Simulating Anything,
as You Imagine It

In my work on LLM simulation, I’ve noticed that when we talk about agents, we often picture a person first: someone who talks, makes decisions, remembers things, and has a personality.

When I picture an agent, I might see a dog, a room, or something that doesn’t exist in the real world.

I asked two people: what is the dog you imagine like?

One wanted a little white dog, soft, with black eyes, preferably one that could talk. Hearing a sound outside, it would look out the window, then curl up in her arms and tell her what was happening.

The other described a dog begging beside the dinner table. A sound outside sends it barking at the door. Its owner walks over, and it barks even louder. “Stop barking, stop barking.” It keeps going.

The agents I want to study could start from either description. They could begin with everyday experience, or with “I just want something like this.” A talking dog, an ordinary room, a room that gradually changes as visitors come and go—all could be things to explore.

We can imagine what it is, how it acts, and how it relates to its surroundings, then explore how LLMs might enact those ideas.

References
[1] Exploring Causal Mechanisms with Generative Agent-Based Models