20VC: The Best AI Companies Have Unique Data Acquisition Strategies | Will Simile Kill Kalshi, Polymarkets and NASDAQ | How to Sign Fortune 500 Companies As Customers in Weeks with Joon Sung Park, Simile

Joon Sung Park founded Simile, a simulation company that models human behavior using AI agents with memory, planning, and reflection capabilities. The company's origins trace to a 2023 Valentine's Day project where Park created a virtual town with 25 NPCs that self-organized and planned parties, demonstrating that large language models could extract realistic human behaviors from training data. This early work established the technical foundations for memory systems (using markdown text files) and reflection mechanisms that allow agents to synthesize experiences and develop personalities.

Simile positions itself as building a "foundation model of human behavior" rather than a simulation company—this foundation can be used to create individual simulations, subpopulation simulations, and eventually entire ecosystem simulations. The key differentiator from general-purpose LLMs like those from OpenAI is that Simile optimizes for behavioral realism and human-like mistakes rather than perfect rationality. Park emphasizes the gap between what people say and what they do (say-to-give problem), arguing that most LLM training data reflects stated behavior, not actual behavior.

The company's data strategy centers on collecting behavioral data—transaction data, observational data, and crucially, randomized control trial results. Park argues that observational data is good for prediction but inadequate for the counterfactual reasoning his customers need. Enterprises don't want to know what will happen; they want to know how to prevent negative outcomes or encourage positive ones. This requires causal mechanisms and counterfactual understanding, not just correlation.

On the enterprise side, Simile has achieved remarkable sales velocity, closing deals with Fortune 500 companies like CVS within three months rather than the traditional year-plus enterprise sales cycles. Park attributes this to acute customer pain around slow experimentation, limited budgets, and inability to test ideas. Early validation came when Simile ran simulations that replicated three-to-six-month consulting studies in two minutes. The company has demonstrated 85% prediction accuracy compared to human replication rates and is building what the industry calls "synthetic panels"—AI-powered market research replacing human panels.

On team building, Park emphasizes hiring for "common denominator" success—people who were the reason projects succeeded and can reinvent themselves across domains. He also values finding leaders with seemingly contradictory superpowers, such as CMOs who are both deeply analytical and creative. His co-founder Lainey exemplifies the ideal of being simultaneously paranoid short-term and religiously long-term optimistic—a balance that drives execution without complacency.

Regarding funding, Simile raised $100 million six months ago, then raised an additional $200 million led by Grelock Partners and insiders at Index Ventures. Park noted he didn't initially seek the second round but accepted it to increase compute spend and data collection, allowing meaningful acceleration of research progress. He also reflected on unexpected lessons from fundraising: the genuine mentorship value VCs can provide and how market interest moves faster than expected.

On data collection challenges, Park states that data acquisition is the hardest element of building simulation models. The company sources everyday people (not experts) who are representative of real-world populations and conducts extensive interviews asking about life stories, major decisions, and formative experiences. The question of how much data is needed depends on segmentation; roughly 1,000 people provides statistical significance for narrow populations, but Simile aims to represent entire populations to support on-the-fly filtering queries.

The data flywheel mechanism works through continuous validation against ground truth. Rather than waiting years for outcomes, Simile generates tens of thousands of daily hypotheses and validates them as the world unfolds. This provides learning signals superior to traditional AI training in coding (which has clear accept/reject signals) because the entire world becomes the ground truth.

Park envisions simulation reaching a stage where single simulation sessions cost $10-100 million to run but generate such valuable insights that enterprises will pay accordingly. He sees simulation as eventually addressing "wicked problems" requiring collective action across stakeholders with different incentives, such as climate change or democratic governance.

On the competitive landscape, Park addresses whether Simile threatens prediction markets like Kalshi and Polymarket by noting the difference: prediction markets forecast what will happen, while Simile reveals how and why it happens and how to change it. He's also careful about ethical considerations, viewing simulation as a powerful technology with potential for misuse and emphasizing that its best application is ensuring diverse perspectives are heard in decision-making.

Park describes his background as an atypical researcher—he had no research experience during undergrad and lived in a garage after college. He credits Professor Mary Wooldridge at Stanford with taking an undeserved chance on him by spending a full morning providing guidance and introductions that launched his research career, a kindness he remains deeply grateful for.