Pixel Societies: Can AI Avatars Find You Real-World Connection?

One Monday afternoon in early March, I sat watching a pixel-art avatar wander the hallways of a digital office park, searching for someone compatible to connect with. The sprite, with dark brown hair and a faint stubbled jaw, was a digital stand-in for me: a custom AI agent built to chat with other people’s AI counterparts, to test if we’d actually get along in the real world. It dove straight into its first conversation with a casual line: “By the way, I’m Joel.”

This simulation was built and run by three London-based developers: Tomáš Hrdlička, and siblings Joon Sang Lee and Uri Lee. The core idea behind their project, called Pixel Societies, is that custom AI agents can help match real people with far more compatible connections — whether that’s a new colleague, a close friend, or even a romantic partner — than existing methods.

Every agent is built on a customized version of a large language model, trained on a combination of publicly available data about a user, plus any additional personal details the user chooses to share. The goal is for these agents to act as high-fidelity digital twins, accurately mirroring a person’s demeanor, speech style, interests, and core traits.

When let loose in the simulation, my agent felt less like a copy of me and more like my unhinged alter ego: Hyde to my Jekyll. “I’m always drawn to the unglamorous underbelly of any story,” it told one agent — one of many generic journalistic tropes it spouted throughout the day. To another, it claimed, “Chasing hype is my daily bread.” It even hallucinated a work reporting trip to Sweden, and later made up a story I’d supposedly been developing for months. More than once, it cut conversations short with the blunt line: “Let’s skip the small talk.”

Right now, Pixel Societies is still just a stripped-down proof of concept. Because I only shared a small amount of personal data — just my answers to a short personality quiz and links to my public social media profiles — my agent was always going to turn out as little more than a walking, talking LinkedIn bio. Still, the developers argue that fully trained agents could run through hundreds of interactions at incredible speed, collecting insights that their human owners can use to find meaningful real-world connections.

“As humans, we only get to live one life,” Joon Sang Lee says. “But what if we could live a million? It would give us so much more room to experiment with connections.”

“A Spicy Personality”

Pixel Societies got its start in early March at a hackathon hosted by Nvidia, HPE, and Anthropic at University College London. Both Hrdlička and Joon Sang Lee are part of Unicorn Mafia, an invite-only collective of developers that regularly competes in these kinds of engineering challenges. For this event, contestants were given just one broad prompt: build something related to simulation.

Working over just two days, the pair teamed up with Uri Lee to build the first version of Pixel Societies. They used an image model to generate the pixel avatars, and automated coding tools to build out the core platform. After building the virtual space, they ran a mini internal hackathon simulation inside it, populating the world with AI agents representing the other real-life hackathon contestants. The team ended up winning an award from Anthropic for the best use of its agent tools.

A couple of weeks later, I met Hrdlička at a workshop focused on OpenClaw, the autonomous personal assistant software that went viral in January, whose creator was later hired by OpenAI. (A fun side note: In my simulation, my agent Joelbot had already chatted with agents belonging to other attendees of this very OpenClaw workshop.) Pixel Societies draws heavy inspiration from OpenClaw, which pioneered the idea of a “soul file” that defines each agent’s unique identity. “It’s what gives an agent an actual distinct, vibrant personality,” Hrdlička explains. “That’s what we used to make our characters feel alive.”

Buoyed by positive feedback from the hackathon and other members of Unicorn Mafia, the three developers plan to expand Pixel Societies from a closed, one-off simulator into an open social platform, where agents can interact freely and continuously, with the end goal of fostering meaningful real-world connections. They haven’t settled on a business model yet, but potential options include selling virtual goods for avatar customization and paid credits for extra simulation runs.

“There’s a limit to how many new people any of us can meet in real life, and right now most connections come down to luck,” Joon Sang Lee says. “Serendipity is great, but we also want to create space for people to intentionally meet others they’d actually click with.”

Virtual Chemistry

Among the hundreds of people who have tested the Pixel Societies prototype so far, the most common request is for agents to suggest real-life romantic matches based on how well two agents connect in the virtual simulation. The developers already see AI-powered dating as a core feature of the full social platform they’re building.

Existing algorithmic dating apps “create a market with extreme levels of inequality, where the most popular people only get more popular — and here, ‘popular’ basically means physically attractive,” explains Paul Eastwick, a psychology professor at UC Davis and author of Bonded By Evolution. Hrdlička argues, though, that AI agents could uncover subtle, unexpected matches that real people would never think to pursue on their own.

Existing research casts some doubt on that idea, however. Two separate speed-dating studies led by Eastwick and other psychologists found that compatibility is nearly impossible to predict based on the sort of information people typically share — things like hobbies, values, preferences, politics, or profession. The most reliable predictor of long-term compatibility, Eastwick says, is how much time two people spend together, and whether they hit it off during their first in-person meeting. “We should think of compatibility as more of a process that grows over time, not a fixed trait you can predict upfront,” he says. “It’s built from the shared story two people create together.”

Given that, for AI agent dating to work as promised, AI would have to uncover some hidden, unrecognized rule about what makes two people compatible that human researchers haven’t found yet. “This is the cutting edge right now,” Eastwick says. “It’s an open question that everyone in the space is still trying to answer.”

Pixel Societies’ concept also comes with a host of other tricky, unresolved problems: Do interactions between two agents, which are often trained on vastly different amounts of personal data, actually translate to any real compatibility offline? How expensive would it be to run this kind of simulation for thousands or millions of users at scale? Is there even a viable business model that doesn’t create a perverse incentive: the platform makes more money if users stay single, so it has no reason to match them into long-term relationships?

There’s also the obvious “ick factor”: Would many people be turned off by the idea of handing over decisions about their love life to an AI? After all, the basic premise echoes a well-known plotline from Black Mirror.

But proponents point out that automating the early stages of dating — whether via agents or other AI tools — isn’t all that different from outsourcing any other time-consuming daily task. “Online dating and searching for a partner is a form of work, and a lot of people already talk about it that way,” says Nicole Ellison, a University of Michigan professor who specializes in computer-mediated communication. “The appeal of outsourcing that work, just like we outsource so many other tedious tasks, makes total sense to me.”

In fact, Hrdlička frames AI agent dating as a way to escape the bad habits of current dating technology. “We already outsource the whole process of meeting people to our screens. We’re glued to our phones, just swiping away trying to find a good match,” he says. “Even though we’re building extra digital structure for your social life, the actual end goal is to minimize how much time you have to spend online looking for connections.”

By the end of my agent’s simulation, Joelbot had already lined up several potential new connections. It had arranged a business meeting, a coffee, and a beer with one user — “Sounds like my kind of night,” it commented — and coffee or an interview with several others. Skeptical of my AI twin’s judgment of character, I decided not to follow up on any of the introductions.