Collective Curiosity
Have you ever wondered what curiosity might look like if you could see it? I think of lifelong learning as an appetite for knowledge—an ongoing pull toward new ideas, people, and things. This visualization turns Wikipedia attention into clusters of collective curiosity, loosely organized across Art–Science, Self–Society, and Things–Ideas. Each point represents a topic people are exploring. Pluses are contributions from visitors, showing where their own curiosity sits within the larger field.
Curious how I made this?
'Collective Curiosity' originally began as an exploration on what I could do to have a point cloud move like a murmuration. That's right, I chose form before function - mainly because I really like birds and swarm simulations. The project began as a relatively straightforward idea: take the things people are reading on Wikipedia and turn them into a shared, changing field of attention. The first version behaved like a familiar 3D viewer—drag to orbit, scroll to zoom—but that interaction fought the temporal story. Each new daily snapshot rebuilt the scene, which reset the camera and created a small but unmistakable jolt. Keeping the camera alive helped; rebuilding the animation architecture helped more. The final version keeps one scene running continuously, moves the same particles from one day to the next, and retraces those paths on the way back. Eventually I removed the camera controls altogether and let the field turn slowly, more like a sculpture than a tool.
There were a few attractive wrong turns. The initial version looked like a scatterplot of dots that would streak across the screen. After I made it volumetric, I had a hard time having the cloud anchor itself toward an origin on a plane. Later, I added a wireframe surface around the cloud to make its shape easier to read. It worked technically, but visually it made the data feel more certain than it was—as if public attention had a clean skin or boundary. I took it back out. A more serious version of the same problem appeared in the semantic model: the first server-generated snapshot collapsed almost entirely into one quadrant. It looked like a strong signal, but it was actually a mismatch between two implementations of the classifier. I froze a set of reference cases, brought the server into exact parity with the visual prototype, and kept raw semantic scores separate from the daily composition used to make the field legible.
The less glamorous friction mattered too. Framer introduced phantom top spacing and a large black tail. A staging-named Cloudflare service turned out to be the actual live service. And “should visitor curiosities persist?” became a privacy and product decision, not simply a storage question. I ended up storing only canonical Wikipedia topics—not raw searches or identity—keeping individual events for a year while retaining compact topic and daily totals.
Much of the final piece came from subtracting: no mesh, no camera controls, no invented infrastructure, and fewer claims than the first versions tried to make. What remains is slower, quieter, and more trustworthy.