用 AI 模拟人类社会：从小镇实验到现实应用

Simulating Human Societies with AI: From Smallville to Real-World Applications

朴俊成 Joon Sung Park · Training Data · 2026-06-16 · 约 39 分钟 · 原视频 ↗

打开互动全文版（中英对照 + 朗读 + 问答）→

本期速览 · Overview

Simile 创始人 June 探讨了生成式智能体在小镇实验中模拟涌现社会行为的过程，以及利用模拟指导社会的愿景。

June, founder of Simile, discusses how generative agents in Smallville simulated emergent social behaviors, and the vision for using simulations to guide society.

要点 · TL;DR

• LLM 能以 85% 的准确率模拟人类行为，但需要真实世界数据弥合言行差距。
  LLMs can simulate human behavior with 85% accuracy, but need real-world data to close the say-do gap.
• 模拟能预测决策的二级影响，超越单问题调查。
  Simulations can predict second-order effects of decisions, beyond single-question surveys.
• 最终目标是打造类似 CERN 的人类社会模拟器，解锁社会科学。
  The ultimate goal is a CERN-like simulator for human society to unlock social science.

核心观点 · Key points

• 大语言模型能从训练数据中编码人类行为，实现逼真的智能体模拟。
  LLMs can encode human behavior from training data, enabling realistic agent simulations.
• 模拟预测人类行为的准确率可达人们自我复现的 85%。
  Simulations can predict human behavior 85% as accurately as people replicate themselves.
• 当前模型在模拟人类多样性上遇到瓶颈；下一个前沿是建模主观价值观。
  Current models plateau in simulating human diversity; next frontier is modeling subjective values.
• 需要真实世界的行为数据来弥合模拟中的言行差距。
  Real-world behavioral data is needed to close the say-do gap in simulations.
• 模拟能测试决策的二级影响，超越单一问题调查。
  Simulations enable testing second-order impacts of decisions, beyond single-question surveys.
• 最终愿景是建造类似 CERN 的人类社会模拟器，解锁社会科学。
  The ultimate vision is a CERN-like simulator for human society, unlocking social science.

反共识 · Contrarian takes

• 为理性优化的前沿模型并不最适合模拟非理性的人类行为。
  Frontier models optimized for rationality are not best for simulating irrational human behavior.
• 模拟可能比在线实验更具代表性，后者偏向特定群体。
  Simulations can be more representative than online experiments, which skew to certain groups.
• 某些模拟即使有误差也会收敛；发散则通过置信区间处理。
  Some simulations converge despite errors; divergence is handled via confidence intervals.
• 耗资一亿美元的模拟运行可能解决根本性的社会问题。
  Simulating a 100-million-dollar run could solve fundamental societal questions.
• 如今的智能体模型能复制完整的人类丰富性，而不仅是简单的红蓝点。
  Agent-based models today can replicate full human richness, not just simple red/blue dots.
• 该领域需要新的统计标准（如 p<0.05）来验证模拟有效性。
  The field needs new statistical standards, like p<0.05, for simulation validity.

本期章节 · Chapters（共 20）

• 科幻灵感与模拟愿景 Inspiration from Science Fiction and the Vision for Simulation
• Smallville：斯坦福实验 Smallville: The Stanford Experiment
• Smallville中的突现行为 Surprising Emergent Behaviors in Smallville
• Smallville起源：从社交计算到生成式智能体 Origins of Smallville: From Social Computing to Generative Agents
• 用角色模拟子版块 Simulating a subreddit with personas
• 与CVS的起源故事 Origin story with CVS
• Simily如何服务客户 How Simily works with customers
• 与自动驾驶模拟对比 Comparison to autonomous vehicle simulation
• 为何收集真实数据而非仅用LLM Why collect real-world data instead of using LLMs alone
• 收集的行为数据类型 Types of behavioral data collected
• 数据收集方式 Modality of data collection
• 访谈培训与数据收集 Interview Training and Data Collection
• 应用AI实验室与模型构建策略 Applied AI Lab and Model Building Strategy
• 用例与模拟范围 Use Cases and Simulation Scope
• 模拟与现实测试对比 Simulation vs Real-World Testing
• 模拟准确性与评估指标 Simulation Accuracy and Evaluation Metrics
• 多智能体模拟中的误差累积 Error Accumulation in Multi-Agent Simulations
• 收敛与发散的数学框架 Mathematical Framework for Convergence vs. Divergence
• 潜在社会用例与宏观经济模拟 Potential Societal Use Cases and Macroeconomic Simulation
• 社会科学模拟 Simulation for Social Sciences

阅读全文双语转录 →