Skip to main content

This is how supercomputers cook: IBM’s Watson dreams up creative dishes

This is how supercomputers cook: IBM’s Watson dreams up creative dishes


Welcome to computer culinary school

Share this story

A master chef can think about a combination of two, maybe three ingredients at a time. Watson, the same IBM supercomputer that won Jeopardy in 2011, can crunch through a quintillion. That's a one and 18 zeroes, as the IBM researchers like to say. But does that make their computer a good cook?

For about two years, IBM's cognitive computing group has been working to apply Watson's vast processing ability to food. The system analyzed about 35,000 existing recipes and about 1,000 chemical flavor compounds, which allows it to make educated guesses about which ingredient combinations will delight and, just as importantly, surprise. From there, it tries to encourage unconventional combinations — like chocolate, coffee, and garlic — in order to produce dishes that have never been made before.

"Creativity is the crowning achievement of human intelligence," says Steven Abrams, director of the Watson group. "Can we make a computer be creative?"

The system asks for an ingredient, a region, and a type of dish, and then spits out a bunch of "recipes"

Chef Watson is certainly creative, although for now it needs a lot of human input. The system asks for an ingredient, a region, and a type of dish, and then spits out a bunch of "recipes," which for now are just lists of ingredients. The system rates each recipe, which typically includes about 12 or 14 ingredients, according to how well the flavor compounds line up, how novel the ingredient matchup is, and how pleasant the computer thinks it will smell and taste. In the future, the team says, Watson will be able to give step-by-step instructions.

On Friday at the South By Southwest Interactive trade show, ICE's director of culinary development James Briscione and creative director Michael Laiskonis were handing out samples from IBM's fire engine red food truck. A vote conducted on Twitter had determined the type of food — kebab — while the chefs decided on the region, proportions, cooking time, and plating. Watson picked all the ingredients: pork, chicken, strawberry, shiitake mushroom, pineapple, apple, curry, green onion, carrot, lemon, lime, and mint. "I have never made anything with this combination of ingredients before," Laiskonis says.

It was delicious, but a little safe, so The Verge tried to come up with something weirder. Using IBM's prototype app, we selected avocado as a base ingredient, the Canary Islands as the regional inspiration, and soup for the type of food. Disappointingly, all the results rated near-zero on the "surprising" scale.

That's because humans have done almost everything there is to do with soup, making it difficult for Watson to come up with something new, explains Florian Pinel, a senior software engineer and a trained chef who has his own food blog. An avocado Napoleon from the Canary Islands, however, got an almost perfect surprise score.

Humans have done almost everything there is to do with soup

IBM's researchers are careful to say that Watson isn't designed to replace chefs or doctors, and it is difficult to imagine a serious chef relying on the machine's recommendations, at least in its early stage incarnation. No doubt there will also be a backlash from those who say a machine can never approximate human ingenuity. For now, Watson seems much better at coming up with long lists of ingredients that will mesh than at introducing an element of surprise. Friday's kebab and Saturday's Caribbean snapper ceviche with fried plantains were both tasty, but the combinations weren’t that unexpected. That makes sense, considering part of the calculus is whether the ingredient has been paired together in existing recipes.

The chefs seemed to find the ingredients more surprising than the eaters, so perhaps Chef Watson is more creative than it seems. A Cuban lobster bouillabaisse doesn’t sound that far-fetched, but Laiskonis and Briscione are pretty sure it’s never been done before.

Cognitive computing systems like Watson that learn and adapt to situations rather than execute programs can be useful everywhere, from music to finance and even healthcare, where it can help doctors make diagnoses. Watson may have made its public debut on Jeopardy, the paragon of useless knowledge, but IBM believes it can change the world.