Version 0.1 the first print from github.com/otaviogood
I have a kid now, and so do a few people I’ve worked closely with for years. We started talking about sampling bias—not in models, but in life. The neighborhoods we’re raising our kids in, places like Palo Alto and Beverly Hills, are extreme outliers. The people our children interact with every day are a narrow slice of the global population. That shared realization eventually turned into a question: what would it look like to generate a more honest sample of humanity—and what if you turned it into a game?
The idea first surfaced when a friend and longtime collaborator told me he’d been trying to make a game for his kids. The motivation was there, but the early results were unsettling. The sampling was honest, but not gentle: one of the first characters it produced was a visibly sad child in war-torn Ukraine. Others leaned uncomfortably into stereotypes. He builds games for kids all the time, and this felt different—not wrong, exactly, but revealing. The system wasn’t misbehaving; it was reflecting something real.
To deal with that, we started talking about history—not as a backdrop, but as a variable. The uncomfortable cards were snapshots, frozen at a single moment in time. But human societies aren’t static. Places that look prosperous today were often miserable in the past, and places struggling now have gone through long periods of stability, creativity, or power. Bringing in historical context let us shift from a flat sample of “now” to something longitudinal. It also opened a way to teach kids that human and cultural development isn’t a snapshot—it’s a time series.
Wikipedia felt like the obvious move. I’ve been reading history there for years, and it’s one of the few places where human knowledge is both deep and structured enough to be programmatically useful. The system starts by crawling Wikipedia’s historical eras portal and generating JSON files of eras across human history. When the sampler picks a location, it finds which eras align and uses those as structured seeds. By recursively walking category graphs and pulling summaries, the system generates people who make sense in their time: era-appropriate names, occupations, interests, stories, and images. Each era becomes a bounded distribution rather than a vibe, constraining technologies, social roles, and narratives. The LLM isn’t inventing history—it interpolates within it. History stops being static and becomes a generator, turning the game into a longitudinal model of humanity instead of a snapshot of the present.