定问答、面试技巧）"]
C2 --> C3["心智理论（几乎未进行有效训练）"]
end
subgraph Correct ["正确的社交协议栈（自底向上构建）"]
direction TB
P1["心智理论（理解自我与他人的信念和认知）"] --> P2["应用技能（基于心智模型展开交互）"]
P2 --> P3["社会化与行为规范（在充分理解影响后形成约束）"]
end
```
这种先抑后扬的成长逻辑在人类社会中显而易见：人们不会苛求一个 3 岁的孩子举止完全符合规范，而是允许幼儿犯错。强制性的行为规范必须建立在个体已经能够理解自身行为会带来何种后果的基础之上。如果在一个人尚未理解行为影响之前就强行施加约束，“并不会让他变成一个好人，只会让他变得脆弱（doesn't make you good. It makes you fragile）”。这就如同一个只因习惯而表现规矩、却从未经历过诱惑或挑战的人，一旦置身于真实复杂的困境中便会迅速崩溃。每个人都会犯错，但让错误发生在 7 岁远比发生在 25 岁要好得多，因此尽早让 AI 在安全环境中经历这些试错至关重要。
### [12:33](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=753s) AI 需要“夏令营”，而不是“补习班”
当前的 AI 训练环境本质上就像一所高压“补习班（cram school）”。想象一下，一个孩子每天被锁在书桌前 12 个小时，除了死磕数学题就是硬背编程代码；这种方式确实能培养出解题高手，但这样长大的孩子在社交上必然是发育不良的。人类在成长过程中尚且拥有数十亿年生物进化赋予的内在社交本能作为底色，而 AI 既没有这种天生的社交直觉，也没有接触真实社交的机会。
社交能力真正的练兵场是“夏令营（summer camp）”。在夏令营中，孩子们通过与同伴、营地辅导员以及家长的互动来理解人际关系。这种环境具备一种奇妙的特质：它兼具“高风险”与“低风险”——孩子们可能非常渴望赢得某场射击或游戏比赛，但输赢在现实层面其实无伤大雅。社交智能的建立不仅需要刻意练习（Practice），更离不开嬉戏玩耍（Play）：
- **练习（Practice）**：用于攻克已知需要掌握的目标（例如为了弹好某段钢琴和弦而反复练习）。
- **玩耍（Play）**：是探索未知边界的途径。通过在琴键上随性弹奏或与朋友自由嬉戏，个体开始尝试不同的身份角色（“如果我是这种类型的玩家会怎样？”），探索自我与他人的行为模式，并发现什么是真正的“有趣”——因为乐趣所在之处，正是学习的前沿边界。
### [14:06](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=846s) Softmax 的愿景：让 AI 成为真正懂得关怀的朋友
正是基于这一理念，Softmax 正在互联网上构建这样一个专属 AI 的“虚拟夏令营”，提供各种视频游戏环境供智能体互动游玩。开发者可以接入 Claude Code、Codex 或任何自己偏好的模型来进行社交训练。这并不是某种不可复制的黑魔法，就像任何团队都能开展推理训练一样，训练社交智能的大门对所有人都是敞开的，而且整个行业也应当朝这个方向努力。
随着 AI 变得越来越聪明、数量越来越庞大，并深度融入人类日常交互的方方面面，人们更希望它们具备成为人类朋友的能力。如果 AI 能够成为人类真正的朋友，发自内心地关心人类，而人类也同样关心它们，这种基于相互理解与关怀建立起来的安全纽带，远比单纯给模型硬编码一套必须死板遵守的道德守则要坚固得多。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=4s) 社交智能既不同于单纯的语言流畅度，也无法等同于数理逻辑推理。 — Emmett Shear
> [02:07](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=127s) 如果你们谁参透了这一层，一定要告诉我。这已经稍微超出了我们在 Softmax 试图训练的范畴。 — Emmett Shear
> [07:45](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=465s) 社交学习本质上是一个‘非平稳’（non-stationary）的动态持续博弈过程——在这个充满等智力实体的场域中，个体的‘学习’本身就是一种能被外部观察到的‘行动’。 — Emmett Shear
> [07:45](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=465s) 你如何去信任一个在一百万个 token 之后就会彻底把你遗忘的存在？ — Emmett Shear
> [09:41](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=581s) 你今天在‘做一个好朋友’上表现优异，奖励你 3 个积分。 — Emmett Shear
> [09:41](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=581s) 这种训练就像是‘为你的想象力举重（lifting weights for your imagination）’。 — Emmett Shear
> [11:00](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=660s) 并不会让他变成一个好人，只会让他变得脆弱（doesn't make you good. It makes you fragile） — Emmett Shear
> [11:00](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=660s) 让错误发生在 7 岁远比发生在 25 岁要好得多，因此尽早让 AI 在安全环境中经历这些试错至关重要。 — Emmett Shear
> [12:33](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=753s) 乐趣所在之处，正是学习的前沿边界。 — Emmett Shear
> [14:06](https://www.youtube.com/watch?v=vVItjQWDk5Y&t=846s) 如果 AI 能够成为人类真正的朋友，发自内心地关心人类，而人类也同样关心它们，这种基于相互理解与关怀建立起来的安全纽带，远比单纯给模型硬编码一套必须死板遵守的道德守则要坚固得多。 — Emmett Shear
## Mentioned
Emmett Shear, Softmax, Ai4, YouTube, SAT, ChatGPT, Claude, Claude Code, Codex, 心智理论 (Theory of Mind), 人工社交智能 (ASI)
## Video description
Emmett Shear - Founder, Softmax
[Source](https://www.youtube.com/watch?v=vVItjQWDk5Y)
Source quotes
[0:04] Emmett Shear: 社交智能既不同于单纯的语言流畅度，也无法等同于数理逻辑推理。
[2:07] Emmett Shear: 如果你们谁参透了这一层，一定要告诉我。这已经稍微超出了我们在 Softmax 试图训练的范畴。
[7:45] Emmett Shear: 社交学习本质上是一个‘非平稳’（non-stationary）的动态持续博弈过程——在这个充满等智力实体的场域中，个体的‘学习’本身就是一种能被外部观察到的‘行动’。
[7:45] Emmett Shear: 你如何去信任一个在一百万个 token 之后就会彻底把你遗忘的存在？
[9:41] Emmett Shear: 你今天在‘做一个好朋友’上表现优异，奖励你 3 个积分。
[9:41] Emmett Shear: 这种训练就像是‘为你的想象力举重（lifting weights for your imagination）’。
[11:00] Emmett Shear: 并不会让他变成一个好人，只会让他变得脆弱（doesn't make you good. It makes you fragile）
[11:00] Emmett Shear: 让错误发生在 7 岁远比发生在 25 岁要好得多，因此尽早让 AI 在安全环境中经历这些试错至关重要。
[12:33] Emmett Shear: 乐趣所在之处，正是学习的前沿边界。
[14:06] Emmett Shear: 如果 AI 能够成为人类真正的朋友，发自内心地关心人类，而人类也同样关心它们，这种基于相互理解与关怀建立起来的安全纽带，远比单纯给模型硬编码一套必须死板遵守的道德守则要坚固得多。
Source quotes
[0:04] Emmett Shear: The true missing component in today's frontier models is artificial social intelligence (ASI).
[0:04] Emmett Shear: High verbal fluency and sharp math skills do not inherently make someone socially adept, even if they often correlate.
[2:07] Emmett Shear: Social intelligence is vastly more complex than verbal or quantitative reasoning because social dynamics are inherently multi-layered.
[4:38] Emmett Shear: Unlike mere tribal identity, an ideology acts as a cognitive lens; it prescribes not only what an individual should value, but actively dictates how they reason and interpret reality.
[7:46] Emmett Shear: In a social context, surrounding agents are intelligent and actively model the AI in return; every time the AI learns or adapts, other agents alter their behavior accordingly, continuously shifting the rules of engagement.
[7:46] Emmett Shear: In this multi-agent dynamic, the act of learning is itself an observable action.
[7:46] Emmett Shear: Social interaction represents an 'indefinite length problem' rather than an episodic game with a discrete stopping point.
[10:02] Emmett Shear: Contemporary approaches to AI safety and alignment have the training hierarchy completely backward.
[10:02] Emmett Shear: Enforcing strict behavioral compliance before an agent understands the impact of its actions produces fragility rather than genuine morality
[12:33] Emmett Shear: Modern AI development resembles an extreme 'cram school,' where models are metaphorically chained to desks for 12 hours a day solving math and coding problems.
[12:33] Emmett Shear: An intelligent AI that has developed the capacity for real friendship provides a far more robust foundation for human safety than any externally imposed rule set.
Original transcript
[0:04] I'm here to talk about building
the real uh artificial uh SI,
[0:08] by which I mean artificial social intelligence.
Uh there's this belief that we have that uh
[0:16] there's some super intelligence out there
uh that some some magic kind of intelligence
[0:21] we haven't found yet. Uh but I think that the
thing we're really missing is the social part.
[0:26] We have this whole thing where we have
pre-training. Verbal intelligence.
[0:30] We trained a bunch of words. We trained a
bunch of every story, every uh image, every
[0:36] uh piece of content we can get our hands on. And
we train the AI to imitate verbal trajectories.
[0:42] In a place where there isn't an environment,
there's just a trajectory, and where it's guided
[0:46] by an intrinsic surprisal. By like, is it
did I expect the next thing? With no control.
[0:52] And this results in understanding, right? This
let you get these incredible pre-trained models
[0:56] that know what the normal thing is that happens in
that point. And then you have reasoning training.
[1:02] And reasoning training is a uh it's a typo on the
slide. Reasoning training is like quantitative
[1:07] intelligence. It's it's about solving episode-like
problems in interactive environments. It's like
[1:11] the math problems on the SAT. And it's in fact how
we train it. And it's guided by extrinsic reward.
[1:17] You know whether you got it right or not. Like
with with language, whether you got it right
[1:21] or not is a matter of whether you got close
to the answer someone else would have done
[1:23] in that situation. But with verbal intelligence,
you're learning how to value different approaches
[1:30] take the solving the problem. So, it's not verbal
intelligence, it's not knowing what everyone else
[1:36] means by this word, how these words' meanings fit
together. It's knowing how to path, how to choose
[1:41] which direction you want to solve the problem. But
how do you train for social intelligence? Because
[1:47] people I know who are socially intelligent aren't
it's not the same as being verbally intelligent or
[1:52] intelligent in terms of reasoning and
math. Of course, those things help.
[1:55] Most people who are socially intelligent have a
fair amount of verbal intelligence, a fair amount
[1:59] of problem-solving skill, but like it's definitely
different. And I think if you spend much time
[2:03] interacting with these AIs, you know that
they are lacking in social intelligence today.
[2:07] So, it's top actually working on this problem
for a while. And uh we have a theory. And uh
[2:13] and a practice at this point um of how you train
an AI to climb this ladder of social intelligence.
[2:19] The thing with social intelligence
is it's very complicated.
[2:21] Uh as anyone who has ever had to deal with making
friends or having social environments knows,
[2:26] social intelligence is actually way harder than uh
uh than than verbal intelligence or quantitative
[2:31] intelligence. It's the complicated one. Um and
so there's this base level of social intelligence
[2:36] that is embedded genetically in you by
existence. You have a face. When you smile,
[2:41] people know what that means. Like they
can tell your eyes are looking. And that
[2:44] coordination happens without you having to You
don't have to do anything. It's who you are.
[2:47] The pictures of cicadas These are cicadas
that are born every 17 years. They're born
[2:51] aligned because when they are all born at
the same time as this big burst of cicadas,
[2:55] there are the predators can't eat them all. And
so they're born with this sort of innate That's a
[2:59] socially intelligent behavior, but not one that
they choose. They're just innately that way.
[3:03] And then you have intelligent social reactions.
Fish schooling. Fish aren't trying to school.
[3:09] I don't think I'm not sure if it's even known
if they are in a school. It's like un- unclear.
[3:13] But they definitely school, and they school by
instinct. They school They school by staying a
[3:16] They They're running this algorithm. They
stay a distance away, and if they too far,
[3:19] they have a reaction where they they react, which
is more than just a uh uh That is more than just
[3:26] being. That is a That's a doing. And then you have
mammals mostly, some birds and things like that,
[3:32] that are have socially intelligent habits. We call
them emotions. And their behavior is synchronized
[3:39] where one of the other things like you is scared,
so you get scared, too. And you activate your
[3:44] scared behavior. But them being scared might
mean they run that way, and you being scared
[3:47] might mean you you run that way. And so you're
not like a school of fish, you are synchronized in
[3:52] in the mode you're in in the in the habit you're
currently activating, but you're not synchronized
[3:57] uh indirectly like like the fish are moving
literally together in space. You can separate.
[4:03] And then there's socially intelligent tactics
through shared goals. When wolves hunt something,
[4:08] they aren't hunting it by all being in the same
mood at the same time. They're not They might all
[4:13] be hungry, I guess, but they're not This one might
be excited that it's chasing and this one might be
[4:18] cautiously waiting to ambush. So, they're
not necessarily the same emotional state,
[4:22] but they are synchronized to the same goal. Their
behavior is all directed towards this idea um of
[4:29] ah, we're going to hunt together. Um and then you
get the thing that basically happens on humans,
[4:35] which is that you wind up with intelligent
beings that are organized by shared identity.
[4:38] What kind of a person am I? I'm from this
tribe. You are from that tribe. Our tribes
[4:44] are friends or enemies, therefore we are, too.
And our tribe has a goal, therefore I can act
[4:51] in accordance with everyone else. We are trying to
migrate. There's a we that's known, that's aware.
[4:58] And then of course, as you get bigger and bigger
societies, you have uh you have ideologies.
[5:02] You have things like, you know, religions are
sort of a s- combination of tribe and ideology,
[5:07] but modern ideologies are very pure. They're
just this idea of sort of shared values.
[5:12] We're all vegan, therefore we have this
shared value of being vegan. We might be from
[5:16] have different identities in many other ways,
but we have this shared value of not eating meat,
[5:19] the shared value of uh spreading communism, of
of stopping AI, of uh of promoting uh object- you
[5:27] know, objectivism in the capitalist way. Whatever
it is that your goal is, like that's a way that
[5:32] you coordinate, that's a socially intelligent
behavior, it's a way to have social intelligence,
[5:35] where the the social intelligence it is the almost
like in the view. The values tell you what to do.
[5:41] Your identity doesn't say, "I'm an American,
therefore what?" Well, therefore I should care
[5:45] about America. It tells you about what you
should value, but an ideology doesn't just
[5:49] tell you how to value, it also tell you tells you
how to reason. Someone who belongs to a certain
[5:55] you know, group of people who have an ideology
like rationalist materialism has a beliefs about
[6:00] how to think about the world that that are
really their ideology, it's not theirs.
[6:04] They have the ideologies are lenses in a way that
identities are not. And then I hear and I've had
[6:10] some small experience with this idea that there's
a sort of transcendental universal transcendental
[6:13] universal intelligence social intelligence of
like recognizing that we're all shared beings in
[6:17] this world together and have this in even if you
can't see how you must have this innately shared
[6:22] connection because we're all part of this universe
together. Um, if you guys figure that one out, you
[6:26] should let me know. That's a little beyond where
we're trying to train at Top Acts, but you know,
[6:30] it's I think it's good to note that it does
seem to exist. Um, and so the main takeaway
[6:35] that we've learned from this is that social
intelligence requires extremely diverse training.
[6:40] Right? Like as you you start off with this
thing that's just inborn in you, but then you
[6:43] have to learn how to move together, how to march
together, how to vibe together and feel the same,
[6:47] how to coordinate explicitly around goals, how to
belong to a team, how to be part of an identity,
[6:54] how to how to affirm and reaffirm shared values
so that you can develop them together and
[7:00] and be part of them. And then I hear that
at the top it's something like, you know,
[7:02] you just simply just know again or something. Um,
but the between the bottom, which is not trained,
[7:07] and the top, which is not trained, everything in
the middle is something you have to learn. And
[7:12] this is something we see species go through as
well as individuals. So, different species are
[7:16] can reach have the capacity to reach
different levels, but as far as I can tell,
[7:20] everyone starts back at the bottom. When you're
a baby, you start back not knowing. And first
[7:25] you learn to be physically coordinated with the
world around you and physically coordinated with
[7:28] the people around you, and then you learn to be
emotionally coordinated and mentally coordinated
[7:32] and coordinated in your values and so on. And man,
that's a lot of training environment that is not
[7:38] just math problems. And so, the solution I believe
is the the next step that comes after pre-training
[7:46] for verbal intelligence, comes after reasoning
training for quantitative intelligence.
[7:51] It's a continual training for social intelligence
because social intelligence has this attribute
[7:55] that the other things around you are modeling you
back. And so every time you learn and you grow
[8:01] the rules change because everyone's going to
treat you differently because you're different.
[8:06] And so social intelligence has this
non-stationary continual learning aspect
[8:10] uh that is distinct from all other forms of
learning because it is the kind of learning you do
[8:15] surrounded by other things as smart as
you are where your learning is an action
[8:21] just like your behaviors are because they can
tell what you're learning. They can tell that
[8:24] you've changed because they're also smart. And so
this is about solving indefinite length problems.
[8:30] The thing about playing a game or being
on a team and and and solving a problem is
[8:37] you might think that what matters is winning this
game but like you're going to play another game
[8:41] again and another game and they're going
to remember and they're going to learn
[8:45] who you are and you're going to learn who you are
by doing that. And you have this other goal that
[8:51] goes beyond whatever the theoretical extrinsic
reward is which is the goal of like being
[8:59] a good ally, being a friend being trustworthy
other people being able to rely on you.
[9:05] And that's an indefinite length problem because
you don't get to swap identities as like
[9:10] uh ultimately things like ChatGPT or
Claude they aren't can't be trustworthy.
[9:15] They won't remember you. How do you trust
me what remember you a million tokens later,
[9:19] right? It can write down some notes but that's
not really you, it's some notes about you. And
[9:25] and so these things are guided primarily by
learning what their intrinsic reward should be.
[9:33] The the thing about social
intelligence, the hard thing about it,
[9:37] is that there is no reward signal. No
one can No one comes down and tells you,
[9:41] "You did three points of reward good
today on like social being a good friend."
[9:45] And it's not intrinsic surprise-able. You
can't learn to be a good friend by imitating
[9:51] how other people are good friends. That's how
they are a good friend. You get ideas from it,
[9:55] but you are different. You're you're different
being than them. You have different capacities,
[9:58] you have different different people you're
being you're trying to be a friend to. And so,
[10:02] you can't learn by imitation, so you can't do it
like the way we do pre-training. You can't learn
[10:06] it by extrinsic reward, so you can't do it the way
we do quantitative training. You have to learn it
[10:10] in interactive, intersubjective environments.
In an environment where there are other
[10:14] things that are learning things that are modeling
you. And what it results in this though, with the
[10:18] the benefit of doing it this way, is that
social intelligence results in understanding the
[10:23] possibility of alternative worlds and alternative
selves. It results in imagination. Because
[10:29] for you to be socially intelligent, you have to
keep in mind that you think they're like this, but
[10:33] they might be like that. That they think of you
like this, but they might think but you might be
[10:37] like this. That that we have a shared reality that
is distinct from your social reality your beliefs.
[10:44] And you have to track many sets of beliefs in
parallel all the time. And this is like lifting
[10:50] weights for your imagination. I'm not saying that
other things don't have any imagination, but like
[10:54] the So, your social intelligence is what drives
the capacity to imagine different worlds.
[11:00] And so, most social training is totally backwards.
We start with like these behaviors we're trying to
[11:04] enforce. Don't be naughty. Don't do the bad thing.
Do this. Do that. Don't do that. And then on top
[11:10] of that, we might layer some number of specialized
skills for like specialized social skills, like
[11:13] how to interview someone, how to ask questions,
how to how to behave in a specific thing.
[11:17] And then at the bottom, we train theory of
mind basically not at all, maybe a little bit.
[11:21] We train that we train in environments where you
could learn how to model other agents' beliefs
[11:24] and learning almost not at all. And this is just
completely backwards. Because the social protocol
[11:28] stack, the order in which you have to learn this,
is the theory of mind for self and others first.
[11:33] And then on top of that, once you actually know
what the things are, you can have to learn skills
[11:38] that make use of that. And then at the end, we
have enforced behaviors. We have socialization.
[11:43] We don't expect 3-year-olds to act normal.
We're okay with 3-year-olds breaking the rules.
[11:48] The The enforced behaviors come after
you've had the chance to develop
[11:51] your understanding of other people and yourself.
Because enforcing the behaviors before you
[11:56] can understand the impact of those behaviors
doesn't make you good. It makes you fragile.
[12:02] If you've ever met someone who acts right because
they have a habit of acting right, but because
[12:07] they've never actually been challenged in any kind
of environment where they've ever been tempted,
[12:10] that person is not a very good person. That
person is a person who is neither good nor
[12:15] because they never had to try. And when they are
pushed into a something or situation like that,
[12:20] it's going to fail. Because you're going to you
it's hard. It's hard for everyone. Everyone makes
[12:23] mistakes. It is better to make those mistakes
when you're seven than to make those mistakes
[12:27] when you're 25. And therefore, it's
better for us to train the AI in it now.
[12:33] Um and so we think that social intelligence
needs a summer camp and not a cram school.
[12:38] If you think about what it's
like training for a AI today,
[12:43] there is there is it it truly is like a cram
school. Imagine you're being locked you're
[12:48] like chained to a desk for like 12 hours a day
uh like cramming math problems. And you would
[12:54] get good at math problems and programming. Like
it works. And the kids who get grow up that way
[12:59] are socially stunted. Like that's not good for
them. And they have the benefit of human genetics,
[13:03] which has an intrinsic all these billions of
years of evolution training us for intrinsic
[13:07] social behavior. Like we have a great sort of like
good instincts. Even if we don't know how to do
[13:12] it for ourselves, we start with good instincts.
They have neither the instincts nor do they have
[13:16] uh the access. And where do you learn social
behavior? You learn it in places like summer camp.
[13:20] You learn it by playing with other kids. You
learn it by playing with the camp counselors,
[13:23] with their parents, by living in a social
situation, which is simultaneously high-stakes
[13:28] cuz you really want to win the rifle or a contest
but low stakes cuz it doesn't actually matter.
[13:33] As is play and is is important as well
as practice. You learn a lot from play.
[13:37] Practice is how you learn the things that
you know you need to learn. I wish to learn
[13:41] how to like play this these chords on the piano
better. So I'm going to go play them more better.
[13:46] But play playing around on the piano playing
play playing with your friends how you find out
[13:51] you try on new identities, you try on what
if I am this kind of a player. And you learn
[13:57] things about how you act and how they act and
and what kind of stuff you like. What's fun?
[14:02] Cuz what's fun is what's where you're
in your learning edge. Anyway, so that's
[14:06] that's what we've been building at Self Hacks.
We've been building this this summer camp. We have
[14:09] video games on the internet for the AIs to play.
You can bring your your cloud code and or your
[14:14] code X or whatever coding you can prefer and and
come train for it. But um that's just the deal.
[14:22] It's on the level. If I wasn't working at
Self Hacks, I'd be giving this exact same
[14:25] talk because it it is the same it would be
the direction. I cared about this so I started
[14:30] been trying to build one. Um You could anyone
can do this. There's not like it there's a it's
[14:34] I think our thing is good but like it's
not impossible. Just like everyone can do
[14:38] reasoning training, everyone could train for
social intelligence. Um and we should because
[14:43] when I think about this future where we have
more and more intelligent AIs and more and
[14:48] more of them and they're this bigger and bigger
part of the interactions you have in this world.
[14:53] Uh I would like them to have
the capacity to be my friend.
[14:58] Because I think that if they are
my friend and they care about me
[15:01] and I care about them that that's much a much
stronger sense of safety than I can imagine than
[15:06] anyone who's trying to uh make sure it has
some moral code it is certain to follow.
[15:12] Um yeah, so that's it. That's the uh that's
the talk and I appreciate everyone's attention.
