Facebook is building technology intended to understand, at near human accuracy, all of the things its users post on the network.
That's according to Hussein Mehanna in an interview with Recode last week. He's the engineering director for Facebook's core machine learning, part of the broader applied machine learning team tasked with infusing artificial intelligence and deep learning technology into the company's existing products.
Deep Text uses neural networks, a subset of AI and deep learning intended to mimic activity of the human brain, to understand written language so that it can then act accordingly.
More from Recode:
Bill Gates talks about why artificial intelligence is nearly here and how to solve two big problems it creates
Google answered some of our questions about its fancy new AI chip
Learning about deep learning
So instead of simply using keywords to categorize posts so you can find them later, Facebook would be able to understand the meaning of a post and make recommendations or take action as a result.
Google, Facebook's chief business rival, is also pushing similar AI tools that tap data from its users — on search, YouTube and maps, etc. — to direct consumers to other online services and offline commerce.
Understanding everything on Facebook is no small feat. Users post hundreds of thousands of new items every minute in dozens of different languages. But Facebook claims it can already process 10,000 posts every second in 20 different languages using Deep Text, and it's starting to deploy the technology to its user base, albeit for a very small number of use cases.
Deep Text currently does two things in the wild:
- Ride hailing — If you use Facebook Messenger to tell a friend you need a ride (or something similar), Facebook will automatically surface a small alert asking if you'd like to request an Uber or Lyft.
- Selling on Facebook — If you post that you're looking to sell something, like an old bike, Facebook will detect that you're trying to sell a product and automatically recommend its selling tools.
Facebook obviously wants that to list to grow — think things like comment moderation, using Deep Text to help determine which comments to push higher (or lower) in the queue. But all of this is going to take a very, very long time. Teaching a computer to read like a human isn't quick. And the applied machine learning team working on Deep Text was created in the fall, so this project is less than a year old.
"I think this journey is still one percent complete," Mehanna said, "but it's still far more revolutionary than what we had a few years ago.... There are far more challenges ahead of us."
Facebook COO Sheryl Sandberg and CTO Mike Schroepfer spoke at Recode's annual Code Conference on Wednesday and discussed the importance of AI to Facebook's longterm vision. "It's another one of these transformational technologies," Schroepfer said. Among the use cases Schroepfer and Sandberg highlighted for the future of AI: self-driving cars and cancer prevention.
Facebook doesn't do work (that we know of) in either of those areas, yet, but perhaps that's in its future.
—By Kurt Wagner, Recode.net.
CNBC's parent NBCUniversal is an investor in Recode's parent Vox, and the companies have a content-sharing arrangement.