Imagine spending two decades carving an entire city out of wood, one building at a time. No blueprints downloaded from the internet. No 3D printer humming in the background. Just you, some balsa wood, and an obsessive attention to detail that would make a Swiss watchmaker weep.
That’s exactly what Joe Macken did. This Queens truck driver spent over 20 years creating a scale model of every building in New York City, carved entirely from balsa wood. His miniature metropolis recently went viral on TikTok with over 10 million views, and honestly? It’s the perfect lens for understanding how AI agents actually work.
The Patient Builder
Macken started his project in 2004 with a single structure: 30 Rockefeller Plaza. From there, he methodically added building after building, neighborhood after neighborhood. No shortcuts. No giving up when the Chrysler Building’s spire proved tricky or when he realized he’d need to carve thousands of brownstones.
This is where the AI agent comparison gets interesting. When people hear “AI agent,” they often picture some magical system that instantly solves problems. But the reality is far more like Macken’s approach: breaking down an enormous task into manageable pieces, then executing each piece with precision.
Task Decomposition in Wood and Code
An AI agent doesn’t wake up one morning and decide to “understand New York City.” Instead, it breaks that goal into smaller objectives: identify buildings, categorize by type, understand spatial relationships, recognize patterns. Macken did the same thing, just with a knife instead of algorithms.
Each building he carved required him to observe, measure, plan, and execute. He had to figure out which details mattered and which could be simplified. Should every window be individually carved? How much texture does a brick facade need? These are the same trade-offs AI agents make constantly: what level of detail is necessary to accomplish the goal?
Memory and Persistence
Here’s what strikes me most about Macken’s project: he had to remember what he’d already built, what still needed work, and how everything fit together. Over 20 years. That’s a staggering feat of human memory and organization.
AI agents need similar capabilities. They maintain context about what they’ve done, what they’re currently doing, and what comes next. They track their progress toward goals and adjust when something doesn’t work. The difference? An agent’s memory is explicit and searchable. Macken’s was probably a mix of physical organization, notes, and sheer determination.
Why This Matters for Understanding AI
When you look at Macken’s completed model, you see New York City. You don’t see 20 years of individual decisions, failed attempts, and gradual refinement. The same thing happens with AI agents. We see the end result—a task completed, a problem solved—and assume it happened through some kind of magic.
But both processes are fundamentally about persistence and iteration. Macken couldn’t carve all of Manhattan in a day. An AI agent can’t solve complex problems in a single step. Both need to work systematically, learn from what works, and keep going when things get difficult.
The Human Element
There’s something beautifully human about spending 20 years on a project that serves no practical purpose. Macken didn’t build his model to improve traffic flow or plan urban development. He built it because he wanted to. Because it fascinated him. Because he could.
AI agents don’t have that motivation. They work toward goals we give them, using methods we design. They’re tools, not artists. But understanding how they break down tasks, maintain context, and persist toward objectives? That helps us use them better.
Macken’s miniature New York now goes on public display, a testament to what focused effort can accomplish over time. Every time someone asks me to explain how AI agents work, I’m going to think of a truck driver from Queens, a pile of balsa wood, and the patience to build a city one building at a time.
That’s not a bad way to understand the future of AI.
đź•’ Published: