By Matt Gross
It was the middle of summer and all I could think about were – tomatoes. They’d just started coming into farmers’ markets en masse and I was eager to start eating them atop toast in the morning, sliced with cucumbers into salads, and chopped into sweet-spicy salsas.
But I wasn’t supposed to be thinking about tomatoes. The IBM and Bon Appétit teams were supposed to be planning the next phase in the development of Chef Watson, our cognitive-cooking system. We’d recently begun a beta test with people pulled from the ranks of our readership in which we issued a challenge. In my mind, there was only one challenge that made sense.
“Tomatoes,” I said. “Tomatoes in anything—anything but salad.”
So began one of the more (surprisingly) difficult competitions in the so-far brief history of Chef Watson with Bon Appétit. Difficult not because the beta-testers didn’t step up—they did in droves, submitting recipes for everything from a tomato pie with blue cheese to a green-tomato consommé—but because the recipe ideas that the Chef Watson app discovered/invented required a lot of human intervention to make them truly work.
In a way, this is by design. Chef Watson is still, for the moment, intended as a guide for reasonably experienced cooks—a spur to creativity, not a be-all-end-all cooking solution. When you tell the system your cooking parameters, you get back a clever, inventive list of ingredients plus cooking directions based on Bon Appétit recipes—which we expect you’ll then play around with until you get the result you want. Chef Watson is an amazing piece of software, but like all computer systems, it’s limited by what we program it to do and what we expect of it.
“Watson doesn’t think visually,” explained Dawn Perry, our senior food editor, who selected, tested, and refined the readers’ recipes. “It doesn’t know that if you blend an avocado into a tomato mixture”—as Watson suggested in one reader recipe for gazpacho—“it’s going to be brown. Or smoked paprika—that’s going to turn everything pink.”
That kind of limitation, however, is exactly what we’re trying to discover in this beta test. And in my conversations with the IBM team, we’ve talked about how to address those issues.
Is it better to tell Watson “Don’t blend avocados!” or to develop a way for it to learn that lesson on its own?
In other words, do we go for expediency or innovation? (Answer: Both! Neither! It depends!) Software development, like cooking, is an ongoing process and whenever you think you have a final product, whether it’s a web app or a tomato pie, you test it fully, enjoy it to the best of your ability—and think about how you’re going to make the next one.
The great thing about this challenge, despite all of its, uh, challenges, is that, thanks to the cooperation of man, woman and machine, we wound up with three deliciously winning recipes—which you can read all about over here at BonAppetit.com