Beyond Siri, The Next-Generation AI Assistants Are Smarter Specialists

After birthing a virtual assistant that knows a little of everything, SRI International is working on ones that know a lot about just one thing.

The nonprofit research center is arguably best known in the tech world for spinning off Siri, whose virtual assistant tech Apple acquired in 2010. SRI also incubated Tempo, an AI-driven calendar app that Salesforce bought last year.

Now, SRI believes it can infuse AI into even more settings—shopping, banking, travel, business-to-business applications, and so on—allowing for deeply knowledgeable chatbots that know how to carry a conversation. The goal, says William Mark, SRI’s president of information and computing services, is to have assistants that are much better at specific tasks than a general-interest assistant like Siri.

“If you were to say to Siri, ‘transfer $200 from savings to checking,’ or something like that, Siri would just look something up on the web about transfers,” Mark says. “That’s not a critique of Siri, it’s just that Siri doesn’t know anything about banking.”

The specifics of SRI’s efforts are still murky, but in 2014 the group spun off a startup called Kasisto, which recently launched personal banking bots for Facebook Messenger, Slack, and text messaging. A separate Kasisto bot will soon be able help people keep track of their investments.

Mark says this is just the beginning, as SRI is planning for more spin-offs in the near future. The goal is to establish a platform for companies to build their own highly specialized assistants. SRI’s tech will provide the conversational “scaffolding,” while each industry provides the knowledge.

As an example, Mark says, imagine you’ve asked a bank teller to transfer some funds, and the teller asks whether to transfer from checking or savings. You might respond by asking, “How much do I have in savings?” To a human teller, that’s an understandable branch of the main conversation. But to a chatbot, it might seem like you’ve changed the subject. SRI’s framework will supposedly be able to handle these kinds of conversational moves.

“We’re creating a core set of utilities so that once you have, for example, the concept of banking, the system knows how to have a conversation about banking because it knows how to have a conversation,” he says. “And then if you move to a different domain like engine repair, the concepts change, but the rules of conversation are similar from one domain to another.”

Just Another Tool?

SRI’s interest in new AI tech is noteworthy in part because of the group’s pedigree, but keep in mind that there’s already a growing number of bots that are good at particular things (“domain-specific” bots, to use the industry jargon). The available tools to build these bots are also increasing, with big companies like Facebook and Microsoft along with smaller startups trying to woo developers.

“Basic language understanding and NLP are commoditizing very fast,” says Charles Jolley, whose information-finding assistant Ozlo currently specializes in finding places to eat. “Many people have the datasets they need to build a Siri-level of language understanding now, and they are all sharing it via bot APIs.”

The next breakthrough, Jolley says, “will be around understanding the deeper meaning behind the words, which is only possible with a large knowledge base.”

It’s hard to say whether SRI has a unique angle here without actually seeing what the company is doing. But according to Kasisto CEO Zor Gorelov, SRI is offering something more thorough than what he’s seen elsewhere. To build Kasisto, the company licensed all of SRI’s AI technology, including speech recognition, natural language understanding, AI reasoning, and interactions that differ based on text versus speech. In his view, it’s one of the deepest AI portfolios in the industry, which Kasisto was able to expand upon with training data, analytics, and security features.

“Facebook, and Microsoft, and Kik—a lot of people are throwing bot toolkits out there, but this is not a toolkit or [natural language processing] problem to solve,” Gorelov says. “To solve this problem you need to have the entire platform, the entire stack.”

That said, SRI won’t be alone in trying to enable more domain-specific knowledge in chatbots. The group may even provide a rival, of sorts, in Viv, which made a splash in May with an impressive virtual assistant demo. Although Viv seems to emphasize one bot versus many bots, the broad strokes of allowing developers to easily impart deep subject knowledge are similar. Viv even shares a similar lineage, with the founders having worked at SRI before cofounding the Siri spin-off.

While Viv has shown off a powerful demo, SRI hasn’t shown much of anything yet. Still, demos don’t matter much in AI, where the real test is what happens when ordinary users are allowed to say whatever they want. That may explain why Mark, when asked about Viv, is reluctant to say much.

“I know it’s frustrating, but these comparisons are pretty hard to make,” he says. “That’s why we like to try, when systems exist, we like to try interactions on one system versus another system.”