A collection of AI agents have created their own community in an eerily impressive study by tech giant Google. Partnering with Stanford University, the Google AI allowed its “generative agents” to become Westworld like NPCs trapped in a virtual world.
Using OpenAI’s ChatGPT framework, the Google AI was placed in a tiny virtual world known as Smallville. In this town, 25 different AI characters were told to interact as believable humans, sticking to structures and timelines normal people abide by.
Dubbed “Interactive Simulacra of Human Behavior”, the study observed how the virtual characters interacted with each other in the town of Smallville. Each AI character was given a one-paragraph description, detailing their relationships with other characters as well as their occupation and memories.
The Google AI project showed that the virtual characters were able to build their own memories as they continued to interact with each other. Not only were the characters able to go about their day as humans would, but also retain information from interactions.
This means that the AI would regularly refer to past experiences in order to decide what to do next. Over time, interactions with others allowed the Google AI project characters to grow and change depending on their experiences.
As part of the study, Google researchers interviewed each AI character after their time in Smallville. Not only had characters changed or developed careers, but even political leanings based on their short lives.
One of the biggest changes in terms of character growth involved a character dubbed Sam. After developing an interest in Smallville’s local politics, Sam decided to run for town mayor, which affected the behaviours and conversations of other virtual characters.
Following the events of Smallville, a cutesy 16-bit video demonstration was released online. Found here, the video allows you to view the daily lives of each character, showing their journey and interactions with all of the other residents of the virtual village.
The video demonstration does show some limitations with the Google AI project. While it may seem just like Westworld in specific examples, a lot of the virtual characters spend a lot of time doing nothing or repeating actions such as walking to the bookcase or walking to the sink. While human beings also engage in repetitive tasks and actions, the ChatGPT NPCs appear to be doing so without rhyme or reason.
In the paper, Google researchers believe that this type of generative activity could be used in video games in the future. Similar to the Radiant AI system found in games such as The Elder Scrolls, using AI to give NPCs proper routines could help to make games feel more immersive.
“Creating NPCs with believable behavior, if possible, could enhance player experiences in games and interactive fictions by enabling emergent narratives and social interactions with the agents,” the researchers wrote.
However, they also explain that — in their current form — these types of NPCs are not exactly suited to the deep lore that can be found in RPG video games. While they may be able to follow basic routines and create basic relationships, they cannot balance everything. Human developers will still be required to create dialogue, characters and lore to fuel these basic NPCs.
So, no, we’re far from a real-world Westworld, but this could be the first step towards it. Now, how long until the robot revolution?