In early 2014, Srikanth Thirumalai met with Amazon CEO Jeff Bezos. Thirumalai, a computer scientist who’d left IBM in 2005 to head Amazon’s recommendations team, had come to propose a sweeping new plan for incorporating the latest advances in artificial intelligence into his division.
He arrived armed with a “six-pager.” Bezos had long ago decreed that products and services proposed to him must be limited to that length, and include a speculative press release describing the finished product, service, or initiative. Now Bezos was leaning on his deputies to transform the company into an AI powerhouse. Amazon’s product recommendations had been infused with AI since the company’s very early days, as had areas as disparate as its shipping schedules and the robots zipping around its warehouses. But in recent years, there has been a revolution in the field; machine learning has become much more effective, especially in a supercharged form known as deep learning. It has led to dramatic gains in computer vision, speech, and natural language processing.
In the early part of this decade, Amazon had yet to significantly tap these advances, but it recognized the need was urgent. This era’s most critical competition would be in AI—Google, Facebook, Apple, and Microsoft were betting their companies on it—and Amazon was falling behind. “We went out to every [team] leader, to basically say, ‘How can you use these techniques and embed them into your own businesses?’” says David Limp, Amazon’s VP of devices and services.
Thirumalai took that to heart, and came to Bezos for his annual planning meeting with ideas on how to be more aggressive in machine learning. But he felt it might be too risky to wholly rebuild the existing system, fine-tuned over 20 years, with machine-learning techniques that worked best in the unrelated domains of image and voice recognition. “No one had really applied deep learning to the recommendations problem and blown us away with amazingly better results,” he says. “So it required a leap of faith on our part.” Thirumalai wasn’t quite ready—but Bezos wanted more. So Thirumalai shared his edgier option of using deep learning to revamp the way recommendations worked. It would require skills that his team didn’t possess, tools that hadn’t been created, and algorithms that no one had thought of yet. Bezos loved it (though it isn’t clear whether he greeted it with his trademark hyena-esque laugh), so Thirumalai rewrote his press release and went to work.
Thirumalai was only one of a procession of company leaders who trekked to Bezos a few years ago with six-pagers in hand. The ideas they proposed involved completely different products with different sets of customers. But each essentially envisioned a variation of Thirumalai’s approach: transforming part of Amazon with advanced machine learning. Some of them involved rethinking current projects, like the company’s robotics efforts and its huge data-center business, Amazon Web Services (AWS). Others would create entirely new businesses, like a voice-based home appliance that would become the Echo.
The results have had an impact far beyond the individual projects. Thirumalai says that at the time of his meeting, Amazon’s AI talent was segregated into isolated pockets. “We would talk, we would have conversations, but we wouldn’t share a lot of artifacts with each other because the lessons were not easily or directly transferable,” he says. They were AI islands in a vast engineering ocean. The push to overhaul the company with machine learning changed that.
While each of those six-pagers hewed to Amazon’s religion of “single-threaded” teams—meaning that only one group “owns” the technology it uses—people started to collaborate across projects. In-house scientists took on hard problems and shared their solutions with other groups. Across the company, AI islands became connected. As Amazon's ambition for its AI projects grew, the complexity of its challenges became a magnet for top talent, especially those who wanted to see the immediate impact of their work. This compensated for Amazon's aversion to conducting pure research; the company culture demanded that innovations come solely in the context of serving its customers.
Amazon loves to use the word flywheel to describe how various parts of its massive business work as a single perpetual motion machine. It now has a powerful AI flywheel, where machine-learning innovations in one part of the company fuel the efforts of other teams, who in turn can build products or offer services to affect other groups, or even the company at large. Offering its machine-learning platforms to outsiders as a paid service makes the effort itself profitable—and in certain cases scoops up yet more data to level up the technology even more.
It took a lot of six-pagers to transform Amazon from a deep-learning wannabe into a formidable power. The results of this transformation can be seen throughout the company—including in a recommendations system that now runs on a totally new machine-learning infrastructure. Amazon is smarter in suggesting what you should read next, what items you should add to your shopping list, and what movie you might want to watch tonight. And this year Thirumalai started a new job, heading Amazon search, where he intends to use deep learning in every aspect of the service.
“If you asked me seven or eight years ago how big a force Amazon was in AI, I would have said, ‘They aren’t,’” says Pedro Domingos, a top computer science professor at the University of Washington. “But they have really come on aggressively. Now they are becoming a force.”
Maybe the force.
The flagship product of Amazon’s push into AI is its breakaway smart speaker, the Echo, and the Alexa voice platform that powers it. These projects also sprang from a six-pager, delivered to Bezos in 2011 for an annual planning process called Operational Plan One. One person involved was an executive named Al Lindsay, an Amazonian since 2004, who had been asked to move from his post heading the Prime tech team to help with something totally new. “A low-cost, ubiquitous computer with all its brains in the cloud that you could interact with over voice—you speak to it, it speaks to you,” is how he recalls the vision being described to him.
But building that system—literally an attempt to realize a piece of science fiction, the chatty computer from Star Trek—required a level of artificial intelligence prowess that the company did not have on hand. Worse, of the very few experts who could build such a system, even fewer wanted to work for Amazon. Google and Facebook were snapping up the top talent in the field. “We were the underdog,” Lindsay, who is now a VP, says.
“Amazon had a bit of a bad image, not friendly to people who were research oriented,” says Domingos, the University of Washington professor. The company’s relentless focus on the customer, and its culture of scrappiness, did not jibe with the pace of academia or cushy perks of competitors. “At Google you’re pampered,” Domingos says. “At Amazon you set up your computer from parts in the closet.” Worse, Amazon had a reputation as a place where innovative work was kept under corporate wraps. In 2014, one of the top machine-learning specialists, Yann LeCun, gave a guest lecture to Amazon’s scientists in an internal gathering. Between the time he was invited and the event itself, LeCun accepted a job to lead Facebook’s research effort, but he came anyway. As he describes it now, he gave his talk in an auditorium of about 600 people and then was ushered into a conference room where small groups came in one by one and posed questions to him. But when he asked questions of them, they were unresponsive. This turned off LeCun, who had chosen Facebook in part because it agreed to open-source much of the work of its AI team.
Because Amazon didn’t have the talent in-house, it used its deep pockets to buy companies with expertise. “In the early days of Alexa, we bought many companies,” Limp says. In September 2011, it snapped up Yap, a speech-to-text company with expertise in translating the spoken word into written language. In January 2012, Amazon bought Evi, a Cambridge, UK, AI company whose software could respond to spoken requests like Siri does. And in January 2013, it bought Ivona, a Polish company specializing in text-to-speech, which provided technology that enabled Echo to talk.
But Amazon’s culture of secrecy hampered its efforts to attract top talent from academia. One potential recruit was Alex Smola, a superstar in the field who had worked at Yahoo and Google. “He is literally one of the godfathers of deep learning,” says Matt Wood, the general manager of deep learning and AI at Amazon Web Services. (Google Scholar lists more than 90,000 citations of Smola's work.) Amazon execs wouldn’t even reveal to him or other candidates what they would be working on. Smola rejected the offer, choosing instead to head a lab at Carnegie Mellon.
“Even until right before we launched there was a headwind,” Lindsay says. “They would say, ‘Why would I want to work at Amazon—I’m not interested in selling people products!’”
Amazon did have one thing going for it. Since the company works backward from an imagined final product (thus the fanciful press releases), the blueprints can include features that haven’t been invented yet. Such hard problems are irresistible to ambitious scientists. The voice effort in particular demanded a level of conversational AI—nailing the “wake word” (“Hey Alexa!”), hearing and interpreting commands, delivering non-absurd answers—that did not exist.
That project, even without the specifics on what Amazon was building, helped attract Rohit Prasad, a respected speech-recognition scientist at Boston-based tech contractor Raytheon BBN. (It helped that Amazon let him build a team in his hometown.) He saw Amazon’s lack of expertise as a feature, not a bug. “It was green fields here,” he says. “Google and Microsoft had been working on speech for years. At Amazon we could build from scratch and solve hard problems.” As soon as he joined in 2013, he was sent to the Alexa project. “The device existed in terms of the hardware, but it was very early in speech,” he says.
The trickiest part of the Echo—the problem that forced Amazon to break new ground and in the process lift its machine-learning game in general—was something called far field speech recognition. It involves interpreting voice commands spoken some distance from the microphones, even when they are polluted with ambient noise or other aural detritus. One challenging factor was that the device couldn’t waste any time cogitating about what you said. It had to send the audio to the cloud and produce an answer quickly enough that it felt like a conversation, and not like those awkward moments when you’re not sure if the person you’re talking to is still breathing. Building a machine-learning system that could understand and respond to conversational queries in noisy conditions required massive amounts of data—lots of examples of the kinds of interactions people would have with their Echos. It wasn’t obvious where Amazon might get such data.
Far-field technology had been done before, says Limp, the VP of devices and services. But “it was on the nose cone of Trident submarines, and it cost a billion dollars.” Amazon was trying to implement it in a device that would sit on a kitchen counter, and it had to be cheap enough for consumers to spring for a weird new gadget. “Nine out of 10 people on my team thought it couldn’t be done,” Prasad says. “We had a technology advisory committee of luminaries outside Amazon—we didn’t tell them what we were working on, but they said, ‘Whatever you do, don’t work on far field recognition!’”
Prasad’s experience gave him confidence that it could be done. But Amazon did not have an industrial-strength system in place for applying machine learning to product development. “We had a few scientists looking at deep learning, but we didn’t have the infrastructure that could make it production-ready,” he says. The good news was that all the pieces were there at Amazon—an unparalleled cloud service, data centers loaded with GPUs to crunch machine-learning algorithms, and engineers who knew how to move data around like fireballs.
His team used those parts to create a platform that was itself a valuable asset, beyond its use in fulfilling the Echo’s mission. “Once we developed Echo as a far-field speech recognition device, we saw the opportunity to do something bigger—we could expand the scope of Alexa to a voice service,” says Alexa senior principal scientist Spyros Matsoukas, who had worked with Prasad at Raytheon BBN. (His work there had included a little-known Darpa project called Hub4, which used broadcast news shows and intercepted phone conversations to advance voice recognition and natural language understanding—great training for the Alexa project.) One immediate way they extended Alexa was to allow third-party developers to create their own voice-technology mini-applications—dubbed “skills”—to run on the Echo itself. But that was only the beginning.
By breaking out Alexa beyond the Echo, the company’s AI culture started to coalesce. Teams across the company began to realize that Alexa could be a useful voice service for their pet projects too. “So all that data and technology comes together, even though we are very big on single-threaded ownership,” Prasad says. First other Amazon products began integrating into Alexa: When you speak into your Alexa device you can access Amazon Music, Prime Video, your personal recommendations from the main shopping website, and other services. Then the technology began hopscotching through other Amazon domains. “Once we had the foundational speech capacity, we were able to bring it to non-Alexa products like Fire TV, voice shopping, the Dash wand for Amazon fresh, and, ultimately, AWS,” Lindsay says.
The AI islands within Amazon were drawing closer.
Another pivotal piece of the company’s transformation clicked into place once millions of customers (Amazon won’t say exactly how many) began using the Echo and the family of other Alexa-powered devices. Amazon started amassing a wealth of data—quite possibly the biggest collection of interactions of any conversation-driven device ever. That data became a powerful lure for potential hires. Suddenly, Amazon rocketed up the list of places where those coveted machine-learning experts might want to work. “One of the things that made Alexa so attractive to me is that once you have a device in the market, you have the resource of feedback. Not only the customer feedback, but the actual data that is so fundamental to improving everything—especially the underlying platform,” says Ravi Jain, an Alexa VP of machine learning who joined the company last year.
So as more people used Alexa, Amazon got information that not only made that system perform better but supercharged its own machine-learning tools and platforms—and made the company a hotter destination for machine-learning scientists.
The flywheel was starting to spin.
Amazon began selling Echo to Prime customers in 2014. That was also the year that Swami Sivasubramanian became fascinated with machine learning. Sivasubramanian, who was managing the AWS database and analytics business at the time, was on a family trip to India, when due to a combination of jet lag and a cranky infant daughter, he found himself at his computer late into the night fiddling with tools like Google’s Tensorflow and Caffé, which is the machine-learning framework favored by Facebook and many in the academic community. He concluded that combining these tools with Amazon’s cloud service could yield tremendous value. By making it easy to run machine-learning algorithms in the cloud, he thought, the company might tap into a vein of latent demand. “We cater to millions of developers every month,” he says. “The majority are not professors at MIT but developers who have no background in machine learning.”
At his next Jeff Bezos review he came armed with an epic six-pager. On one level, it was a blueprint for adding machine-learning services to AWS. But Sivasubramanian saw it as something broader: a grand vision of how AWS could become the throbbing center of machine-learning activity throughout all of techdom.
In a sense, offering machine learning to the tens of thousands of Amazon cloud customers was inevitable. “When we first put together the original business plan for AWS, the mission was to take technology that was only in reach of a small number of well-funded organizations and make it as broadly distributed as possible,” says Wood, the AWS machine-learning manager. “We’ve done that successfully with computing, storage, analytics, and databases—and we’re taking the exact same approach with machine learning.” What made it easier was that the AWS team could draw on the experience that the rest of the company was accumulating.
AWS’s Amazon Machine Learning, first offered in 2015, allows customers like C-Span to set up a private catalog of faces, Wood says. Zillow uses it to estimate house prices. Pinterest employs it for visual search. And several autonomous driving startups are using AWS machine learning to improve products via millions of miles of simulated road testing.
In 2016, AWS released new machine-learning services that more directly drew on the innovations from Alexa—a text-to-speech component called Polly and a natural language processing engine called Lex. These offerings allowed AWS customers, which span from giants like Pinterest and Netflix to tiny startups, to build their own mini Alexas. A third service involving vision, Rekognition, drew on work that had been done in Prime Photos, a relatively obscure group at Amazon that was trying to perform the same deep-learning wizardry found in photo products by Google, Facebook, and Apple.
These machine-learning services are both a powerful revenue generator and key to Amazon’s AI flywheel, as customers as disparate as NASA and the NFL are paying to get their machine learning from Amazon. As companies build their vital machine-learning tools inside AWS, the likelihood that they will move to competing cloud operations becomes ridiculously remote. (Sorry, Google, Microsoft, or IBM.) Consider Infor, a multibillion-dollar company that creates business applications for corporate customers. It recently released an extensive new application called Coleman (named after the NASA mathematician in Hidden Figures) that allows its customers to automate various processes, analyze performance, and interact with data all through a conversational interface. Instead of building its own bot from scratch, it uses AWS’s Lex technology. “Amazon is doing it anyway, so why would we spend time on that? We know our customers and we can make it applicable to them,” says Massimo Capoccia, a senior VP of Infor.
AWS’s dominant role in the ether also gives it a strategic advantage over competitors, notably Google, which had hoped to use its machine-learning leadership to catch up with AWS in cloud computing. Yes, Google may offer customers super-fast, machine-learning-optimized chips on its servers. But companies on AWS can more easily interact with—and sell to—firms that are also on the service. “It’s like Willie Sutton saying he robs banks because that’s where the money is,” says DigitalGlobe CTO Walter Scott about why his firm uses Amazon’s technology. “We use AWS for machine learning because that’s where our customers are.”
Last November at the AWS re:Invent conference, Amazon unveiled a more comprehensive machine-learning prosthetic for its customers: SageMaker, a sophisticated but super easy-to-use platform. One of its creators is none other than Alex Smola, the machine-learning superstar with 90,000 academic citations who spurned Amazon five years ago. When Smola decided to return to industry, he wanted to help create powerful tools that would make machine learning accessible to everyday software developers. So he went to the place where he felt he’d make the biggest impact. “Amazon was just too good to pass up,” he says. “You can write a paper about something, but if you don’t build it, nobody will use your beautiful algorithm,” he says.
When Smola told Sivasubramanian that building tools to spread machine learning to millions of people was more important than publishing one more paper, he got a nice surprise. “You can publish your paper, too!” Sivasubramanian said. Yes, Amazon is now more liberal in permitting its scientists to publish. “It’s helped quite a bit with recruiting top talent as well as providing visibility of what type of research is happening at Amazon,” says Spyros Matsoukas, who helped set guidelines for a more open stance.
It’s too early to know if the bulk of AWS’s million-plus customers will begin using SageMaker to build machine learning into their products. But every customer that does will find itself heavily invested in Amazon as its machine-learning provider. In addition, the platform is sufficiently sophisticated that even AI groups within Amazon, including the Alexa team, say they intend to become SageMaker customers, using the same toolset offered to outsiders. They believe it will save them a lot of work by setting a foundation for their projects, freeing them to concentrate on the fancier algorithmic tasks.
Even if only some of AWS’s customers use SageMaker, Amazon will find itself with an abundance of data about how its systems perform (excluding, of course, confidential information that customers keep to themselves). Which will lead to better algorithms. And better platforms. And more customers. The flywheel is working overtime.
With its machine learning overhaul in place, the company’s AI expertise is now distributed across its many teams—much to the satisfaction of Bezos and his consiglieri. While there is no central office of AI at Amazon, there is a unit dedicated to the spread and support of machine learning, as well as some applied research to push new science into the company’s projects. The Core Machine Learning Group is led by Ralf Herbrich, who worked on the Bing team at Microsoft and then served a year at Facebook, before Amazon lured him in 2012. “It’s important that there’s a place that owns this community” within the company, he says. (Naturally, the mission of the team was outlined in an aspirational six-pager approved by Bezos.)
Part of his duties include nurturing Amazon’s fast-growing machine-learning culture. Because of the company’s customer-centric approach—solving problems rather than doing blue-sky research—Amazon execs do concede that their recruiting efforts will always tilt towards those interested in building things rather than those chasing scientific breakthroughs. Facebook’s LeCun puts it another way: “You can do quite well by not leading the intellectual vanguard.”
But Amazon is following Facebook and Google’s lead in training its workforce to become adept at AI. It runs internal courses on machine-learning tactics. It hosts a series of talks from its in-house experts. And starting in 2013, the company has hosted an internal machine-learning conference at its headquarters every spring, a kind of Amazon-only version of NIPS, the premier academic machine-learning-palooza. “When I started, the Amazon machine-learning conference was just a couple hundred people; now it’s in the thousands,” Herbrich says. “We don’t have the capacity in the largest meeting room in Seattle, so we hold it there and stream it to six other meeting rooms on the campus.” One Amazon exec remarks that if it gets any bigger, instead of calling it an Amazon machine-learning event, it should just be called Amazon.
Herbrich’s group continues to push machine learning into everything the company attempts. For example, the fulfillment teams wanted to better predict which of the eight possible box sizes it should use with a customer order, so they turned to Herbrich’s team for help. “That group doesn’t need its own science team, but it needed these algorithms and needed to be able to use them easily,” he says. In another example, David Limp points to a transformation in how Amazon predicts how many customers might buy a new product. “I’ve been in consumer electronics for 30 years now, and for 25 of those forecasting was done with [human] judgment, a spreadsheet, and some Velcro balls and darts,” he says. “Our error rates are significantly down since we’ve started using machine learning in our forecasts.”
Still, sometimes Herbrich’s team will apply cutting-edge science to a problem. Amazon Fresh, the company’s grocery delivery service, has been operating for a decade, but it needed a better way to assess the quality of fruits and vegetables—humans were too slow and inconsistent. His Berlin-based team built sensor-laden hardware and new algorithms that compensated for the inability of the system to touch and smell the food. “After three years, we have a prototype phase, where we can judge the quality more reliably” than before, he says.
Of course, such advances can then percolate throughout the Amazon ecosystem. Take Amazon Go, the deep-learning-powered cashier-less grocery store in its headquarters building that recently opened to the public. “As a customer of AWS, we benefit from its scale,” says Dilip Kumar, VP of Technology for Amazon Go. “But AWS is also a beneficiary.” He cites as an example Amazon Go’s unique system of streaming data from hundreds of cameras to track the shopping activities of customers. The innovations his team concocted helped influence an AWS service called Kinesis, which allows customers to stream video from multiple devices to the Amazon cloud, where they can process it, analyze it, and use it to further advance their machine learning efforts.
Even when an Amazon service doesn’t yet use the company’s machine-learning platform, it can be an active participant in the process. Amazon’s Prime Air drone-delivery service, still in the prototype phase, has to build its AI separately because its autonomous drones can’t count on cloud connectivity. But it still benefits hugely from the flywheel, both in drawing on knowledge from the rest of the company and figuring out what tools to use. “We think about this as a menu—everybody is sharing what dishes they have,” says Gur Kimchi, VP of Prime Air. He anticipates that his team will eventually have tasty menu offerings of its own. “The lessons we’re learning and problems we’re solving in Prime Air are definitely of interest to other parts of Amazon,” he says.
In fact, it already seems to be happening. “If somebody’s looking at an image in one part of the company, like Prime Air or Amazon Go, and they learn something and create an algorithm, they talk about it with other people in the company,” says Beth Marcus, a principal scientist at Amazon robotics. “And so someone in my team could use it to, say, figure out what’s in an image of a product moving through the fulfillment center.”
Is it possible for a company with a product-centered approach to eclipse the efforts of competitors staffed with the superstars of deep learning? Amazon’s making a case for it. “Despite the fact they’re playing catchup, their product releases have been incredibly impressive,” says Oren Etzioni, CEO of the Allen Institute for Artificial Intelligence. “They’re a world-class company and they’ve created world-class AI products.”
The flywheel keeps spinning, and we haven’t seen the impact of a lot of six-pager proposals still in the pipeline. More data. More customers. Better platforms. More talent.
Alexa, how is Amazon doing in AI?
The answer? Jeff Bezos’s braying laugh.