Artificial intelligence is the topic of choice at most big tech conferences these days, and last week's Fortune Brainstorm Tech was no exception. It was the focus of a number of sessions, both on the main stage—where topics included AI and self-driving vehicles—and in many of the conversations at various roundtables and meals at the conference.
While nearly everyone at these conferences talks up the huge potential benefits of AI, getting it deployed and working correctly seems to be a difficult and often lengthy process replete with challenges.
For instance, at a breakfast roundtable covering what you can do with AI, a number of participants seemed both very hopeful as to the potential benefits of AI and equally clear on the issues that companies face as they try to deploy it today. Speakers at the breakfast included two representatives of companies with major customer-facing operations—eBay and OpenTable—and two others that focus on enterprise applications—Box and Oracle. Each described different sets of challenges.
EBay Chief Strategy Officer Kris Miller talked about how AI can help to suggest products as part of "an end-to-end customer journey," as well as about the company's effort to build a central database containing all the relevant information about a single customer—which could be used to provide real-time customization and personalization. It's hard to create this, Miller said, and they are still working out latency issues, among other things.
Another goal is to make it possible for customers to take a picture of something—a handbag or pair of shoes, for instance—and have the eBay app immediately show customers similar items for sale on the site. This involves ingesting a huge number of images, tagging over 1 billion items, and then training the AI on these images.
OpenTable CEO Christa Quarles talked about how AI is helping the company add new criteria into its search ranking, resulting in better search, and hopefully, additional sales.
The goal, Quarles said, is to create the "ultimate recommendation engine." The problem is that users have different needs at different times, so context is highly important, as is recognizing both implicit and explicit signals.
Quarles also talked about using Alexa or other voice assistants for "conversational commerce," though she said at present Alexa isn't very good as a "browse experience."
On the enterprise side, Box Chief Product Officer Jeetu Patel said that although he's convinced that in the long term AI "will fundamentally change how people will interact with content," he's afraid of over-hyping the technology, and cautioned that such change will take time.
Patel said Box is interested in AI in three main areas. One is the "Box Graph," which aims to understand the relationship between two pieces of content—content
Patel said it's important to be clear that Box doesn't own this data, but rather the individual business customer does. He noted that with AI, "beta cycles take longer," in part because you need to be very careful about unintended consequences and you need to make sure you are not doing something wrong, such as exposing sensitive information. Patel also said that compared to things like the big image databases used in consumer applications, enterprises just don't have as much training data, and training has to be done "per tenant" (in other words, for each enterprise individually), so we need better algorithms that require less data. Patel also said organizations will need a "chief ethics officer" to make sure the data is used properly.
Kyle York, General Manager, Business & Product Strategy for Oracle, who joined the firm as part of the Dyn acquisition, noted that Oracle has added AI components to many of its applications, in areas such as Enterprise Resource Planning (ERP), Human Resources (HR), and Customer Relations Management (CRM), and mentioned the company's recent acquisition of DataScience.com. York also noted that only 10 to 15 percent of enterprise workloads have moved to the cloud, and he said Oracle aims to make AI and machine learning safe and secure.
York said there is both platform data, which Oracle can use to improve its products, as well as customer data, which remains the customers', and said there are opportunities for improved "governance tooling"—helping enterprises understand what data is valuable, what data is risky, what data customers could object to you collecting, and more.
The two enterprise vendors later discussed the issues surrounding propriety
York agreed that proprietary data sets are a real issue, and said that while anonymized data sets can be put together, you have to move carefully because these can turn out to disrupt privacy and potentially be detrimental to future business models. He said Oracle is "trying to democratize data," through things such as its Internet Weather Map, which brings together lots of data sets, including aggregate and anonymized data. We're still in early days when it comes to data openness, in his view, and "a lot goes back to human context."
In other conversations I had at the show, I heard some perspectives on other issues facing AI. Cliff Justice, who heads KPMG's Innovation & Enterprise Solutions practice, told me that cultural issues are in fact the biggest problem for most companies when deploying AI. Justice noted that to properly implement today's systems, you need to first tag a lot of content; then, once a model has been created, employees need to accept that it will make mistakes, and determine what the correct answer should have been; and finally, create a new model and repeat. But this all depends on line employees understanding that the system won't be perfect, and committing to take the time to identify and implement corrections. This is a big change from the kind of activities these employees are accustomed to, in many cases, and making the change is not easy, Justice said.
Self-Driving Cars and the Larger Impact of AI
On the main stage, speakers discussed self-driving cars and some of the larger issues facing AI.
Transportation Secretary Elaine L. Chao repeated the mantra that the department intends to be "tech neutral" and not pick winners and losers, emphasizing that "safety is always paramount."
Asked when we will see self-driving cars, Chao responded that these will happen "a lot faster than what some people think but not as fast as others."
Chao noted that the department issued a roadmap for such vehicles last fall, but as things are moving faster than expected, new guidelines will be issued later this year. In particular, she said, everyone is moving quickly to "Level 2" autonomy—in which a human being still has to touch the steering wheel—and that the department only has one application for a waiver for a "Level 3" vehicle, from General Motors. A procedure for granting that waiver hasn't yet been determined.
Chao noted that there is a perception problem; 74 percent of Americans say they would be uncomfortable getting into a self-driving car. She referenced an accident in which an autonomous vehicle powered by Uber killed a pedestrian in Arizona, which "showed how fragile public trust is," she said.
Chao said it's important that we do not end up with a patchwork of state regulations, but said she is unsure whether regulations should be assembled by the federal government or by states working together under federal guidance.
Another session featured General Motors VP of Strategy Mike Abelson and Diveplane CEO Mike Capps, formerly of Epic Games. They discussed "what impact AI will have on humanity" in a conversation moderated by Marissa Mayer, co-founder of Lumi Labs and former Yahoo CEO.
"AI will impact everything," Abelson
AI will change how people interact with all sorts of devices, Abelson said, and voice interfaces "will feel a lot more like Star Trek really quickly." Capps said he's more afraid of the Twilight Zone. "A black box scares the hell out of me," he said, and to that