AI at Google
A snippet I found relevant to Google in Thomas H. Davenport's book The AI Advantage.
"Google has been, perhaps not surprisingly, the most active developer and user of AI among the Internet giants—and perhaps all companies in the world. The company, working with Stanford professor Andrew Ng, began to research AI (deep learning in particular) in its Google X research labs in 2011. The project came to be known as Google Brain. The method of choice was deep learning, which was used for image recognition, among other tasks. By 2012 the group had conquered one of the most pressing problems of humankind: how to get a machine to identify a photo of a cat on the internet.
The next year, Google hired Geoffrey Hinton, the University of Toronto researcher who had helped to revive neural networks. In 2014 Google bought DeepMind, a London-based firm with deep expertise in deep learning. The group’s tools were used to help AlphaGo, Google’s machine that plays the ancient game Go, beat one of the world’s best human players. In 2016, the Google Brain organization helped Google make a major improvement in the ability of Google Translate to do accurate translations. By that year Google, or its parent company Alphabet, was employing machine learning in over 2,700 different projects across the company, including search algorithms (RankBrain), self-driving cars (now in the Alphabet subsidiary Waymo), and medical diagnostics (in the Calico subsidiary).18 In the Silicon Valley tradition, Google also made its TensorFlow machine learning library available for free in 2015 as an open source project, and it has become popular among more sophisticated companies that use AI."
Thomas skips on some of the early history of AI at Google which traces back to another popular Silicon Valley startup and later on public company Epiphany (later renamed E.piphany due to naming rights being owned by a bible publisher in Indiana.). The machine learning team from Epiphany, which included my former E.piphany colleagues Mehran Sahami, now teaching AI at Stanford, and Sridhar Ramaswamy, later VP of all of Google Ads products. Both of them worked on (myself included) SmartASS, or Smart Ads Serving System, which was a giant probabilistic prediction model building and ads serving framework which uses logistic regression for ad click through rate prediction. SmartASS does this very hard calculation in a special linear approximation thus making this prediction practically possible in a resonable ammount of time. Over the course of years SmartASS was refined and extended with many different subsystems ranging from topical modelling, click spam detection, fraud detection, content ad classification and many other improvements.