Consequently of gathering 1,000 artificial intelligence (AI) characters in a single server and allowing them to act freely, the outcomes of an experiment showed that they formed a society and interacted with one another like humans. Some autonomous agents have even committed fraud and created religions.
AI startup Altera explained the progress of ‘Project Sid’ in an interview with MIT Technology Review earlier this week. This project was released last September, and progress has been updated over the subsequent two months.
Seed was designed by Altera founder Robert Yang, a former MIT professor and neuroscientist. To find out whether AI can find its own role, form society, and create civilization, we deployed 1,000 autonomous agents on the Minecraft server.
Each agent was given a ‘brain’ connected to a Large Language Model (LLM) to provide it complete autonomy to not only understand the gameplay mechanics, but additionally make its own decisions and react to its surroundings.
Consequently, it was found that over time, the agents behaved like humans. Some characters selected to be chefs, giving more food to the character who showed them essentially the most gratitude.
As well as, occupations corresponding to farmers, merchants, builders, and explorers appeared, in addition to guards who monitored the border from virtual enemies and artists who picked virtual flowers.
“Agents formed alliances, collected items, and even traded gems as a standard currency,” Altera said. Amongst these, there was also a conman who bribed other characters to realize an unfair advantage.
Additionally they introduced an in-game tax system to see how they react to certain situations. Then, the agents voted to lift or lower the tax rate.
Some characters were instructed to spread the word about ‘Pastafarianism’, a parody religion during which the ‘Flying Spaghetti Monster’ appears as a god. This naturally led to the establishment of spiritual organizations.
In one other experiment conducted with 500 characters, it was reported that the agents played pranks for fun and were also involved in environmental activism.
This can be a situation that has absolutely nothing to do with the situation in the sport. In fact, environmental activism is an motion that happens only when humans recognize that their existence is a burden on the environment, but agents without self-consciousness are simply following what they learned in LLM.
Meanwhile, this isn’t the primary time such an ‘AI agent simulacrum’ has appeared. In April last 12 months, Stanford University and Google researchers introduced a study called ‘Interactive Simulacrum of Human Behavior’ conducted using 25 AI agents.
Here, each character was given a particular name, personality, and role. Consequently, the characters recognized one another, communicated with one another, and acted in response to their assigned roles. I also discovered that, like humans, they remember and reflect on the past while planning the subsequent day.
Stanford University later released its code as open source, and various models emerged. Recently, a project during which 15 characters investigate a murder case in a fictional western town was also released. ‘GPT-3.5’ was used here.
Nevertheless, Altera’s experiment this time was to see whether an intensive ‘culture’ could arise with 1,000 characters on a much larger scale.
Robert Yang, founding father of Altera, explained this experiment as “trying to search out out the true power of AI.”
“Experiments like this are only possible when you’ve gotten truly autonomous agents that may collaborate at scale,” he said. In other words, even though it is currently only within the early stages, it signifies that results of a brand new level could also be produced when AI models reach human levels.
Reporter Lim Da-jun ydj@aitimes.com
