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

Summary

In this episode, Joon Sung Park explains why he thinks AI moats will increasingly come from unique data strategies, and why Simile focuses on modeling human mistakes, values, preferences, and taste rather than superhuman reasoning. He walks through the company’s origin in early agent research, including a 25-NPC town simulation that used memory, planning, and reflection, and describes how that work evolved into a broader human-behavior foundation model. Park says Simile’s training assets include transaction data, observational data, customer/vendor data, and randomized controlled trials or A/B tests, because customers care less about raw prediction than about shaping future outcomes. He also discusses enterprise go-to-market, claiming some Fortune 500 deals closed in three months, and says the company’s $300 million raise will fund more data, compute, and spend to accelerate progress. Looking ahead, he predicts simulation sessions could become extremely expensive to run but still worth up to $100 million to customers, and frames simulation as a potential “GPU of intelligence” alongside large language models as the “CPU.”

Notes

• Simile’s core thesis is that AI moats come from defensible data acquisition, not just model scale.
• Park frames simulation as a high-value product category with potentially massive willingness to pay.
• The company’s roots are in early agent systems, before ChatGPT, using GPT-3.5/text-davinci.
• The 25-NPC town simulation produced emergent social behavior like self-organized parties and Valentine’s planning.
• Memory, planning, and reflection were the key architectural pieces in those early agents.
• Reflection was used to compress many memories into higher-level beliefs and preferences.
• Simile’s objective is to model human-like errors, bias, values, preferences, and taste.
• That puts Simile in a different lane from frontier LLM labs focused on coding, math, and science.
• Simile trains on behavior data, transaction data, observational data, and customer/vendor data.
• He argues that web text captures what people say, while Simile needs what people do.
• Randomized control trials and A/B tests are core parts of Simile’s training asset.
• The company is focused on causal mechanisms and counterfactuals, not just correlation.
• Park says customers care about shaping outcomes, not merely forecasting them.
• He illustrates the point with a Starbucks example: the useful question is how to prevent a bad outcome.
• Simile’s enterprise wedge is market research, because it has budget and a tight validation loop.
• Enterprise customers also help Simile validate when its simulations are right or wrong.
• Fortune 500 board members and C-suites immediately saw the market potential in the Smallville demo.
• Simile says its models predict behaviors and attitudes 85% as accurately as people replicate their own.
• That work, released at the end of 2024, helped seed the synthetic panels category.
• Synthetic panels may overtake the current human panel market within three years.
• He says current research workflows answer only a small fraction of possible questions.
• A production model was reduced from roughly 100x higher run cost to much cheaper inference.
• Inference-time efficiency improvements are part of Simile’s productization path.
• Park says the world is the ground truth, enabling tens of thousands of hypotheses per day.
• He describes a closed-loop validation system where hypotheses can be checked against real-world outcomes over time.
• Simile says some of the largest customers moved with unusual speed for enterprise.
• Park claims Simile can close deals in roughly three months with major enterprise buyers.
• Simile’s value proposition is explaining how outcomes happen, not only what will happen.
• Customers want the causal path so they can prevent or encourage a future outcome.
• Park says strong teams are balanced and built around gaps being filled ahead of time.
• He says a great team preserves consistent values and rigor as it scales.
• He uses top CMOs as an example of rare dual-skill leadership profiles.
• His cofounder Lainey combines short-term paranoia with long-term conviction.
• Park argues paranoia is useful because it drives preparation, urgency, and execution.
• Top research talent can command compensation in the tens of millions.
• Researchers are often motivated more by vision and impact than by salary alone.
• He says leadership should create a platform where each person can express their superpower at maximum degree.
• He is proud that his core team members have stayed with him for six years.
• He advises looking for researchers married to impact, not just a narrow problem.
• Simile raised a $100 million round about five months earlier and then raised another $200 million.
• That brought Simile’s total funding to $300 million in roughly six months.
• He says the new capital is meant to increase data, compute, and spend to accelerate research.
• Park says research is about controlling inputs and process, not guaranteed outcomes.
• He says simulation workloads get expensive as the decision becomes more complex and higher stakes.
• More expensive simulations can still produce higher ROI for customers.
• He predicts a single simulation session could cost $10 million to $20 million to run.
• The same session could be valuable enough that customers pay $100 million.
• The main challenge is bridging proof of concept to productionizable technology.
• Park sees the hard part as productization, not initial simulation feasibility.
• He frames today’s large language models as the “CPU of intelligence.”
• He thinks simulation could become the “GPU of intelligence.”
• Simile wants models that are as smart as humans, not superintelligent.
• The long-term vision is representation at scale, with a simulated twin for each person.
• If perfect simulation exists, even the stock market could change materially.
• He argues society may no longer need human representatives as approximations of everyone’s perspective.
• He thinks a representation layer for collective intelligence is plausible.
• He is personally bullish on more organic human relationships, even if simulation could improve matching.
• Large-lab researchers who do not publish may be underrated.
• He sees unclear-vision neoclouds as potentially turning into research projects rather than viable companies.
• Future AI companies should have access to hard-to-collect data.
• He highlights robotics, the inference layer, chips, and hardware as interesting AI investment areas.