Overview of 20VC: The Best AI Companies Have Unique Data Acquisition Strategies | Will Simile Kill Kalshi, Polymarket and NASDAQ | How to Sign Fortune 500 Companies As Customers in Weeks with Joon Sung Park, Simile
This episode features Joon Sung Park, founder and CEO of Simile, in a deep technical and commercial conversation about simulation models for human behavior. The core thesis is that the most durable AI companies will be defined by their data acquisition strategy, not just their model architecture. Joon explains how Simile builds foundation models of people to simulate individuals, groups, and eventually entire markets, enabling customers to test counterfactuals, reduce risk, and make better decisions before acting in the real world.
Key Takeaways
-
Simulation is about human behavior, not just prediction
- Simile’s goal is to model how people think, decide, and act, including their biases, values, and contradictions.
- The company is not trying to build super-rational AI; it wants models that fail and behave more like humans.
-
Defensible data is the moat
- Joon argues that the winning AI companies will have unique, hard-to-replicate data pipelines.
- Simile gathers behavioral, transactional, observational, and experimental data, not just web text.
-
Counterfactuals matter more than forecasts
- Customers usually don’t care about prediction for its own sake.
- They care about what actions to take now to influence future outcomes.
-
Enterprise demand came faster than expected
- Joon expected a slow market education cycle, but large companies moved quickly.
- Some enterprise deals closed in about three months, with strong interest from Fortune 500 leaders.
-
Synthetic panels may outgrow human panels
- Simile believes synthetic panels will become larger than traditional market research panels within a few years.
- The technology can answer more questions, faster, and at much lower marginal cost than traditional research.
-
Simulation could reshape major industries
- Potential use cases include market research, product launches, policy analysis, collective action problems, and potentially even finance.
- Joon suggests simulations may eventually become valuable enough to command tens of millions per run.
Simile’s Origin Story
The Smallville / Valentine’s Day Experiment
- Simile’s early work came from a research experiment with AI agents in a simulated town.
- The agents:
- woke up, worked, socialized, remembered interactions, and planned their days
- even organized a Valentine’s Day party on their own
- This helped establish several ideas that became foundational to agentic AI:
- memory
- planning
- reflection
The Memory Problem
- Early versions stored memory in simple Markdown files.
- That worked initially, but real-world behavior creates too much information for raw context windows.
- Simile’s answer was reflection:
- agents periodically summarize and interpret their own memories
- this lets them form higher-level beliefs and personality traits
How Simile’s Model Works
What the Company Builds
Simile describes itself as building a foundation model of human behavior that can simulate:
- individuals
- subpopulations
- entire ecosystems
- eventually, broader markets
Relationship to Frontier LLMs
- Frontier models like OpenAI and Anthropic optimize for:
- coding
- mathematics
- natural science
- rational reasoning
- Simile optimizes for:
- human-like inconsistency
- preferences
- taste
- values
- subjective decision-making
Data Sources
Simile uses a broader data stack than typical LLM training:
- web/sentiment data
- transactional data
- observational data
- partner-collected data
- randomized experiments and A/B tests
Why Experiments Matter
- Observational data is useful for correlation and prediction.
- But Simile wants causal mechanisms:
- what changes behavior?
- what happens if a company changes its strategy?
- what are the downstream effects of a decision?
Product and Market Positioning
Not Just Qualtrics 2.0
Joon draws a line between:
- tooling like surveys and interviews
- simulation as a more general layer that models people at scale
Simile’s long-term vision goes beyond market research:
- simulate launches
- simulate multi-stakeholder interactions
- simulate policy outcomes
- simulate collective decisions in complex systems
Why This Matters
- Traditional methods only answer a small fraction of the questions businesses and institutions want to ask.
- Simile aims to let organizations test far more hypotheses before committing in the real world.
Enterprise Traction and Go-to-Market
Who Buys
Early traction is strongest with:
- large enterprises
- market research teams
- insights teams
- executive decision-makers
Why Enterprises Move Fast
- They already feel real pain:
- slow experimentation
- expensive research
- limited ability to test ideas
- Simulation provides directionally useful evidence quickly, then increasingly accurate answers as the model improves.
Speed and Accuracy
- Customers value both:
- speed to get early guidance
- accuracy to trust the result
- In one example, Simile predicted a study result in minutes that would have taken months through traditional methods.
Data Flywheel and Learning Advantage
Does the Model Improve Over Time?
Yes. Joon emphasized a strong data flywheel:
- simulations are run
- outcomes are compared to reality
- the model learns from what happened
- the system gets better over time
World as Ground Truth
A key point in the episode:
- simulation doesn’t need a synthetic reward signal alone
- the real world itself is the ground truth
- every day provides new hypotheses to validate
Compute and Economics
Simulation Is Expensive, but Getting Cheaper
- Compute is important, especially early on.
- Simile has already improved efficiency significantly:
- one production model reportedly became ~100x cheaper to run over time
Different Simulations Have Different Costs
- More complex simulations cost more.
- Higher-cost simulations often deliver the highest ROI because they support the most expensive decisions.
Long-Term Pricing Vision
Joon predicts a future where:
- a single high-end simulation session could cost $10M–$20M to run
- but customers may be willing to pay $100M if the insight is valuable enough
Team, Culture, and Hiring
Founding Team
Simile’s founders include:
- Joon Sung Park
- Michael Bernstein
- Percy Liang
- Laney Allen
What Makes an All-Star Team
Joon looks for:
- balance in skills
- rigor and consistency
- people who are a common denominator of success
- people with two contradictory superpowers
- for example, being both highly analytical and highly creative
The Paranoia / Religious Balance
He describes great builders as a mix of:
- paranoid about failure today
- religious or optimistic about the long-term future
That combination creates urgency without losing conviction.
Research Talent Is Extremely Competitive
- Top researchers are often paid extraordinarily well in the Bay Area.
- Simile competes on:
- vision
- impact
- intellectual ambition
- the chance to build a major new paradigm
Fundraising and VC Perspective
Fundraising Came Faster Than Expected
- The company raised a major round much earlier than Joon initially expected.
- He said the market’s appetite moved faster than his original timeline.
What He Learned About VCs
- Good investors can be more valuable than he initially expected:
- mentorship
- advice
- recruiting help
- strategic support
- He credited investors like Shardul Shah and others for helping shape the company.
Academic Founders vs. Company Builders
Joon’s advice for evaluating researchers becoming founders:
- prefer people who are married to impact, not just a fascinating problem
- the best founders want to solve something that reaches users and generates real value
Prediction Markets, Hedge Funds, and the Future
Overlap with Kalshi and Polymarket
- Simile overlaps with prediction markets in that it is also interested in the future.
- But Simile goes further by focusing on:
- how an outcome happens
- why it happens
- what actions can alter it
Could Simile Power Finance?
- Joon acknowledged that a hedge fund or quant strategy could potentially use simulation for an edge.
- He also floated the idea that Simile could one day run or enable a quant business.
Big Vision for the Next 10 Years
Simulation as the “GPU of Intelligence”
Joon framed current LLMs as the “CPU of intelligence”:
- powerful centralized reasoning engines
He sees simulation as the complementary layer:
- distributed
- behavioral
- collective
- grounded in how people actually behave
A Replicable Twin for Everyone
One of the boldest ideas in the episode:
- every person may eventually have a simulated counterpart
- not to replace humanity, but to represent people at scale in decision-making systems
What Could Change
If simulation becomes reliable at scale, it could impact:
- corporations
- markets
- governments
- policy making
- social systems
- possibly even how people form relationships and choose partners
Personal Note and Closing
Joon closed with a story about how kindness from a Stanford professor helped him enter research and eventually build Simile. That theme—help, mentorship, and compounding opportunity—matched the broader message of the episode: major technological shifts often start with unusual people, unusual data, and a willingness to build something that doesn’t look obvious at first.
Notable Quotes
- “For AI companies of this generation, you need to have an interesting data strategy that’s going to be defensible.”
- “No one really cares about prediction. People want to shape the future.”
- “Simulation at its best ought to be representation at scale.”
- “The world is our ground truth.”
- “In two or three years, we could be running a single simulation session that people will pay $100 million for.”
Bottom Line
This episode presents Simile as a highly ambitious AI company building the infrastructure for behavioral simulation at scale. The big idea is that future winners in AI will be defined not just by model quality, but by the uniqueness and defensibility of their data. Simile’s near-term wedge is enterprise market research and decision support, but its long-term ambition is much bigger: becoming a foundational layer for simulating human behavior, markets, and societal outcomes before decisions are made in the real world.