In the summer of 2014, a graduate student in Montreal built a fix for a stubborn bug in a translation model, one that would outlive the paper it was written for. He didn't call it attention. For a while, neither did his supervisor.
In 1957, a linguist wrote the sentence that word embeddings would spend the next fifty-six years trying to compute. The breakthrough wasn't a new idea. It was making an old one fast enough to run.
Before AlexNet made headlines in 2012, two graduate students in Toronto had already proven the same idea worked — on something harder than pictures: sound.
Before Rumelhart, Hinton, and Williams got the credit in 1986, a lonely doctoral thesis had already cracked backpropagation — and almost no one noticed.
Months before Expert Systems ever hit a balance sheet, the hardware behind them was already fracturing — over a disagreement that had nothing to do with technology.
