It’s something halfway between The Sims and Westworld. A team of researchers from Stanford University and Google have unleashed 25 artificial intelligence (AI) agents on a virtual city called Smallville. The objective was to analyze if they could simulate human behavior in a credible way and develop some skills, such as memory.
Avatars with different identities were created to make the study. The scientists they used GPT 3.5, the model behind OpenAI’s ChatGPT. They entered a paragraph for each agent, describing their occupation, their relationship to the other “characters” and the memories they had. With this ready, they activated the entire display, which they called “Interactive Simulation of Human Behavior.”
Smallville was equipped with a dormitory, a school, a cafeteria, a park, a bar, various houses, and shops, where the AIs could interact with each other. What did you do? Well, they slept, had breakfast, drank coffee. But they also had a party and got involved in politics.
Avatars named Isabella Rodríguez and Tom Moreno, for example, debated the upcoming city elections. “To be honest, I don’t like Sam Moore,” Tom opined. Sam was another Smallville “inhabitant” AI, who decided to run for mayor. after being “involved in local politics for years.” The candidate had already started talking about his plans with other avatars. “I think he’s out of touch with the community and doesn’t keep our best interests at heart,” Tom explained to Isabella.
Smallville Proves How Human-like AI Can Be
The investigation concluded that these Smallville AI agents had the ability “produce credible individual and emergent social behaviours”. The characters were even able to respond to the contingencies of their environment. Isabella, for example, turned off the stove and made a new breakfast when she was warned that she was burning her food.
Other avatars, such as John Lin, held spontaneous conversations without being ordered to do so throughout the day. Some put together their own routines and even organized, on their own initiative, a meeting to celebrate Valentine’s Day. Isabella was in charge of the party and invited several friends that she had met in the cafeteria. They all arrived together, on time, at five in the afternoon.
The researchers “interviewed” each character after the simulation had been running for a while. Thus they were able to verify that these characters were capable of recovering information from their “memory” and using their memories to decide how to act.
Agents could also reflect. Those responsible for the study discovered that some of them had developed careers and interests of their own during the time the experiment lasted.
AI in the future of video games
The analysis argues that the ability to create credible simulations can enhance the development of various virtual spaces, such as video games. Develop something like The Sims, but with a considerably greater level of independence of the characters and a richer and more unpredictable development.
The team designed an animated representation of avatar interactions that can be viewed check online. It looks like a 16-bit pixelated video game, similar to harvest moon. Smallville’s AIs are represented by their initials in this simulation, but their actions can be seen in more detail by clicking on them.
The specialists highlighted within the conclusions they recorded “a series of emergent behavior patterns”. That is, actions carried out by these agents that had not been pre-configured. For example, sharing information between them spontaneously or creating new links.
The possible jump to the real world
These types of increasingly realistic representations offer an idea of what can be achieved in the real world. AI developers could hone the cognitive capabilities of such agents without, for example, needing to implement robotics in the real world, the study authors explained in their report.
There are already those, of course, who are looking at ways to implement this type of simulated reality in the real world. Artur Sychov, the founder of a metaverse company called Somnium Space, is creating a project called “Live Forever”. His idea is that people can talk to their relatives in the metaverse, even after their death. Sychov has also integrated ChatGPT technology.
The Smallville AI experiment raises some ethical red flags. There is a risk of “that humans form parasocial relationships with generative agentseven when such relationships may not be appropriate,” the Stanford and Google team noted. They also warned against the production of misinformation and other malicious content, one of the big complaints about generative AI.