在20VC的一次引人入胜的采访中，Simile的创始人兼首席执行官朴俊晟（Jun Sung Park）深入探讨了他的AI公司背后的宏伟愿景，该公司旨在模拟并预测未来人类行为。朴俊晟的旅程始于2023年，当时他开展了一个独特的项目，利用大型语言模型（LLMs）创建一个由25个非玩家角色（NPCs）居住的模拟小镇。这些NPCs由GPT 3.5提供支持，并具备记忆、规划和反思能力，展现出令人惊讶的逼真人类行为，甚至自行组织了一场情人节派对。这项名为“Smallville”的实验凸显了LLMs在建模复杂人类互动方面的潜力，并为Simile奠定了基础。 朴俊晟解释说，虽然传统的LLMs专注于为编程或科学等任务打造理性、智能的机器，但Simile的目标却截然不同。他们旨在建立一个“人类行为基础模型”，复制人类的偏见、价值观、偏好乃至错误。他指出，挑战在于弥合“人们所说的”（通常反映在网络数据中）与“人们所做的”之间的鸿沟。为了解决这个问题，Simile不仅收集观察数据，还收集交易数据，并开展大量的随机对照试验和A/B测试。对因果机制和反事实的关注至关重要，因为客户不只想要预测；他们更希望能够 *塑造* 未来。例如，像星巴克这样的公司不仅仅满足于知道销售额会下降；他们想知道 *为什么* 以及 *可以采取哪些行动* 来阻止它。 Simile的数据策略是其核心竞争力。朴俊晟强调获取有代表性的人类数据的重要性，不限于专家，而是包含普通大众，并就他们的生活经历提出深入的探究性问题。他断言，合成样本组在规模和能力上很快将超越人类样本组，这将开启大量目前因预算、时间或实验限制而无法解答的问题。这一转变将允许组织在做出实际决策之前测试几乎所有关于世界的假设。 尽管Simile拥有深厚的研究基础——联合创始人包括斯坦福大学研究员迈克尔·伯恩斯坦（Michael Bernstein）和珀西·梁（Percy Leung，‘基础模型’一词的提出者）——它已在企业领域找到了强大的产品与市场契合度。大型企业面临切实的痛点，迅速采纳了Simile的技术。朴俊晟举例说明Simile如何能在几分钟内预测出数月之久的研究结果，这展示了其在防止代价高昂的错误和优化策略方面的强大能力。他将模拟视为企业的“止痛药”，能够带来显著的投资回报率。 该公司最近的2亿美元融资，使总资本达到3亿美元，由现有投资者提前认购，并由Greenoaks领投。朴俊晟表示，融资速度之快出乎意料，但反映了市场对Simile技术进步的兴奋和认可。他承认研究人才的高昂成本，但强调研究人员被宏伟愿景和潜在影响力所吸引。 展望未来，朴俊晟将LLMs比作“智能的CPU”，将模拟比作“智能的GPU”。Simile旨在创建“与我们一样聪明”的模型，复制人类的多样性以及集体互动中涌现的现象。他设想，未来单次模拟会话可能耗资数千万美元，但能为大型企业或政府带来数亿美元的价值。最终，Simile最宏伟的目标是创建一个社会“表征层”，实现对个体视角进行细致入微、可扩展的表征，从而为政策制定、产品开发和社会决策提供依据。 朴俊晟还分享了关于团队建设的个人见解，强调需要平衡的团队和作为“成功公分母”的个人。他重视那些拥有“两种看似矛盾的‘超能力’”的人，比如一位既具备深度分析能力又直观富有创造力的首席营销官（CMO），或一位对短期保持“偏执”，对长期则“笃信”成功的领导者。这种科学严谨性与创业精神的独特融合，支撑着Simile旨在从根本上改变我们理解和塑造未来的方式的宏伟征程。

In a captivating interview on 20VC, Jun Sung Park, founder and CEO of Simile, delves into the ambitious vision behind his AI company, which aims to simulate and predict future human behavior. Park's journey began in 2023 with a unique project involving large language models (LLMs) to create a simulated town populated by 25 non-player characters (NPCs). These NPCs, powered by GPT 3.5 and enhanced with memory, planning, and reflection capabilities, exhibited surprisingly realistic human behaviors, even self-organizing a Valentine's Day party. This "Smallville" experiment highlighted the potential of LLMs to model complex human interactions and laid the foundation for Simile. Park explains that while traditional LLMs focus on creating rational, intelligent machines for tasks like coding or science, Simile's goal is different. They aim to build a "foundation model of human behavior" that replicates human biases, values, preferences, and even mistakes. The challenge, he notes, is bridging the gap between "what people say" (often reflected in web data) and "what people do." To address this, Simile collects not just observational data but also transaction data and conducts extensive randomized control trials and A/B testing. This focus on causal mechanisms and counterfactuals is crucial, as customers don't just want predictions; they want to *shape* the future. For instance, a company like Starbucks isn't merely interested in knowing sales will drop; they want to know *why* and *what actions* they can take to prevent it. Simile's data strategy is central to its defensibility. Park stresses the importance of sourcing representative human data, going beyond experts to include everyday people and asking deeply probing questions about their life experiences. He asserts that synthetic panels will soon surpass human panels in size and capability, unlocking a vast array of questions currently unanswerable due to budget, time, or experimental limitations. This shift will allow organizations to test virtually every hypothesis about the world before making real-world decisions. Despite its research-heavy foundation—with co-founders including Stanford researchers Michael Bernstein and Percy Leung (who coined "foundation model")—Simile has found strong product-market fit in the enterprise sector. Large corporations, facing acute pain points, have quickly adopted Simile's technology. Park recounts instances where Simile predicted outcomes of multi-month studies in minutes, demonstrating its power in preventing costly errors and optimizing strategies. He views simulation as a "painkiller" for businesses, offering significant ROI. The company's recent $200 million funding round, bringing total capital to $300 million, was preempted by existing investors and led by Greenoaks. Park shares that the rapid pace of fundraising was unexpected but reflects the market's excitement and Simile's technological progress. He acknowledges the high cost of research talent but emphasizes that researchers are drawn to ambitious visions and impact. Looking to the future, Park draws an analogy between LLMs as the "CPU of intelligence" and simulation as the "GPU of intelligence." Simile aims to create models that are "as smart as we are," replicating human diversity and emergent phenomena from collective interactions. He envisions a future where single simulation sessions could cost tens of millions but yield hundreds of millions in value for large enterprises or governments. Ultimately, Simile's most audacious goal is to create a "representational layer" of society, enabling granular, scalable representation of individuals' perspectives to inform policies, product development, and societal decisions. Park also shares personal insights into team building, emphasizing the need for balanced teams and individuals who act as a "common denominator of success." He values those who possess "two contradictory superpowers," like a deeply analytical yet intuitively creative CMO, or a leader who is both "short-term paranoid" and "long-term religious" in their belief in success. This unique blend of scientific rigor and entrepreneurial drive underpins Simile's ambitious journey to fundamentally change how we understand and shape the future.