The IBM computer known for demolishing humans in chess and Jeopardy is now trying its hand in the kitchen.
But every cook has culinary experiments go wrong once in awhile. Even with a successful cookbook on shelves, the Chef Watson app has turned up some less than scrumptious results.
The Chef Watson app takes the user's requested ingredients and throws together a recipe based on its memory of flavors that work together.
A Reddit AMA on August 25 with IBM, Bon Appetit magazine and the Institute for Culinary Education, revealed a Chef Watson original recipe for Chicken Breast Taco that might be delicious but is well, physically impossible.
"I've played around with the Chef Watson app but sometimes it leads to hilarious results," wrote bemused Reddit user ZipBoxer. "For example, [one] calls for wasabi powder (never used), shelled green peas (2 1/2 cup shelled green peas) cut into 3/4 pieces, then placed on a barbecue."
Grilled 3/4 pieces of peas might taste amazing with chicken tacos, but no reasonable human chef would go through the trouble of trying to grill them. IBM Watson researcher Patrick Wagstrom chalks the anomaly up to a natural language glitch, a consequence of a supersmart machine that lacks common sense reasoning and a hearty understanding of language.
In other words, the computer doesn't really understand what it's reading and recommending.
"It's probably trying to substitute green peas in for a similar ingredient," Wagstrom wrote. "Likely the original ingredient needed to be unwrapped and then sliced, so the natural parallel was to suggest to shell the green peas."
Indeed, the original recipe that Chef Watson was riffing on required "cut vegetables crosswise into 3/4-inch pieces." From Chef Watson's point of view, peas are a vegetable, so it works.
Chef Watson may seem clueless but it's endured a unique form of culinary training, called machine learning. According to the Washington Post, Chef Watson "ingests a huge amount of unstructured data — recipes, books, academic studies, tweets — and analyzes it for patterns the human eye wouldn't detect."
For the webapp that produced this pea-loving recipe, Chef Watson analyzed 10,000 Bon Appetit recipes. It then looked for statistical correlations among ingredients that tended to appear in recipes together.
This isn't the first time Chef Watson recommended an impossible step. Wagstrom said an earlier version of the software advised him to "One: Refrigerate the goat. Two: Skewer the tequila."
You take care of the goat, Chef Watson. I'll do the tequila.
