Researchers put 25 ChatGPT-like AIs into a game & the results are wild
A research paper has put 25 AI agents into a game, with surprising findings, as the AI agents began to create a unique social structure and learn from each other.
As AI begins to disrupt the entire tech industry, one group of researchers from Stanford and Google have tasked themselves with interesting AI agents into a single instance where they all coexist, with different areas and motivations. A research paper named “Generative Agents: Interactive Simulacra of Human Behavior” set up a life-sim style sandbox, which is free to view right now, in addition to the finished research paper.
Each of the 25 AI agents is tasked with different tasks or jobs, in order to see how they interact with each other. The small town of 25 is outfitted with various areas, with each AI NPC to fulfill certain goals. Each NPC in the artificial society sleeps, wakes up, brushes their teeth, gets coffee, and does other things which we know as “normal” human behavior.
With the advancement of natural language processing through applications like ChatGPT, we have seen autonomous versions, and AI agents, materialize over the past several months. Unlike ChaosGPT, these agents have goals and tendencies related to their own personalities. Many of them sound like people we all know.
Party planning is much more complex than you think
One, for example, was tasked with throwing a Valentine’s Day party. Through the daily interactions and social structures that the AIs have created for themselves, word spreads throughout the small town, and people who were never initially invited to the party learn of it, too.
This implies that the 25 generative AI agents are able to perceive events, retrieve past memories (as one states that they cannot meet with another NPC due to going to a part), and ultimately act on those “thoughts” and speak to others about them completely autonomously.
Additionally, our example of the Valentine’s Day event planner, Isabella showcases that the AI was able to break down all of the necessary elements of the party, such as ordering decorations and researching ideas for the party too. They are also able to reflect on their emotions.
There are limitations
The AI agents also come with their own limitations, which can cause some issues, these included going to a bar at midday or walking into a bathroom that was already occupied. These errors come from a larger memory set being unable to appropriately learn relevant information about the type of place they are going, or the appropriate spaces to execute the actions they are currently thinking about.
The paper also states that there is a fanger about users who “may anthropomorphize” them, or form parasocial relationships with AI agents in the future.
Could this be the future for video games?
You only have to have a read of the paper in order to understand how monumental this paper is. Generative AI agents applied to a sandbox video game like Skyrim could totally change the ways in which games are made. However, the technology is still early, and it’s likely that generative AI models won’t exactly be able to move you in the same way that human-written content or characters can.
High-concept understandings about dialogue and narrative are still fairly abstract for generative AI, as demonstrated by their ability to not understand that two people cannot enter a single toilet. So no, we wouldn’t say to expect a town full of AI NPCs to be included in The Elder Scrolls 6 any time soon, though some games are generating AI dialogue for certain NPCs already.