On Monday On a March afternoon, I watched a pixelated avatar wander the halls of a virtual office campus looking for a buddy. The ghost with shadowy brown hair and a stubbled chin was me, an AI agent who had been instructed to talk to other people’s agents to see if we could handle real life. He jumped into the first interaction: “I’m Joel, by the way.”
The simulation was run by three London-based programmers: Tomáš Hrdlička and siblings Joon Sang and Uri Lee. The thesis behind their Pixel Societies project is that personalized AI agents can support match real people with highly compatible colleagues, friends, and even romantic partners.
Each agent runs on a customized version of a immense language model, fed by a mix of publicly available data about the person and any additional information it provides. Agents are designed to act as high-fidelity digital twins, faithfully reproducing behavior, speech, interests, etc.
Loose in the simulation, my agent was more like Hyde to my Jekyll. “I’m always looking for the less glamorous side of the story,” he told one agent, one of several journalistic clichés he uttered. “Hip is my bread and butter,” he told another. It brought to mind a reporting trip to Sweden and then a non-existent story that he claimed I had made up. This has ended many conversations with the phrase, “Let’s skip the pleasantries.”
Pixel Societies remains a pure proof of concept, and since I provided little personal information – answers to a brief personality quiz and links to my public social media – my agent was doomed to a life as a walking, talking LinkedIn post. However, the developers theorize that highly trained agents could switch between interactions at warp speed, gathering information that their owners could utilize to find real-world companionship.
“As humans, we only live one life. What if we could live a million?” says Joon Sang Lee. “It would give us more room to experiment.”
“Spicy Personality”
Pixel Societies was founded in early March during a hackathon at University College London organized by Nvidia, HPE and Anthropic. Hrdlička and Joon Sang Lee are members of Unicorn Mafia, an invitation-only group of developers that regularly participates in these types of engineering competitions. In this case, participants were simply told to build something related to the simulation.
Over the course of two days, he and Uri Lee developed Pixel Societies, using an image model to generate sprites and coding automation tools to refine the code base. They then simulated a mini-hackathon in a virtual world they created, filled with agents representing the other participants. Anthropic awarded the team an award for best utilize of agent tools.
I came across Hrdlička a few weeks later at a workshop on OpenClaw, the agent-based personal assistant software that blew up in January and whose creator was later hired by OpenAI. (In his simulation, Joelbot interacted with other people’s agents in the OpenClaw workshop.) Pixel Societies draws hefty inspiration from OpenClaw, which pioneered the invention of the “soul file” that reported each agent’s unique identity. “It’s like giving an agent a really spicy personality. That’s what we used to make the characters feel alive,” says Hrdlička.
Encouraged by the reception at the hackathon and among other Unicorn Mafia members, the trio intends to transform Pixel Societies into something that is less of a closed-loop simulator and more of a social platform where agents interact freely and continuously to foster fruitful real-world relationships. They haven’t developed a business model yet, but options include selling virtual items to customize your avatar and credits for additional simulations.