We can carry these algorithms around on our smartphones and smartwatches
Science fiction tells us one day we’ll talk to our phones and computers like they’re people, as suggested in movies like "Her" and "2001: A Space Odyssey." But we’re now facing a different reality: Our digital lives aren’t managed by one entity, but rather a collection of artificial intelligence that gather information about how we type, what kind of media we like to see and what we should be doing throughout the day.
Right now, a lot of these algorithms require too much computing power to run on phones and other smart devices—that’s part of the reason why the cloud has been so revolutionary. But a coming wave of advances in AI research and deployment will bring algorithms that require far less compute power, meaning we can carry them around on our smartphones, smartwatches and smart belts, without sending personal information like texts back to the servers of Facebook and Google. This technology could allow a fresh start for privacy within messaging apps, especially now that users are beginning to understand the importance of end-to-end encryption on their personal messages.
Google has announced one such optimization—the Smart Reply feature found in its popular Inbox app is coming to the new Android Wear 2.0 (a smartphone), and will be able to run locally on any application on the device. That means suggestions will populate faster, but more important, the data about what you type will never leave your device. (The algorithms are built and trained in the cloud, however.)
To make this on-device prediction possible, Google researchers dumbed down how the algorithm sees each message. In the full Smart Reply feature, each snippet of text is analyzed and turned into a complex string of numbers that indicates information about each word and how the words work together as a phrase.
On the watch, Google uses a much simpler code, consisting of only 8 bits, or ones and zeros. Similar messages are grouped together (like “How are you” “how are you doing”), making far fewer kinds of messages the watch needs to remember. This simplification, along with a few other tweaks, led to the algorithm requiring 100 times less compute power, according to a Google spokesman, which made it possible to bring Smart Reply to the Wear 2.0.
When tested by humans against the original Smart Reply feature, Google found the new suggested messages were “surprisingly more rich and diverse,” its spokesman said.
Other tech companies like Facebook are also thinking about this problem, but from a different angle. The way cloud-based artificial intelligence algorithms work today is incompatible with any strong encryption—for the AI to decrypt the message, analyze it and encrypt it again would cause a host of vulnerabilities and defeat the point of encrypting it in the first place.
The AI Facebook is researching to someday deploy in Messenger is text analysis that can suggest actions (like calling an Uber or buying dinner) based on your conversations with others. When asked by Quartz about ensuring security while providing these features, Facebook’s director of applied machine learning, Joaquin Quiñonero Candela, said the solution would probably be personal AIs that live inside each user’s app on their phone.
“The answer is going to be bringing AI to run locally on your phone, and that’s an active area of research,” Candela said, noting Facebook was already working on this for its AI camera filters, which run without connecting to the cloud.
Facebook has long been motivated to streamline its AI in the interest of speed: Most people use the social network on mobile devices, and having phones continuously pinging Facebook servers takes time and data, and depends on signal strength. Last October, Chief Product Officer Geoffrey Fowler showed off the AI system within Facebook’s mobile app, called Caffe2Go, and its ability to process real-time, Snapchat-like camera filters.
“We perceive the world in real-time,” Facebook engineer Hussein Mehanna told Wired when the internal tool was announced. “Why wouldn’t you want the same thing from your AI?”