Google has always used its annual I/O conference to connect to developers in its sprawling empire. It announces new tools and initiatives, sprinkles in a little hype, and then tells those watching: choose us, and together we’ll go far. But while in previous years this message has been directed at coders working with Android and Chrome — the world’s biggest mobile OS and web browser respectively — yesterday, CEO Sundar Pichai made it clear that the next platform the company wants to dominate could be even bigger: artificial intelligence.
For Google, this doesn’t just mean using AI to improve its own products. (Although it’s certainly doing that). The company wants individuals and small companies around the world to also get on board. It wants to wield influence in the wider AI ecosystem, and to do so has put together an impressive stack of machine learning tools — from software to servers — that mean you can build an AI product from the ground up without ever leaving the Google playpen.
The heart of this offering is Google’s machine learning software TensorFlow. For building AI tools, it’s like the difference between a command line interface and a modern desktop OS; giving users an accessible framework for grappling with their algorithms. It started life as an in-house tool for the company’s engineers to design and train AI algorithms, but in 2015 was made available for anyone to use as open-source software. Since then, it’s been embraced by the AI community (it’s the most popular software of its type on code repository Github), and is used to create custom tools for a whole range of industries, from aerospace to bioengineering.
“There’s hardly a way around TensorFlow these days,” says Samim Winiger, head of machine learning design studio Samim.io. “I use a lot of open source learning libraries, but there’s been a major shift to TensorFlow.”
Google has made strategic moves to ensure the software is widely used. Earlier this year, for example, it added support for Keras, another popular deep learning framework. According to calculations by the creator of Keras, François Chollet (himself now a Google engineer), TensorFlow was the fastest growing deep learning framework as of September 2016, with Keras in second place. Winiger describes the integration of the two as a “classic tale of Google and how they do it.” He says: “It’s another way that making sure that the entire community converges on their tooling.”
But TensorFlow is also popular for one particularly important reason: it’s good at what it does. “With TensorFlow you get something that scales quickly, works quickly,” James Donkin, a technology manager at UK-based online supermarket Ocado, tells The Verge. He says his team uses a range of machine learning frameworks to create in-house tools for tasks like categorizing customer feedback, but that TensorFlow is often a good place to start. “You get 80 percent of the benefit, and then you might decide to specialize more with other platforms.”
Google offers TensorFlow for free, but it connects easily with the company’s servers for providing data storage or computing power. (“If you use the TensorFlow library it means you can push [products] to Google’s cloud more easily,” says Donkin.) The search giant has even created its own AI-specific chips to power these operations, unveiling the latest iteration of this hardware at this year’s I/O. And, if you want to skip the task of building your own AI algorithms all together, you can buy off-the-shelf components from Google for core tasks like speech transcription and object recognition.
These products and services aren’t necessarily money-makers in themselves, but they other, subtler benefits. They attract talent to Google and help make the company’s in-house software the standard for machine learning. Winiger says these initiatives have helped Google “grab mindshare and make the company’s name synonymous with machine learning.”
Other firms like Amazon, Facebook, and Microsoft also offer their own AI tools, but it’s Google’s that feel pre-eminent. Winiger thinks this is partly down to the company’s capacity to shape the media narrative, but also because of the strong level of support it provides to its users. “There are technical differences between [different AI frameworks], but machine learning communities live off community support and forums, and in that regard Google is winning,” he tells The Verge.
This influence isn’t just abstract, either: it feeds back into Google’s own products. Yesterday, for example, Google announced that Android now has a staggering two billion monthly active users, and to keep the software’s edge, the company is honing it with machine learning. New additions to the OS span the range from tiny tweaks (like smarter text selection) to big new features (like a camera that recognizes what it’s looking at).
But Google didn’t forget to feed the community either, and to complement these announcements unveiled new tools to help developers build AI services that work better on mobile devices. These include a new version of TensorFlow named TensorFlowLite, and an API that will interface with future smartphone chips that have been optimized to work with AI software. Developers can then use these to make better machine learning products for Android devices. Google’s AI empire stretches out a bit further, and Google reaps the benefits.