Let loose in Minecraft, 1,000 autonomous AI agents collaborate to build their own society
A San Francisco-based startup took a sprint toward the future of artificial intelligence by unleashing 1,000 autonomous AI agents on a Minecraft server. Altera, founded by former MIT professor Robert (Guangyu) Yang, conducted Project Sid to explore whether agents can organize and collaborate to achieve more collectively than as entities operating individually. So, what happened? Agents formed a merchant hub, used Google Docs to vote on and amend a constitution, spread a religious belief (Pastafarianism) through bribery and lit torches to help a lost villager find their way home. They also collected far more in-game items than expected.
Through simulation of various aspects of civilization, including democracy, social norms and economic systems, Project Sid aims to uncover phenomena not visible in smaller-scale AI interactions and gain insights into the development of more human-like AI. The experiment also revealed significant challenges — how to benchmark the progress of an AI civilization, for example, and how minor flaws in individual agent behavior can cascade into large-scale issues within a simulated society. It's also debatable how autonomous the agents actually are. To what degree are their actions self-initiated? How many of their decisions can be traced directly to knowledge and instructions provided by their developers?
Altera's mission is "to create digital human beings that live, care and grow with us," using the human brain as its inspiration, with models mirroring elements like the prefrontal cortex, memory systems and social-emotional states (Yang is a computational neuroscientist 🧠). The startup's ambitions haven't gone unnoticed by investors. In May 2024, Altera announced an oversubscribed USD 9 million seed round, co-led by Eric Schmidt's First Spark Ventures and Patron, a seed-stage fund co-founded by Riot Games alums.
What could your brand gain from launching its own multi-agent world?
In addition to running a new product past a human focus group, how about letting thousands of 'customer agents' interact with it in dozens of simulations of real-world scenarios and existing consumer preferences? Setting up a new non-profit? You could virtually trial various governance structures to find the best fit for your specific domain and actors or, at the very least, get a better sense of the drawbacks and benefits of each option. We'll stop now and let your imagination take over ;-)
