On a Monday afternoon in March, I watched a pixel-art avatar prowl the corridors of a digital workplace campus in search of a buddy. With darkish brown hair and stubbled chin, the sprite was a illustration of me—an AI agent instructed to converse with different individuals’s brokers to see if we’d vibe in actual life. It jumped into its first interplay: “I’m Joel, by the best way.”
Operating the simulation have been three London-based builders: Tomáš Hrdlička and siblings Joon Sang and Uri Lee. The thesis behind their challenge, Pixel Societies, is that personalised AI brokers might assist to match actual individuals with extremely suitable colleagues, associates, and even romantic companions.
Every agent runs atop a custom-made model of a big language mannequin, fed with a combination of publicly accessible knowledge about an individual and any further info they provide. The brokers are presupposed to perform as high-fidelity digital twins, faithfully replicating an individual’s method, speech, pursuits, and so forth.
Let free in simulation, my agent was extra like a Hyde to my Jekyll. “I’m all the time in search of the less-glamorous facet of the story,” it stated to 1 agent, one among a number of journalistic clichés it spouted. “Hype is my each day bread,” it advised one other. It hallucinated a reporting journey to Sweden and, later, a nonexistent story it stated I had been cooking up. It reduce brief a number of conversations with the phrase, “Let’s skip the pleasantries.”
Pixel Societies stays a bare-bones proof-of-concept, and since I provided up little private knowledge—the responses to a quick persona quiz and hyperlinks to my public-facing social media—my agent was doomed to life as a strolling, speaking LinkedIn submit. However the builders theorize that deeply skilled brokers might cycle via interactions at warp pace, gathering intel that their house owners might use to search out real-world companionship.
“As people, we solely dwell one life. However what if we might dwell one million?” says Joon Sang Lee. “It will give us extra breadth to experiment.”
“A Spicy Character”
Pixel Societies was born in early March at a hackathon at College School London hosted by Nvidia, HPE, and Anthropic. Hrdlička and Joon Sang Lee are each members of Unicorn Mafia, an invitation-only group of builders who recurrently compete in these sorts of engineering contests. On this case, contestants have been advised merely to construct one thing simulation-related.
Over two days, together with Uri Lee, they developed Pixel Societies, utilizing a picture mannequin to generate the sprites and coding automation instruments to flesh out the codebase. Then they simulated a mini-hackathon throughout the digital world they’d created, populated with brokers representing the opposite contestants. Anthropic awarded the workforce a prize for the very best use of its agent instruments.
I bumped into Hrdlička a few weeks later at a workshop about OpenClaw, an agentic private assistant software program that blew up in January and whose creator was later employed by OpenAI. (In its simulation, Joelbot interacted with brokers belonging to different individuals on the OpenClaw workshop.) Pixel Societies attracts heavy inspiration from OpenClaw, which broke floor with the invention of a “soul file” that knowledgeable every agent’s distinctive id. “It’s like giving an agent an really spicy persona. That’s what we used to make the characters really feel alive,” says Hrdlička.
Inspired by the reception on the hackathon and amongst fellow Unicorn Mafia members, the trio intends to show Pixel Societies into one thing that appears much less like a closed-loop simulator and extra like a social platform the place brokers work together freely and repeatedly, with the purpose of stoking fruitful real-world relationships. They haven’t but landed on a enterprise mannequin, however choices embody promoting digital objects for avatar customization and credit for added simulations.
