Early in 2015, artificial-intelligence researchers at Google created an obscure piece of software called TensorFlow. Two years later the tool, which is used in building machine-learning software, underpins many future ambitions of Google and its parent company, Alphabet (see "50 Smartest Companies 2017.")
TensorFlow makes it much easier for the company’s engineers to translate new approaches to artificial intelligence into practical code, improving services such as search and the accuracy of speech recognition. But just months after TensorFlow was released to Google’s army of coders, the company also began offering it to the world for free.
That decision could be seen as altruistic or possibly plain dumb, but nearly two years on, the benefits to Google of its great AI giveaway are increasingly evident. Today TensorFlow is becoming the clear leader among programmers building new things with machine learning. “We have significant usage today, and it’s accelerating,” says Jeff Dean, who led TensorFlow’s design and heads Google’s core artificial-intelligence research group. Once you’ve built something with TensorFlow, you can run it anywhere—but it’s especially easy to transfer it to Google’s cloud platform. The software’s popularity is helping Google fight for a bigger share of the roughly $40 billion (and growing) cloud infrastructure market, where the company lies a distant third behind Amazon and Microsoft.
The head of Google’s cloud business, Diane Greene, said in April that she expects to take the top spot within five years, and a core part of Google’s strategy for catching up is to appeal to the sudden enthusiasm about artificial intelligence in industries from health care to autos. Companies investing in the technology are expected to spend heavily with cloud providers to avoid the costs and complexity of building and running AI themselves, just as they pay today for cloud hosting of e-mail and websites. Customers like insurer AXA—which used TensorFlow to make a system that predicts expensive traffic accidents—also get the benefits of the same infrastructure Google uses to power their own products. Google says that means better performance at competitive prices. S. Somasegar, a managing director at venture fund Madrona who was previously head of Microsoft’s developer division, says TensorFlow’s prominence poses a genuine challenge to Google’s cloud rivals. “It’s a fantastic strategy—Google is so far behind in cloud, but they’ve picked an area where they can create a beachhead,” he says.
Inside Google, TensorFlow powers products such as the Google Translate mobile app, which can translate a foreign menu in front of your eyes when you point your phone at it. The company has created specialized processors to make TensorFlow faster and reduce the power it consumes inside Google’s data centers. These processors propelled the historic victory of software called AlphaGo over a champion of the ancient board game Go last year and are credited with making possible a recent upgrade that brought Google’s translation service close to human level for some languages.
TensorFlow is far from the only tool out there for building machine-learning software, and experts can argue for hours about their individual merits. But the weight of Google’s brand and its technical advantages make its package stand out, says Reza Zadeh, an adjunct professor at Stanford. He originally built his startup Matroid, which helps companies create image recognition software, around a competing tool called Caffe, but he dumped it after trying TensorFlow. “I saw it was very clearly superior in all the technical aspects, and we decided to rip everything out,” he says.
Google’s tool is also becoming firmly lodged in the minds of the next generation of artificial-intelligence researchers and entrepreneurs. At the University of Toronto, an AI center that has schooled many of today’s leading researchers, lecturer Michael Guerzhoy teaches TensorFlow in the university’s massively oversubscribed introductory machine-learning course. “Ten years ago, it took me months to do something that for my students takes a few days with TensorFlow,” says Guerzhoy.
Since Google released TensorFlow, its competitors in cloud computing, Microsoft and Amazon, have released or started supporting their own free software tools to help coders build machine-learning systems. So far, says Guerzhoy, neither has as broad and dedicated a user base as TensorFlow among researchers, students, and working coders.
The future of open source is still very much in flux
Free and open software have transformed the tech industry. But we still have a lot to work out to make them healthy, equitable enterprises.
What’s next for the world’s fastest supercomputers
Scientists have begun running experiments on Frontier, the world’s first official exascale machine, while facilities worldwide build other machines to join the ranks.
The beautiful complexity of the US radio spectrum
The United States Frequency Allocation Chart shows how the nation’s precious radio frequencies are carefully shared.
How ubiquitous keyboard software puts hundreds of millions of Chinese users at risk
Third-party keyboard apps make typing in Chinese more efficient, but they can also be a privacy nightmare.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.