20VC: The Best AI Companies Have Unique Data Acquisition Strategies...

Host states the fundamental thesis for AI companies: needing a defensible data strategy.

Guest describes the creation of a simulated town populated with 25 NPCs, showcasing complex behaviors and self-organization, which served as an early demonstration of Simile's capabilities.

Guest explains two fundamental contributions of their earlier work: the creation of agents with explicit memory, planning, and reflection components.

Guest details the approach to solving the memory problem for agents by using reflection to make sense of vast amounts of memory.

Host asks for a definition of a simulation model, and the guest clarifies it is the creation of agents to produce activities showing a simulated future.

Guest describes Simile as a company building a foundation model of human behavior for various simulations, not just a simulation model company itself.

Guest explains Simile's relationship with foundation model providers, emphasizing their focus on human behavior, biases, and subjective preferences, rather than pure rationality.

Host inquires about the chasm between what people say and do and its impact on Simile's models.

Guest addresses the say-do chasm by explaining Simile's collection of behavior data, including transaction and observational data, to build causal mechanisms and counterfactual reasoning.

Guest discusses why clients care more about shaping the future than just predicting it, emphasizing the need for cause, mechanism, and counterfactual reasoning.

Guest explains the data collection challenge for Simile, focusing on sourcing representative everyday people and asking the right questions to uncover their core nature.

Guest reiterates the thesis that AI companies need defensible data strategies and outlines Simile's two primary data collection challenges: sourcing representative people and asking the right questions.

Guest elaborates on the type of data collected, including personal life stories and challenges, to understand individuals' core nature.

Host questions the assertion that people don't want to predict the future, using Starbucks as an example of how prediction drives actionable changes.

Guest clarifies that the value of prediction lies in enabling people to *shape* the future and react to potential dips, emphasizing counterfactuals and how to prevent negative outcomes.

Host asks if Simile is a next-generation Qualtrics, and the guest explains their focus on generalizable models of people for simulations, beyond just surveys.

Guest outlines the future potential of simulation, including simulating people interacting, product launches, and addressing "wicked problems" like climate change.

Guest uses the example of democracy failing to illustrate how simulation can model complex decision-making processes and predict outcomes.

Host asks about balancing work with potentially problematic entities, and the guest emphasizes the importance of principles and simulation as a technology for representation at scale.

Guest explains the data requirements for accurate predictions, emphasizing the need to represent the entire population for effective segmentation.

Guest discusses the "data flywheel" effect, where more simulated results lead to better predictions and a compounding advantage.

Guest draws an analogy to coding agents to explain how the data flywheel works in simulation, emphasizing clear reward functions and validation against real-world outcomes.

Guest explains how simulation learns from the world by generating and validating hypotheses daily, making it the best way to learn about the world.

Host asks about the compute required for simulations, and the guest explains the focus on efficiency and technological breakthroughs.

Guest mentions a model that used to cost 100 times more, highlighting the efficiency gains achieved through better modeling.

Guest discusses the serviceable and total addressable market, including enterprise customers and regular consumers.

Guest explains the timing for Simile, serving enterprise customers initially for validation and feedback, with a long-term vision for broader societal use.

Guest describes how they knew they had product-market fit when Fortune 500 board members and C-suites reached out after seeing demos.

Guest details spending a year demonstrating the accuracy and predictive power of their models, leading to the field of synthetic panels.

Host asks if synthetic panels will be larger than human panels in three years, and the guest confidently says yes.

Guest explains that simulation unlocks the limitation of not being able to test every hypothesis, enabling better decision-making.

Guest discusses balancing profit maximization with research purity, emphasizing that Simile is both a product company and a research company.

Guest explains the balance between research and product by highlighting the alignment between technological promise and market demand.

Guest discusses the rapid adoption by enterprise customers, with deals closing in three months, contrary to traditional expectations.

Host asks about the importance of speed vs. accuracy for customers, and the guest states that both are crucial for making informed decisions.

Host asks about the compelling customer conversation of predicting campaign effectiveness, and the guest confirms it's a powerful selling point.

Guest discusses value extraction, focusing on preventing damaging positions that could cost hundreds of millions.

Guest addresses whether Kalshi and Polymarkets can still exist, noting Simile's focus on "how" and "why" events will happen, not just prediction.

Guest discusses the craft of team building, emphasizing balance and recognizing team gaps.

Guest uses a painting analogy to describe how team members should share a similar vibe or essence.

Guest elaborates on team building, looking for individuals who are the "common denominator of success" and can reinvent themselves.

Guest highlights the research backgrounds of co-founders and their transition from academia to entrepreneurship.

Guest discusses seeking co-founders with "two superpowers that don't coexist in one person," like data rigor combined with creative artistry.

Guest describes the "paranoid but religious" archetype in a team, balancing short-term urgency with long-term belief.

Guest explains the difficulty in balancing paranoia and religious belief, and how it drives success.

Guest addresses the high cost of research talent and the war for talent in the Bay Area.

Guest discusses how to attract researchers by focusing on vision and impact rather than just salary.

Guest acknowledges retention challenges in the Valley and the role of leadership in creating a platform for talent.

Guest shares a personal anecdote about their core team members never leaving over six years, emphasizing trust and shared vision.

Host asks about advice for researchers considering starting a company.

Guest advises looking for researchers married to impact, not just problems, who can generate revenue.

Guest discusses the recent $200 million funding round, explaining the preemptive nature and insider interest due to market traction and technological progress.

Guest details the investors in the latest round, including Index Ventures, Green Oaks, and others, bringing total funding to $300 million.

Host asks if the company needed the money, and the guest explains the decision was driven by the opportunity to significantly accelerate progress by increasing input like data and compute.

Guest reflects on what they didn't know about fundraising, initially being skeptical of VC roles but realizing their potential as partners and mentors.

Guest emphasizes the unexpected speed of fundraising rounds and the need to be prepared.

Host asks about a frothy market and hubris, and the guest acknowledges market frothiness but stresses the importance of fundamentals.

Guest confirms that simulation costs vary based on complexity and desired outcomes, with more complex simulations yielding higher ROI.

Guest envisions a future where single simulation sessions cost millions but are worth hundreds of millions due to their value.

Guest discusses the challenge of bridging the gap between proof-of-concept and productionizable technology in AI.

Guest describes the "time machine game" exercise that motivated the development of the simulated town.

Guest compares simulation to the GPU of intelligence, focusing on creating models that are as smart as humans, including their flaws.

Host asks about a crazy prediction for 10 years out, and the guest suggests the possibility of a "replicable twin" for everyone.

Guest discusses the idea of Similee owning a hedge fund due to the insights its simulations can provide.

Guest discusses the potential for simulations to create a representation layer of society and collective intelligence.

Host asks what we assume to be true today that won't be in five years.

Guest believes the assumption that it's impossible to get everyone's perspective will change with simulation.

Host asks if the future of love is similar to simulation in its efficiency.

Guest shares a romantic perspective on love, emphasizing shared journeys and organic growth over efficiency.

Host asks a quick-fire question about the most underrated AI researcher.

Guest discusses the most overheated AI area, cautioning against NeoLabs without clear impact.

Host asks about under-invested AI areas, and the guest highlights simulation, robotics, and the hardware/inference layer.

Guest shares the kindest act they received: a professor's mentorship and connection that opened doors in research.