We Learned to Forecast the Weather. Can We Forecast Society?
Simile just raised $100M to find out.
In 1950, a team at the Institute for Advanced Study in Princeton ran the first numerical weather prediction on the ENIAC computer. It took 24 hours to compute a 24-hour forecast. The results were mediocre. Critics called it a waste of time. Why bother simulating the atmosphere when you could just look out the window?
Seventy-five (76 now! dang, 2026) years later, weather forecasting is arguably the most successful prediction system humanity has ever built. We simulate the entire planet’s atmosphere on supercomputers, and the results save thousands of lives per year. A 5-day forecast today is more accurate than a 1-day forecast was in 1980.
Last week, a startup called Simile emerged from stealth with $100M in Series A funding to attempt something structurally similar — but instead of simulating air molecules and pressure systems, they’re simulating people. Their bet: if you can build high-fidelity digital twins of real humans, ground them in behavioral data, and run them forward in time, you can forecast social outcomes the way we forecast storms.
The Architecture That Started It All
To understand why this matters, you need to understand what Simile’s founders actually built before they started a company.
In April 2023, Joon Sung Park and his collaborators at Stanford published a paper called “Generative Agents: Interactive Simulacra of Human Behavior”. It won Best Paper at UIST and became one of the most discussed AI papers of the year. Andrej Karpathy called it out. The New Yorker covered it. For good reason — it introduced something genuinely new.
The paper described an architecture for AI agents that don’t just generate plausible single-turn responses (which any LLM can do), but maintain coherent behavior over extended periods of time. The trick was three interlocking components: a memory stream that logs every experience, a reflection mechanism that synthesizes raw observations into higher-level self-knowledge, and a planning system that generates and dynamically updates multi-level schedules.
They tested it by dropping 25 agents into a sandbox world called Smallville — think The Sims but powered by GPT. Each agent was initialized with just a single paragraph of backstory. Then they let the simulation run.
What happened was remarkable. One agent was given the intent to throw a Valentine’s Day party. Without any scripting, the agents autonomously spread invitations through organic conversation, decorated the venue, asked each other on dates, and showed up at the right time. Information about a mayoral candidacy diffused through the social network from one agent to thirteen over two simulated days. Agents formed new relationships, recalled past conversations in contextually appropriate ways, and made plans that reflected their personality and experience.
In controlled evaluations, the full architecture beat every ablation and — critically — beat human crowdworkers who watched replays of each agent’s life and tried to role-play their responses. The effect size against the baseline (representing the prior state of the art) was Cohen’s d = 8.16. Eight standard deviations.
The follow-up paper in November 2024 scaled this to 1,000 agents grounded in real-world interview data, validating simulated attitudes against actual survey responses at 85% accuracy. That paper is the direct scientific foundation for Simile.
Back to the Weather
Weather forecasting works because we identified the fundamental dynamics (Navier-Stokes equations, thermodynamics), built instruments to observe current conditions at scale (satellites, weather balloons, ground stations), and developed computational infrastructure to run simulations forward in time (supercomputers). The output is probabilistic, not deterministic — we forecast that there’s a 70% chance of rain, not that it will rain at exactly 2:37 PM. And the models improve continuously as we get better data, better models, and better compute.
Simile’s approach maps onto the same framework with surprising fidelity. The fundamental dynamics are human behavioral patterns — how people form opinions, make decisions, respond to stimuli. The observation layer is the data they train on: interviews with real people about their lives, transaction histories, behavioral science literature, and now a partnership with Gallup for nationally representative panel data. The compute layer is the generative agent architecture running population-scale simulations forward under different scenarios.
And just like weather models, the output is probabilistic. You’re not predicting that Jane Smith will buy Product X at 3 PM on Tuesday. You’re predicting that a population segment with these characteristics has a 73% likelihood of responding positively to this product positioning. That’s the same epistemic structure as forecasting rainfall probability over a region.
The trajectory is similar too. Early weather models were crude. They got the big patterns roughly right and the details badly wrong. Nobody uses them to plan a picnic, but they’re useful for knowing whether a hurricane is coming. Simile is probably at a similar stage — useful for big-picture directional insights, not yet reliable for predicting precisely what your specific customer will say.
The Deeper Question
Strip away the company-specific analysis and there’s a genuinely profound question here: is human behavior simulable in the same sense that weather is?
Weather is governed by known physical laws. The equations are right — it’s the initial conditions and computational resolution that limit accuracy. Human behavior might not have “equations” at all, at least not at the individual level. We have heuristics, biases, cultural norms, emotional states, and free will (or its functional equivalent). Weather is chaotic but deterministic. Human society is chaotic and potentially non-deterministic.
But here’s the counterargument: you don’t need to simulate every molecule to forecast the weather. You need to simulate enough of the dynamics at the right level of abstraction to make useful predictions. And the same might be true for society. You don’t need to predict what every person will do. You need to predict aggregate patterns with enough accuracy to beat the alternative — which is focus groups, surveys, intuition, and gut feel.
My Take
Weather forecasting started as a curiosity and became essential infrastructure. Social simulation might follow the same arc — but on a longer timeline and with a lower ceiling of accuracy. The question isn’t whether Simile can predict the future. It’s whether their simulations are consistently useful enough that decision-makers would feel naked without them.
In my opinion, The dream endgame looks something like this: prediction markets provide the calibration benchmarks and behavioral training data, Gallup provides the demographic grounding, and generative agents synthesize it all into a simulation that can answer questions none of those individual tools can answer alone.