New hardware debuts from Alphabet Inc.'s Google in San Francisco last week showed how the acquisition of London-based artificial intelligence company DeepMind might start to generate revenue rather than just research papers.
Alphabet bought DeepMind for a reported $400 million in 2014. The firm has produced a steady stream of machine learning research -- from teaching software to play the strategy game Go better than any human on the planet to creating artificial intelligence that mimicked the human brain's ability to dream and even imagine future actions. Most of these breakthroughs, however, have seemed academic, with potential commercial uses seemingly years or even decades away. In the first year Alphabet owned the company, DeepMind reported zero revenue.
But when Google showcased its new digital assistant recently in its flagship Pixel phone, laptop computer, and Google Home devices, the company noted how much more natural its computer-generated voice had become. That more human-sounding speech, at least when the assistant speaks in English and Japanese, is thanks to an algorithm DeepMind invented last year and which it took from research to full-scale commercial application inside 12 months.
When DeepMind first published a 2016 paper on WaveNet, a new way of using software to generate speech, tests showed that human listeners rated it more natural-sounding than existing technologies by a 50 percent margin. But the method was too computationally intensive, requiring a lot of computation power. Even DeepMind's own researchers said the system was "not something we could deploy in the real world."
Over the past year it found a way to make the WaveNet algorithm 1,000 times faster while also allowing it produce even higher fidelity sound. This results in computer speech that human listeners judge to be almost indistinguishable from a real person's voice. These efficiency improvements -- run using data centers equipped with Google's new computer chips -- have allowed the company to deploy WaveNet into its new assistant.
This clarity will be useful for DeepMind's bottom line. Since joining in 2015, Google's Chief Financial Officer Ruth Porat has been on a mission to drive down costs across Alphabet, forcing some of Google's more left-field projects to show a path to profitability.
DeepMind continues to remain independent from its parent company, but its contribution to Google's product launch is well timed. It reported its first-ever revenues -- $30.3 million in 2106 -- from products and services it supplied to other Alphabet companies, according to filings made public on the United Kingdom business registry Companies House.
This is just one example of how DeepMind is starting to help Google. Some others that DeepMind has been willing to talk about include supplying algorithms that have helped Google increase the energy efficiency of its data centers by 15 percent, and also improvements to Google's core ad words product that DeepMind says it cannot detail.
Of course, hiring all the brainpower to produce those algorithms means that DeepMind is still shelling out far more cash than it takes in. The company reported that "staff and other related costs" equaled nearly $136 million in 2016 and the company reported an overall loss almost $123 million, nearly double the amount it reported the year before.
SundayMonday Business on 10/09/2017