Hunch.com: Not Ready for Prime Time

I’ve been following Hunch.com for a while since their site offers “personalized recommendations,” a concept I’ve been exploring with Trybe. Hunch’s “topics” are user-generated and they use AI (Artificial Intelligence) to make recommendations based on members’ answers to a series of questions like “Can you do 10 pull-ups?” and “Do you live in a city, the suburbs or a rural area?” Hunch creates a profile for each user and alleges they’re creating a “taste graph” based on the preferences of thousands of users.

‘The ultimate goal of the company is to map every person on the Internet to every object on the Internet, be that a product, a service, or a person,’ Co-founder Caterina Fake says.” Apart from the hubristic nature of this boast, is it possible or, more importantly, is it useful? It certainly doesn’t sound “personal.”

I used to think Hunch’s model made sense. Curated recommendations aren’t scalable and their relevance is a hit or miss proposition. Same with social network based recommendations. My friends don’t share all my tastes. Other AI recommendation services don’t work because they have an incomplete picture of the user. I’ve described in a previous post how Amazon keeps sending me emails about power tools because I bought one power saw—after buying hundreds of books and CDs.  So why not ask the customer for more information? Surely if you described the benefits of a truly personalized service—filtering, relevance, etc.—people would see it as a valid trade. I give you personal information, you save me time and effort. eHarmony users accept this. But people don’t like to give information about themselves and, when they do, there is a problem with the accuracy of the information they provide. A surprisingly large percentage of dating site users lie about their age.

Hunch is well funded and has a team of “machine learning” experts. But I’ve found the experience of using their site to be decidedly lacking. The overall impression is chaotic and lacking in relevance. They’re a mile wide and an inch deep. They rely too much on user-generated content. It’s hard to see the connection between your answers to their questions and the site’s recommendations. The Hunch lexicon, containing words like “taste graph” and “THAY” (Tell Hunch About Yourself), smacks of serious geekiness. Like the parents of a newborn, they seem to have a love affair with their own creation: “If you use an electric toothbrush, you’re 28% more likely to favor an aisle seat when flying.” The rest of us don’t find the baby quite as cute as they do.

There are several problems with Hunch’s approach. One is the accuracy of the data, mentioned above. Two is whether it’s predictive or not. In other words, I may know you like “The Sopranos” but how do I use that information to predict something else that you’ll like? Three is the sheer scale of building a system that can automate recommendations for millions of people based on millions of data points. They promise, “Hunch gets smarter the more you use it.” But how much delayed gratification will “normal” consumers tolerate? Four is the limitations of “machine learning” systems.

Jaron Lanier wrote about these limitations in a recent Op Ed column in the NY Times:

What all this comes down to is that the very idea of artificial intelligence gives us the cover to avoid accountability by pretending that machines can take on more and more human responsibility. This holds for things that we don’t even think of as artificial intelligence, like the recommendations made by Netflix and Pandora. Seeing movies and listening to music suggested to us by algorithms is relatively harmless, I suppose. But I hope that once in a while the users of those services resist the recommendations; our exposure to art shouldn’t be hemmed in by an algorithm that we merely want to believe predicts our tastes accurately. These algorithms do not represent emotion or meaning, only statistics and correlations.

What makes this doubly confounding is that while Silicon Valley might sell artificial intelligence to consumers, our industry certainly wouldn’t apply the same automated techniques to some of its own work. Choosing design features in a new smartphone, say, is considered too consequential a game. Engineers don’t seem quite ready to believe in their smart algorithms enough to put them up against Apple’s chief executive, Steve Jobs, or some other person with a real design sensibility.

I’m sure Hunch has a lot of great improvements in the works. And maybe they’ll be the next Yahoo as the service gets “smarter.” But they’ve disappointed a lot of early adopters, which isn’t a good sign.