Stanford Spinout Simile Simulates Human Behavior with 85% Accuracy

斯坦福衍生公司 Simile 以 85% 准确率模拟人类行为 前沿模型在社会模拟方面表现不佳，因为它们是针对高度理性而非人类的非理性进行训练的；要建模复杂的经济与社会动态，需要能够捕捉主观价值观和行为差距的专用架构。 Joon Sung Park 探讨了从其里程碑式的斯坦福“Smallville”研究到创立 Simile 的历程，Simile 是一家专注于模拟人类行为的应用 AI 实验室。尽管当前的前沿模型扮演着极度理性的“智能 CPU”，但 Simile 致力于构建“智能 GPU”来捕捉人类的非理性、价值观以及二阶社会影响。通过将深度行为数据与智能体架构相结合，Simile 相较于人类重测基准达到了 85% 的准确率，支持从企业概念测试到复杂宏观经济现象建模的广泛应用。 基于斯坦福“Smallville”虚拟小镇实验中 25 个生成式智能体涌现自主社交行为的研究，Joon Sung Park 创立了应用 AI 实验室 Simile，专注于实现大规模人类行为与社会模拟。 (0:00) Park 将擅长客观确定性问题的前沿模型比作“智能的 CPU”，而 Simile 旨在打造编码人类多样价值观与主观偏好的“智能的 GPU”，并在 1,000 名美国代表性样本测试中达到了人类自身重测一致性 85% 的行为预测准确率。 (10:00) 为解决传统大模型因训练数据偏向线上态度而导致的“言行差距”，Simile 引入社会科学随机对照试验（RCT）与深度生命叙事数据，并与 Gallup 合作校准模型，目前已获 CVS 等企业用于概念测试及财报电话会模拟。 (15:00) Park 将社会模拟比作观测人类社会的“哈勃望远镜”，终极目标是构建“人类社会的 CERN”，未来通过高算力模拟探索银行挤兑、气候合作乃至宏观政策等重大复杂社会议题。 (35:00)

Joon Sung Park argues that while frontier AI labs train hyper-rational models for objective problem-solving, simulating human society requires architectures tailored to subjective human behavior, diverse values, and the gap between what people say and do. The Say-Do Gap and Architectural Divergence (7:12) Park asserts that mainstream foundation models represent a rational "CPU of intelligence" built for technical problems with singular correct answers. Because human behavior is inherently irrational and driven by personal preferences, scaling hyper-rational frontier models plateaus when predicting human social dynamics. Furthermore, conventional large language models trained on web text primarily capture attitudinal speech rather than actual actions, creating a significant "say-do gap." To bridge this divide, Park's applied AI lab, Simile, positions its platform as a distributed "GPU of intelligence" designed to reflect diverse subpopulations. Simile gathers deep narrative life stories using voice and text interviews conducted by reinforcement-learning-optimized agents, alongside discrete choice surveys. The architecture pairs these personal histories with a foundation of social science randomized controlled trials (RCTs) and pricing studies to reliably translate attitudes into behavioral predictions. Validation Benchmarks and Enterprise Use Cases (10:00) Simile evolved from Park's academic research at Stanford, which included the 2022 "Social Simulacra" paper and the 2023 "Smallville" study of 25 generative agents. Before commercializing alongside co-founders Percy Liang and Michael Bernstein, the team simulated 1,000 individuals representative of the U.S. population. The benchmark demonstrated that their agent architecture predicted human behavior 85% as accurately as human subjects replicate their own answers over time. Enterprise deployment began with clients like CVS, which partnered with Simile to bypass the temporal and financial bottlenecks of traditional surveying. While corporate clients often start with rapid concept testing—such as testing 1,000 ideas across 1,000 subpopulations—they also deploy multi-agent earnings calls and evaluate second-order market impacts, such as how an electric vehicle rollout influences perceptions of an automaker's broader portfolio. To ensure outputs remain anchored, Simile partners with Gallup to ingest ongoing real-world human panel data. Convergence, Divergence, and Simulation Evaluation (25:00) Addressing compounding errors in multi-agent networks, Park divides scenarios into convergent and divergent simulations. Convergent simulations, such as modeling human communication networks, naturally align with structural patterns like scale-free networks, meaning agent-level micro-errors do not derail overall macro outcomes. Park notes that Simile measures population-level accuracy using Total Variance Distance (TVD), viewing a TVD score below 0.15 as strong evidence for organizational decision-making. Conversely, divergent simulations involve path-dependent environments where small variations yield radically different futures, such as election outcomes. In these probabilistic settings, researchers run simulations repeatedly—such as 100 iterations—and apply bootstrap resampling to establish statistical confidence intervals. Park compares the current state of agent simulation to the formative era of inferential statistics, where foundational thresholds and methodological standards must still be established. Macroeconomic Modeling and High-Compute Simulations Beyond commercial product testing, Park and host Sonya Huang discuss applying generative agents to macroeconomics and central banking, areas where human psychological factors dominate. Park contrasts rich generative agent models with classical agent-based models, such as Thomas Schelling's segregation studies that relied on simple threshold rules for binary dots. Economists have approached Simile to examine macro phenomena, including bank runs, climate cooperation, democratic decline, and the emergence of monetary systems. Looking long-term, Park envisions high-compute simulations that could cost $100 million and run for months to resolve complex civilizational questions. Citing Percy Liang's observation that scientific leaps rely on novel measurement tools, Park frames large-scale social simulation as an observational instrument akin to the Hubble Space Telescope, aiming to establish a "CERN of human society" for empirical social science. While Simile establishes mathematical bounds like TVD and bootstrap resampling for probabilistic runs, formalizing the exact mathematical rules governing when multi-agent simulations converge or diverge remains an active, unresolved research frontier.

Joon Sung Park 及其联合创立的 Simile 提出，当前前沿人工智能实验室训练的超理性模型难以捕捉人类主观的非理性行为；该团队主张通过结合深度访谈、社会科学随机对照试验与真实人口数据构建“智能 GPU”，在大规模仿真中预测群体行为并评估决策的二阶效应。 范式转变：从理性“CPU”到映射人性的“GPU” (7:12) Park 的研究始于斯坦福大学的“Smallville”虚拟小镇实验，该实验利用具备记忆、规划与反思机制的 25 个生成式智能体，观察到了自主筹备情人节聚会等自发涌现的社交动态。然而，Park 指出目前主流大模型研发存在核心错位：OpenAI、Anthropic 等机构致力于打造面向通用人工智能（AGI）和超智能的“超理性机器”，这类模型擅长解决具有客观标准答案的技术问题，但在模拟人类社会时遇到瓶颈。 由于真实人类受主观价值观、偏好与情感驱动，单纯放大追求理性的模型规模无法提升对人类行为的拟合度。Park 将当前大模型比作追求单一超强能力的“智能 CPU”，而 Simile 则旨在构建“智能 GPU”——这种架构不以超越人类智力为目标，而是通过分布式的子单元，精准表征不同细分人群的多样性、品味与非理性特征。 弥合言行鸿沟：结合长尾生活史与基准校准 (15:00) 针对直接通过提示词让大模型扮演特定人口画像的做法，Park 强调通用大模型存在严重的“言行鸿沟”（say-do gap）：由于主要基于网络文本训练，模型只能反映人们“说了什么”（态度），无法真实映射人们“做了什么”（实际行动）。为构建态度与行为之间的转换层，Simile 采用经强化学习优化的 AI 访谈员，通过约 15 分钟的对话提取受访者成长背景、艰难决策等长尾生活故事，并结合离散选择调查获取系统性数据。 在底层数据支撑上，Simile 将大量社会科学随机对照试验（RCT）与定价研究作为行为基础模型进行训练，并与 Gallup 合作引入真实民调面板，同时协助 CVS 等企业客户结合内部行为数据定制专属智能体。在模拟 1,000 名具有美国人口代表性样本的基准验证中，该架构预测行为的准确率达到了人类自身重测基线的 85%。 企业落地演化：从概念测试扩展至二阶效应推演 (20:00) 在商业应用中，企业客户通常从具体的产品概念测试切入，尝试在数千个细分人群中即时验证上千种创意方案。随后，应用场景逐步扩展至高维动态交互，例如让智能体连续使用数字化产品 10 分钟以反馈体验，或在正式财报发布前模拟包含管理层、分析师与投资者的多智能体财报电话会议。 Park 说明，相比在社交平台上投放大量测试广告，仿真不仅能消除点击人群所固有的样本偏差，提供真正具备人口统计代表性的受众，更能推演复杂决策的长远连锁反应。例如，汽车厂商可以通过模拟评估推出一款热销电动车如何反向冲击消费者对其燃油车产品线的认知，从而在方案落地前评估整体产品组合的二阶影响。 系统动力学与长期愿景：迈向“社会科学的 CERN” 在评估模型可靠性时，Simile 采用全变差距离（TVD）量化模拟与真实分布的差异，并将 TVD 低于 0.15 视为决策参考门槛。针对多智能体交互是否会引发误差滚雪球式放大的问题，Park 将系统划分为两类：收敛型仿真（微观扰动会被宏观结构吸引力吸收，自然形成无标度网络等稳定形态）与发散型仿真（如选举等高度依赖路径的场景）。对于发散场景，可通过 100 次重复运行与自举重采样（bootstrap resampling）建立统计置信区间，将分歧本身作为风险分析的依据。 在更广阔的研究维度上，Park 认为生成式智能体已大幅超越托马斯·谢林（Thomas Schelling）时代简单的二元阈值模型。他预测未来可能出现单次耗资 1 亿美元、计算周期达数月的大规模社会仿真，犹如观测宇宙的哈勃望远镜或物理学界的 CERN 一样，成为量化分析宏观经济波动、银行挤兑机制及重大公共政策长期后果的基础科研仪器。 Simile 尝试将社会科学转化为可计算、可实证的研究工具，但多智能体仿真在面对高度发散、路径依赖的复杂系统时，其数学收敛边界与行业评估标准仍处于早期探索阶段，在宏观政策及复杂社会议题上的长远有效性仍需依赖高算力试验与更广泛的现实校准。

## Background 该内容由红杉资本（Sequoia Capital）于 2026 年 6 月 16 日在播客及网络平台发布，由 Sonya Huang 主持。节目聚焦 Simile 创始人兼 CEO、斯坦福生成代理小镇（Smallville）开创者 Joon Sung Park，探讨当今主流模型作为超理性“智能 CPU”的局限，以及 Simile 如何打造编码人类价值观与品味多样性的“智能 GPU”，进而实现对真实人类社会的大规模精准模拟。 ## TL;DR 访谈梳理了 Joon 从斯坦福 Smallville 生成代理实验走向创立 Simile 的历程，指出纯理性大模型在互联网态度数据训练下无法弥合人类“言行脱节”的核心瓶颈。Simile 结合长尾人生叙事与高质量真实行为数据训练行为基座模型，使人群模拟预测准确率达 85%，已被企业客户广泛用于并行概念测试与财报电话会预演。Joon 进一步解析了以全变差距离（TVD）量化模拟精度以及收敛与发散模拟的机理，并描绘了构建“人类社会 CERN”的宏大蓝图，期望借助模拟技术解开银行挤兑、集体行动困境与制度演化等社会科学深层命题。 ## Atmosphere & style 对话洋溢着前沿人工智能与经典社会科学交相辉映的智性张力，既有严密的技术推导，又蕴含深沉的硬核科幻情怀。Sonya Huang 的提问敏锐直接，紧扣工程瓶颈与商业本质；Joon 则以“智能的 CPU 与 GPU”等精妙隐喻深入浅出地拆解认知架构与统计规律。整场交流在务实的产业落地探讨与宏大的文明级科学构想之间层层推进，让听众深刻体会到一种宛如借“哈勃望远镜”窥探人性幽微与复杂社会的震撼视野。 ## Bulletin - `10:00` 斯坦福 Smallville 生成智能体之父 Joon Sung Park 创办 Simile，通过对 1,000 名美国人构建数字孪生，在行为预测准确率上达到了人类自身复现答案的 85%。 - `20:00` Simile 将主流前沿大模型定义为擅长客观逻辑的“智能单元 CPU”，并将自身定位为捕捉多元价值观、偏好与品味的“智能单元 GPU”。 - `15:18` 针对人类普遍存在的“言行脱节（Say-Do Gap）”，Simile 联合盖洛普采集真实人类长尾叙事，并利用强化学习训练 AI 访谈官，打造专注于行为预测的基座模型。 - `34:40` 商业落地已从 CVS 等财富 500 强企业的概念测试和财报电话会预演，拓展至探索推演银行挤兑、气候合作等宏观难题的“人类社会 CERN 对撞机”。 ## Key takeaways - `00:00` 成熟高等文明的科幻构想往往依赖两大支柱：形态成熟的 AGI 与指引并辅助复杂社会运行的大规模模拟系统（Simulations）。 - `00:44` 斯坦福小镇（Smallville）通过记忆流、长远规划与反思总结三大认知架构，成功证明大模型结合 Agent 架构能自发涌现真实社交动力学与群体事件。 - `04:00` 社会计算在系统面对海量用户时难以预测复杂的群体涌现效应，Social Simulacra 证明利用虚拟多智能体沙盘推演是破解实地测试高昂代价的核心路径。 - `07:22` 主流大模型以“超理性机器”为北极星，但在模拟和预测充满偏见、情绪波动与非理性特质的真实人类行为时已出现明显分歧与停滞瓶颈。 - `10:00` 学术研究本质是“广度优先搜索”的发散探索，而创办公司则是调动全部资源翻越单一愿景高山的“深度优先搜索”机器。 - `15:18` 互联网语料充斥修饰过的“态度数据”，导致纯提示工程无法跨越人类的“言行脱节”，唯有真实的长尾人生叙事才能构建态度向具象行为转化的翻译层。 - `17:15` Simile 正在利用 RCT 实验数据库训练人类行为基座模型，并运用强化学习训练 AI 访谈系统，以最小时间成本最大化挖掘受访者的深层决策图谱。 - `20:00` 系统架构采取分工协同路线：前沿大模型充当“智能单元 CPU”负责编排研究方案，Simile 自研专有模型作为“GPU”负责细分子单元的人类价值观表达。 - `22:15` 企业客户需求迅速从单点概念测试演进到引入时间维度的产品体验测试，甚至利用多智能体网络推演财报电话会议中投资人与市场的连锁反应。 - `23:45` 真实线上广告测试存在样本偏差且无法 A/B 测试多米诺骨牌式的长尾后果，多智能体模拟能够以具备统计代表性的人群推演复杂的二阶与长期效应。 - `26:40` 群体行为预测采用全变差距离（TVD < 0.15）作为商业决策的北极星指标，并严格区分具备内在聚集收敛拉力与初始扰动导致分岔的发散型模拟。 - `30:00` 针对发散型模拟，利用 Bootstrap 重采样等方法量化置信区间，当前的模拟科学正处于类似于推断统计学早期确立统计显著性规范（如 p < 0.05）的拓荒阶段。 - `32:15` 托马斯·谢林曾凭借极简规则的多主体模型解释隔离现象并获诺奖，生成式智能体赋予个体记忆与心理机制，大幅提升了宏观经济与社会实验的精度上限。 - `34:40` 模拟科学的终极愿景是打造类似 CERN 的大型科研基础设施，通过单次耗资巨大、耗时数月的终极模拟，解开银行挤兑、民主韧性与气候合作等文明级谜题。 - `36:20` 正如哈勃太空望远镜彻底重塑人类对宇宙的观测，生成式社会模拟有望成为观测人性与社会幽微机制的测量仪器，推动社会科学从经验主义走向精确科学。 ## Detailed walkthrough ### `00:00` 科幻双支柱：AGI 与社会模拟 Joon Sung Park 坦言自己深受科幻小说的启发。在那些展现科技发展进入高度成熟阶段的文明作品中，往往能看到两大不可或缺的技术支柱：一个是某种形态的通用人工智能（AGI），另一个则是真正能够引导、辅助人类社会运转的**大规模仿真系统（Simulations）**。在 Joon 看来，五年前或许没人敢断言时机成熟，但随着近年来在底层研究上的不断深耕，如今终于迎来了首次真正动手构建此类社会模拟系统的历史契机。红杉资本的 Sonya Huang 在开场中介绍，Joon 现为 Simile 的创始人兼 CEO，该团队正致力于打造一个聚焦于模拟人类行为与社会形态的应用 AI 实验室。 - 科幻文明高度成熟的两大标志：AGI 与指引社会运行的大规模模拟系统。 - Simile 的核心定位：构建一个专注于人类行为与复杂社会模拟的前沿应用 AI 实验室。 --- ### `00:44` 斯坦福小镇（Smallville）的诞生与情人节派对 时间回溯到 2023 年 4 月，在加州斯坦福大学，Joon 主导了一项轰动一时的研究——斯坦福生成代理小镇（Smallville）。这项研究源于一个关键观察：大型语言模型在吸收了互联网海量语料与社交媒体数据后，已经隐式编码了极其丰富的人类行为模式。研究团队发现，只要切入的角度和 Prompt 设计得当，就能从模型中诱导激发出大量细致入微的微观行为（例如：在极其具体的场景描述下，特定的人会做何反应）。这正是学界长久以来构建复杂自主代理所缺失的核心拼图。 在 Smallville 实验中，研究团队将生成式 AI 模型与三大认知架构结合——**记忆流（Memory）**、**长远规划（Planning）**和**反思总结（Reflection）**，打造出具备鲜活生活经历的“生成代理”（Generative Agents）。小镇里生活着 25 位拥有不同人设（Persona）的虚拟居民，他们早上起床、执行日常例行日程、出门上班、建立并维系人际关系，甚至自发涌现出各种复杂的社会群体活动。 ```mermaid flowchart TD accTitle: Generative Agent Cognitive Architecture accDescr: The cognitive loop of a Smallville agent showing perception, memory, reflection, planning, and execution. subgraph Perception_Storage["感知与认知存储"] A["环境感知<br/>(观察周边、他人行为与对话)"] --> B["记忆流 Memory Stream<br/>(时序经验与微观记录)"] end subgraph Reasoning_Engine["反思与规划决策"] B --> C["反思机制 Reflection<br/>(提炼高层见解与认知)"] B --> D["规划机制 Planning<br/>(生成日/周日程与动态调整)"] C --> D end subgraph Action_Execution["行动执行"] D --> E["小镇行为执行<br/>(开店、对话、邀约聚会)"] E -.->|"产生新事件"| A end ``` 最令团队震撼的是实验中完全自然涌现的情人节派对事件： - 模拟设定的起点是情人节前一天。咖啡店老板 Isabella 产生了一个自主想法：“如果能办一场情人节派对，邀请朋友和顾客过来就太棒了。” - 随后，Isabella 在前一天四处奔走采购派对物资，并向进店的每一位顾客发出邀请。 - 到了情人节当天，小镇居民竟然自发聚集到咖啡馆，形成了一场热热闹闹的派对。 - 派对中还涌现出了戏剧性的真实细节：有的人虽收到了邀请却因为日常琐事彻底遗忘；居民 Klaus 则在接到邀请后，借此机会鼓起勇气约上了自己暗恋的女孩，双双出席了派对。整个社交动力学的演进逼真得近乎超现实。 --- ### `04:00` 从社会计算到社交拟象（Social Simulacra） 谈及做 Smallville 的研究初衷，Joon 回忆起自己 2020 年刚加入斯坦福攻读博士时的经历。当时正值 GPT-3 问世前夕，Joon 与联合创始人 Percy Liang（现斯坦福基础模型研究中心主任）等学者共同撰写了奠基性长文《基础模型的机遇与风险》（Opportunities and Risks of Foundation Models）。当同行们都在惊叹于大模型做分类任务或简单文本生成的零样本泛化能力时，出身人机交互（HCI）与**社会计算（Social Computing）**背景的 Joon 看到的却是另一面：模型已经内化了人类行为的底层模式。 在人机交互与社会计算领域，最艰难的挑战并非设计前端 UI/UX，而是在系统面对成千上万乃至数十亿用户时，如何预测并治理人群聚集带来的复杂涌现效应（无论好坏）。以往唯一的检验方式只有直接将原型推上线做实地测试（Field Testing）。这种方法代价惨重——不仅消耗巨额的人力和时间，一旦产品设计存在结构缺陷（例如某种推荐 Feed 流会放大恶意或极端情绪），其负面后果将直接由现实社会承担。 为此，团队在 2022 年（即生成代理论文发表前一年）发布了前序研究 **Social Simulacra（社交拟象）**： - 针对 Reddit 社区版主与系统设计者的痛点，用上千个由大模型驱动的角色（当时称之为 Personas）去模拟一整个 Subreddit 论坛。 - 设计者可以预先设定版规、治理策略与社区目标，让虚拟居民在其中自由发帖互动，从而沙盘推演社区形态。 - 在一个模拟“匹兹堡旅游攻略”的社区测试中，虚拟用户不仅自发交流小众景点，甚至在线协作规划行程，邀请彼此在线下组团出行。这让团队坚信：模拟不仅可行，更是洞察宏观社会交互的杀手锏。 --- ### `07:22` 模型演进的瓶颈：超理性的“CPU”与非理性的人类 随着底层模型的发展，社会模拟的技术底座经历了剧烈跃迁： 1. **GPT-3 时代（Social Simulacra 时期）**：模型非常生硬粗糙（Janky），缺乏指令微调（Instruction Tuning），根本不听指挥。研究者必须依赖极其晦涩的提示词技巧才能勉强引导其生成符合人设的内容，但即便如此，人类行为数据的火种已经显露。 2. **早期指令微调模型时代（Smallville 时期）**：模型具备了严格遵从系统指令的能力，使得构建能够调用记忆、执行多步推理并做深层规划的复杂 Agent 架构成为可能。 3. **前沿基座模型时代（当下）**：模型的综合能力已经足以支撑真正严肃的模拟系统落地应用。 然而，Joon 指出了当前 AI 研发范式中一个极其尖锐且根本性的矛盾： > “当今的主流大模型公司——无论是 OpenAI、Anthropic 还是新兴的前沿实验室——它们的北极星目标都是制造某种**超级智能机器（Superintelligent Machines）**。这些机器被刻意训练得极度理性，在解决具有唯一客观标准答案的技术与逻辑难题上表现出超越人类的能力。” 但这恰恰偏离了人类社会的真实本质。**人类本质上是非理性的，充满主观价值观、偏见、情绪波动、多变喜好与个人品味。** Joon 观察到，随着前沿模型的参数规模不断攀升、推理逻辑越来越“超级理性”，模型在预测与模拟真实人类行为方面的能力反而出现了**分歧与停滞（Plateaued）**。依靠现有的纯理性模型范式，无法逼真再现复杂的现实社会生态。行业正亟需一场新的前沿突破——打造能够容纳非理性、多维度价值观的“拟人化”新型模型体系。 ### `10:00` 从学术研究到创立 Simile：广度优先与深度优先的转变 Joon Sung Park 回顾了自己从斯坦福大学的学术探索走向商业创业的心路历程。他坦言，早期在从事“社会拟象”（Social Simulacra）和“生成智能体”（Generative Agents / Smallville）研究时，就深受其应用前景的启发。然而随着时间推移，他深刻体会到学术研究与开办公司有着截然不同的功能逻辑：研究本质上是一种“广度优先搜索”（Breadth-First Search），身处顶尖实验室的学者们各自认领一个小命题进行发散探索，许多想法会开花结果并产出令人惊艳的学术成果，但科研人员通常不擅长“把工作彻底做完”，也很少亲自把研究成果的现实影响力落到产业界；相反，公司则是一台纯粹的“深度优先搜索”（Depth-First Search）机器——当团队对某一方向形成笃定信念并找到一座想要翻越的高山时，公司体制能让你调动全部资源和顶尖人才，毫无迟疑地奔向单一愿景。 这种坚定信念大约是在《Generative Agents》论文发表半年后形成的。当时，团队收到了海量的主动咨询：起初是渴望在虚拟平台上运行随机对照试验（RCT）的社会科学家们；紧接着，许多世界 500 强企业的 CEO、董事会成员在访问斯坦福时看到了 Smallville 的演示，纷纷跑来询问：“我们每年在线下花费巨资做问卷调查和市场测试，但市场中仍有太多无法解答的未知数，我们能否直接在模拟环境中运行这些决策？”这让 Joon 意识到学术研究通往现实世界影响力的清晰路径。为了验证模拟的有效性，团队对 1,000 名具代表性的美国人进行了数字孪生建模，并做出了令人振奋的定量验证：该模型与架构预测人类行为的准确率，达到了人们自我复现答案一致性的 85%（"85% as accurately as people replicate their own"）。这一成果让 Joon 坚信该技术已经足够成熟，可以作为商业决策的模拟平台交付给客户。随后，Joon 与自己的两位斯坦福博士导师——Percy Liang 和 Michael Bernstein 坐到了一起。师生三人当时已紧密合作了五年（如今已达六年），共同开启了将研究转化为 Simile 公司的创业历程。 - **研究与公司的范式差异**：研究是“广度优先搜索”（发散、探索、但不负责彻底收尾落地）；公司是“深度优先搜索”（聚焦、调动资源、不遗余力翻越特定高山）。 - **产业界痛点驱动**：世界 500 强高管面临大量传统调研无法覆盖的市场盲区，强烈需要沙盘模拟能力。 - **关键验证数据**：对 1,000 名美国人口进行模拟，预测人类行为的准确率达到个体自身复现答案的 85%。 - **创始团队渊源**：Joon Sung Park 与其斯坦福导师 Percy Liang、Michael Bernstein 拥有长达五六年的深入学术合作背景。 --- ### `12:15` 客户落地范式：CVS Health 与 Gallup 的真实数据驱动模拟 在谈及 Simile 目前如何与标杆客户合作时，Joon 以全美最大医疗健康企业之一的 CVS Health 为例，梳理了端到端的交付流程。Simile 在 CVS 的核心买手是负责“人类洞察”（Human Insights）的高级副总裁 Sri。Sri 在阅读了 Joon 验证智能体模拟精度的学术论文后，深感必须将这项技术引入 CVS。因为传统市场测试严重受限于能够实地发放的问题数量，更受限于“人类社会的物理法则”——在真实世界中做问卷或小规模测试是一回事，但要模拟整个市场动态、推演政策与产品决策给管理层带来的二阶甚至多阶连锁反应，是完全不同的维度。恰巧 Sri 的表亲与 Joon 相识，便促成了双方的牵线合作。 实际交付中，Simile 的工作流与企业客户习惯对接的传统民调或样本库公司（Polling/Panel Companies）表面上非常契合：企业客户提出他们希望深度理解的目标人群画像（如特定健康状态或消费习惯的群体）。此时，Simile 会依托战略合作伙伴（如知名民调机构盖洛普 Gallup）接触真实人类，采集高效且具备强泛化性的真实底层数据。Joon 强调，这绝非凭空捏造，而是将数字孪生牢牢锚定在真实数据之上：“设想你只有 15 分钟时间，你能向这些真实的人提出哪些最具魔力的问题？”Simile 采集到这批高价值数据后，将其注入系统生成对应人群的 Agent 模拟副本，这些智能体能够回答远超原始提问领域的各类复杂问题。最终，这些智能体被部署到 Simile 的 SaaS 平台中，企业客户登录后即可针对感兴趣的受众群体进行无上限的自由提问与策略推演。 ```mermaid flowchart TD accTitle: Simile 与企业客户交付全流程 accDescr: 展示从客户需求输入、真实数据采集、基座模型智能体构建到 SaaS 交互推演的完整交付路径。 subgraph Input["1. 客户需求输入"] A["企业客户 (如 CVS Health)<br/>提出目标人群画像与研究议题"] end subgraph Data["2. 真实人类数据采集"] B["战略民调伙伴 (如 Gallup)<br/>触达真实人类样本"] C["15 分钟核心人生叙事与行为采集<br/>长尾人生故事 + 关键抉择节点"] B --> C end subgraph Modeling["3. 智能体构建与对齐"] D["Simile 行为基座模型<br/>连接态度与行为的翻译层"] E["高保真人群数字孪生 (Agents)<br/>具备超越原始领域的泛化预测力"] C --> D D --> E end subgraph Output["4. SaaS 决策推演平台"] F["SaaS 交互平台<br/>无上限沙盘推演 / 二阶连锁效应评估"] E --> F end A --> B ``` - **典型买手画像**：企业内部负责消费者洞察（Customer/Human Insights）的 VP 或 SVP 级别高管。 - **传统调研瓶颈**：问卷容量有限，无法推演复杂的社会二阶效应（Second-order impacts）。 - **外部生态合作**：与 Gallup 等专业民调机构达成战略合作，确保底层 Agent 的构建基于高质量的真实人类样本。 - **产品形态**：以 SaaS 云平台交付，企业可直接与生成的人群智能体群组进行交互式概念测试与策略推演。 --- ### `15:18` 破解“言行脱节”：为什么不能直接用大模型 Prompt 模拟人类？ 主持人 Sonya 提出疑问：这听起来很像自动驾驶汽车（AV）的开发逻辑——先在真实道路采集海量数据，再通过模拟仿真扩充 corner case；但如果是模拟人类，为什么不能直接在 Claude 或 GPT 这样的超大语言模型里输入一段 Prompt（例如“你是一位生活在东西海岸大都市的 34 岁女性”），而必须大费周章找 Gallup 去采集现实世界的数据？Joon 解释说，自动驾驶的类比非常贴切，两者都要求仿真系统植根于真实世界的客观规律，同时模型必须具备超出训练数据之外的泛化能力（能在不同天气与路况下稳定运行）。但纯大模型在模拟人类时存在一个致命缺陷：模型根本无法处理人类普遍存在的“言行脱节”（Say-Do Gap）。 Joon 直言：“人们嘴上说的，和他们实际去做的，往往是两码事，这之间的鸿沟是真实存在的。”（"There are things that people say, and then there are things that people actually do. And the gap there is real."）现有的大型语言模型主要是基于互联网上的文本语料训练出来的，本质上充斥着人们表达观点和立场的“态度数据”（Attitudinal Data）。而真实世界的模拟必须弥合这种态度与实际行为之间的断层。因此，Simile 采集的核心数据在本质上是“行为数据”（Behavioral Data）。更反直觉的是，Simile 往往会向受访者抛出极其朴素的问题：“请跟我讲讲你的人生故事”（"Just tell me the story of your life."）。Joon 指出，这类人生叙事包含了每个人身上的“长尾信息”（Long-tail Information）——它无关受访者此时此刻的具体动作，也不是泛泛询问宏观政治立场，而是关于你成长于何处、人生中曾做过哪些艰难抉择。这种深度的长尾叙事正是构建“态度向行为转化”翻译层（Translational layer）的最佳黏合剂。 - **自动驾驶类比的共性**：均需遵循现实物理/社会法则，必须具备强泛化力而非简单过拟合。 - **大模型模拟人类的核心缺陷**：互联网文本多为“态度数据”，充满理想化或修饰过的观点，无法反映真实的执行力与行动决策。 - **“言行脱节”（Say-Do Gap）**：人类在被问及意图与真正付诸行动时存在普遍偏差，单纯 Prompt 大模型只会复现陈词滥调。 - **长尾人生叙事的作用**：通过家庭背景、逆境选择与成长经历，构建出连接受访者抽象观念与具象行为模式的核心翻译层。 --- ### `17:15` 人类行为基座模型与强化学习 AI 访谈系统 针对 Simile 所沉淀的数据类型，Joon 进一步透露，团队正在训练一个字面意义上的“人类行为基座模型”（Foundation Model of Human Behavior）。Simile 拥有庞大的随机对照试验（RCT）数据库，涵盖了社会科学实验以及大量的定价策略研究（Pricing Studies）。团队的核心目标，就是将所有这些来自 RCT 的纯净行为信号直接编码进模型之中，最终打造出一个能够精确预判各类未知 RCT 实验结果的 AI 模型。此外，客户在见识到模拟能力后，往往主动提出更激进的想法：例如 CVS 自身拥有多达 9,000 万名客户的交易与交互数据，如何以负责任且合乎伦理的方式将企业自有的私域行为数据与 Simile 的基座模型相融合？这构成了针对特定客户群体的定制化、增强版微调模型。 在访谈形式与模态方面，Simile 结合了文字问卷与语音访谈。Joon 强调，深入访谈是捕获人类长尾信息的无价手段。在团队 2024 年发表的奠基性研究中，他们正是通过询问“讲讲你的人生故事”来提取高维度画像。而在工程实现上，Simile 正在利用强化学习（RL）闭环训练专有的 AI 访谈官。该访谈系统的优化目标函数非常清晰：如何在消耗受访者最短时间的前提下，最大化获取关于这个人的深层能见度与行为特征？这套 AI 访谈官不会去纠缠死板的事实清单，也不会机械地询问用户对某特定软件功能的体验反馈，而是像一位极富洞察力的倾听者一样，高效引导受访者袒露最核心的人生脉络与决策图谱。 - **人类行为基座模型**：以海量社会科学与定价研究的 RCT 数据为底座，旨在直接预测人类面对新场景时的实验行为。 - **私域行为数据融合**：探索与企业第一方数据资产（如 CVS 的 9,000 万客户群）安全合规结合，微调定制高精度行业垂类智能体。 - **强化学习 AI 访谈官**：通过 RL 优化“时间成本 vs. 信息可见度”的目标函数，以最高信息增益挖掘真实个体的深层人生长尾叙事。 ### `20:00` 自研模型与“智能单元的 GPU”：捕捉人类价值观的多样性 在探讨数据收集方式时，Joon Sung Park 解释说，事实性或离散选择题（如问卷调查）具有极高的时间和数据效率，受访者能在短时间内回答大量问题。Simile 会充分利用这类数据，以宏观掌握大众在特定议题或政策上的态度。针对 Simile 被定位为“应用 AI 实验室”（Applied AI Lab）这一角色，Sonya Huang 提问：团队究竟在哪些层面选择自主构建模型，又在哪些层面依赖现有的前沿模型？ Joon Sung Park 给出了一个生动的技术隐喻：当前的前沿大模型类似于“智能单元的 CPU”（CPU of intelligence unit）——它们是单一模型，在高度理性的数据上训练而成，极擅长解决具有确定标准答案的复杂客观问题；而 Simile 正在打造的模型更接近“智能单元的 GPU”（GPU of intelligence unit）。在 Simile 的体系中，模型的目标不是达到“超人类”（superhuman）的智力，而是尽可能地“像人”（as human as possible）。核心假设在于：必须存在一种专门的模型来编码人类价值观、偏好和品味的多样性，而这是纯粹的理性模型所无法做到的。 ```mermaid flowchart TD accTitle: Simile 与前沿基础模型的分工架构 accDescr: 对比前沿基础模型的 CPU 角色与 Simile 模型的 GPU 角色及其分工协作流程 subgraph FrontierLLM ["前沿大模型 (智能单元 CPU)"] F1["超强理性与逻辑推理能力"] F2["制定并协调研究规划 (Research Plan)"] end subgraph SimileCore ["Simile 专有模型 (智能单元 GPU)"] S1["编码人类价值观、偏好与品味"] S2["模拟不同亚人群的细分真实视角"] S3["高度并行化的人类行为子单元"] end F2 -->|"编排协调"| SimileCore ``` - **架构分工**：Simile 并不是全盘抛弃前沿模型。在研究协调方面，前沿大模型是制定研究方案（research plan）的绝佳工具，因此 Simile 会使用它们负责编排协调；而在各个子单元需要表达不同亚人群真实视角的核心领域，团队则专注于开发自己的专有模型。 --- ### `22:15` 客户应用场景演进：从单点概念测试到多智能体财报电话会模拟 当 Sonya 问及客户是否带着新产品发布、营销方式、定价等问题找上门时，Joon 确认客户的需求涵盖了所有这些维度。然而，典型的客户旅程往往始于极具体、最直接的需求——概念测试（concept testing）。企业通常带着新想法、新产品或新市场宣传文案，希望了解目标受众对其的真实看法。过去，企业在传统实践中最多同时测试 5 到 10 个方案；但在 Simile 提供的愿景中，企业能够立即在 1000 个不同亚群中并行测试 1000 种不同想法。 随着客户深入体验，需求会迅速延伸并引入“时间维度”（temporal dimension）： 1. **动态产品体验**：客户不再满足于向智能体展示一张静态图片，而是要求“让这些智能体去完整体验该产品 10 分钟，然后再反馈所见所感”。 2. **多智能体复杂互动（Multi-agent Simulation）**：一个令 Joon 起初感到意外、但如今极其普遍的需求，是模拟企业的财报电话会议（earnings call）。企业 CEO 和董事会成员必须反复推敲财报电话会议的沟通设计和投资者反应，Simile 可以利用多智能体网络预演不同投资人与市场的连锁反应。 --- ### `23:45` 模拟 vs 真实广告投放：规模、代表性与二阶效应评估 面对 Sonya 提出的尖锐问题——“为什么不直接在 Facebook 上投放 1000 组真实广告来测试点击率（CTR）？真实世界的数据难道不比模拟数据更直接有用吗？”——Joon 从规模、人群代表性和二阶效应三个层面做出了剖析。 首先是规模与代表性（representativeness）。虽然 Facebook 广告测验能提供点击率，但线上实验天然存在样本偏差，因为只有特定画像的人群才会对线上广告做出反应，许多企业根本无法通过线上广告触达其真正关心的细分受众。而在模拟环境中，受众规模完全不受平台可用人群的限制，Simile 承担了前期繁重的清洗与采样工作，确保其建立的数字人群具备真正的统计代表性。许多企业最核心的痛点并非“该向受众问什么”，而是“根本找不到能够代表目标受众的人”。 更深远的技术壁垒在于对“二阶效应与长期影响”（downstream implications / second-order impact）的模拟预测： - 传统调研或广告测试只能回答单点问题（如“你喜不喜欢这款产品”、“愿不愿意为此掏钱”）。 - 以汽车制造商为例：当车企在市场上推出一款极其成功的电动车（EV）时，它不仅需要测试该 EV 的营销方案，更需要预判该车型的推出是否会改变市场对非电动车型的认知感知，以及这会如何波及整个燃油车产品线的销量与品牌价值。 - 在现实世界中，企业根本没有办法去 A/B 测试这种多米诺骨牌式的长尾后果，但在多智能体模拟环境中，这种复杂的系统性连锁反应能够被全面推演。 --- ### `26:40` 预测准确度评估：TVD 核心指标与“收敛 vs 发散”模拟论 谈及评估人类行为预测的准确性，Joon 指出人类行为本身具有天然的随机性与理论极限——即便是同一个人，面对同一个问题在不同时间也会给出略微不同的回答。然而，在群体层面，分布预测依然存在巨大的优化空间。 对于量化问答与随机对照试验（RCT）类场景，Simile 采用全变差距离（Total Variation Distance, TVD）作为衡量真值分布与模拟分布差异的“北极星指标”： | 指标 | 评测目标 | 决策阈值 (Threshold) | 业务意义 | |:--- |:--- |:--- |:--- | | **TVD (Total Variation Distance)** | 衡量群体回答概率分布与真实世界真实样本（Ground Truth）分布的拟合度 | **TVD < 0.15** | Simile 认为低于 0.15 即可作为企业制定高风险商业决策的有力依据 | 随后，Sonya 提出了关于多智能体交互中误差传递的隐忧：如果单体智能体的预测准确率是 85%，在多智能体像雏菊链（daisy chain）般逐级传递信息时，系统误差是否会不断累积并失控？ Joon 提出了一个深刻的核心观点：在模拟多智能体系统时，必须严格区分“收敛型模拟”（converging simulations）与“发散型模拟”（diverging simulations）： ```mermaid flowchart LR accTitle: 收敛型与发散型模拟机制 accDescr: 展现收敛型模拟对抗误差复合与发散型模拟因微小扰动分岔的对比 subgraph Converging ["收敛型模拟 (Converging Simulations)"] direction TB C1["单体误差微幅复合"] --> C2["系统内在收敛引力 (Pull towards convergence)"] C2 --> C3["涌现稳定结构 (如无标度网络 Hub 节点、PageRank 效应)"] end subgraph Diverging ["发散型模拟 (Diverging Simulations)"] direction TB D1["微小扰动与随机输入"] --> D2["历史路径依赖与分岔"] D3["多结果并存 (如一战是否不可避免)"] end ``` - **收敛型模拟**：在这类系统中，微小的误差复合并不致命，因为系统存在强大的“收敛拉力”（pull towards the convergence）。例如在模拟人群社交网络时，系统总会涌现出核心集线器节点（Hub），形成网络科学中的“无标度网络”（scale-free network）。这与 Google 的 PageRank 算法基石一致——无论个体网络连接细节如何微调，总会有少数页面汇聚指数级的大量链接。只要单体模拟达到一定精度，这种宏观收敛就会稳定发生。 - **发散型模拟**：这类问题本质上具有分岔特性，类似历史学经典追问“第一次世界大战是否不可避免”。在多轮模拟中，系统可能会因为初始扰动走向截然不同的结局，这也使得这类设计与决策的评估必须采用完全不同的范式。 ### `30:00` 模拟的发散与收敛：推断统计学初期的度量标准构建 如果运行一场选举的模拟，同一个人每次都会获胜吗？当模拟涉及长链条的级联决策时，微小的初始扰动往往会导致结果大幅发散。朴俊成（Joon Sung Park）解释道，Simile 当前并不一定聚焦于选举预测，但在发散性模拟中，核心评估标准在于“置信度”（Confidence）。例如将同一种情境运行 100 次，结果呈现为某一状态的频率是多少？研究人员正是借助类似自举重采样（Bootstrap Resampling）的统计学方法，来计算模拟结果的置信区间。展现发散性恰恰是模拟的核心力量之一——它不追求唯一的确定解，而是向决策者完整展示所有可能演变的多样性，让人们理解导致不同结果的底层因果机制，从而提前制定预案。 面对索尼娅（Sonya Huang）关于系统“收敛还是发散是否有严格数学表征”（例如取平均值的函数倾向收敛，二分切分决策倾向发散）的“直觉型数学猜想”（Vibe Mathing），朴俊成给予了肯定。他指出，这正是 Simile 深入钻研的前沿科研课题。朴俊成将当前的社会模拟领域比作推断统计学（Inferential Statistics）诞生的第一天：正如当年统计学家们经过漫长的论证与实践，才达成“$p < 0.05$ 具有统计显著性并足以支撑科学结论”的共识一样，Simile 如今也在尝试为整个人类行为模拟领域建立类似的置信度门槛与数学规范，界定系统何时收敛、何时不可逆地发散。 - **置信度量化机制**：通过多次重复运行模拟（如 100 次实验）并使用 Bootstrap 重采样方法，估算复杂长程互动演变中的统计置信区间。 - **发散的核心价值**：暴露小概率极端路径及分支节点的因果机制，让组织在不可逆的现实后果发生前洞悉潜在盲区。 - **范式对比**：当下的模拟研究正处于类似推断统计学早期确立显著性阈值（$p < 0.05$）的制度化确立阶段。 --- ### `32:15` 超越财富500强：宏观经济学与托马斯·谢林的诺贝尔级设想 索尼娅将话题引向更宏大的命题：既然诸多财富500强（Fortune 500）企业已经开始采用 Simile 进行概念测试与财报电话会预演，那么这套系统是否能解答人类社会最深层的未解之谜？在她看来，宏观经济学中常常充斥着“没人真正懂任何事”（Nobody knows nothing）的混沌，许多经济危机纯粹源自群体心理学；即便在风险投资行业，大家也常常争论 AI 价值链中哪些层级具有持久性，而这些本质上都是群体行为的大规模博弈。如果真有一个完美的人类行为模拟器，它的用武之地绝不仅限于商业服务，更应扩展至国家层面的决策。 朴俊成深表认同，并分享了当年在斯坦福立项时与导师迈克尔·伯恩斯坦（Michael Bernstein）和梁 Percy（Percy Liang）的对话。他当时直截了当地告诉两位导师：“如果我们做对了这件事，这里面孕育着一个诺贝尔奖（There's a Nobel Prize to be won there）。”这种信心源于对经济学历史的回顾。三十年前，学者托马斯·谢林（Thomas Schelling）通过极度简单的多主体模型（Agent-Based Models, ABM）探讨了居住隔离（Segregation）的形成机制，并最终获得了诺贝尔经济学奖。 当时的经典模型极其简陋，每个个体仅仅是一个“红点”或“蓝点”，规则也完全是机械式的——每轮博弈中，个体只观察身边邻居同色的比例，一旦低于某个简单阈值便搬迁到新网格。然而即便如此单薄的模型，都揭示了宏观社会偏好的惊人自组织现象。而今天，借助生成式智能体，研究者终于能赋予每个主体以丰满的记忆、价值取向、情感偏好与复杂的心理机制，从而在宏观尺度上展开前所未有的精准实验。 ```mermaid flowchart TD accTitle: 传统ABM与现代生成式智能体模拟对比 accDescr: 对比谢林时期的简易规则多主体模型与Simile采用的基于大模型的丰满智能体模拟 subgraph Classical["经典多主体模型 (托马斯·谢林时期)"] A["简单离散个体 (红点/蓝点)"] --> B["机械式硬编码规则"] B --> C["局部邻居颜色阈值判断"] C --> D["机械移动位置 (解释隔离现象)"] end subgraph Modern["现代生成式智能体模拟 (Simile 范式)"] E["丰富人设与认知模型"] --> F["LLM驱动的记忆、反思与规划"] F --> G["多维社会心理互动与自然语言交流"] G --> H["涌现复杂宏观现象 (银行挤兑/气候合作)"] end ``` --- ### `34:40` 人类社会的“CERN”：耗资上亿的终极模拟与社会科学质变 当宏观经济学家和学者前来与 Simile 交流时，他们提出的正是这类影响人类命运的大命题：银行挤兑（Bank Run）究竟在何种心理临界点下瞬间爆发？在应对气候变化时，各国深陷的集体行动困境（Collective Action Problem）如何通过机制设计来打破？一个民主政体在走向崩溃前，早期微弱的预警信号是什么？货币系统最初究竟是如何从易货交易和信用关系中自发起源的？ 朴俊成强调，解答这类重大问题应当成为模拟科学的“北极星”（North Star）。今天的商业模拟追求轻量、快速，几分钟或几小时即可完成，但他坚信未来必定会出现一种全新的科学基础设施——类似物理学界的高能粒子对撞机。 > “在未来的某一天，可能会出现单次运行耗资高达 1 亿美元、耗时数月的大规模模拟。但只要它跑完一次，就能解开我们人类社会最根本的核心谜题之一。” 索尼娅补充道，政治学领域也将因此迎来彻底变革：如今每逢政策推行，不同利益群体总是在为“某项政策究竟会带来何种影响”争执不休；但若能通过数字孪生社会提前运行模拟，不仅能看到当下的涟漪，更能看清未来 5 到 10 年内所有连锁的下游后果。 - **核心研究议题**：银行流动性危机与挤兑心理演变、跨国气候合作机制设计、民主制度崩溃的早期预警、货币演进的历史因果还原。 - **超级模拟构想**：建立规模类似欧洲核子研究组织（CERN）的“人类社会对撞机”，以单次千万乃至上亿美元的算力成本求解文明级宏观规律。 - **政策制定范式转变**：从基于主观臆断和政治利益的说辞博弈，转变为依靠可观测、可回溯的全尺度动态模拟实验。 --- ### `36:20` 观测人性的“哈勃望远镜”：科幻蓝图与社会科学的终极解答 在对谈尾声，谈及究竟是什么让他对未来保持最大的兴奋时，朴俊成坦言自己深受科幻小说的启发。在几乎所有描绘成熟高等文明的硬核科幻构想中，推动文明稳定前行的基石永远有两个：一个是某种形态的通用人工智能（AGI），另一个则是引导社会前行、预测文明风浪的全局社会模拟系统（Simulations）。五年前他或许还不敢轻言此愿，但随着近年的研究深入，这一信念已变得无比坚定。 Simile 联合创始人梁 Percy 常说，人类科学史上的每一次伟大跃迁，往往不是始于理论构想，而是始于一台前所未有的测量仪器——正如哈勃太空望远镜（Hubble Space Telescope）的升空彻底重塑了人类对宇宙和星系的认知一样，生成式模拟系统有望成为观测人类社会内部幽微机制的“哈勃望远镜”。 长期以来，科技界的大规模算力与顶尖智力大多被倾注在自然科学与客观工程难题上，而人类社会与行为本身的复杂性却常被视作不可测的玄学。朴俊成坚信，模拟技术提供了一次“震撼人心的机会”（Breathtaking chance），让人类社会科学与行为经济学从经验主义走向精确科学。索尼娅深表共鸣，并向朴俊成表达了感谢，盛赞从斯坦福的虚拟小镇 Smallville 到如今 Simile 所开启的全新篇章，不仅为理解人类行为提供了全新标尺，更有望重塑未来社会的决策根基。 ## Notable quotes > `07:22` 当今的主流大模型公司——无论是 OpenAI、Anthropic 还是新兴的前沿实验室——它们的北极星目标都是制造某种超级智能机器（Superintelligent Machines）。这些机器被刻意训练得极度理性，在解决具有唯一客观标准答案的技术与逻辑难题上表现出超越人类的能力。 — Joon Sung Park > `10:00` 85% as accurately as people replicate their own — Joon Sung Park > `12:15` 设想你只有 15 分钟时间，你能向这些真实的人提出哪些最具魔力的问题？ — Joon Sung Park > `15:18` There are things that people say, and then there are things that people actually do. And the gap there is real. — Joon Sung Park > `15:18` Just tell me the story of your life. — Joon Sung Park > `32:15` There's a Nobel Prize to be won there — Joon Sung Park > `34:40` 在未来的某一天，可能会出现单次运行耗资高达 1 亿美元、耗时数月的大规模模拟。但只要它跑完一次，就能解开我们人类社会最根本的核心谜题之一。 — Joon Sung Park > `32:15` Nobody knows nothing — Sonya Huang ## Mentioned Joon Sung Park, Sonya Huang, Simile, Stanford University, Smallville, Sequoia Capital, Percy Liang, Michael Bernstein, Social Simulacra, GPT-3, CVS Health, Gallup, Thomas Schelling, OpenAI, Anthropic, Reddit, CERN, Hubble Space Telescope ## Video description The race to build superintelligence is producing models that keep getting better at objective problems, but not at behaving like actual people. Joon Sung Park, founder and CEO of Simile and creator of Stanford's "Smallville" generative agents study, argues that simulating human society requires a fundamentally different kind of model. He frames today's frontier models as the "CPU of intelligence"—rational, superhuman at problems with right answers—and Simile as creating the "GPU of intelligence," built to encode the diversity of people's values, preferences, and tastes. It simulated 1,000 Americans and predicted their behavior 85% as accurately as people reproduce their own answers. CVS uses it for concept testing; some customers simulate their own earnings calls. Joon's larger bet: a "CERN of human society" that could one day model bank runs, climate cooperation, or the early signals of a collapsing democracy. Hosted by Sonya Huang, Sequoia Capital [Source](https://pscrb.fm/rss/p/traffic.megaphone.fm/CPUAI1964871992.mp3)

[7:22] Joon Sung Park: 当今的主流大模型公司——无论是 OpenAI、Anthropic 还是新兴的前沿实验室——它们的北极星目标都是制造某种超级智能机器（Superintelligent Machines）。这些机器被刻意训练得极度理性，在解决具有唯一客观标准答案的技术与逻辑难题上表现出超越人类的能力。 [10:00] Joon Sung Park: 85% as accurately as people replicate their own [12:15] Joon Sung Park: 设想你只有 15 分钟时间，你能向这些真实的人提出哪些最具魔力的问题？ [15:18] Joon Sung Park: There are things that people say, and then there are things that people actually do. And the gap there is real. [15:18] Joon Sung Park: Just tell me the story of your life. [32:15] Joon Sung Park: There's a Nobel Prize to be won there [34:40] Joon Sung Park: 在未来的某一天，可能会出现单次运行耗资高达 1 亿美元、耗时数月的大规模模拟。但只要它跑完一次，就能解开我们人类社会最根本的核心谜题之一。 [32:15] Sonya Huang: Nobody knows nothing

[0:00] Isabella: it would be great if I can do a Valentine's Day party where we invite a lot of friends, customers. [3:22] Joon Sung Park: these models can actually encode human behavior. [3:42] XYZ places are amazing [10:00] Joon Sung Park: 85% as accurately as people replicate their own [12:18] Joon Sung Park: all the second order impact of the decisions [15:00] Joon Sung Park: Turns out, if we understand the person's story of your life, the kind of data you get from it is what we consider to be the long tail information about this person... It's about where you grew up, what were some of the difficult decisions you had to make in life. And what's interesting about this data is it's an amazing way to build a translational layer between attitudes and behavior. [20:00] Joon Sung Park: ...Right now we're very much in the practice of testing 5 to 10 different ideas at most. But what does it look like for us to test instantly 1,000 different ideas across 1,000 different sub-populations? That's the initial vision they see. [25:00] Joon Sung Park: A TVD of, let's say, less than 0.15, we believe is actually quite strong evidence for making a decision. [25:00] Joon Sung Park: Was World War I inevitable?

[0:00] I am somebody who is quite inspired by science fiction. And when you read science fiction that covers societies that have progressed far enough in its technological maturity, you always see two pillars. You have some version of AGI and you have some version of simulations that really help guide the society. I do see an opportunity today to really take the first crack at building the simulation. I would not have said that even five years ago. But that is a conviction that we have built up over the years as we're going deep into this research. Today, we're delighted to have Jun, founder and CEO of Simile. Simile is building an applied AI lab, simulating human behavior and societies. And I'm very excited to have you here to discuss what you're building. Same here, thank you for having me. Okay, take me back to April, 2023, Stanford California specifically Smallville Stanford California What was that So Smallville was a project that we were running at Stanford where the idea was that we made this observation that large models can now encode a lot of human behavior that is embedded in its training data from the web and social media and so forth that if you sort of prove at the right angle, you can actually get a lot of micro behaviors out of these models. So given a very specific demonstration or description of a situation, what would person X do? And it would actually generate really interesting behaviors. We found that to be so interesting. And we found that to be the ingredient that we have been waiting for, for creating really complex agentic behaviors. So Smallbit actually was an experiment where we decided that if we push this as far as possible, what would a society that is created by these agents look like? So we basically created generative agents that is paired with generative AI model. with memory, planning, and reflection to basically create this lived experience of agents living in this small town. So Smallville was basically a game town of 25 agents living in it. Individual agents had a description of persona, but they would actually wake up in the morning, do their routines, go to work, actually have relationship, sort of like people would, and they would actually have emergent phenomena like having parties and so forth So that was the experiment that we ran What was the most surprising things to come out of the experiment So one of the surprising things was so the experiment, the simulation itself actually sets place the day before a Valentine's Day. So you actually see these agents, one of the agents actually thinking, well, I run a cafe, so she's a cafe owner. Her name's Isabella. She goes and thinks, it would be great if I can do a Valentine's Day party where we invite a lot of friends, customers. So you actually see her on the day before Valentine's Day going around actually gathering materials for the party, actually telling our customers, hey, we're going to have this party, please come. And on the day of Valentine's, you actually see this immersion party that actually gets formed with all these agents coming to the basically cafe. Did anyone not get invited? Well, some of the people did get invitation, but they forgot. That's one thing that did happen. Some of the agents did not explicitly get invited, But we had one agent who got the invite, Klaus, who decided to ask his crush out on a date. So he would actually bring in the date. They would actually have a party at this cafe. So quite surreal. So how did you end up building Smallville in the first place? Were you studying kind of human psychology and social behavior? Or was this coming from the kind of customer back Or was it coming from the technology out So my protocol team has been excited about simulations and we saw the vision of simulation failure early on So my career as a researcher at Stanford really started back in 2020. That was the year when GPT-3 was about to come out. It wasn't quite there yet, but it was just about to come out. We started to get its first demos. And my first year, we wrote this paper called Opportunities and Risks of Foundation Model. alongside many of the Stanford researchers, and was led by one of my co-founders, Percy Leung, who is now the head of the Stanford Foundation model at Stanford. And when we were writing that, the part that I was really focused on was, well, here's a new class of models that we have not seen in the past, that these models that can be very generalizable in ways we didn't quite have in the past. And I got into thinking, well, if we can imagine the kind of interaction we can create with these models, what would that be? And many of my colleagues back then were surprised that these agents or these models can do classification or a simple generation. And that was really incredible to see because these models didn't really know or didn't really, wasn't really taught to do that. But the part that was surprising to me wasn't that these models can do that because from interaction perspective, we've known how to do this for a long time. The interesting part was, well, these models can actually encode human behavior. [5:00] What does that mean if we were to push this as far as possible? So part of the tradition that comes from research included what we call social computing. And social computing within human computer interaction really has to do with this idea of how can we build a better technological platform that would enable social interactions and collaboration. One of the most difficult challenges of building a social platform is not necessarily testing the UI UX of the system, but it's more about when you have tens of people, millions of people, and down the line, billions of people, how do all these people come together to create the immersion phenomenon that's both good and bad, and how can we design for a scale? And so far, we didn't really have a tool that would enable us to test for that. The only way we test it today is you basically field test it. You release your prototype, see what happens, and sometimes it actually comes at a real cost. Obviously, it's high cost in terms of human hours and the time it takes. But at the same time, if you have a bad design, imagine you have a feed on social media that is more likely to propagate certain emotion that is negative. And obviously that is something that we want to avoid But this now gets tested in the field So we wanted to see whether we can actually create a simulation that would actually let you test for this So 2022, this was actually a year before we generated agents, we worked on a paper called Social Simulacra, which actually really was the precursor to the agent paper that we ended up writing. The core thesis was, imagine you're building a subreddit. You're a designer on a subreddit. You want to see what people might do in the subreddit, which is surprisingly hard task, even for practice designers. And we basically decided, hey, we have this model, seems unique. Let's use this model to create simulations of the entire subreddit. So you define the goal, you define the moderation strategies, and you populate it with thousands of, back then we didn't call them agents, but we call them personas, but populate it with thousands of personas. This is basically 22 version of notebook, which is quite interesting that it actually came back. And when we saw that, we actually got a lot of really important insights out of this. What are the good behaviors? We actually simulated a community where the entire idea was for people to discuss with each other, the sites, places to sightsee in Pittsburgh. And all of a sudden you start to see this personas actually collaborate to actually discuss hey XYZ places are amazing Do you want to actually go to a trip together and actually plan those trips live in the simulated subreddit So that how we got excited So we saw the vision and the excitement and the potential applications fairly early on. But then the work that we had to do was then demonstrating how can we go beyond simple personas to create complex agents that actually can think over time because we want to simulate the longitudinal aspect of our society, and then actually validating that these simulations are actually accurate in practice. Was there a point of model evolution at which you felt like, okay, we're there, the models are good enough for us to actually have a, you know, faithful representation of human society? So GPT-3, when it came out, and Social Simulacra was built with GPT-3, and it was very janky. It didn't do any instruction tuning. It did not follow your instructions. So just to have it to listen to you and do what you wanted to do, you had to do some weird tricks with prompting and so forth. Well, you could actually see the promise. The model actually have encoded a lot of human behavior, and you could actually see the trajectory. And when we had the generative agents paper, it wasn't quite chat GPT, but we now had instruction tuning. So we could actually build much more complex agents that can reason about its memory That wasn really possible when we did social simulacra And since then of course the models have improved So where we are today is the models at its foundational level have reached the point where we can actually imagine building these kind of applications. Now, the part that actually I do think, however, that's quite interesting here, today, if you look at many of the large language model companies, whether it's OpenAI, Anthropic, and many of the neolabs that are getting formed, the models they are creating are models that I would consider to be, they're no start to be, something that is similar to, let's build a super intelligent machines. These machines are meant to be rational. And these machines are supposed to be really amazing at technique problems that have an objective answer. So maybe that's not even the best simulation of true human society then. Turns out, people are irrational. We have a lot of subjective values, preferences, and taste. So you actually start to see divergence in model size going up and the performance in its ability to predict and simulate human behavior. So we have sort of plateaued with current modeling paradigm, our ability to really simulate humans. So it is sort of at the starting good foundational level. But to make it really amazing, we do need the next frontier that is more humanized. [10:00] towards actually modeling people's diversity. Very interesting. At what point did you realize that, you know, what you did with Smallville could become a company? Right. So again, the promise of application was something that I was very much inspired by early on by simulation with social simulacra and so forth. But the part that I realized over time is research and a company have very different function. Research is an amazing vehicle. if you want to basically do breadth-first search. You are in a lab surrounded by a really smart set of people, and each of the researchers own a small piece of thesis, and they go explore some of those thesis blossoming to amazing research product. But we're not necessarily known for finishing our job. We're not usually the one to bring that research impact to the real world. Company is a machine for depth-first search. You have a conviction on an area. you find a hill that you want to climb. This is the vehicle that let you put together resources and an amazing group of people to go after a singular vision without hesitation. And we got that conviction I would say about half a year after Generative Agents After the original Generative Agents paper we got so much inbound interest initially from actually social scientists who wanted to run their experiments and all the RCTs on our platform Then very soon after, many of the Fortune 500 companies who saw this demo and their board members and CEOs who sometimes visit Stanford saw that. And they start asking, well, we go run all of these surveys and experiments. And there's so many research questions about the market that we cannot answer today. Can we run that in simulation? That started to really intrigue me because that showed a clear line towards a real world impact for research, which is not always the case that we have that kind of opportunity. So that is when we decided we actually want to validate the simulations are accurate. So we went out and actually created simulations of 1,000 people of the U.S. population. We demonstrated that using our architecture and the models, we can actually predict people's behaviors 85% as accurately as people replicate their own. When we saw that, we thought, okay, this is something that we feel comfortable providing to our users as a platform for simulating their really important decisions. So that when the co myself Percy as well as Michael Bernstein was researcher and my advisor at Stanford Both of them were actually my advisors So the three of us have been working together for five years And now at this point, it's seemingly six years. But that's when we got together to have the initial conversation of can this be a company? Got it. Amazing. Maybe walk me through a customer engagement end to end today. Like who's a canonical customer and which department? And they come to you. What are they asking you? and what product or service do you deliver to them? Right. So maybe an example that I can give to make this concrete. So CVS has been partnering with Simile for the past, I would say, nearly half a year. And they've been an amazing partner. The way we initially got in touch with, so our main buyer at CVS is the lead, is a senior VP who leads human insights. And original story there was he was, he basically read my paper that validated the agent simulations and thought we have to bring this to cbs because today we are bottlenecked by the number of questions we can field test and we also bottlenecked by truly the physics of human society it one thing to ask surveys and experiments totally different thing if down the line you actually want to simulate the entire market and actually map out all the second order impact of the decisions you suggest to your leadership So he's been looking around for the solution and his cousin happened to know me and basically told our buyer, Sri, that the authors of the paper are actually looking to start something. So that's how we got connected. And in this particular engagement, usually the way this goes is our customers are very much used to working with polling companies or panel companies today. And there they go and basically ask these companies, XYZ are the populations that we're interested in better understanding. Can we go run a research study of these topics? That initial stage looks very similar for a simile. So our buyers come and they tell us, we want to better understand XYZ population. Then Simile goes out and we have, through our partnership with vendors, we have a strategic partnership now with Gallup, for instance, who is a polling and panel company, where we go out, work with our vendors to actually reach out to real humans. So these simulations are grounded in real data, but reach out to those people, collect data that we believe are efficient and generalizable. [15:00] about that person. So imagine you have 15 minutes. What are the magical questions you can answer or you can ask these people during that time? We collect that data, use that data to create agents or simulations of these people that can basically be used to answer a large number of questions that goes way beyond the original domain. We load that onto our platform and it's basically a SaaS product. Our customers come and they can basically ask any questions about the group of people of their interest. So interesting. It reminds me of in autonomous vehicles, you know, you go and collect a bunch of data from the road and then you're able to augment it with simulation. Is this a similar concept or are there big differences to what you're doing? It is similar concept in the sense that, of course, you with the self-driving vehicles, you want to create model that is based on real world physics, but you want to create a model that is generalizable beyond your training data. It needs to be generalizable in two different locations with different weather conditions. Very similar concept, where what we want to create is we want to reach out to real people and for these people want to understand something fundamental about these people in a way that we can encode into the model I would have thought that the large language models would be such a good representation of the whole world that you could almost narrow it down You could tell Claude, you are a 34-year-old woman living in a bi-coastal metropolitan area, and it would be able to have a faithful representation. So I'm actually surprised that you go out to Gallup. Maybe can you just explain why you have to go out and collect any real-world data at all? Yeah. One of the big questions here is the question around say-do gap. There are things that people say, and then there are things that people actually do. And the gap there is real. And a lot of the large language models are trained on attitudinal data. Fundamentally, it is the things that people have said online that does cover a large quantity of its training data. So one of the things a simulation platform does is actually closing that gap. So a lot of the data that we end up collecting by nature are behavioral. It also includes data that actually goes into literally questions like, just tell me the story of your life. Turns out, if we understand the person's story of your life, the kind of data you get from it is what we consider to be the long tail information about this person it not about what you done in this particular moment it not about very broad questions like what your view on politics It about where you grew up what were some of the difficult decisions you had to make in life. And what's interesting about this data is it's an amazing way to build a translational layer between attitudes and behavior. So we combine these kind of data sets, but fundamentally that's the gap that we want to close. What sort of behavioral data do you have? So Simile does run a lot of experiments. So kind of models that we have trained, for instance, we have a huge repo of RCTs. So randomized control trials that were run in social scientific context, that were run around pricing studies. So one of the models that we are training is basically the foundation model of human behavior in quite a literal sense. We have all the behavioral signal from RCTs. Can we actually encode that into the model so that the end outcome is a that can basically predict the results of any RCTs. At the same time, one of the conversations that we keep on having with our customers that we're very excited by is our customers then come in, see that potential, and their mind goes to, wow we have 90 million customers let say here at CVS How can we leverage this kind of data to create better simulations So there also conversation around how can we in a responsible and ethical way leverage existing data that is also in-house for our customers, then use that to create augmented version of Simuli's model. So that, of course, is going to be more fine-tuned, specific to the population of these customers. But that's the kind of data that we will be be leveraging. I see. And are you doing these interviews typically by voice? Is it a survey that you fill out? What's the modality? So it's a huge breath. The quick answer here is it's both. Interviews are fantastic if you want to get the long tail information about people. So we actually do in the original study that I conducted back in 2024, we literally ask a question, tell me the of your life. Now, the way we do it is we are training our own model. So it's a reinforcement learning loop, but basically imagine the objective function here is how can you spend the minimum amount of time to get the maximum amount of visibility about this person? So that is one of the things that we do. So basically training an interviewer that is not really asking for factual information or an experience about a particular platform, but just what are the life story that people have? [20:00] that can be used to train our own model for these agents. And then for the more factual or sort of more discrete choices, choice questions, surveys, and so forth, these are also very efficient. These are time and data efficient because people can fill out many of the questions in a short period of time. So for those, we actually do leverage them. For instance, if you want to just have a broad understanding of people's viewpoints on certain topics, certain policies, and things like that. You describe yourself as an applied AI lab. How do you think about where you want to build your own models versus where you want to rely on other existing models? So in terms of building our own model, really the core thesis here is there is an amazing model to be built that really uncodes the diversity of people's values, preferences, and taste in ways that simply a rational model cannot do. So one way I actually post this, we're sort of building, so imagine the current today's model are akin to the CPU of intelligence unit. It's a single model trained on amazingly rational data that is amazing at solving very complex objective questions. Similarly this model is much more akin to developing something that is closer to the GPU of the intelligence unit where the idea here is we don actually need a model that is superhuman at Simile. In fact, we want a model that's as human as possible, but we want to make sure that these models at the sort of individual subunits can represent the real viewpoints of different subpopulations. So where we see that gap, that's when we go develop our own model. But at the same time, we do leverage frontier models, for instance, as a way to coordinate the research. Frontier models are amazing at coming up with a research plan. So that's where those models actually do get leveraged. Very interesting. Are people typically coming to you with questions around new product launches, how they should be marketing their companies, pricing, all of the above? So it is all of the above. Our customer journey usually does, however, start with a very concrete use cases and problems they are trying to solve. Concept testing is a big one. It's also a very straightforward one. So they have a new concept, new product idea, new market message they want to test, and they want to hear from their users what they would think about X, Y, Z. This is one way for them to quickly test those ideas And then the promise they quickly see is well right now we very much in the practice of testing five to 10 different ideas at most But what does it look like for us to test instantly thousand different ideas across thousand different sub-populations? That's the initial vision they see. Then we really get into the nitty-gritty details of, well, where does the simulation go from here? They then pretty soon start asking, well, Can this be used to do product testing, but not just simply submitting an image, but imagine basically asking these agents, go experience this product for 10 minutes and tell us about what you experienced, what you saw. So you're basically adding temporal dimension. And then you go into things like multi-agent simulation. Some of our customers very routinely actually ask us to simulate their earnings call. This is actually a use case that both surprised me at first, but this is also surprisingly a common ask Because of course the CEOs and board members always need to think about, hey, how are we going to design our earnings call? How would the audience react? So that is something that we also do. And this is very much a multi-agent simulation. You know, it seems like there's so many use cases that could potentially be tested once you have like a simulated almost customer population right Yeah I curious the value of research and testing in sim versus just like let say you have a new product concept that you want to test Why not just go run a thousand Facebook ads and like you actually get the click through rates on this stuff? Isn't that real world data almost more useful than the simulated data on how people might behave that you then correct for with your own models? So it's a great question. And I think to some extent here, the answer has to do with initially scale and then down the line, truly the new capability that comes because you can simulate interactions. The scale question here is actually quite straightforward, where, yes, you can actually run Facebook ads and Facebook testing. But the kind of experiments that you can run in simulation is actual behavior simulation at scale. So you can basically pull in any number of users. It doesn't even have to be bounded by the number of population that's available on Facebook. And it's also much more representative because only certain groups of people will actually respond to the online experiments. But similarly, the model that we are creating, one of the key promises is that it is representative. We do the hard work of actually getting the representative set of people and then collecting the data that would actually represent them properly. So the scale represented to [25:00] is something that many of our users do not have easy access to. This is actually one of the common asks also that we do get or common sort of pain points that we have heard, where the question that many of these people have isn't about like what questions do we ask these people, but it's about in the first place, how can we get to the population that we're excited to talk to? That's a huge bottleneck. Then down the line, you can actually really start to imagine, and this is something that our customers and some of the most forward-looking customers are now going into, which is what are all the downstream implications of the decisions that you make it's not just about whether imagine you have this particular product will do you like it or do you not like it would you pay for this not pay for this it's not necessarily the just that initial questions that we want to answer and finish but we want to understand imagine you're a car company you launched an electric vehicle in this market maybe the electric vehicle does really really well So we can help you do concept testing around marketing and the product around the electric vehicle. But what does that do to the perception of, let's say, non-electric vehicle? Does it change the market perception? Then what does it mean for the rest of the product line And how do you balance those kind of second order impact of your decision in a way that is more evidence Today there no way to test for this You can run this in simulation So really going beyond simply asking one question at a time, but then to think about what are the long-term implications of your decisions is something that our customers are quite excited by. I'd love to understand how you think about how predictive your model is in actually simulating real human behavior. I imagine you have a lot of evals on this. I guess, what is your North Star metric? How do you guys do on that? And what do you think is the theoretical limit? It's a great question. So theoretical limit, and let me just start from there, certainly does exist in the sense that humans are genuinely, there's a lot of randomness. That if you ask me the same question, I'll actually answer the question slightly differently. So there is certainly that degree of randomness in human behavior. However, there's a lot of gains in performance that we can have even today in the way we are predicting people. So the measurement that we do is so at the level of population, we measure the distribution of responses if it is more quantitative. So we actually measure total variance distance which basically shows how close are the distributions of the ground truth versus the simulated information And that is a metric that we run across all the use cases that our customers have And we have certain threshold that we believe is good enough for decision So a TBD of, let's say, less than 0.15, we believe is actually quite strong evidence for making decision. So that is a North Star state that we want to hit for this class of use cases that are more quantitative, that's more question and answers. This also does cover RCTs, which is many of the core use cases our customers have. Now, there's actually a really interesting question to ask around, well, what about multi-agent simulation? What about all the downstream implications that we're going to be simulating? What does the evaluation of those look like? Yeah, and then do you daisy chain errors as you kind of, you know, if this one is 85% accurate and then this agent is telling another agent something, do you accumulate errors as you go towards multi-agents? Exactly. And one of the core thesis here is we basically see two categories of simulations. One simulation is what I would consider to be simulations that converge. The other categories of simulations are the simulations that diverge. And sometimes they actually coexist. And it's really about what research questions do you have. Questions that converge doesn actually matter if you have a little bit of error Now the error cannot be obviously so dramatic that it sort of is completely detached from reality but you actually are okay even if the errors do compound over time because the pull towards the convergence is strong enough that you'll actually understand where everything would fall. A good example here actually is if you simulate a network of people, then that network will always have a hub that gets formed. This is what sort of network scientists would call the scale-free network, for instance. This is actually what powered Google too. One of the core observations of PageRank was it doesn't matter how these networks actually get formulated. You actually see some webpages that get exponentially more links that are attached to it. This is a very fundamental behavior in humans. That we also see in simulated networks. And that convergence always happens as long as you are replicating human behavior with certain threshold accuracy. Now, there are then questions that generally do diverge. It's like your classical questions like, was World War I inevitable or was it not? And there, it is sometimes difficult to run the same simulation over time and get the same exact outcome. Imagine you're running a design simulation [30:00] not something that necessarily similarly right now is going into, but imagine you're running a simulation of an election. With the same person win the election every time, there are a lot of downstream implications of every single decision that does happen. So it does diverge. There, the core evaluation is around confidence. So imagine you run the simulation 100 times. How many of those times do the results come out to be X? And how can we actually use that to basically create a bootstrap resampling to calculate the confidence around the simulations? Those are some of the questions that we do ask. And a huge part of this also, of the power of simulation, is then to show when it diverges, to show the diversity of possible outcomes so that people can actually look, understand the cause or mechanism of how we got to those outcomes, and prepare for those features. So those are some of the implications of divergence in simulations. Are there any mathematical descriptions of why something would converge or diverge? like I imagine if you have like an average function maybe you converge and then if it's like a you know you're splitting outcomes to a binary then you might maybe diverge but yeah so the intuition I think is close and technically this is also a research topic so similarly is a company where we do go deep into this research topic in the sense that I see simulation as a field as akin to developing your day one of inferential statistics You know, inferential statistics scientists actually had to do a lot of discussion and research over time to decide that P less than 0.05 is actually evidence that is strong enough for science. similarly is working on setting the same kind of threshold and standards for the rest of the field so those are the intuition i think that's exactly the right intuition in terms of actually how to make a robust mathematical equation around when it's going to like what's going to happen when it is a real research frontier for simulations thank you for being nice but my vibe mathing um i'm curious you know it seems like so a lot of force 500 is coming to you So I'm wondering whether there are non-existing corporate use cases that might, you know, there are like great mysteries of our society that might become solved. And for example, I'm wondering about economics, you know, central bank decisions. Oftentimes, like, you know, I personally believe in macro. Nobody knows nothing. And oftentimes a lot of the issues come about from human psychology So to me macroeconomics is a function of simulating human behavior at scale You know I thinking even in the venture capital use case we often debate internally, you know, does value accrue to this company or not? You could run the simulation of all the different layers of the AI stack and almost figure out where durability and value accrues. Like if you had a kind of perfect simulator of human behavior, there's so much more you could do than serving the Fortune 500? Do you agree with that? And then if so, are you serving governments, you know, the like? Yeah. So it's interesting. When we were still researching in this area, the way I actually got back then my advisors, Michael and Percy, excited about this was I basically told them, look, we do this right. There's a Nobel Prize to be one there. And I truly believe that. And it's also not surprising in that your classical economics simulations, things like agent-based models, that really pioneered our understanding of back in the day, the kind of topic they studied was, how does segregation happen? What are the causal mechanisms for segregation So scholars like Thomas Schelling would actually build agent models that are extremely simple and rudimentary but that showed something deep about human macro behaviors And he, of course, went on to win an award prize. I see the same opportunity here, but in an augmented way, where back in the day, the agent-based models were very much deterministic in some sense, where you basically, in this simulation of, let's say, like motor segregation from 30 years ago, individual agent was simply red dot or blue dot. And every game iteration, they would look around its corner, see how many of its neighbors are of the same color. And if that threshold goes below certain threshold, then they would decide to move to a new location. That was it. But now we can actually create real agents that replicate the full richness of individuals and run the same kind of simulations. So the kind of questions that we can ask that goes beyond simply the commercial use cases, For instance, in the context of macroeconomics, actually the questions that I actually did get asked from economists were things like, when does bank front happen? Or questions like climate change. One of the sort of core blocker of climate, like solving that issue is the collective action problem of many nations. Can we actually simulate that? Or what are the signals of a democracy that is about to collapse? Can we understand the, [35:00] origin story of the monetary system. These are the kinds of simulations that I do believe ought to be the North Star state of this field. And it is sort of interesting to imagine like what that would actually look like in practice, right? Because these would involve very large scale simulations with many agents interacting with each other. I do see a future where today, this is something not the case. Today, a simulation is quick and fast to run. But what about simulation that takes actually $100 million to run once and could take many months to run. But when we run it, it solves one of the fundamental questions of our society. That I do think is genuinely a very exciting possibility for this field. I'm even thinking like politics, for example, could be forever changed. Today, everyone has an agenda of how they say some policy change will impact things. Well, why don't we just run the simulation? And understand all the downstream implications. Yeah. And not just what's going to happen this year, but what does it mean in the next five to ten years? Exactly. Fascinating. I was going to close by asking you what makes you excited about the future? Is it what we just talked about or is it something else? I am somebody who is quite inspired by science fiction. And when you read science fiction that covers societies that have progressed far enough in its technological maturity, you always see two pillars. You have some version of AGI and you have some version of simulations that really help guide the society. I do see an opportunity today to really take the first crack at building the simulation. I would not have said that even five years ago but that is a conviction that we have built up over the years as we are going deep into this research And what exciting is there a clear use case today that can serve our users But then there a lot of innovation that is yet to come that I do think will build up to actually building simulator that's akin to discern human society. And one of the things that one of my co-founder Percy sometimes say, is you look at the greatest scientific innovation, they often start from an amazing measurement. Hubble Telescope really changed the trajectory of how we understand the universe. Simulation can be that for human society. So the thing that does excite me, there's a lot of focus on natural sciences, but how can simulation really unlock our understanding of humanity and social sciences? And how can we actually use it to make our society be a better place? That's exciting. I remember reading somebody was excited about, you know, So there's a small but breathtaking chance that the field of economics, as we know it, may actually become solved by simulation. And I extend that not just to be economics, but kind of everything that deals with human behavior and social sciences, which ultimately is everything around us. Truly. Wonderful. Thank you so much for joining today and sharing the story of both Smallville and what you're now up to at Simile. I really enjoyed the conversation. Same here. Thank you for having me. you