Social Village Simulator Smallville Advances AI Research

With all the talk and buzz surrounding AI and LLMs, many people are either fascinated and incorporating them as much as they can into their lives or fear the possibility of them taking over the world, the group known as “AI doomers.”

When imagining a world completely filled or dominated by AI, dystopian pictures like The Matrix or Blade Runner might be what comes to their mind because many assume that AI would take over and diminish the chance of peaceful coexistence. But what if someone simulated a world entirely populated by AI, and instead of getting scary and doomsday-y, we get functional and wholesome?

This is what the research team from Stanford University and Google, consisting of Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein, have managed to present by setting up their game/simulator, Smallville.

Simulated Sandbox

Last August, Stanford and Google released a paper titled “Generative Agents: Interactive Simulacra of Human Behavior,” filled with observations and analyses regarding their simulated world, Smallville.

Smallville is a small, 15-minute city consisting of several locations like a library, a cafe, a college with dorms, a couple of shops, and several houses, as well as a co-living space for 25 generative agents. These agents, or the city inhabitants, if you will, were able to simulate believable and realistic human behaviors as they were assigned their own identities, goals, and roles as “characters” in the game/simulator.

The inhabitants had the ability to interact with each other as well as perceive the environments they were set to live in. Though images of Smallville that showed inhabitants interacting or moving from one location to another were shared by the research team, the agents didn’t actually move from place to place as characters in Stardew Valley would.

The world can be deemed as a sandbox, an isolated environment that enables researchers to conduct creative and safe program and code experiments with a metaphorical boundary imposed. In the case of Smallville, the inhabitants’ actions were done by having the research team give direct prompts to and set up conversations between various instances of ChatGPT. The inhabitants would also act based on the pre-programmed goals that the research team had given them.

For example, one of the inhabitants, named Isabella Rodriguez, was given the goal of planning a Valentine’s Day party at “Hobbs Cafe.” What is fascinating about the simulator is the behaviors between agents that emerge, behaviors that were not at all expected nor pre-programmed by the research team.

Agent Interaction and Behavior

The Valentine’s Day prompt led to sequences of events that presented new, autonomous behaviors, as if the agents had minds of their own. These include what the researchers call “information diffusion,” a situation where agents tell each other information and socially spread said information across town, and “coordination,” where the agents manage to plan and attend the party together.

Isabella was able to invite friends and customers, which included Maria, who also assisted Isabella in decorating the cafe. Maria, who, in the initial programming, wasn’t asked to be involved in the party, invited her crush to attend the party. At the end of the simulation sequence, a total of 12 agents heard about the party, and five ended up attending the party.

The Valentine’s Day sequence becomes one example of how complex social interactions in this virtual world can cause unexpected situations, just as it is in real life.

A third behavior that emerged was “relationship memory,” which is the memory of past interactions and events between agents and the remembrance of them moments later. This is shown through an interaction between a family of three: husband John Lin, wife Mei Lin, and son Eddy Lin.

John was pre-programmed with a character description made by the researchers, elaborating his identity, his interests and preferences, and his relationship with other agents.

In a simulated sequence, John got up, brushed his teeth, showered, got dressed, and ate breakfast at the dining table. His son Eddy follows not long after and notices his dad at the table. This is followed by a conversation Eddy initiated regarding working on a music composition for class later today. As Eddy leaves for school, Mei joins John, and he tells her about the music composition by saying, “He’s working on a music composition for his class… I think he’s really enjoying it!”

Once again, these behaviors suggest that the agents/inhabitants have developed their own minds, and this has only been partly made possible due to generative AI’s ability to perform natural language processing (NLP) tasks, which we once discussed in the Demystifying LLM Technology deep dive.

This simulation runs deeper, however, as it ran for 2 whole simulated days with all the agents maintaining memories of their identities and relationships. If we were to simply use ChatGPT, for example, it wouldn’t be able to remember a conversation shared the day prior. I tried checking by asking myself.

So, what makes this simulation different?

“Human” Memory

Though we know how vast LLMs can be, they have limited “memory” due to their limited context window. The context window is the limited number of tokens, or word chunks, that ChatGPT can process at a time.

Creating a virtual, simulated society where inhabitants can’t remember each other’s names or permanently perceive the relationship between them and their families wouldn’t be too good, would it? This is why the research team designed a system they’ve named the “Memory Stream."

This system allows agents to perceive their environment and keep a comprehensive record of all their experiences. The record retrieves relevant memories when an action needs to be taken by the agent, which will also be used to form longer-term plans and higher-level reflections for future use.

Possibilities and Risks

The realism the researchers have managed to create with their town filled with immersive human experiences, behaviors, and interactions had to be evaluated.

The team invited human evaluators to watch replays of the simulation and conduct interview questions with the agents to assess how well the agents produced believable human behavior. The team sneakily included humans to role-play as AI agents during the interview as well, and evaluators perceived the real AI agents as more believably human than the role-playing humans.

The agents’ proven ability to understand and act out the nuisances of human behavior presents the technology’s immense potential. Numerous industry applications and breakthroughs can be imagined, for instance, as the means of understanding complex human behavior in psychology and behavioral science or social relationships in sociology through recreation and simulation. Financial markets may be able to utilize the technology to collect trader and investor behavior for future economic scenarios or policy changes. Security enforcement can use generative agents to simulate criminal scenarios to study potential crowd behaviors and train for crime prevention.

The results of this simulation are quite phenomenal, especially for those whose field of expertise and focus is artificial intelligence. However, the team still expresses worry regarding the ethical impacts and risks of this technology. These include the risk of inappropriate “parasocial relationships” being formed between humans and computers and the risk of unchecked incorrect and biased information.

The research team also argues that developers looking to utilize the technology need to stay ethical and socially responsible, one way by explicitly disclosing the computational nature of agents if used in any product designs, especially ones consumer-facing.

You can read the paper on Smallville by clicking the image below.

Meme & AI-Generated Picture

Job Posting

• Canva - Template Designer, 6 Month Contract - San Francisco, CA (Remote)

• HungryRoot - Junior IT Systems Administrator - United States (Remote)

• Tempus - Senior Product Manager, Data Governance - Peninsula (Hybrid)

• The Aerospace Corporation - 2024 Astrodynamics and Mission Design Graduate Intern - Washington DC (Hybrid)

Promote your product/service to Digger Insights’ Community

Advertise with Digger Insights. Digger Insights’ Miners are professionals and business owners with diverse Industry backgrounds who are looking for interesting and helpful tools, products, services, jobs, events, apps, and books. Email us [email protected]

Your feedback would be greatly appreciated. Send it to [email protected]