The laboratories

Google Brain was born in 2011 as a secret project inside Google X, the lab for extravagant ideas. The premise was simple: what happens if you give a neural network access to Google's infrastructure?

The answer came quickly. In June 2012, Andrew Ng and Jeff Dean's team published a paper that became famous for an unexpected reason: their neural network had learned to recognize cats in YouTube videos. Without anyone telling it what a cat was.

"We didn't show it the word 'cat'. We didn't tell it what to look for. The network found the concept on its own, because cats appear a lot on the internet." — Jeff Dean, 2012

The cat paper was a turning point. Not because recognizing cats was useful, but because it demonstrated that neural networks could discover abstract concepts without human supervision. If they could find cats, what else could they find?

While Google was experimenting, another lab was being born in Montreal. Yoshua Bengio, along with a group of students that included Ian Goodfellow and Aaron Courville, was building MILA — the Montreal Institute for Learning Algorithms. Their approach was different: less engineering and more basic science.

The tension between these two approaches — Google's engineering at scale and academia's fundamental research — would define the next decade of AI. And it still defines it today.