Why is all the good Vietnamese food in the North West of Adelaide? Are they racist?

I like living close to my friends. I also make friends easier with people who have similar life experiences to me. I don’t want everyone to be just like me, but I’d like maybe 1 in 3 people to have a similar value system to me (e.g. political views).

Schelling's Segregation Model

A simulation showing how individual preferences lead to neighborhood segregation

About this simulation:

Based on Thomas Schelling's 1971 model, this demonstrates how even mild individual preferences (e.g., wanting just 30% of neighbors to be similar) can lead to highly segregated neighborhoods at the population level.

Controls

Number of different groups in the simulation

Minimum % of similar neighbors desired

Metrics

Iteration: 0
Agent Happiness: 0.0%
Segregation Index: 0.0%

% of neighbors that are the same ethnicity

Ethnicities in Area: 2

Thomas Schelling’s 1971 model shows something weirder: even when people only want 30% similar neighbors, you get 70-90% segregated neighborhoods. Individual preference, collective outcome.

This is model-based reasoning. Instead of arguing about what “causes” segregation, you build a simple model and see what happens.

What’s interesting is I have a concrete idea of what is a reasonable “similar neighbours preference”, but before running the model I had no idea what type of segregation that would cause.

Why this matters now

LLMs make building models trivial. Before, you’d need to:

  • Learn a programming language
  • Figure out visualization libraries
  • Debug for hours
  • Maybe give up

Now you can describe what you want and get a working simulation in minutes.

When to use models

Models are useful when:

  • Intuition fails at scale. What feels right for 10 people breaks at 1000.
  • Feedback loops exist. A causes B causes C causes A. Your brain can’t track this.
  • You’re arguing about mechanisms. “Does X cause Y?” is better answered with “here’s a world where X exists, does Y happen?”

Models are useless when:

  • You just need data. Don’t simulate customer behavior, talk to customers.
  • The model is more complex than reality. If it takes longer to build than to test, just test.
  • You’re using it to confirm what you already believe. Models are for exploring, not proving.