The search for the killer bot
Bots are here, they’re learning — and in 2016, they might eat the web
By Casey Newton | Illustrations by Peter Steineck
The first bot I ever befriended went by the name of GooglyMinotaur. The Minotaur appeared in 2001 to promote Amnesiac, the latest album from Radiohead, which was and still is my favorite band. I happily chatted with the Minotaur about Radiohead history, information about the band’s tour, and the MP3s it offered for download. The Minotaur was popular among fans like me: 1 million people added it as a friend, and in its lifetime it sent more than 60 million messages.
But the Minotaur died a few months after it appeared, along with the rest of the era’s bots. The entire field seemed dormant for more than a decade. And then a couple years ago, the bots tentatively came back to life.
For XOXCO, it started with tacos. By the fall of 2013, employees of the boutique software development company found themselves facing a problem familiar to office workers in Austin, Texas: total midday what’s-for-lunch paralysis. Should they order from the beloved Veracruz All Natural Food Truck? From ubiquitous Torchy’s Tacos, with three nearby locations? Or from El Primo, makers of what XOXCO co-founder Ben Brown calls "the best fucking tacos" — conveniently located in the company’s parking lot?
Brown, a developer and technologist, was determined to end the staff’s daily lunchtime indecision. In the year 2000, someone in Brown’s position might have simply Googled their options. In 2010, they might have opened the Yelp or Foursquare apps on their phones. But in the fall of 2013 Brown chose to write a bot — a simple piece of software that, when sent a message, returned a single lunch option from among the 20 or so restaurants and food trucks that Brown entered into its database. Lunchbot, as Brown called it, was a simple technology that soon grew more sophisticated. Other employees added restaurants to the program; later, an updated version accounted for places the team had recently ordered from, preventing consecutive visits to Torchy’s.
In the proud tradition of stupid internet toys before it, Lunchbot evolved into a real business. Last October, XOXCO announced it had raised $1.5 million and would henceforth be known as Howdy, a bot company devoted to automating common workplace tasks. Its product lives in Slack, the fast-growing team-communication service. Howdy operates in the background, listening for the keywords and questions that will activate its powers. In its first iteration, Howdy automates meetings, asking what people are working on, collating their answers, and distributing them to the team. (And yes, Howdy will also take your lunch order.)
In the proud tradition of stupid internet toys before it, Lunchbot evolved into a real business
In 2015, a host of trends converged to put the focus of investors and entrepreneurs squarely on messaging interfaces, and the growing number of bots that live inside them. On smartphones, WeChat, WhatsApp, and Facebook Messenger emerged as some of the most popular apps in the world. Meanwhile, Slack put messaging and conversations at the center of work — and opened itself up to accommodate developers’ bots. Advancements in natural language processing made it easier to build software that understands our requests and personalizes its responses.
Growing frustration with the web over the last half-decade — both the slow-growing desktop web, and the just-plain-slow mobile web — has created a market for an alternative. In recent years, the alternative has been native apps. But most apps had a terrible 2015. The average person spends 80 percent of their time on mobile devices using just three apps, according to ComScore; for developers, buying new users with ads is prohibitively expensive — averaging $4.73 per installation, according to AdParlor, a social advertising company.
Enter the message bots. As 2016 dawns, there’s a sense in Silicon Valley that the decades-old fantasy of a true digital assistant is due to roar back into the mainstream. If the trend in past years has been assistants powered by voice — Siri, Alexa, Cortana — in 2016 the focus is shifting to text. And if the bots come, as industry insiders are betting they will, there will be casualties: with artificial intelligence doing the searching for us, Google may see fewer queries. Our AI-powered assistants will manage more and more of our digital activities, eventually diminishing the importance of individual, siloed apps, and the app stores that sell them. Many websites could come to feel as outdated as GeoCities pages — and some companies might ditch them entirely. Nearly all of the information they provide can be fed into a bot and delivered via messaging apps.
That said, there are bot skeptics. One venture capitalist I spoke to said bots could turn out to be "the Bitcoin of 2015" — a seemingly irresistible idea that, after tens of millions of dollars of venture capital invested in related businesses, finds itself mired in a niche. Silicon Valley is always chasing the next big thing: bots may simply be the latest technology to enter the hype cycle. Other technologists told me that the technical challenges of building and scaling text-based virtual assistants cannot be overcome with the current technology.
But that hasn’t deterred most of the entrepreneurs I spoke with. "Messaging is going to be the interface — or the anti-interface — of the next phase of the internet," says Robin Chan, CEO of Operator, an app that uses a mix of artificial intelligence and human workers to let you shop through text-based conversations. "This is such a mega-trend that almost every large application is moving toward this."
Just this week, Mark Zuckerberg, announced he would spend 2016 building an Iron Man-style artificial intelligence to help him run his household and help him with work. He had been inspired, he said, by the work his team is doing on Facebook Messenger — and its quest to build an "AI to answer any question you have."
If the smart money has only recently turned to bots, the technology itself has a long history. In a landmark 1950 paper, computer scientist Alan Turing proposed a test to determine whether it was possible for machines to mimic human intelligence: analyzing a text-based conversation between a computer and a person, could an observer determine which was which? Any bot that sufficiently confused the human evaluator could be said to pass the test.
A bot passed the Turing test for the first time in 1966. ELIZA was a program that reacted to users’ responses to its scripts; most famously, it mimicked a psychotherapist. ELIZA would ask you to describe your problem, scan your response for keywords, and formulate an appropriate response.
Today we call lots of things "bots." There are bots that crawl the web to make it searchable; bots that control the behavior of characters in video games; "botnets" of computers that have been organized by hackers to email spam or defraud advertisers or launch denial-of-service attacks on websites. But ELIZA pointed toward the emergence of one particular kind of bot: a virtual assistant that you access through text.
Bots bubbled up again in 2001, when a company named ActiveBuddy introduced SmarterChild: an ELIZA-style chatterbot inside AOL Instant Messenger. You could ask SmarterChild to give you the news, weather, or movie times, among other kinds of information. Thirty million people added SmarterChild to their Instant Messengers.
But AIM eventually declined, along with AOL, and bots like SmarterChild and GooglyMinotaur went with it. In the meantime, Google emerged as the primary access point for every piece of information we desired. For almost a decade, the bots disappeared.
Then came the mobile device. Today there are more than 1.5 million apps on iOS, and 1.6 million on Android. The apps have given rise to a host of new services: for hailing rides or for ordering takeout, for booking travel, or for messaging co-workers.
In the desktop era, the glue that bound everything together was the search engine, routing us from Wikipedia to Orbitz to Priceline to Yelp. In the mobile era, that glue is the application programming interface — the API — a bit of software that allows apps to talk to each other. It’s an API that lets you upload a photo from your phone to Facebook, or order an Uber from the Google Maps app.
But for developers, getting us to download their apps is increasingly difficult and expensive. Outside of games, we spend the vast majority of our time in apps built by Facebook and Google: they make eight of the 10 most-used apps, according to ComScore. To anyone paying attention, it’s becoming apparent that the golden age of apps is coming to a close.
But the bot makers say the mobile era has produced another big opportunity: a meta-app, or a layer that links everything together. To many, the interface for that app looks a lot like SMS. Text messaging is the single most popular smartphone feature, according to a Pew Research report. "Text is often more comfortable even if it’s less convenient," investor Jonathan Libov wrote last year about the medium’s superiority to graphical and voice-based interfaces. "I believe comfort, not convenience, is the most important thing in software, and text is an incredibly comfortable medium. Text-based interaction is fast, fun, funny, flexible, intimate, descriptive and even consistent in ways that voice and user interface often are not. Always bet on text."
If "instant messaging" was once a niche behavior on desktop computers — used primarily by young people and, later, office workers — on mobile devices, text dominates. Asian mega-messengers like WeChat and Line have become giant portals in their own right, connecting users to businesses, on-demand services, games, and more. The effect of all this messaging is to make us feel suddenly comfortable with what Silicon Valley has taken to calling "conversational UIs" — user interfaces that you can access through text. User interfaces that scan for keywords and message you back.
User interfaces, in other words, that look a lot like ELIZA.
"Hi, Slackbot here!" So begins every user’s experience of Slack. Opened to the public in February of 2014, Slack makes an app for desktop and mobile devices that lets you send instant messages to your co-workers. Because most businesses have never communicated in this way, Slack took it upon itself to teach them how, using a friendly script named Slackbot.
"To make things easier for your teammates, I can set up a few personal details for you," Slackbot tells you, in a private message, when you first sign on. It goes on to ask for your last name, a photo, and your phone number. All the bot is doing is building a simple profile for you. But in the process, it teaches you how Slack works.
And because Slack was built with external services in mind, it’s easy for developers to start building bots of their own. XOXCO started with Lunchbot. In time, though, the tools built for Slack have grown much more powerful — one of the reasons why Slack’s base of daily active users doubled in the second half of 2015.
On one hand, Slack is a business tool — its potential audience looks much smaller than say, Facebook M, a virtual assistant inside Messenger that the social network could someday make available to more than 1 billion people. But Slack could become at least as important to productivity as Microsoft Office once was — and the bots that are built there could very well influence bots built everywhere else.
Last month, Slack announced an $80 million fund to invest in companies that build apps that run on top of Slack. In its announcement, Slack showcased bots: there was Howdy, the meeting-running app from Brown’s Lunchbot team; Birdly, which makes filing expense reports conversational; and Awesome, which uses natural language processing to summarize the Slack discussions you missed while you were away.
Along with the fund, Slack announced BotKit, a tool that Howdy built for building other bots. It takes a set of tools and information useful to almost any bot — that "yeah" and "yup" and "yep" can all stand in for "yes," say — and packages them together.
Brown is a Star Wars fan — on the day we meet, he is wearing socks branded with the logo of the Rebel alliance — and he likens Slackbots to the astromechs of the Star Wars universe. BotKit will build generic bot skeletons; developers will infuse with them knowledge, memory, and personality. "There are R2 units, and there’s R2-D2," Brown says. "He’s different from the other units because of what he learned along the way." At the same time, Slackbots aren’t pursuing true artificial intelligence — not yet, anyway. "We’re not trying to reach consciousness here," Brown says. "We’re just trying to expose certain functionality through language."
For the most ambitious bot makers, reaching true intelligence — or something resembling it — is among their long-term goals. But that sort of AI remains many technological breakthroughs away. In the meantime, some entrepreneurs see big potential in keeping it dumb. "We believe within five years, every business will be programming its own bots," says Shane Mac, co-founder of Assist, which launched this week. Mac and his co-founder, Robert Stephens, envision Assist as the front end to all manner of web services, accessed in whichever way you prefer.
Tell Assist where you want to go, and it will tell you your cheapest transportation option, after first determining whether Uber or Lyft have surge pricing turned on. It will tap into data from its partner Olset and recommend a hotel room based on your past lodging preferences. You can access Assist through Slack, if it has been enabled on your team’s account, or through a Facebook message. You can even send it a good old-fashioned SMS.
No humans ever see your Assist queries. "We are 100 percent bots," Mac says. This is a strategic calculation: it’s easier for a small team to build and maintain a fully automated service than to recruit an army of human contractors to handle more nettlesome queries. But it also reflects just how far automated technology has come.
Inevitably, websites fall into disrepair; native apps break with the next iteration of the operating system
Mac and Stephens say that businesses are going to love the bot era. For years they have invested in building and maintaining expensive, time-consuming websites, Facebook pages, and native apps in an effort to reach their customers. Inevitably, websites fall into disrepair and the information they contain becomes outdated. Native apps break with the next iteration of the operating system. And all we ever wanted to know was whether they were open on New Year’s Day!
So what if they could use a Slackbot-style messaging system to update their hours of operation, current menu, or inventory, and so on? "It’s gotta work for the coffee shop on the corner," says Stephens, who previously founded Geek Squad, which he sold to Best Buy. "We would like to be a clearinghouse, where it doesn’t require that you download an app to use a service."
The Assist founders are among the bot builders who see big implications for the web itself. If we flock to bot-driven messaging apps to handle more of our customer service needs, what happens to the big chunk of the internet devoted to business information and e-commerce? "Most websites already aren’t being updated," Mac says. "It’s too hard to update them. If there’s an easier way that lets them communicate with their customers, they’re going to continue to not update their sites, and the sites are all going to die."
While some apps are starting at the dumber end of the bot spectrum, virtually everyone assumes bots will grow smarter over time. Operator is among those trying to speed up the process. The app, which is now in an invite-only beta, was founded by former Zynga executive Robin Chan and Uber co-founder Garrett Camp. Its aim: to be "the most exciting shopping app to come across your phone, ever."
"Uber brought you a car," Chan says. "We’ve always thought of [Operator] as another one-button journey. That’s the interface utopia that we’re striving to achieve." Open up Operator and you see a big blue button marked "send a request." Tap it, and you’ll see options for clothing, home decor, electronics, and a handful of other categories. (There’s also a catch-all "something else" button.) Pick one, and a bot will ask you a little bit about what you want. Some transactions can be handled entirely by the bot; others require human intervention in the form of contractors, which the company calls "operators."
Answering a series of questions about what kind of vacuum cleaner you’re looking for may not represent any real improvement over browsing Amazon listings, but the idea is that over time Operator’s software will improve, enabling the system to handle more and more queries without human intervention. "I think the only way you build this business is to build a network around humans and AI," Chan says, "refining your own combination to identify and aggregate consumer demand, and then automate as much as possible."
Facebook M, a virtual assistant that lives inside Messenger, is taking a similar approach. A base layer of machine learning automates as many tasks as possible, while contractors do things that machines can’t (make phone calls, argue on your behalf). It can order burritos for takeout; it can negotiate with Amazon customer service. For now, M is available only to a small number of users in California.
Magic, a company that aims to fulfill all of your requests via SMS message, is playing a similar game. So is Fin, co-founded by former Facebook executive Sam Lessin. The company is still in stealth mode, but according to a source who has access to the app, it functions similarly to M. Until recently, these services, like Google, were free to use. These apps sit at what business types call "the top of the sales funnel" — like Google, these messaging interfaces are places where customers can describe their intent. A business that fully understands your intent is a powerful partner to advertisers — and potentially a very lucrative enterprise. Little wonder, then, that Google itself is now reportedly building a smart, bot-based messenger of its own.
But that isn’t the only way to profit. This week Magic announced that it would begin charging $100 an hour for an advanced version of its services, to charter private helicopters, buy out-of-stock items, or even complete "an item on your ‘bucket list.’"
In time, says April Underwood, Slack’s VP of product, large enterprise software companies are going to invest in building smart bots. And smaller developers will be able to rely on those companies’ expertise in machine learning, so that they don’t all have to build their own engines for interpreting our texts. "I predict by the end of 2016, we’re going to see more really great examples of household-name companies creating great bot experiences," she says.
Others speculate that the promise of bots may be exaggerated: the most ambitious bots of the day, Operator and Facebook M, are still in closed beta. And for the most part, bots still depend on platforms they don’t control (iOS, SMS, Slack) — which some venture capitalists believe limits their growth potential.
Benedict Evans, a partner at Andreessen Horowitz, told me that bots face a big challenge as they grow: the number and volume of human desires might always outpace our ability to write software that can address them. "You’re writing recipes," he says. "And how many recipes can you write?"
There’s an interface challenge, too. Give people a blank box, and they may not know what to ask for. "People try to bolt AI onto every new user interface model. But we don’t actually have HAL 9000, and may be 50 Nobel Prizes away from that," Evans posted on Twitter a few days after we spoke. "The problem with ‘no user interface’ is that, since you don’t really have HAL 9000 behind it, it’s almost as opaque as a DOS prompt."
"We don’t actually have HAL 9000, and may be 50 Nobel Prizes away from that."
But with Facebook and Google now investing in building virtual assistants, bot makers are betting that people will learn. "Inevitably, they’re going to get educated on this interface," says Operator’s Chan. And in many ways, they’re educated already. "The messaging interface feels very familiar to anyone who’s ever sent a text message," Slack’s Underwood says. You’ll text your boyfriend to pick up pizza on his way home — why wouldn’t you text the same thing to Facebook M?
The fact is, as investor Semil Shah has written, messaging has usurped the browser on mobile devices: it’s where most of our activity takes place. And once you’ve dethroned the browser, which empires will crumble? Could a new e-commerce channel rise to challenge Amazon? Could a bot outdo Google when it comes to understanding what you’re looking for?
Those companies are rich enough to buy up competitors before they become existential threats. But could you disrupt struggling Foursquare or Yelp with a conversational UI? (See Luka.) What about fitness apps? (See Lark.) What about banks? (See Digit, or Penny.) Which media companies are poised to succeed in a world where we consume more information through text messages? (In the bot era, Facebook’s Notify app looks a lot like a media company.) What does a media company look like when it optimizes for Slack?
Does Siri add a text interface? Does Alexa?
And the web? The web won’t die; mediums never do. That said: "It’s going to continue the erosion of the power of the home page, and even the power of the search bar," Brown says of bots’ rise. "The more software can notice the signals we’re sending it constantly, and deep-link us all the way to the answer, the less I have to go browse or search."
Of course, all this depends on a still-rickety infrastructure of services and machine learning actually coming together. It depends on an automated messaging interface that feels as trustworthy as a message to your boyfriend, rather than a mysterious DOS prompt. And it depends on bots living up to their billing as friendly, powerful, all-knowing assistants. "Building an AI that can be ‘everything to everyone’ is obviously appealing, yet we've seen pitches of this sort fail dozens of times over," says Maran Nelson, founder of Clara Labs, which is building a virtual assistant that schedules meetings. "The technical challenges to delivering on this promise at scale make this almost inconceivable."
For now, using Facebook M to order takeout burritos feels like a novelty. What will it take to feel like a superpower? Is it coming within reach, as so many entrepreneurs in Silicon Valley now believe? Or are we still several advances in computer science from getting there?
We’re about to find out. Existing bots are getting smarter every day, through the sheer volume of data passing through them. Across Silicon Valley, an armada of new intelligences is now under construction. If there’s a killer bot to be found on our phones, 2016 may be the year it says hello.
Correction: a previous version of this article incorrectly stated that Cortana has no text interface, which it does.
Edited by Michael Zelenko