Optimus Prime—the software engine, not the Autobot overlord—was born in a basement under a West Elm furniture store on University Avenue in Palo Alto. Starting two years ago, a band of artificial-intelligence acolytes within Salesforce escaped the towering headquarters with the goal of crazily multiplying the impact of the machine learning models that increasingly shape our digital world—by automating the creation of those models. As shoppers checked out sofas above their heads, they built a system to do just that.
They named it after the Transformers leader because, as one participant recalls, “machine learning is all about transforming data.” Whether the marketing department thought better of it, or the rights weren’t available, the Transformers tie-in didn’t make it far out of that basement. Instead, Salesforce licensed the name of a different world-transforming hero—and dubbed its AI program Einstein.
The pop culture myths the company has invoked for its AI effort—the robot leader; the iconic genius—represent the kind of protean powers the technology is predicted to attain by both its most ardent hypesters and its gloomiest critics. Salesforce stands firmly on the hype side of this divide—no one cheers louder, especially not in AI promotion. But the company’s actual AI program is more pragmatic than messianic or apocalyptic.
This past March, Salesforce flipped a switch and made a big chunk of Einstein available to all of its users. Of course it did. Salesforce has always specialized in putting advanced software into everyday businesses’ hands by moving it from in-house servers to the cloud. The company’s original mantra was “no software.” Its customers wouldn’t have to purchase and install complex programs and then pay to maintain and upgrade them—Salesforce would take care of all that at its data centers in the cloud. That seems obvious now, but when Salesforce launched in 1999 it sounded as revolutionary as AI does to us today.
Talkin’ revolution has been good for Salesforce. The firm now has 26,000 employees worldwide, and it has pasted its name on the city’s new tallest skyscraper. Its founder, Marc Benioff, is a philanthropist who has put his own name on hospitals and foundations. Despite all this, in its own world of B2B (business-to-business) software, Salesforce still holds onto its scrappy upstart self-image.
So naturally, when the AI trend took off, the people inside the company and the experts they recruited coalesced around an idealistic mission. The team set out to create “AI for everyone”—to make machine learning affordable for companies who’ve been priced out of the market for experts. They promised to “democratize” AI.
That sounds a bit risky! Can we trust the people with such awesome powers? (Cut to chorus of Elon Musk, Stephen Hawking, and Nick Bostrom singing a funeral mass for humanity.) But what Salesforce has in mind isn’t all that subversive. Its Einstein isn’t the guy who overthrew centuries of orthodox physics and enabled the H-bomb; he’s just a cute brainiac who can answer all your questions. Salesforce’s populist slogan is simply about making a new generation of technology accessible to mere mortals. Other, bigger companies—Microsoft, Google, Amazon—may outgun Salesforce in sheer research muscle, but Salesforce promises to put a market advantage into its customers’ hands right now. That begins with the mundane business of ranking lists of sales leads.
“What do I work on next?” Most of us ask that question many times every day. (And too many of us end up answering, “Check Facebook” or “See if Trump tweeted again!”) To-do apps and personal productivity systems offer some help, but often turn into extra work themselves. What if artificial intelligence answered the “next task” question for you?
That’s what the Salesforce AI team decided to offer as Einstein’s first broadly available, readymade tool. Today Salesforce offers all kinds of cloud-based services for customer service, ecommerce, marketing and more. But at its root, it’s a workaday CRM (customer relationship management) product that salespeople use to manage their leads. Prioritizing these opportunities can get complicated fast and takes up precious time. So the Einstein Intelligence module—a little add-on column at the far right of the basic Salesforce screen—will do it for you, ranking them based on, say, “most likely to close.” For marketers, who also make up a big chunk of Salesforce customers, it can take a big mailing list and sort individual recipients by the likelihood that they’ll open an email.
But wait, what qualifies this as artificial intelligence? Anyone can tell a spreadsheet to sort a list based on different factors. The machine learning difference is simple but profound: The program studies the history of the data and figures out for itself which factors best predict the future—and then it keeps adjusting its model based on new information over time. The more data, the subtler and more powerful the answers, which is why Einstein can work not only from columns of basic Salesforce data but also from information like sales email threads that it parses and images that it reads.
Salesforce director of product marketing Ally Witherspoon uses the example of a solar-panel sales outfit using the machine learning tool to discover that a key factor in predicting a customer’s chances of saying “yes” is whether the house’s roof is pitched in a solar-friendly way. Further down the road, a different deep learning-style program could check satellite photos of different properties and automatically tag homes by roof geometry.
This roof info might start out as a major ingredient in how the machine learning program sorts its list—and, in one of Einstein’s nifty design flourishes, users can click to reveal which factors shaped each priority scoring. If users are going to trust the tool, that kind of transparency helps. But what happens when all the sales reps have learned to ignore the folks whose roofs are flat?
As Salesforce President of Technology Srini Tallapragada explains, “At a certain point, a column of data can become useless—it becomes a best practice, so it loses predictive value. The model has to keep changing.”
That is cool. It’s also pretty standard-issue machine learning tech for 2017. But to get it up and running at your company, you’d need to spend a ton of time and effort building a model that understands what’s important in your business, and then cleaning up your data to get good results. That’s the reason your bank, your insurance company, and your doctor aren’t all using AI already, explains Vitaly Gordon, who left LinkedIn in 2014 to become one of Salesforce’s machine learning pioneers. Ironically, for a field predicated on the ideas of automating human work, “It’s an access to people problem,” Gordon says. These companies probably know more about you than Facebook or Google, but they can’t compete for the data scientists who know how to mine the mountains of information.
Right now, the demand for these experts is like the run on internet routing gurus in the ’90s or SEO experts in the 2000s—even crazier than the Bay Area housing market. If you’re the likes of Facebook, Google, or Amazon, you can hire the field’s leading lights and put them to work optimizing algorithms and inventing new ways of serving billions of customers with more artificial intelligence. If you’re anyone else, you’re pretty much screwed. You’ll either pay a fortune to a giant consultancy to custom-build a machine learning system, or you’ll watch from the sidelines. What Salesforce is selling is the idea that if your business is in its hands, you’re going to get the benefit of AI without fighting for that talent to customize it for you. It all comes in the box—or would, if there were a box. (Our metaphors need to keep changing, too.)
Salesforce has 150,000 customers, most of whom have customized the system for their own needs and kinds of data. The Salesforce “multi-tenant” approach means that each company’s data is kept separately, and when a customer adds a custom data field, Salesforce doesn’t even know the nature of the information.
To bolt Einstein onto each of these businesses’ unique software configurations, Salesforce’s AI braintrust realized that it needed a new approach. “There aren’t enough data scientists in the world to build all the predictive models we need,” says John Ball, Salesforce Einstein’s general manager. Just as AT&T realized a century ago that if it stuck with manual operators, everyone in the US would end up sitting at a switchboard, Salesforce saw that automation was inevitable.
This is where Optimus Prime comes in. (Inside Salesforce, developers still use that name.) It’s the system that automates the creation of machine learning models for each Salesforce customer so that data scientists don’t have to spend weeks babysitting each new model as it is born and trained to deliver good answers. Optimus Prime is, in a sense, an AI that builds AIs—and a tool whose recursive nature is both beautiful and unsettling.
“Normally a data scientist studying one problem might take several weeks to a month to come up with a good model for a problem,” explains Shubha Nabar, Salesforce’s director of data science. “With this automated layer, it takes just a couple of hours.”
Today, the fruits of Optimus Prime are chiefly available in neatly packaged features of the Salesforce cloud applications that customers can turn on by checking a box. Next, Salesforce plans to open up the technology by steps. First, users will be able to extend Einstein’s capabilities more widely to more of their customized data. Then, a point-and-click interface will let non-programmers build custom apps for users. “We want to allow an admin—not a data scientist, not even really a developer—to predict any field in any object,” Ball says. Even further down the line, Salesforce intends to expose more of the guts of its machine learning system for external developers to play with. At that point, it will be competing directly with all the AI heavyweights, like Google and Microsoft, to dominate the business market.
Salesforce recently released research that claims AI’s impact through CRM software alone will add over $1 trillion to GDPs around the globe and create 800,000 new jobs. The company has gone all-in on AI since it first announced Einstein in 2016. Benioff said then, “AI is the next platform—all future applications, all future capabilities for all companies will be built on AI.”
Benioff even told analysts on a quarterly earnings call that he uses Einstein at weekly executive meetings to forecast results and settle arguments: “I will literally turn to Einstein in the meeting and say, ‘OK, Einstein, you’ve heard all of this, now what do you think?’ And Einstein will give me the over and under on the quarter and show me where we’re strong and where we’re weak, and sometimes it will point out a specific executive, which it has done in the last three quarters, and say that this executive is somebody who needs specific attention.”
That may sound a little Big Brother-ish, but everyone I spoke with at Salesforce is careful to keep the AI talk friendly. Einstein isn’t after your job—it just wants to help you work smarter. Nonetheless, the AI universe is mined with vague fears about the future of work and questions about bias, privacy, and data integrity. As Salesforce expands its AI projects, it will inevitably tangle with them.
One of Salesforce’s advantages in attracting talent in the field is that, under Benioff’s command, the company has built a strong reputation for having a social conscience. It’s the anti-Uber. That was one of the factors that mattered to Richard Socher, an AI hotshot whose company, MetaMind, was acquired by Salesforce a year ago.
Socher, who now leads Salesforce’s research efforts, specializes in deep learning techniques that help software understand natural language and images. He teaches a wildly popular AI course at Stanford, and co-publishes papers with titles like “Pointer Sentinel Mixture Models” and “Your TL;DR by an AI: A Deep Reinforced Model for Abstractive Summarization.”
With his unruly straw mop of hair, Socher still looks like the grad student he was not that long ago—and he has a youthful enthusiasm for testing the limits of what we think AI can handle.
“I want to be able to have more and more of a real conversation in the future with a system that clearly has tackled a diverse range of intelligent capabilities,” he says. For now, that means building learning routines that can “read” arbitrary paragraphs and then correctly answer questions about them, and exploring new methods of building AI systems that can do more than one thing at a time.
As the technology grows more powerful, Socher says, we can’t put off the conversations about its ethics. “AI is only as good as the data it gets,” he says. “If your data has certain suboptimal human biases in it, your AI will pick it up. And then you automate it, and it makes that same mistake hundreds of millions of times. You need to be very careful.”
Sales work can be painfully hard, and salespeople have to stay positive even when their work gets ugly and desperate. Salesforce has thrived for two decades by embracing that optimism. Where Google’s AI efforts are all about perfecting information access and Facebook’s aim to connect people more intelligently, Salesforce wants to make the world a better place by helping customers smarten up their work days.
At times, Salesforce’s portrait of a future powered by its AI sounds too good to be true. For a down-to-earth assessment of the company’s plans from an outsider, I turned to Pedro Domingos, an AI expert at the University of Washington and author of The Master Algorithm.
Domingos says Salesforce is “a bit of a latecomer” to the field and may find it harder than it expects to integrate AI fully at deeper levels of its products. But he thinks the company is on the right track: At this stage in AI’s evolution, there’s more to be gained from putting basic tools in more people’s hands than from squeezing an extra few percentages of efficiency from an algorithm.
Domingos also says that Salesforce’s relatively tardy entrance to AI—compared with, say, IBM or Google—shouldn’t necessarily hold it back. “They’re still a small player in this space. But other companies came from behind and got pretty far pretty quickly—look at Facebook. Just because you’re a late starter doesn’t mean that in a few years you can’t become a leader.”
Salesforce faces a crowded field in the fight to put AI tools to work on behalf of the warm-handshake crowd. Competitors include giants like Microsoft (with its LinkedIn Sales Navigator) and Oracle, as well as smaller rivals like SugarCRM and startups like Conversica (the latter of which uses AI to automate conversations with incoming sales leads). If Salesforce does succeed in moving to the front rank of today’s crazy corporate AI race, company insiders point to one advantage as its not-so-secret weapon: its well-tended warehouses of consistently labeled and organized customer data.
Those much-competed-for, highly paid data scientists everyone is trying to hire? They spend enormous amounts of time today “preparing data,” which means figuring out how to prep piles of information so that it can be digested by machine learning programs and produce good results. There is a whole lot of grooming and massaging of information that has to take place before most AI systems can even begin to start making predictions.
This represents an ironic breakdown in the ethos of automation that underlies AI. Too often today, Domingos points out, the IBMs and Accentures of the world are just throwing armies of experts at their customers’ problems. “What they do at the end of the day is, they actually have human labor do this stuff,” he says, “That makes money but is not scalable.”
But Salesforce customers have all already entered their data into a single software platform, even if many of them have added their own custom flourishes. “People put everything in there,” says Salesforce technology president Tallapragada. Salesforce doesn’t look at the content of its customers’ data, but it does know how a lot of it is organized. “The Salesforce advantage is the metadata. That lets us automate stuff,” says data science director Nabar.
For all the utopian dreams and Skynet nightmares that today’s advances in artificial intelligence provoke, the winners and losers in this transition will probably be determined by what computer scientists call “data hygiene.” In other words: No matter how smart our programs get in the AI future, tidiness still counts. Clean up after your work, and remember to wash your files before you leave.
Let others conquer Go and solve knotty theorems. Salesforce could achieve victory through neat power.
Source: Google News