The 49,000 People Under the Breakthrough

In 2012, two graduate students at the University of Toronto trained a neural network on two gaming graphics cards. Consumer Nvidia GTX 580s, the kind you'd buy to run video games. The story goes that one of them, Alex Krizhevsky, ran the training in his bedroom at his parents' house. He wrote the GPU code himself because the network didn't fit on a single card and he had to split it across two.

The thing they built, later called AlexNet, won a competition that year by a margin nobody had ever seen. Top-5 error of 15.3 percent against 26.2 percent for the next team. For years the field had been crawling forward a point at a time, stuck around 25 or 26 percent. Then in one year it dropped more than ten points at once. That result is the moment people point to when they say modern AI started. The deep learning boom traces back to two grad students and a couple of gaming cards in a bedroom.

It's a great story. It's just not the story.

Because none of it would have worked without something that got built first. Something that took four years, tens of thousands of people, and almost no recognition at all.

Six years earlier, in 2006, a researcher named Fei-Fei Li made a bet that the rest of the field thought was a waste of time. Everyone was chasing better algorithms. Better models. The smarter math. Li bet the opposite. She bet the bottleneck wasn't the model. It was the data. In her words, "Pre-ImageNet, people did not believe in data."

So she set out to build a dataset. Not a clever one. A massive one. The plan was to label millions of images, organized into tens of thousands of categories, so that a machine could actually learn what things looked like by seeing enough examples. The final version held over 14 million labeled images across roughly 22,000 categories. They called it ImageNet.

This is the part that gets skipped. A mentor told her she'd taken the idea too far. The advice was to grow with your field, not leap so far ahead of it. The reasonable, experienced read on ImageNet was that it was a four-year detour into grunt work that wouldn't pay off.

And here's the thing about that read. It wasn't stupid. It was right, by the only measure the field rewarded at the time. The foundation doesn't bring glory. It doesn't get published as a breakthrough. A lot of what we celebrate in any field is built on recognition and marketing, and the boring layer builds neither. Not at first. It just sits there looking like a detour, right up until the moment everything else is standing on top of it.

Understanding doesn't get downloaded. It gets ground out, past the point where most people quit. Li ran an early estimate on the labeling. A few people doing the work by hand would have needed nearly two decades to finish. She did it anyway. That is what it actually takes to lay a foundation. You commit to the grind before you have any proof it will matter.

Now, the honest part. Li wasn't a lone tinkerer grinding away in a bedroom. The bedroom belongs to the other half of this story. She had Princeton, then Stanford. She had grant money and a lab. And that matters, because the foundation was too big to be heroic. You cannot label 14 million images by being stubborn and staying up late. It needs coordination and time at a scale that has to be held by something. Historically that something has been the institution. The research university that can bankroll four years of work with no payoff to show for it.

But the institution wasn't the only thing doing the work. Here's how the labeling actually happened. They used Amazon Mechanical Turk, a platform where anyone could sign up to do small online tasks for small amounts of money. Roughly 49,000 people across 167 countries did the labeling. They looked at candidate images, checked them against a definition, and confirmed whether the picture showed the thing it was supposed to show. The nearly-two-decades estimate was for a small team doing it the old way. Distributed across tens of thousands of people, it took about two.

That's worth sitting with. The breakthrough underneath the breakthrough wasn't a smarter algorithm. It was coordination. Tens of thousands of strangers each doing a sliver of a job no single team could finish.

And almost none of those people were AI researchers. They were ordinary people who needed a little extra money. They happened to be a part of something much bigger than themselves, even if they didn't know it. All of their contributions led to what we've got now with this AI boom that has completely changed the world.

So when you line it up, the famous part is the smallest part. The neural network design wasn't new. Convolutional networks went back to Yann LeCun's work in the 1980s. The gaming GPUs weren't built for AI. They were built for video games. The dataset wasn't a model at all. What made 2012 work was the whole stack arriving at once. Old idea, repurposed hardware, and underneath both of them, a foundation that 49,000 people had quietly spent years building.

Here's what I keep coming back to, and it's the reason this old story feels urgent to me right now.

That same structure, many ordinary hands on one big thing, is available to almost anyone today. You used to need to be a coding whiz or a trained researcher to be part of building something like this. You don't anymore. The tools collapsed the barrier. The open-source projects shaping AI right now run on exactly the model that built ImageNet. More hands. More perspectives. More people who don't yet know what they're a part of.

But it doesn't run on goodwill. It never did. The Turkers got paid. That's the part people skip when they tell this story. The foundation didn't get built because people loved the mission. It got built because someone designed a structure where ordinary self-interest and a massive collective good pointed in the same direction.

The infrastructure is there to get massive amounts of people to take on projects like this. If we can figure out the incentive structure.

And we already have the proof. ImageNet showed it works. The only real question left is what you pay people in. Money is one answer. Status, ownership, belonging, learning, a cut of what gets built, those are answers too. The incentive doesn't have to be cash. It just has to be real.

That's not a problem to wait on. It's a thing to design.