LAS VEGAS — Amazon’s Alexa virtual assistant will soon get savvier about juggling its thousands of skills — starting with arranging all the elements for a night out.
Cross-skill action prediction is one of the upgrades for Alexa announced here today at Amazon’s re:MARS conference.
Rohit Prasad, Amazon’s vice president and head scientist for Alexa, laid out a scenario where a user of Echo Show could engage in a seamless dialogue to choose a showing of “Dark Phoenix,” reserve seats through Atom Tickets, find a nice Chinese restaurant nearby, make a dinner reservation through Open Table, set up an Uber ride and watch a movie trailer.
“We’ll be bringing this experience to our customers soon,” Prasad said during today’s morning keynote.
Dave Limp, senior vice president for Amazon Devices and Services, said after Prasad’s presentation that planning a night out was the first scenario addressed by Alexa’s developers, using an upgraded AI model that takes advantage of Amazon’s work with recommendation engines.
Instead of figuring out which movies or books you might like, based on what you’ve checked out previously, the model anticipates which skills are most likely to be needed to follow up on the skills Alexa has already employed.
“We started with the ‘night-out’ experience,” Limp told reporters. “We’re doing a couple of other ones internally right now. We’ll roll those out in the coming months. But over time, that is where you can see Alexa going. It has been the holy grail of voice science, which is how you can make a conversation string together when you didn’t actually programmatically think about it end to end.”
Limp said being able to knit multiple skills together should reduce the time required to arrange a complex series of tasks, such as the reservations for a night out, from as much as 40 minutes to just a few minutes of fluid conversation.
In addition to the cross-skill training, Prasad said Amazon is providing a private preview of a software upgrade known as Alexa Conversations, which will let developers create natural voice experiences with “less effort, fewer lines of code, less training data.” Hand coding of the dialogue flow is replaced by a recurrent neural network that models the conversation from the developers’ input.
Alexa Conversations should make developing a structured voice experience “about a half an order of magnitude easier,” Limp said.
Amazon is also making Alexa seem less like a robot and more like a person.
“We have an entire dedicated team, single-threaded, that wake up every day and all they think about is Alexa’s personality,” Limp said. “And they’re replicated around the world. So her personality in Germany is slightly different from her personality in Japan. She has some common threads. She’s incredibly intelligent, she’s incredibly empathetic. But if you were to get answers from her in Japan, they’d be slightly different.”
He said Amazon goes so far as to provide Alexa with opinions. “If you’ve ever gone to a dinner party devoid of opinions, you wanted to run, not walk away from that dinner party. It’s a boring dinner party,” Limp said. “So you can ask Alexa what her favorite beer is. And her favorite beer in the U.S. is different from her favorite beer in Germany. It’s a global effort, with lots of people at work on it all the time.”
Limp almost made it sound as if it’d be worth scheduling a night out … with Alexa.