PAY NO ATTENTION TO THE EIGHT PEOPLE BEHIND THE CURTAIN

 

Meet Max Ventilla and Damon Horowitz, technologists with a vision to build a new type of search software designed to answer the kinds of questions that befuddle state-of-the-art companies such as Google. Google befuddled? Think about it. Google and its peers excel at answering factual questions: What is the tallest mountain in the world? Who was the twenty-third president of the United States? But for more subjective questions, Google struggles. Ask, “What’s a good place to go out for a drink after the ball game in my city?” and the technology flails. What’s interesting about this class of queries is that they are relatively easy for a person to answer. Imagine being at a cocktail party surrounded by friends. How likely would you be to get a high-quality answer to your subjective question? You almost certainly would get one. Unlike factual queries, because these subjective questions have no single right answer, today’s technology struggles to answer them. Such questions depend on the person answering them, his or her personal experience, taste, and assessment of what you’re looking for.

To solve this problem, Max and Damon created a product called Aardvark. With their deep technical knowledge and industry experience, it would have been reasonable to expect them to dive in and start programming. Instead, they took six months to figure out what they should be building. But they didn’t spend that year at the whiteboard strategizing or engage in a lengthy market research project.

Instead, they built a series of functioning products, each designed to test a way of solving this problem for their customers. Each product was then offered to beta testers, whose behavior was used to validate or refute each specific hypothesis (see examples in sidebar).

广告：个人专属 VPN，独立 IP，无限流量，多机房切换，还可以屏蔽广告和恶意软件，每月最低仅 5 美元

The following list of projects are examples from Aardvark’s ideation period.7

Rekkit. A service to collect your ratings from across the web and give better recommendations to you.

Ninjapa. A way that you could open accounts in various applications through a single website and manage your data across multiple sites.

The Webb. A central number that you could call and talk to a person who could do anything for you that you could do online.

Web Macros. A way to record sequences of steps on websites so that you could repeat common actions, even across sites, and share “recipes” for how you accomplished online tasks.

Internet Button Company. A way to package steps taken on a website and smart form-fill functionality. People could encode buttons and share buttons à la social bookmarking.

 

Max and Damon had a vision that computers could be used to create a virtual personal assistant to which their customers could ask questions. Because the assistant was designed for subjective questions, the answers required human judgment. Thus, the early Aardvark experiments tried many variations on this theme, building a series of prototypes for ways customers could interact with the virtual assistant and get their questions answered. All the early prototypes failed to engage the customers.

As Max describes it, “We self-funded the company and released very cheap prototypes to test. What became Aardvark was the sixth prototype. Each prototype was a two- to four-week effort. We used humans to replicate the back end as much as possible. We invited one hundred to two hundred friends to try the prototypes and measured how many of them came back. The results were unambiguously negative until Aardvark.”

Because of the short time line, none of the prototypes involved advanced technology. Instead, they were MVPs designed to test a more important question: what would be required to get customers to engage with the product and tell their friends about it?

“Once we chose Aardvark,” Ventilla says, “we continued to run with humans replicating pieces of the backend for nine months. We hired eight people to manage queries, classify conversations, etc. We actually raised our seed and series A rounds before the system was automated—the assumption was that the lines between humans and artificial intelligence would cross, and we at least proved that we were building stuff people would respond to.

“As we refined the product, we would bring in six to twelve people weekly to react to mockups, prototypes, or simulations that we were working on. It was a mix of existing users and people who never saw the product before. We had our engineers join for many of these sessions, both so that they could make modifications in real time, but also so we could all experience the pain of a user not knowing what to do.”8

The Aardvark product they settled on worked via instant messaging (IM). Customers could send Aardvark a question via IM, and Aardvark would get them an answer that was drawn from the customer’s social network: the system would seek out the customer’s friends and friends of friends and pose the question to them. Once it got a suitable answer, it would report back to the initial customer.

Of course, a product like that requires a very important algorithm: given a question about a certain topic, who is the best person in the customer’s social network to answer that question? For example, a question about restaurants in San Francisco shouldn’t be routed to someone in Seattle. More challenging still, a question about computer programming probably shouldn’t be routed to an art student.

Throughout their testing process, Max and Damon encountered many difficult technological problems like these. Each time, they emphatically refused to solve them at that early stage. Instead, they used Wizard of Oz testing to fake it. In a Wizard of Oz test, customers believe they are interacting with the actual product, but behind the scenes human beings are doing the work. Like the concierge MVP, this approach is incredibly inefficient. Imagine a service that allowed customers to ask questions of human researchers—for free—and expect a real-time response. Such a service (at scale) would lose money, but it is easy to build on a micro scale. At that scale, it allowed Max and Damon to answer these all-important questions: If we can solve the tough technical problems behind this artificial intelligence product, will people use it? Will their use lead to the creation of a product that has real value?

It was this system that allowed Max and Damon to pivot over and over again, rejecting concepts that seemed promising but that would not have been viable. When they were ready to start scaling, they had a ready-made road map of what to build. The result: Aardvark was acquired for a reported $50 million—by Google.9