Generative AI could start its own civilization

Can artificial intelligence beings start their own civilization? Researchers from Stanford University are taking the first step in this direction.

In the simpsons-episode Treehouse of Horror VII, from 1996, Lisa dips her recently fallen tooth into a Coke-filled Beatrice dish for a school science project, which also accidentally causes an electric shock. To her surprise, microscopic creatures spring to life on a tooth and quickly build a civilization. Almost the same idea is recognizable Generative Agents: Interactive Simulation of Human Behaviora study from Stanford University, which examines whether large language models (LLMs), such as OpenAI’s GPT-4, can convincingly simulate human behavior in a dynamic and realistic way.

Village with AI inhabitants

The research explores the potential for generative AI to create artificially intelligent agents that remember their interactions, reflect the information they receive, and plan long- and short-term goals. In other words, the ability to mimic the behavior that people display in their daily lives. Furthermore, the study investigates more complex social behaviors that emerge from interactions between individuals of a large group of AI “objects.”

Artificial intelligence agents created by researchers are released into their own online world… Smallville Baptized, an online environment similar to the popular computer game The Sims, with environments such as a restaurant, school, and café. 25 innovative AI “objects” have now been dropped in there, with real-time visibility into the data they generate about their behavior and relationships. They speak to each other through a built-in language model, based on memories and past interactions. The researchers also allowed human users to interact with AI creatures through “narrative voice,” or by directly controlling an agent.

Population dynamics

Also in game worlds such as Grand Theft Auto V And Doctrine killer Computer-controlled characters often respond in a complex way to what the human player character does, but these reactions are programmed by game designers. The Stanford researchers’ online world is populated by completely autonomous “non-playable characters,” as these types of characters are called in video games.

The university says the study offers many possibilities, especially in simulating population dynamics, and provides valuable insights into behavior and social interactions. The experiment shows, among other things, that AI-powered agents are able to put plans together, without being explicitly asked to do so. For example, they actually threw a party on Valentine’s Day, without any input from the researchers

Very polite

The experiment had some drawbacks. Like all artificial intelligence, the residents of Smallville have been found to occasionally hallucinate, remembering things that never happened. They are also exceptionally polite and helpful, with no conflict ever arising. But there is great interest in the project. For example, venture capital firm Andreessen Horowitz has its own version of it operating in its own R&D lab.