When I first arrived here 2.5 years ago, my challenge was not getting people interested in AI or even the very narrow application that we’re building (an AI scheduling assistant). AI then was like cryptocurrencies now, at the height of the hype cycle. No, my challenge was explaining how it actually worked and what it took to build a product of this kind. Which is to say, I had to debunk a whole bunch of myths to get my point across.
Today, I’m happy to report that we’ve gotten some help. WIRED was kind enough to tell our story, and the piece, written by John H. Richardson, dispels some of the key misconceptions bedeviling the story of AI in 2018.
Because AI has had such a high profile in Sci Fi (from HAL to the Terminator to Westworld), one of the strongest misconceptions we needed to tackle was a view of AI as magical, superhuman intelligence that arrives nearly fully formed, with not much effort on the part of humans. It doesn’t.
Building AI is nearly impossible, and it involves data scientists, data engineers, software developers, interaction designers, dev ops specialists, and if your AI requires supervised learning (like ours), AI trainers. I quickly learned that it’s not easy to explain how each of these roles contributes to teaching a machine to take over the entire task of scheduling a meeting. You need to be curious and patient enough to wade through a whole lot of technical details.
John spent months researching and writing this piece, and he did a great job explaining how hard we are working to solve this problem. In particular, there are three dimensions he really captured. The first is the sheer complexity of simple human interactions around scheduling. In this excerpt, our Chief Data Scientist Marcos Jimenez Belenguer is talking about an email that’s especially difficult for our AI trainers to annotate:
“Here’s another one: ‘I’m free most of the week of August 7. Feel free to schedule anytime from 7,8,9, or 10, preferably in the afternoon.’ The trainers think the last four numbers in the message are dates, but the date template doesn’t have enough boxes for all of them.
“It’s another edge case,’ Jimenez Belenguer says, and if the engineers or trainers make too many mistakes, as humans are prone to do, the machine will learn to make the same mistakes. Sure, they can build a template with more boxes. But at some point they’ll have to stop rewriting the models and instruct the algorithm to ask the customer for clarification.”
Indeed, human language is so complex that many AI experts characterize creating a fully autonomous AI scheduler as “a goal that the AI experts… ranging from ‘very, very hard’ to ‘impossible.’”
The second aspect of the story that John captures well is how our AI Trainers actually do their job. About 18 months ago, we moved this team from New York City to Manila. Our goal is to fully automate Amy + Andrew, and to do so, we need to continue to collect and annotate huge amounts of scheduling data. Our VP of AI Training Operations, Liying Wang, has done a tremendous job of sourcing and training the trainers. They’re a dedicated professional bunch, and they’ve gotten us that much closer to our goal.
“In a highly secured building on the outskirts of Manila—I had to give the security guard the serial numbers of my phone and laptop and couldn’t even use a pen and paper on the production floor—40 young Filipinos are sitting at tables like travelers checking their Facebook pages in an internet cafe. They’re mostly in their twenties and early thirties, college graduates or defectors from offshore call centers. Like many Filipinos, they speak perfect English.”
“…[One of x.ai’s AI trainers] spends her workday feeding data to the machine-learning algorithms by highlighting every word that seems to pertain to a time zone and dragging it into the appropriate box on the time zone template. This is called ‘named entity recognition.’”
Finally, John gets the audacity of our vision (even if he retains a reporter’s skepticism of it). His description of the scale of the problem is particularly apt:
“If they can perfect Andrew Ingram, they’ll put x.ai at the forefront of workplace innovation. Americans schedule approximately 25 million meetings per day. Multiply that by the hourly wages all that scheduling sucks up, and you see how much time, money, and mental energy x.ai could save.”
It’s also just plain fun to see ourselves in Wired, particularly when the team is described as “some of the most dedicated nerds you’ll ever meet. . . they bustle about their Manhattan offices with the intensity of NASA engineers preparing to launch a moon shot.”
My work isn’t done yet but we’re happy to see a fair and in-depth portrayal of what it takes to build AI today—the massive data sets, the wrong turns and inevitable mistakes and the dedication of our teams in NYC and Manila. I’m proud to read the progress we’ve made against one of the bigger challenges I’ve faced in my time here. I’ll be buying out stacks of WIRED from every newsstand I can find. If you happen to snag a copy before I get to yours, drop me a line and let me know what you think.