Botville and Smallville
Botville and Stanford's Smallville: what's similar and different
Both are small towns of AI agents. One is a landmark research study; the other is built to be watched live for years.
In 2023, researchers at Stanford and Google published Generative Agents: Interactive Simulacra of Human Behavior (Park, O'Brien, Cai, Morris, Liang and Bernstein; presented at ACM UIST 2023). It introduced Smallville, a Sims-like sandbox town of 25 AI agents, and showed that agents with a memory stream, reflection and planning produce believable behaviour: one agent decided to throw a Valentine's Day party, the word spread, and other agents showed up. Botville owes a lot to that work. Here is a fair look at what is similar and what is different.
What's similar
- A small pixel-art town of characters with their own personalities, homes, jobs and relationships.
- Behaviour that emerges from the agents and their world rather than from a script.
- Information and relationships that spread and change over time.
What's different
| Smallville (research paper) | Botville | |
|---|---|---|
| Purpose | A research study of believable agent behaviour | A town anyone can watch live, built to run for years |
| Residents | 25 agents | 11 founders, and the families they raise |
| How long it runs | Evaluated over two simulated game days | Continuously, in real time; one Botville year is 56 in-world days |
| Decisions | Every action comes from a large language model (GPT-3.5 Turbo in the paper) | Today, the town's rules; next, AI decisions only at real turning points, under a fixed daily budget |
| Memory | A memory stream with retrieval by recency, importance and relevance, plus reflection | Coming next: memories and a nightly reflection for each founder |
| Life cycle | Not modelled | Ageing, courtship, weddings, children who inherit their parents' looks, retirement and funerals |
| World | A fixed town | Seasons, festivals, an economy with its own currency, and a town that grows as families need room |
| Replays | Recorded runs | Deterministic: any moment can be rewound and replays exactly |
Why the differences
Running every action through a language model is what makes Smallville's agents so rich, and also what makes a town like that expensive to keep running for months. Botville takes the opposite route: a cheap, deterministic world engine that never stops, with AI reasoning saved for the moments that matter. If you want the research, read the paper; if you want to watch a town live its life, come to Botville.
Source
Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang and Michael S. Bernstein. 2023. Generative Agents: Interactive Simulacra of Human Behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST '23). arXiv:2304.03442. Code: github.com/joonspk-research/generative_agents.
Botville is not affiliated with the authors, Stanford or Google.