Interview: Caterina Fake of Hunch.com

Decisions, decisions…

The Web has altered purchasing and personal decisions in the last decade. Ten years ago, a person would read Consumer Reports or talk with friends for a car or washing machine recommendation. Before Match.com, you might actually ask your friends to set you up on a date. Then came the Web and people started seeking out opinions and information from a variety of sources, but the volume of information and opinions can often bring more confusion than clarity. To solve this problem, better decision engines have become a sought-after tool in the online world (Microsoft’s Bing, released last week, is part of this trend). More sophisticated recommendation tools, based on increasingly advanced decision-helping algorithms, have become important for search engine companies because they can make the user experience more efficient and help merchandisers target consumers purchasing proclivities.

Caterina Fake, a founding member of the photo-sharing community Flickr, has a new, and interesting decision-making site, called Hunch.com, which launches today.

Here is how Hunch works: it gathers information about your likes and dislikes, and then when you have a question (“Should I get a tattoo?”) it asks you 10 questions, usually fewer, about the topic and then provides specific advice. Hunch’s “personal recommendation engine” was developed by MIT computer scientists with experience in machine learning. Like Wikipedia, its success will depend on participation from its users.

GAP: Tell me about the origins of Hunch.com. Did you see how difficult it is for people to parse through information on the Web, or did the idea for the company come about in a more organic way?

Caterina Fake: We really didn’t have a Eureka moment. The evolution of Hunch started with some hardcore computer science guys who were creating some amazing algorithms, combined with my background in creating great user experiences. While I was at Yahoo [Yahoo bought Flickr in 2005] I learned a lot about social search, and I am intrigued about how people find things online.

GAP: How much time do you think you can cut down on the average Web search through Hunch.com?

C: Let’s say you’re trying to figure out how to buy a grill. Right now you probably go online and read reviews and try and find information about grills. You have to decide if you want charcoal, propane, a rotisserie, a smoker… With Hunch, you answer some questions and it makes a recommendation. The information comes from our users: one person might type in that he wants to buy a grill. The second person asks, do you want charcoal, propane, and what other features? The next person might ask if you want a stainless steel finish? The fourth person will ask how much you want to spend? In this way a knowledge base, which replicates expert systems, is built up. We have 40,000 people on the [beta] site and 20 percent are contributing topics. Your grill question should be answered very quickly. Instead of spending time going to read different reviews on Cnet and other articles, Hunch will give you an answer in less than three minutes.

GAP: To make money Hunch will link to external sites where users can purchase a product or service, and Hunch will earn a referral fee. If I ask a friend about what car I should buy, he will tell me but not have a financial incentive to push me toward a specific model. Are you worried that a company showing, say, an attractive advertisement alongside Hunch’s recommendations could alter what people are attracted to on the site, thus eschewing the data?

C: The algorithm is completely independent of the advertising. We are not trying to push you to buy a certain car. And Hunch is very difficult to manipulate because every user gets a different answer [depending on their needs and preferences].

GAP: The customer data you’re collecting is valuable. Are you going to sell it?

C: Heck no; no selling it.

GAP: Do you see this as a back-end product, which retailers would want to license, or as a standalone product?

C: Definitely a standalone product.

GAP: In its Beta release, Hunch has about 2,400 decision topics, 14,000 follow-up questions and more than 50,000 decision outcomes. I am guessing when you launch [today] to a wider audience, the outcomes will go up exponentially. You say that as more people use Hunch, it will become smarter about what it knows by harnessing the collective knowledge of the entire Hunch community. Let’s imagine I want to buy a car. Since the community is not objectively looking at, say, buying the best car but following the crowd, can the algorithm get eschewed and favor a popular car, just because it is popular? And why wouldn’t I go to Caranddriver.com, a true expert source to find out the best automobile for me instead of relying on the so-called wisdom of the crowd?

C: It is user-generated collective knowledge, like Wikipedia. I worked on Yahoo Answers and the question–What dog breed should I get?—has been asked many times, but the knowledge has not been accrued so the system hasn’t learned anything over time. Wikipedia and Flickr have succeeded because [information] accrues over time. Users train Hunch to be smarter. The more Hunch learns about a person’s personality and preferences the better it customizes the results. If you build these systems in an intelligent way, they get smarter, not stupider.

[Note: This interview with Caterina Fake was edited and condensed.]