What if you could predict how your customers responded to releases before they went out into the real world? Or if you could get a signal on who will win an election?

We simulated the population of San Francisco distributed across real US Census data and achieved near-parity with these types of real-world events.

It now spans five US cities — San Francisco plus neu york (New York City), synth la (Los Angeles), cybercago (Chicago), and simami (Miami) — each its own Census-sampled population. Switch between them from the title in the top-left.

How do we know if our simulation of SF is accurate? The honest test needs a model whose knowledge cutoff predates the event, so it can't have just memorized the outcome. The older Claude models with a 2023 cutoff have since been retired — so for this leakage-free backtest we ran GPT-4o (knowledge cutoff October 2023) and asked the twin city to predict results it had never seen. (The live app you're using runs the current Claude Sonnet 4.6.)

All of decision-making is predicated on our understanding of causality. How will the choices we make influence the world? How will people react?

To enable (first) SF, and then the world, to understand the causal nature of reality, we present sim francisco.