One big area where virtual assistants like Amazon Alexa and Google Assistant fall short of real human assistants is their inability to contextualize and anticipate what you’ll want next.
Currently, requests made to virtual assistants are often siloed, and go something like this:
“Alexa, how long should I steep tea?”
Alexa answers, with something like “Five minutes,” and then:
“Alexa, set a timer for five minutes.”
In a corporate blog post today (hat tip Geekwire), Amazon announced that Alexa is now getting better at bundling those types of requests together. Amazon refers to this as figuring out your “latent goal,” and actually provides tea steeping as an example. Asking Alexa how long to steep tea could have Alexa guess that your latent goal is to make tea. This, in turn would trigger an immediate and automatic follow up response from Alexa like “Would you like me to set a five minute timer?”
While this seems straightforward, as with so many AI-related tasks, understanding what people want isn’t exactly easy. From Amazon’s blog post:
The first step is to decide whether to anticipate a latent goal at all. Our early experiments showed that not all dialogue contexts are well suited to latent-goal discovery. When a customer asked for “recipes for chicken”, for instance, one of our initial prototypes would incorrectly follow up by asking, “Do you want me to play chicken sounds?”
Beyond tea, it’s not hard to think of how identifying latent goals could be useful in a smart kitchen. In the case of asking for chicken recipes, Alexa could follow up with offers to pre-heat an oven, or, more relevant to Amazon, offer to order you the necessary groceries for delivery that day (preferably from an Amazon grocery store).
Amazon says this latent goal capability is available now in English in the U.S. And while it doesn’t require any additional work from developers to activate, they can make their skills more discoverable with the Name-Free Interaction Tool Kit.
FWIW, I tried asking Alexa the tea steeping question, and it did not follow up with a timer suggestion. So its latent goals capabilities seem to still be, well, latent.