iacsf: What’s Next? Shaping the Future

Mark Yahiro, Intel
Timothy Tuttle, Expect Labs
Liesl Capper, IBM
Roberto Pieraccini, Jibo

Intel offers platform and foundation to use speech, motion, who you are, to use data in intelligent ways.

Jibo (a male character) is not humanoid, but has stereo camera, mic, speaker identification, motion or facial expression detection, display, touch points. Just got big crowned sourced funds. Japanese trend is to more humanoid, is creepy, uncanny valley. farthing can express feelings (teapot in beauty and beast).

Ongoing conversations? Watson uses experts systems, known data sets, to develop ranked diagnosis of medical conditions.

Movie Her and anticipatory agents? Tim suggests we’re not as far off as we think. We are going to starts seeing intelligence in new devices. Recent breakthroughs from IBM in deep learning remarkably reduce error rate within a couple of years.

What is the future? Cognitive glue working across other agents. Holy grail is all human interaction. There are already a lot of agents, interacting between and across them is learning a new language. Filter the right data specific usage, depending on the usage do we filter before or after? Where is context? What makes sense.

Speech recognition is about 60, 65 years old. Big problem is that once you understand how to did it, depends on the language and what you are talking about. Can’t create a closed loop.


iacsf: Executive Summit

Brett Beranek, Nuance
Fred Brown, NextIT
Brian Garr, LinguaSys (21 languages, today’s process release)
Phil Gray, interactions
Tom Lewis, Smart Action

Dan: people turn to others as they escalate customer care, going to people as last resort. Human interaction? Phil: path is invisible to customers. Highly conversational experience does take people, but systems learn too. Looking for same level of service. Fred says his work is getting very close, accurate. Brian works with semantics. Look at Google translate that works with swashes of words, vs LinguaSys which gets the right fidelity of the conversation. They build first 21 language maps, will add next 20 languages at no cost. Brett: great digital experience for gen y or millennials feel that if they have a to talk an agent, you’ve failed.

Dan: how much is automated and how soon? Fred: it’s not simple implementation. Must build it well first. Sargent Star is a good model, how much effort to build it up. Tom: $3.5 billion investments, but consumer has been programmed to think that IVR stinks. Call center manager perspective for 2050 will not be same as today, so where is tipping point? Investments in AI, also in voice recognition in HD vs calling a copper line device. Plumbing, standards, telcos.

How to resolve the zero out problem? Phil: from Hyatt, don’t remove people, maximize their expertise. Is about making things easier for customers. Tom: reflective interaction is key. Fred: start on smart device. Brett: consumer expectations are changing quite a bit, experience with e.g. Dom makes consumers aware that this experience may be elsewhere. Its up to enterprises to make that work. Brian: conversation needs to expand to consistent global experience.

Plumbing: when will this happen for smaller enterprises? Brian: we are doing this now, will change the price structure for NLU. Phil: call centers have different adoption curves, different technologies. Where are priorities? Today this is scalable. Brett: is accessible technology, kaspersky software implemented, got Roi within 3 months. Fred: it matters what consumers need, getting the natural language understanding, and how to hook that to back end. Brett: business models need to evolve, esp in contact center space. Dan: models, APIs, will happen, is in a parallel experience.

Brian: going for statistical to semantic application, coder I can build a system.

Brett: lot of enterprises waking up but not a good comprehension of where this might deliver good business value. Where are your cost drivers? What are your shopping cart use cases that cause problems? Build virtual assistants to address these problems. Fred: look at conversations, drop AI in after examining. Roi will be proof. Tom: need to present in a way that minimizes risk, adoption of technology has decreased over time. Business model includes understanding risk.


iacsf: Dennis Maloney, Domino’s

Four years ago, brand started going through a reinvention (pizza tastes like cardboard), by listening, going digital to ordering online. Goals: increased customer satisfaction, increased revenue and profit, and better products and trials. About 45% of their orders come for digital, >$1B in business. Consumer experience drives results.

From order, you can track your pizza’s process from order to delivery, including who made the pizza. “FedEx of pizza.”

Pizza profiles enables simple order with 4 clicks. Now voice ordering in iPhone and android. 25 SKUs on their site, some items are customizable.ordering by voice is common, allowed them to break some constraints. This is a foundational technology, adds to a strong brand message. Their persona Dom is a character, fully integrated into ordering process. Nuance is partner. Huge learning curve. Was harder, took longer, was more expensive, but was worth it.

Demo done live (worked pretty well) for ordering pizzas. Knew from store what selection to offer, dealt with ambiguity, has a sense of humor and is conversational. “Easy order” would place order in one step. Consumers like it: appears faster, simpler (less steps), easier process, more intuitive, positive brand impact.

Problem: consumer needs to choose store first, then limited by what that store offers, not by what consumer wants.


iacsf: Norman Winarsky and Bill Mark, SRI

SRI started work with Calo, funded through DARPA, probably largest AI project in 1953. Focus on learning in the wild, the way a real human personal assistant would learn. Still quite a frontier. Biggest challenge is user experience, when things in the interface change, he’s out of control. Wants something that does the right thing every time. Calo was most major project before Siri. People who worked on Siri came out of Calo team.

sRI is very financially pragmatic about what they fund, must be relatively certain. We start with a market problem, not technology.

How important was natural speech? Global conversations with telephone ecosystem, most people don’t talk to a phone for service but Bill thought that wasn’t a future predictor. Is rare that customer expresses their intent in first sentence. W are learning what we want as we talk. One Question might lead to another.

Tempo, a smart calendar, offers insights into your daily life, predictive intent. Destiny is a travel company, bought by Nokia. Quantum game about math. Open question do we need verticals or can we use one? Different kinds of personal assistants. On kind is general, other is really about me.

There will be millions of assistants, but two layers. APIs to talk to each other? Th hard part is how to have this multi-term part conversation between all. Won’t stay mobile device centered, internet of things will help with simple intents that happen all the time. Which decisions are contingent on emotional responses or other timely options? Language, interacting, is not the future that is learning and understanding. Mobile will be center of interactions related to where, when…

Genderization is also a complex issue. Real world: not homogeneous, will be a variety in personas.

Personalization in retail, for example, can recommend, but puts you in a category. Difference between, and we need both little data as well as big.

Eliza, from 1960s, tried to emulate a psychologist. People really engaged, some went on for hours. Touring test of thinking machines… Some people want a human interaction, some dont.

What technologies need to improve before it’s conversational? Deep knowledge, represents fundamental things that are known to us (e.g., bank account), also context for human interactions and how we remember things. The way we deal with this now is through verticals, where context and intents are more established. Across market verticals not possible today. Shallow vs deep.

Tooling and analytics are useful to people buying them. Problems with deep knowledge, structuring the representations, and how it all works together.