Stanford Smallville is officially open-source!

Imagine a digital realm where 25 AI agents live in a simulated Westworld, unaware they’re in a simulation. They work, gossip, socialize, make friends, and even fall in love. Each has a unique personality and backstory.

In 2023, Smallville is a standout AI experiment. We often discuss the emergent abilities of single AI, but the complexity of multi-agent emergence on a larger scale is even more captivating. A population of AI could simulate the evolution of an entire civilization.

This marks the start of endless possibilities, with gaming likely being the first to experience the impact.

In this game, 25 ChatGPT AI agents interact with each other in natural language, plan their days, form relationships, and even engage in self-reflection. They discuss real-world topics such as upcoming elections, and some of them even throw a Valentine’s Day party!

What’s remarkable is that the agents’ interactions resemble those of real humans. They ask each other’s opinions, spread invitations to the party, and coordinate to show up at the right time. The agents are informed citizens and reflect on their own personalities and interests.

But here’s the kicker: when scientists at Google asked a separate set of evaluators to compare the AI agents’ behaviour with that of humans who played the game, the evaluators found the AI agents’ behaviour more human-like than the humans themselves! Yes, you heard it right.

The AI agents’ behaviour was deemed more human-like because they displayed more diversity and richness in their interactions. Humans, on the other hand, tended to stick to the same topics and actions.

It’s amazing to see how far we’ve come in AI research, and it’s exciting to imagine the possibilities that lie ahead.

The paper “Generative Agents: Interactive Simulacra of Human Behavior” is available here: https://arxiv.org/abs/2304.03442

The pre-computed replay of the simulation that accompanies the paper: https://reverie.herokuapp.com/arXiv_Demo/

Github: https://github.com/joonspk-research/generative_agents