If researchers create artificial intelligence agents that behave like humans, then it stands to reason that they are able to populate virtual worlds with these AI simulators and let them loose to interact with one another. A recent experiment conducted by Stanford and Google researchers proved that such an endeavor could be successfully achieved.
The research is centered around machine learning models that generate “generative agents” which are capable of taking in their environment and formulating realistic actions in response. In the experiment, 25 “ChatGPT” instances were set up, each one equipped with their own backstory. For example, one such avatar was John Lin, a pharmacy shopkeeper at the Willow Market and Pharmacy. He had a wife, Mei Lin, and was living with his son Eddy, who was studying music theory at college.
The environment was then generated based on the specific circumstances for each agent. For instance, when John awoke at 8 AM, he went through his morning routine of brushing his teeth and getting dressed before he went to the kitchen, while Eddy – completely independent to John – simultaneously went through the same kind of process.
The agents could then interact with one another come into contact through the provided virtual environment. Here’s a sample conversation between John and his son Eddy:
John: Good morning Eddy. Did you sleep well?
Eddy: Good morning dad. Yeah, I slept great.
John: That’s good. What are you working on today?
Eddy: I’m working on a new music composition for my class. It’s due this week, so I’m trying to get it finished. But I’m having so much fun with it!
John: That sounds great!
It was shown that the AI agents were able to respond to events from the environment, such as the dripping faucet and planning a party, with realistic responses. ChatGPT was also capable of synthesizing observations into memories, for instance understanding that Eddie and Fran were friends once he saw them talking together in the park. All of this is befitting of characters in a text adventure, except it’s all accomplished thanks to the power of machine learning models.
This bodes well for the future of AI-driven simulations. If a chatbot can imitate an average person and actually learn from its environment, then it’s only a matter of time before these AI agents become even more realistic. And outside of gaming, this technology could be applied to a number of different industries, such as customer service, where providing a human-like experience is paramount.
The lead researchers of this experiment are at the forefront of machine learning models, with Google and Stanford at their back. Google has long been one of the biggest proponents and users of AI technology in the world, while Stanford is a world-class research university with blockchain and AI experts from all over the globe. Together, they have come up with an experiment that the world should be watching closely.
