For decades, there has been a vision of creating believable simulacra of human behavior through computer generative behavior. This vision involves creating behavior that is so compelling and human-like that it provides an illusion of life. Today, I’m going to introduce a new way of simulating human behavior in fully general computational agents using a language model like CHBT with a novel agent architecture that can remember, plan, and reflect based on constantly growing memories and acting social dynamics.
These agents can not only plan and lead a believable day in life where they wake up in the morning, do their routines, and go to work as individuals in a sandbox game environment without any hardcoded scripts, but they can also come together to give birth to an entire artificial society of their own. This concept has been brought to life in a game world developed called Smallville, which features all the common affordances found in a typical small village, such as houses, apartments, stores, a café, bar, and school, as well as the sub-areas and objects that make the space functional.
Smallville is populated with 25 generative agents, each with distinct identities. For example, Isabella is told that she is the café owner of Hops Café, who loves to make people feel welcome and is planning to host a Valentine’s Day party. With these identities as the only human input, we are able to simulate believable lives for these agents.
A typical day in the life of Isabella involves waking up early, completing her morning routine around 6:00 a.m., and opening her café by 8 a.m. Throughout the day, she interacts with her customers, perhaps even inviting them to the party she’s organizing. Her working day comes to an end around 6:00 p.m. as she heads to a local store to buy supplies. In close-up, these broad stroke behaviors are composed of smaller sequences of actions that can impact and alter the game world environment.
One example of interaction between two agents, Latoya and Sam, demonstrates how they form a relationship over time based on their interactions. The first time they meet, Latoya tells Sam that she’s there to take photos for a project she’s working on, and the next day, Sam remembers her and asks about her project. It’s not just one-on-one interactions that take place in Smallville; the agents remember and form relationships, leading to events such as Valentine’s Day parties and other gatherings.
What makes this agent behavior possible is the development of language models and a unique agent architecture. The perceptual stimuli of the agents are translated into natural language sentences, stored in a long-term memory called the memory stream, and retrieved to respond to the agents’ current situation. Using this architecture and a language model, the generative agents’ behavior is significantly more believable than both chat PT and even human crowd workers role-playing as these agents.
The potential applications of this research are vast, extending beyond AI research to simulations of real-world scenarios, such as predicting the behavior of employees and customers in a large corporation or simulating the consequences of societal changes. This technology could also be used to create more realistic and believable characters in video games, making the gaming experience more immersive than ever before.
In conclusion, the development of fully general computational agents with the ability to simulate human behavior opens up a world of possibilities in AI research, gaming, and real-world simulations. The work done by researchers like Junun Park has laid the foundation for a future where artificial intelligence can create believable and realistic human-like behavior. As the technology continues to advance, we can look forward to even more exciting developments in this field.
