Can AI simulation of 8 billion people help solve climate change and other wicked problems?
要点 · TL;DR
模拟人类行为需要建模真实行为，而非仅依赖自我报告的态度，并利用随机对照试验数据理解因果关系。 Simulating human behavior requires modeling actual behavior, not just self-reported attitudes, using RCT data for causal understanding.
使用随机对照试验数据对模型进行后训练，可将个体反应复现的预测准确率提升至 85%。 Post-training models with RCT data improves prediction accuracy to 85% in replicating individual responses.
模拟是辅助人类决策的工具，其市场远超市场研究，可影响所有关于人类决策的制定。 Simulation is a tool for human decision-making, with a market larger than market research, informing every decision about humans.
核心观点 · Key points
模拟人类行为需要建模人类的“暗知识”——真实行为，而不仅仅是网上自我报告的态度。 Simulation of human behavior requires modeling the 'dark knowledge' of humanity—actual behavior, not just self-reported attitudes online.
来自随机对照试验的行为数据对于理解因果机制至关重要，使模拟能够塑造未来，而不仅仅是预测未来。 Behavioral data from randomized controlled trials is crucial for understanding causal mechanisms, enabling simulations to shape the future, not just predict it.
使用随机对照试验数据对模型进行后训练，显著提高了其预测人类行为的能力，在复制个体反应方面达到了85%的准确率。 Post-training models with RCT data significantly improves their ability to predict human behavior, achieving 85% accuracy in replicating individual responses.
模拟是人类决策的工具，其市场远大于市场研究，可能为每一个关于人类、为了人类做出的决策提供信息。 Simulation is a tool for human decision-making, with a market far larger than market research, potentially informing every decision made about humans for humans.
愿景是模拟80亿人，从而能够回答像气候变化这样复杂的“棘手问题”，这些问题涉及相互竞争的激励因素。 The vision is to simulate 8 billion people, enabling answers to complex societal problems like climate change, which are 'wicked problems' with competing incentives.
反共识 · Contrarian takes
前沿模型是“超级理性的客观机器”，会忽略人类的偏见和错误；模拟必须建模“愚蠢”的人类，而不是超级智能的。 Frontier models are 'super rational objective machines' that miss human biases and mistakes; simulations must model 'dumb' humans, not superintelligent ones.
简单的基于人设的模拟（如十亿人设论文）之所以失败，是因为它们依赖已知统计，忽略了关于个体的丰富、小众知识。 Simple persona-based simulations (like the billion-persona paper) fail because they rely on known statistics, missing the rich, niche knowledge about individuals.
模拟就像绘画：它突出主题的本质，而不是完美的再现，这正是它的价值所在。 Simulation is like painting: it highlights the essence of the subject, not a perfect representation, and that's what makes it valuable.
模拟的成本不是障碍；它比现实世界的研究更便宜，并且通过提供更好的决策信息可以节省数百万美元，即使其成本与训练基础模型相当。 The cost of simulation is not a barrier; it's cheaper than real-world studies and can save millions by informing better decisions, even if it costs as much as training a foundation model.
我们是否生活在模拟中并不重要；我们的体验是真实的，所以我们应该关注模拟对社会可能产生的影响。 Whether we live in a simulation or not doesn't matter; our experience is real, so we should focus on the impact simulation can have on society.
本期章节 · Chapters（共 31）
开场愿景Opening Vision
赞助商信息Sponsor Message
Joon的人生故事Joon's Life Story
生成式代理论文Generative Agents Paper
时间机器练习The Time Machine Exercise
个人代理改进期望Desired Improvements in Personal Agents
行为建模数据类别Data Categories for Behavior Modeling
获取行为数据Acquiring Behavioral Data
客户用例与定制Customer Use Cases and Customization
概念测试与用例Concept testing and use cases
政治与盖洛普合作Politics and strategic partnership with Gallup
介绍与类似项目Introduction and Similar Projects
预注册与数据源Pre-registration and data source
模型粒度与训练Model granularity and training
人类问题与模型局限Human problems vs model limitations
期望数据集与价值Desired datasets and value
与组合方法比较Comparison with combinatorial approaches
模拟规模与涌现Simulation Scale and Emergence
基于代理模型的机会Opportunity in Agent-Based Models
其他案例与WealthfrontOther case studies and Wealthfront