Loading slide
Imagine you've spent ten years improving a method, and you get it to 26% error. Then someone walks in and gets 15%.
Not 24%. Not 22%. Fifteen.
Researchers in the room assumed it was a mistake. A reporting error, a miscalculation, something. It was too far outside the range of what normal progress looked like. Normal progress was half a percentage point a year.
When it held up, the response wasn't celebration exactly. It was more like recognition. Something had crossed a line. The conversation that had been happening in computer vision for a decade, what's the best way to do this?, had just been answered. And the answer came from the direction most people in the room hadn't taken seriously: instead of humans hand-designing what the system should look for, you let a deep network find those features for itself, just by training on labeled images.
Some researchers started asking whether what they'd been working on was still worth doing. Others immediately started retraining as deep learning researchers. Within a year, the conference had a different shape.
That's what a turning point feels like from inside: not like triumph, but like realizing the map was wrong. The field had believed progress would come from humans engineering ever-cleverer features by hand. It came from scale instead: a deep enough network, fed enough labeled data, left to work out for itself what mattered.