For the last 13 years, Alexandre Lebrun has been teaching machines to understand humans.
As humans, we think machines understand us. We type in a search term into Google or the name of a city into the weather app, and the information is displayed before us.
But, we’re not really using what is considered “natural language.”
It’s the difference between searching for the weather in a city to see if you need an umbrella versus just asking the machine “Can you remind me in the morning if I need to bring an umbrella to work?”
In one, you’ve learned how to search for the answer you need. In the second, you’re just asking what you really are trying to accomplish.
When Facebook bought Lebrun’s startup Wit.Ai in January, the company brought him in to tackle how people can communicate in a new way not only with computers, but also with businesses.
“People hate to call or hate to waste their time doing search. So the question was, if they need something to get done, how can we bring that experience inside messenger with something as simple as a text?” Lebrun said.
The answer is supposed to be M
Facebook’s answer is M, a personal digital assistant tucked into its Messenger app.
Facebook began testing the digital assistant three months ago, giving beta access to more than 10,000 testers in the San Francisco Bay Area.
M is supposed to replace your phone calls, web searches, and apps with one simple interface — the genderless, avatar-less M. But behind M are some humans.
Rather than testing it internally for years, Lebrun says Facebook made the decision to launch the nascent program into beta so it could learn as it grows. Facebook doesn’t allow M to have access to any of your data either — not even your public Facebook profile.
“M is incredibly ambitious and hard to do,” Lebrun said. “There’s no way we can do it without a huge amount of data and training, so it was not a good idea to make fake data and then one day launch a super AI and hope it works.”
Instead, the company employs more than 30 contractors as “human trainers” to work at Facebook’s campus and train the program. As questions come in, M, the AI part, gives its best guess as to how to answer it. For now, Facebook’s human trainers validate the answers in what Lebrun describes as a close partnership.
“Let’s say somebody says, ‘Tell me a joke.’ The AI will read it and the AI will understand that it wants a joke. The AI will fetch a joke from the joke database, then go to the trainer and say ‘hey, is this a good joke? should I validate this?’ And the trainer will say ‘yes,’ and it’s done,” Lebrun explains.
It sounds easy, but the AI is learning along the way.
First, by having the human trainer confirm it chose the right answer, it means the AI recognised that the “Tell me a joke” meant to retrieve jokes from a website. Second, the joke it chose came from a good source of jokes and was funny.
Over time, M will use this knowledge build a database of jokes and joke sources that it can use, and rule out what it shouldn’t.
Asking M to tell you a joke is a relatively easy. The harder requests are the transactions, like “send flowers to my wife.” They’re also the most common.
For restaurant delivery orders, the human trainers actually pick up the phone and call to place the order, Lebrun said. But before they ever get to that point, the AI accomplishes most of the small parts ahead of it, from recommending restaurants to telling you what’s on the menu and the pricing.
The human trainers remain involved in every question to make sure M is answering correctly, and as it grows, Facebook will have to employ more and more trainers.
Without the trainers, there are only a few things that the AI could probably do correctly on its own, although Lebrun cautions that the Facebook experiment is only three months old.
“The AI is really good at things like weather, traffic, and set reminders because they are relatively simple and very frequent,” Lebrun said.
Lots left to learn
Facebook isn’t the first to try to build an AI assistant.
ChaCha developed a Q&A company in 2006 to give people a way to search the internet via a text. The company exploded in popularity in an age before easy smartphone internet access, but it came with a giant problem of users abusing ChaCha — and that was only with hundreds of thousands of users. Half of the company was devoted to managing unruly guides, contractors similar to M’s trainers, and the unruly users who just wanted to ask it crazy questions to see if ChaCha could answer.
“The whole goal was to break the system. You have guides who come in and try to do foul play, and then you have users who come back in over and over again,” said ChaCha’s CEO Scott A. Jones in an interview with Business Insider. “It was only 1-2% of the users or the guides, but they could do damage quickly because they were intent on it. They would spend hours a day just pummelling the system or answering the system with bad answers.”
Facebook and M will have to learn how to deal with this as it goes. Lebrun said abusive users haven’t been a problem yet. Even the craziest requests, like hand-drawn pictures, have gone answered.
“We don’t have the problem yet, but if you clearly abuse M or just play with it, you’ll have slower response times or you won’t be a higher priority,” Lebrun said. “Then on the trainer, unlike ChaCha, it’s not crowdsourced. They work just for us.”
Avoiding the creep factor
Lebrun is also taking pains to make sure M never gets creepy.
“We have to start with zero information about the user. So right now, M just knows what you tell M. Nothing else. Not even your public profile is used. But if you give M your phone number, for instance for a restaurant reservation, so you can slowly provide information just when it’s useful. In general I think it has to be very carefully managed balance between how much data we get, and how much data we return to you,” Lebrun said.
The challenge ahead will be for Facebook to make M accessible for everyone. The company will have to hire more trainers as it adds more people to the platform and the questions get broader and more complicated.
For example, M has so far been used mostly as an executor of tasks rather than a base of knowledge, so it hasn’t dealt with many illegal, hypothetical, or open-ended questions, Lebrun said.
It does know that it shouldn’t deliver illegal content, but it still remains up to the trainer to verify what’s illegal.
“We’re still talking about it. It’s very hard to define what is porn, what is not porn, but when I see it, I know it. For M, it’s very hard to define exactly, so we took the approach to start with to be very open to everything that’s legal,” Lebrun said.
“The only thing that are legal that we don’t do is anything involving medical operations, prescriptions, calling a doctor.”
Business Insider Emails & Alerts
Site highlights each day to your inbox.