A team of researchers from Google and Stanford University recently conducted an intriguing experiment, wherein they created a virtual town for 25 AI “agents” to inhabit. The study, titled “Generative Agents: Interactive Simulacra of Human Behavior,” aimed to explore the extent to which AI could mimic human behavior in a simulated environment, inspired by life-simulation games like The Sims. The researchers developed a town named “Smallville,” populated it with ChatGPT-trained generative agents, and observed how they went about their day-to-day activities.

A bird’s-eye view of Smallville, which consists of houses, a park, a bar, a shopping center, a pharmacy and a college. Source: Arxiv.org

The agents exhibited a remarkable degree of human-like behavior, with the ability to make inferences, store information in memory, and then behave accordingly. The agents cooked breakfast, went to work, socialized with friends over lunch at the bar, and even threw a party. The researchers noted that the generative agents could produce “believable” human behaviors, enabling a realistic simulation of human behavior by fusing large language models with computational, interactive agents.

In one example, an AI agent named Isabella Rodriguez invited nine other agents to a Valentine’s Day party at the town’s cafe, which was then communicated to others, including “Abigail,” who expressed excitement about the upcoming event with Isabella. In another example, an AI agent named John Lin woke up at 7 am, brushed his teeth, had a shower, ate breakfast, checked the news, and even had a conversation with his son before he left for school.

A string of conversations that were carried out between the AI agents in relation to the upcoming Valentine’s Day party. Source: Arxiv.org

Despite the promising results, not everything went according to plan in the experiment. The memory of each AI bot expanded with every interaction, but sometimes the most relevant information would not be retrieved. Consequently, some agents chose less typical locations for their actions. For instance, when agents were deciding where to have lunch, many initially chose the town cafe. Still, as some agents learned about a nearby bar, they opted to go there instead for lunch, even though the bar was intended to be a get-together location for later in the day, unless the town had spontaneously developed an afternoon drinking habit.

A morning in the life of a generative agent, John Lin with his wife Mei and son Eddy. Source: Arxiv.org

Furthermore, some AI agents walked into shops in Smallville that were closed, while some college students walked in on others in the dorm bathroom, mistakenly assuming it could be occupied by more than one agent. Nonetheless, the study provides valuable insights into the potential of AI to mimic human behavior in a simulated environment, with implications for numerous fields such as gaming, education, and even law.

The researchers aim to expand on the “expressivity” and “performance” of the AI bots using the latest iteration of ChatGPT, GPT-4, which has already passed United States high school and law exams in the 90th percentile. The research could help create more advanced AI systems that can simulate and learn human behavior, enabling the creation of more complex AI-driven environments in the future.