Jun Song Park 探讨 Simily 如何利用 AI 模拟预测人类行为，以及未来单次模拟可能价值 1 亿美元的潜力。
Jun Song Park discusses how Simily uses AI simulations to predict human behavior and the potential for a $100 million simulation session in the future.
要点 · TL;DR
模拟应建模人类的价值观和偏见，而非超理性智能，以预测和塑造行为。 Simulation should model human values and biases, not super-rational intelligence, to predict and shape behavior.
数据策略，尤其是行为数据和实验数据，是 AI 模拟的最大挑战和关键。 Data strategy, especially behavioral and experimental data, is the biggest challenge and key for AI simulation.
模拟能进行反事实推理，测试因规模或成本而无法实现的假设，从而塑造未来。 Simulation enables counterfactual reasoning and testing hypotheses impossible at scale or cost, shaping the future.
核心观点 · Key points
模拟模型应代表人类的价值观、偏好和品味，而非超级理性智能。 Simulation models should represent human values, preferences, and taste, not super-rational intelligence.
数据策略对AI公司至关重要；收集行为数据和实验数据是关键。 Data strategy is crucial for AI companies; collecting behavioral and experimental data is key.
模拟可以解锁当前因规模或成本而无法进行的假设测试。 Simulation can unlock testing of hypotheses that are currently impossible due to scale or cost.
模拟的价值在于反事实推理和塑造未来，而不仅仅是预测。 The value of simulation lies in counterfactual reasoning and shaping the future, not just prediction.
合成面板将在三年内超过人类面板的规模。 Synthetic panels will surpass human panels in size within three years.
反共识 · Contrarian takes
人们并不真正关心预测；他们关心的是塑造未来。 People don't really care about prediction; they care about shaping the future.
模拟模型应该像人类一样有偏见，而不是优化准确性。 Simulation models should be biased like humans, not optimized for accuracy.
获得反馈的最佳方式是让人们付钱给你。 The best way to get feedback is to ask people to pay you.
研究人员应该与影响力结合，而不仅仅是问题。 Researchers should be married to impact, not just a problem.
几年后，一次模拟会话可能耗资1亿美元。 A single simulation session could cost $100 million in a few years.
本期章节 · Chapters（共 35）
引言与背景Introduction and Background
情人节模拟实验The Valentine's Day Simulation
解决记忆难题Solving the Memory Problem
今日模拟模型Simulation Models Today
Simility 简介Introduction to Simility
与前沿模型的关系Relationship with Frontier Models
言行之间的差距The Gap Between Saying and Doing
数据是最大挑战Data as the Biggest Challenge
预测与反事实的价值The Value of Prediction vs. Counterfactuals