A Hunky-Dory Week at Hunch—Questions and Answers with Caterina Fake, the Only “West Coasty” in a Roomful of MIT and Harvard Grads
It was a huge week for Hunch, the New York-based startup that’s experimenting with a new approach to answering life’s big and little questions—from “Which savings account should I use?” to “Which science fiction author would I like?” After more than two months in invitation-only preview mode, the site threw open its doors to the public on Monday, and doubled its user base in a single day to some 80,000 registered users.
An excited Caterina Fake, Hunch’s co-founder and chief product officer, told me, “There’s been a ton, a ton, a ton of activity” on Hunch since the public launch—including an unexpected number of users who are contributing to Hunch’s signature collection of questions or “topics.” As I explained in a column on Hunch back in April, each topic steers users toward concrete results after asking 10 or fewer multiple-choice questions. Designing topics—which take the form of decision trees, with each answer leading to a different set of possible outcomes—is fairly challenging, which is why Fake is pleasantly surprised to see as many as 20 percent of visitors trying their hands at it.
Fake, who’s famous as the co-founder of the photo-sharing service Flickr, calls Hunch a “collective knowledge system.” She thinks the high early participation rates can be attributed at least in part to the fact that audiences have been primed by similar online collections of user-generated advice and information, such as Yahoo Answers and Wikipedia.
But my bet is that Hunch is also tapping into something deeper—the fact that there are a lot of people who love giving advice. I’m talking about the congenitally opinionated; the people who, based on the scantiest information, seem to be able to supply definitive recommendations about where their acquaintances should shop or how they should dress or who they should marry. In other lives, these people might have been therapists, advice columnists, or astrologers—but now they can become topic authors on Hunch.
Actually, I’m being a bit tongue in cheek—in reality, as you can see from visiting Hunch and looking at the profiles of contributors, the service is attracting people for all sorts of different reasons. For now, according to Fake, the company isn’t too focused on making money—its staff of 10 is working mainly to make sure that it’s easy to build topics, and that users find useful answers.
While Hunch is based in New York, it’s practically an outpost of Boston, considering that General Catalyst Partners of Cambridge supplied most of its $2 million Series A funding. New York’s Bessemer Venture Partners, which has a Boston office, also participated. Then there’s the fact, in Fake’s words, that “We’re pretty much an MIT shop.” Co-founders Tom Pinckney and Matt Gattis, product designer Hugo Liu, and software engineer Peter Coles all have degrees from MIT. CEO and co-founder Chris Dixon and software engineer Will Gaybrick are from Harvard, which is close enough. And several of the Hunch principals worked together at SiteAdvisor, a Boston-based anti-spyware company acquired by McAfee in 2006. “I’m the only West Coasty type,” Fake says. [Correction 9:20 a.m. June 18, 2009: Polaris Venture Partners is not a Hunch funder, as this paragraph previously stated.]
I spoke at length with Fake on Tuesday. Here’s a condensed version of our conversation.
Xconomy: What’s the activity been like at Hunch this week?
Caterina Fake: Yesterday was obviously a huge day, because of the launch. We got a ton more contributions. We launched with 500 topics, and by the end of the preview period we had 2,500, and I’m going to venture a guess that we had roughly a bajillion added yesterday. We had roughly 40,000 members at the end of the trial period and that roughly doubled in one day yesterday.
X: What unexpected lessons have popped up since you opened up Hunch for preview in March?
CF: The main thing is that we didn’t anticipate we would have as many contributions as we got. That was huge, actually. We weren’t quite ready for the influx of topics, questions, and results. We had 20 percent of our users during the preview period turn out to be contributors. Usually, whenever you build a social media or collective knowledge system like this, like a Yelp or a Wikipedia or a Flickr, you get a different percentage of contributors, and our estimate was that less than 5 percent of Hunch users would become contributors, because [contributing] is fairly complex. It’s a different idea. And when I say 20 percent, I mean that 20 percent of people are contributing topics and questions, which is just a crazy number.
X: What do you think you’re doing with Hunch that draws people in and get them contributing?
CF: I think we’ve hit on something that is at large on the Internet today. We were well set up for this by previous sites such as Yahoo Answers and Wikipedia and the quizzes on Facebook. They’ve all set us up well for people understanding what we’re doing and accepting the idea that collective knowledge systems can be beneficial.
X: Do you think your 20 percent contribution rate will continue now that you’ve opened up Hunch to everyone?
CF: From my experience with Flickr and other products I’ve worked on, including Yahoo Answers, it tends to diminish over time. The contributors at the outset tend to be early adopters, people who read TechCrunch and spend a lot of time online. They are much more familiar with how you use this kind of software. They are contributor types. Later on in the process, as the topics get built out and people are finding Hunch through search engines and coming at it from a different direction, you will have a less participatory type of crowd who are more consumers than contributors. When Flickr was acquired by Yahoo, we had about a quarter of a million users, and 20 percent of the users were creating 80 percent of the content. At Wikipedia, there is no longer any need to build out the articles about gazelles or orang-utans. So we anticipate that the contribution percentage at Hunch will slow down.
On the other hand, there’s no limit to human questioning. There are a lot of people who need to make the same decisions—where should I go to college, what video camera should I buy, how should i get my hair cut—but there are also lots of questions that I didn’t even know people had. For example, some really wonderful things have shown up in Hunch in the past 24 hours, one of which was the question, “Which monster truck should I root for?” I knew monster trucks existed, but somebody built out this fantastic topic where they listed all the types of monster trucks, and the teams, and paint styles. Somebody else made another topic—what pen-and-pencil game should I play? Before that, I could only name two or three, tic-tac-toe, hangman, and pen-the-pig. So there’s a topic I would never have though of. That’s one of the things that is wonderful about building user-generated content sites. There is no end of interesting things that people will add to the system.
X: That’s true, but doesn’t it actually take a bit of skill and practice to come up with a topic where the decision tree includes good questions, and leads to useful outcomes? One of the problems I noticed with Hunch when I was reviewing it back in April was that there wasn’t much guidance about how to create good topics.
CF: This is exactly the kind of thing we’re working on codifying. There are already some very successful topics—for example, “Which hotel should I stay at in London?” It starts out by asking whether you want to stay in this or that neighborhood, how much you’re looking to pay, do you want something that’s fancy and decorated or more contemporary, a boutique or a bit hotel, et cetera. That’s a very successful topic, because it gives people answers that they say they like. We should be able to replicate the best practices for the “hotels in London” topic for hotels in Dallas, Paris, Tokyo, et cetera. That’s something we’re working on, but we haven’t actually come up with a design for how you present this as a best practice to someone who creates a topic. This is something we’re acutely aware of.
X: What else goes into ensuring a positive outcome for users?
CF: There are a bunch of different algorithms and machine-learning aspects to this problem. For example, there is a question selection algorithm that figures out which questions most frequently lead to a positive outcome. For example, I was helping to write a topic on “What young adult book should I read?” It had all these recommendations like The Red Pony and The Black Pearl and Of Mice and Men and Lord of the Flies and all these things you read when you’re a teenager. One of the questions that had been asked by the original question-maker—and this is not to embarrass this person—but the question was, “Have you ever read a young adult novel before?” Which is kind of a lame question for this topic, because the answers were either, “Yes, I have,” or “No, but I’m willing to try.” So what will end up happening is that the algorithm will discover over time that this is not a very predictive question. People answer yes or no, but it doesn’t actually contribute to a positive outcome. So the algorithm will mothball that question after a period of time.
Then there’s also this body of “Teach Hunch About You” questions, like “Do you believe in angels?,” which are fun and, interestingly, very predictive. So we also try to figure out, when we’re giving replies, how “taste-oriented” the question is. An example would be, if I’m a punk-rock chick hanging out in Harvard Square, or if I’m a housewife in suburban Connecticut, I might own the same TV either way and it wouldn’t have much to do with my taste. On the other hand, punk rock chick would probably read a very different set of magazines from suburban Connecticut housewife. So there’s an algorithm that tries to figure out, “Is this a taste-driven topic or not?”
X: Well, these algorithms that influence the outcome are another element that actually bugs me a little bit about Hunch. If I’m going to take the trouble to write a really good topic, I don’t want algorithms interrupting and sending people off in different directions than the ones I planned.
CF: Yeah, there is a kind of balance between these two things. If you are a hard-core contributor, there are all of these levers you can use. You can set a question’s importance—you can say that where a hotel is located is significantly more important than its price, and set that manually. And the algorithms tend not to override human input, if there has been a significant amount of it. It will, however, trigger something on the inside of the system if the topic is not giving people good answers. If we’re finding that there is a low success rate, the question creator could potentially be wrong [about how they designed the decision tree]. So the algorithm doesn’t override human input, but it does detect when those discrepancies take place.
X: You were saying earlier that there’s an unlimited number of questions that humans might have. But on the other hand, isn’t there a fairly finite number of frequently asked questions? And in that case, how do you handle it when a contributor wants to create a topic but it turns out that somebody has already written a topic for their area of interest?
CF: When you’re creating a topic that’s already in the system, the system does de-duping, and it’s up to our content team which topics to de-dupe. We prefer that people flow through to existing topics rather than create duplicates.
X: Okay, but wouldn’t the de-duping discourage people who are fired up about contributing? They might want to write a topic about “Which macro lens should I buy for my camera,” but if it’s already been written, they’ll just go away. And if you have algorithms that can measure the success of a topic, why not just let duplicate topics compete, and keep the ones that produce better ratings?
CF: I totally get what you’re saying. There is a constant battle here between the “Let a thousand flowers bloom” model and the constrained, “We’re only going to have one Hotels in London topic” model. This is an art, not a science, and whenever we have content meetings, these are exactly the questions that come up. I’m very much of the “Let a thousand flowers bloom” school, so I’m more liberal in terms of the number of topics I think we should have. But the reason there’s such a strong tension around these very questions is that if you’re going to create a good topic, you need to concentrate the training as much as possible around one topic, so that it’s well-trained and gives people good answers, rather than spreading the algorithm’s attention over a vast number of topics. If there were 12 topics all about hotels in Paris, none of them would be very well-trained. But this is a constant source of debate internally, and the care and feeding of topics is one of the most important decisions we have to make behind the scenes.
X: I love all these technology and philosophy questions, but I also have to ask the revenue question. I don’t see any advertising on the site yet—how are you going to make money?
CF: If you do product-related searches, you’ll see that the recommended products have links to e-retailers. So affiliate revenues, commissions on sales, is our model right now. But right now we’re not really focused on revenue. We are a 10-person team, and we’ve got enough cash in the bank to last us until we get some kind of revenue.
X: Do you think Hunch has the potential to become as big as Flickr?
CF: I hope so. I think it’s a very different model. There’s a very different read/write ratio on Hunch. At Hunch, you can create a topic that thousands of people will use. Whereas on Flickr, you can upload a photograph, and only three people will ever see it, and that’s still a successful Flickr interaction. So the systems are very different in that way. Obviously, Hunch is not a social network; you don’t hang out on Hunch, whereas you can hang out on Flickr. Hunch is more similar to Wikpedia or Yahoo Answers; it’s used episodically. Although you could also argue that Flickr has become, in some ways, the world’s infinite National Geographic, a vast photo encyclopedia. I do think that Hunch has the potential to have that wide a use.