杏吧原创

I LIKE to think I鈥檓 a pretty good cook. I鈥檓 adventurous, have a decent understanding of the techniques required and I haven鈥檛 poisoned anyone 鈥� yet. But I still find myself cooking the same tried-and-trusted favourites, using the same old ingredients.

To shake things up, I ask to test the prototype of a cooking app that uses the brain of IBM鈥檚 Watson supercomputer to invent new dishes.

The key to the ChefWatson app is the supercomputer鈥檚 ability to devour large amounts of information and make links between chunks of it. It has already proved its mettle by winning the quiz show Jeopardy! and is being used to help doctors make cancer diagnoses at Memorial Sloan Kettering Hospital in New York. Now it is attempting to do something computers usually find difficult: using creativity to invent recipes people will actually want to eat.

To provide the data, IBM teamed up with US recipe website . This has a database of more than 9000 recipes, tagged according to their ingredients, type of dish and cooking style 鈥� Cajun or Thai, for example. Watson creates a statistical correlation between ingredients, styles and recipe steps in the database and uses this to work out which ingredients usually go together and what each type of food requires. 鈥淭hat鈥檚 how it knows that a burrito, a burger and a soup all need different elements,鈥� says IBM鈥檚 . 鈥淚t knows a burrito always needs a wrapper of some sort, whereas soup always needs liquid. That鈥檚 how you don鈥檛 get a runny burrito.鈥�

To use the app, the cook first types in an ingredient they want to use. Next, they decide the amount of experimentation they want Watson to take, ranging from 鈥淜eep it classic鈥� to 鈥淪urprise me鈥�. Watson then offers the cook further ingredients, styles and dishes that it thinks usually go well with the initial ingredient. They can promote or exclude ingredients by clicking the 鈥淟ove it鈥� or 鈥淗ate it鈥� buttons. Finally, the cook hits search, and Watson analyses its database to come up with a range of basic recipes that can then be further tweaked to make them more or less experimental.

If the cook wants to ramp it up, a database of flavour compounds found in a wide range of foods is also consulted and used to combine ingredients that should theoretically fit together, such as vodka and Brussels sprouts, or cauliflower and chrysanthemum. Psychological research into what flavours people find more or less pleasant is taken into account (PLoS Computational Biology, ), as is a 鈥渟urprise鈥� score for combinations of ingredients: the higher the score the less often they are found in recipes together.

鈥淎s you move to the experimental end of the scale, you鈥檙e looking less at what ingredients go nicely together and more at the flavour compounds that the ingredients have in common,鈥� says Abrams. 鈥淚t works out what things might go together that you鈥檇 never think of.鈥�

Well, that鈥檚 the theory at least. From my culinary forays, it鈥檚 easy to see why the app is not yet deemed ready for launch. In a creamy pasta dish (I had opted for the 鈥渆legant鈥� style), cr猫me fraiche had been replaced by a glass of milk. Another time, with a high experimental setting, Watson told me my tuna bake needed half a kilo of goose meat. I declined.

It was occasionally inspired. As I ramped up the surprise element, Watson suggested I replace clam juice in a classic gumbo dish with the Japanese soup base dashi. It worked 鈥� giving the dish a deep savoury flavour that I would never have thought to try out.

, creative director of the Institute of Culinary Education in New York, was one of the first professional chefs to help guide Watson. 鈥淚t lets you organise creativity in a more structured way, rather than just going in the kitchen and experimenting,鈥� he says.

The plan is for ChefWatson to become more sophisticated and to suck in data from more sources, says , the project鈥檚 lead software engineer. It already scours Wikipedia鈥檚 world cuisine pages and looks up nutrition facts in a US Department of Agriculture database for help on determining ingredient proportions.

Apart from the occasional odd ingredient combination, there are a few other bugs in the way ChefWatson conjures creative recipes out of pure data. One tester, Matt McSherry, says it seems to struggle with portion size. Another found that a recipe called for precisely 554 juniper berries, and she was also told to 鈥渟kin and bone鈥� her tofu.

Overall, McSherry is loving the experience. 鈥淚 get to muck about in the kitchen with a thinking, reasoning supercomputer. That鈥檚 pretty exciting right there,鈥� he says. 鈥淚t鈥檚 the future we were promised as children.鈥�

To try ChefWatson, go to

This article appeared in print under the headline 鈥淐ooking by numbers鈥�