Tech | Google sprints ahead in AI building blocks, leaving rivals wary

There’s a high-stakes race under way in Silicon Valley to develop software that makes it easy to weave artificial intelligence technology into almost everything, and Google has sprinted into the lead.
Google computer scientists including Jeff Dean and Greg Corrado built software called TensorFlow, which simplifies the programming of key systems that underpin artificial intelligence. That helps Google make its products smarter and more responsive. It’s important for other companies too because the software makes it dramatically easier to create computer programs that learn and improve automatically. What’s more, Google gives it away.
But for some competitors, there’s a big downside to adopting Google’s standard. Using TensorFlow will help Google recruit more AI experts by training them on the same tool it uses internally, spotting their code, and hiring the best contributors. It could also let the search-engine provider exert outsize influence over the burgeoning AI ecosystem. If the internet giant dominates in this field, it could gain an advantage in the fast-growing cloud-computing business, turning the popularity of its software into real revenue.
“It’s the next big area, and people are worried Google’s going to own the show,” said Ed Lazowska, a computer science professor at the University of Washington who has served on the technical advisory board of Microsoft Corp.’s research lab. “There is a network effect, and it’s a really excellent system.”
Google initially used TensorFlow internally for products like its Inbox and Photos apps. The company made it available for free in November. Technology companies like Microsoft Corp., Amazon.com Inc. and Samsung Electronics Co. rushed to give away their own versions, hoping to get the most outside developers using their standards.
The company that wins will benefit from the collective efforts of thousands of developers using, but also updating and improving, its system. That’s an advantage when it comes time to make money from the new asset. Whoever has the most popular software will have the best chance of creating commercial cloud services for AI because potential customers will already know how to use it.
Success in these types of open-source projects sometimes yields big rewards. Google released Android for free in 2008, and it’s now the most widely used mobile operating system with over 400,000 developers and more than a billion users. Google generates billions of dollars a year from ads shown on Android devices and the cut it gets from revenue app developers make through the operating system.
Since emerging, TensorFlow has become the most popular AI programming project on software code sharing service GitHub, leapfrogging well-regarded systems created by universities and corporate rivals, according to data gathered by Bloomberg.
On launch day, TensorFlow had around 3,000 “stars” on GitHub, meaning that number of programmers had bookmarked the code, indicating interest. As of July 13, it had 27,873. Two other popular AI software projects, Theano and Torch, have less than a fifth of that following. In 2014, Torch was the leader. A Microsoft tool called CNTK, released for free in January, and Amazon’s free DSSTNE, which rolled out in May, have so far failed to dent Google’s lead much.
Linux, an open-source operating system launched in 1991, now helps run everything from supercomputers to phones to airplanes and helped turn Red Hat Inc. into a USD13 billion enterprise software company. Linux has 33,967 stars on Git-
Hub. “It’s kind of crazy,” said Dean, a top Google engineer and one of the main developers of TensorFlow. “We’re almost to Linux level.”
Google will soon begin generating revenue from this lead. It plans to offer a version of TensorFlow that runs on its Google Cloud Platform service, letting people and businesses pay to run their AI software in Google’s data centers.
Google made the software free so it could give the community a useful tool “and everyone could standardize on that,” said Corrado, a senior research scientist at Google. “In a giant green field, trying to build a fence around the next blade of grass is really absurdist. It’s really better to help everybody run into that field.”
That openness, and continual Google updates, have lured developers like computer-vision startup Matroid, which re-wrote its software to work with TensorFlow, after building on another free AI tool called Caffe. Kindred, a robotics startup, made a similar switch.
Not everyone is so keen. As TensorFlow’s usage grows, some companies are realizing an increasingly important part of the technology toolkit is controlled by Google, and they don’t want to exacerbate that trend.
They’re “skeptical about using a language backed by another large company,” said Soumith Chintala, a Facebook Inc. artificial intelligence researcher and one of the people behind Torch.
The unease stems from the fact Google can tweak TensorFlow to suit its own purposes, he said. If the company changes the software too much, then other companies that have adopted it will need to make a copy of the software and rewrite it to suit their own needs – an expensive and time-consuming process known as forking.
That’s led some to look elsewhere. Skymind, which makes free AI software, has had more than five customers tell it they are wary of using TensorFlow, said CEO Chris Nicholson. He declined to name any of the companies, citing non-disclosure agreements.
Since TensorFlow launched, designers of other AI programming projects have been inundated with queries from companies that don’t want to rely on Google. Several reached out to the creators of Theano, developed mainly at the University of Montreal, to see if they can donate resources to the project, according to Yoshua Bengio, a professor who leads AI research at the school.
The same happened with Torch, said Chintala. Facebook does much of its AI research with Torch, and Chintala helps guide development of the project in his spare time. He and other backers moved Torch into a non-profit organization called SPI Inc. in May to make it easier for more people to work on the language and donate to it.
“One of the reasons we want to stick with Torch is to create a strong counterpoint” to TensorFlow, said Clement Farabet, who helped develop Torch. He now works at Twitter Inc., which uses Torch to run AI systems that analyze images and select tweets people may want to read. It’s better for the community if there’s a choice of AI software, he said.
Google could solve some of these problems by donating TensorFlow to a neutral third-
party, said Bengio, who has discussed his ideas with the company. This structure could “provide neutral software for all,” he said.
Google has no plans to do that, but it’s open to letting outside people have a say in what code gets merged into the main software, said Jason Freidenfelds, a spokesman for Google.
Google’s strategy may be dictated by past failings, said Reza Zadeh, chief executive officer of Matroid, who worked at Google a decade ago. Back then, Google developed the Google File System and MapReduce to store and analyze lots of data. It published research papers on them, but no code. Some employees at Yahoo! Inc. used the research to create Hadoop, technology that underpins public company Hortonworks Inc. and larger private rival Cloudera Inc.
“They’ve learned from that,” said Zadeh.  Jack Clark, Bloomberg