There’s been a lot of press coverage about Facebook’s launch of their new Messenger app virtual assistant M. M is a human-assisted artificial intelligence. Users communicate with it primarily via texting, not by voice. The feature that differentiates Facebook’s M from Apple’s Siri, Google’s Google Now, and Microsoft’s Cortana is the system’s reliance on humans.
The humans who assist the M algorithms are called trainers. They’re given that designation, because their job is to do what the intelligent assistant should be able to do, but doesn’t yet know how to do. Every step the trainers execute to complete a task is recorded and goes into the vast database that will become new fodder for M’s deep learning algorithms. Eventually, when someone asks M to call the DMV to set up an appointment for a driving test, M will know the steps it needs to carry out and everything it needs to say along the way to get the task completed.
M is designed to offer assistance with a broad range of activities, from suggesting and buying the perfect birthday gift for a loved one, to planning and booking a vacation, to making dinner reservations.
In a recent article about M written by Cade Metz for Wired, Facebook’s Alex Lebrun is quoted using the term “bootstrapping.” The human trainers bootstrap M by filling in for the gaps in its knowledgebase. They are only present to help the AI grow smarter, until such time as their assistance is no longer required.
Are you scared yet?
In the Wired interview, Lebrun seems to be saying that you shouldn’t be, because humans will be required into the foreseeable future to aid the AI in learning how to carry out ever more complex tasks. The other comforting factor for those who are nervous about the future of humanity: we don’t know if Lebrun’s plan will work. There’s no certainty that the trainers will follow repeatable, or easily duplicated, steps for many tasks. Recommending a great birthday gift isn’t as easy as learning how to recognize cats after watching a million cat videos.
And who’s to say that DMV employees won’t just hang up when an intelligent assistant calls to schedule your driving test appointment. Unless, that is, the DMV employee is an AI too. Hmmm.
Another question people are asking: is the model of augmenting the AI’s weaknesses with human trainers scalable? If people start to rely on M and the number of users increases, how many trainers will Facebook need to hire? And how many people want the job of intercepting people’s text messages and pretending to be their virtual assistant? Apparently enough people do. A recent article in TechCrunch lists a slew of apps that let users text “expert shoppers” who make recommendations and purchases on a user’s behalf.
The strategy that Facebook is taking with M signals that a couple of trends have staying power. Firstly, texting is becoming ever more entrenched as the preferred way to communicate using mobile devices. Will wearables change that, forcing a shift to voice? That remains to be seen. Secondly, the novelty of mobile personal assistants that can tell us the weather, do math, recite facts from Wikipedia, and offer the occasional joke has passed. People want assistants to do more for them and pure AI isn’t up to the challenge yet.
Will Facebook’s experiment be successful? If it is, the more important question might be: what does it mean for the future of intelligent assistants?