Everyone’s obsessed with training models on massive datasets. We’re burning through GPU clusters, scraping the internet, and calling it progress. Meanwhile, a truck driver named Joe Macken spent over two decades building something that puts our entire approach to data collection and model building to shame.
Macken created a scale model of every building in New York City. Not a 3D render. Not a procedurally generated simulation. An actual physical model, carved from balsa wood, piece by piece, over 20 years.
The Anti-Automation Manifesto
Here’s what gets me about this story: Macken’s approach is the exact opposite of how we build bots today. We optimize for speed, scale, and automation. We want our models to learn faster, train on more data, and produce results instantly. Macken did none of that. He spent two decades on a single project, working with his hands, one building at a time.
And the result? Something that went viral on TikTok with 10 million views. Something that’s now going on public display. Something that people actually care about.
When was the last time anyone got excited about your training dataset?
What Bot Builders Can Learn From Balsa Wood
I build bots for a living. I write tutorials about architecture and code. I’m supposed to be all-in on automation. But Macken’s project makes me question some fundamental assumptions about how we approach our work.
First, there’s the data collection problem. We scrape, we aggregate, we pull from APIs. We treat data like a commodity to be extracted as quickly as possible. Macken treated each building as something worth studying, understanding, and recreating with precision. He wasn’t trying to capture every building in the world. He focused on one city and did it right.
Second, there’s the iteration problem. We A/B test, we deploy updates, we ship fast and break things. Macken started with 30 Rockefeller Plaza in 2004 and kept going for 21 years. No pivots. No rebranding. Just consistent, focused work on a single vision.
Third, there’s the validation problem. We measure success with metrics: accuracy scores, response times, user engagement rates. Macken’s validation was simpler: does this look like New York City? Can someone recognize the buildings? Does it capture something real?
The Human-in-the-Loop We Forgot About
The AI community loves talking about human-in-the-loop systems. We design interfaces for human feedback. We build tools for human oversight. We create workflows where humans can correct our models.
But Macken’s project is human-in-the-loop taken to its logical extreme. Every decision, every cut, every detail came from a human who spent his days driving a truck through Queens and his nights carving buildings from wood. He wasn’t supervising an automated process. He was the process.
And somehow, that produced something more compelling than most of what we’re building with our sophisticated pipelines and neural networks.
Speed Isn’t Everything
The tech industry worships speed. We want faster training, faster inference, faster deployment. We measure ourselves in milliseconds and celebrate when we shave off a few percentage points of latency.
Macken spent 20 years on his model. That’s not a bug. That’s a feature. The time investment is part of what makes it meaningful. The dedication is part of what makes it impressive. The patience is part of what makes it valuable.
I’m not saying we should all spend 20 years on our next bot project. But maybe we should stop treating speed as the only metric that matters. Maybe some problems are worth sitting with for a while. Maybe some solutions require the kind of sustained attention that can’t be automated away.
Macken’s miniature metropolis is now going on display, a physical artifact of two decades of focused work. Most of the bots I’ve built will be deprecated and forgotten in two years. That’s the trade-off we make when we optimize for speed over everything else.
Sometimes the best model is the one that takes 20 years to build.
🕒 Published: