There are 318,979,564,000 possible ways to play the first four moves of chess. That may seem like a lot, but in reality, there are only six different pieces on the board that can each move a finite number of ways. And the goal is simple: Trap the king.
The average chef coming up with a nightly special may be considering hundreds of different ingredients, each with the potential to be paired and prepared in a vast multitude of ways. The possibilities quickly reach the quintillions. And the goal is incredibly vague: Taste good.
Almost two decades after the supercomputer Deep Blue beat Kasparov at chess, and two years after Watson beat Jennings at Jeopardy, IBM Research has begun work on an unnamed cyberchef, an AI system designed to create new dishes that can delight our palates at their theoretical peaks of enjoyment.
“If you look at chess, it’s a deductive problem. All the pieces are on the board and you deduce what to do,” project lead Lav Varshney tells Co.Design. But cooking isn’t chess. A chef must choose their own pieces to construct not an objective goal (a checkmate), but a complex and highly subjective interplay of flavor, texture, and presentation to delight our senses. It’s inductive reasoning, something IBM began to explore with Watson (a system that had to reason Jeopardy answers that weren’t on a chess board).
“We’ve been interested in pushing computing to a new direction, computational creativity. We’re trying to draw on data sets, not just to make inferences about the world, but to create new things you’ve never seen,” Varshney says.
And somewhere amidst the seemingly infinite possibilities, sheer numeric processing gives way to a seemingly magical, entirely human process: Creativity.
In the case of the flavorbot, these “new things” IBM is after range from spotting underrated, highly flavorful ingredients (like black tea, bantu beer and cooked apples), strange-but-tasty flavor pairings (like white chocolate and caviar, jamaican rum and blue cheese, or even bell pepper and black tea), and even whole recipes, complete with basic preparation steps.
To generate these food leads, if you will, AI cross references three databases of information:
- A recipe index containing tens of thousands of existing dishes that allows the system to infer basics like “what makes a quiche a quiche”
- Hedonic psychophysics, which is essentially a quantification of whether people like certain flavor compounds at the molecular level
- Chemoinformatics, which sort of marries these two other databases, as it connects molecular flavor compounds to actual foods they’re in
IBM tells me that these data sets–stored both locally and in the cloud–are actually “fairly small.” The AI’s job is to juggle each, decomposing and recomposing plates digitally while weighing cultural norms and our chemical perception of food.
“You generate a million new ideas, but of course a million isn’t useful,” Varshney says. “You want to rank them by which will be perceived as flavorful and which are novel.”
This ranking is essentially a means to maximize creative potential. So a meal of saffron and bear meat flavored with sandalwood–a recent pitch of IBM’s flavorbot–is inherently prioritized over another new recipe that features a more common set of ingredients like sweet potatoes, butter and rosemary. Bear meat may be a tough sell, and that’s entirely the point. If we were all already eating that dish, the computer would have had no reason to propose it. The team’s goal is novelty, or even more specifically, “changing one’s beliefs as to what food is.”
Once the dish idea is complete, the computer turns to a human counterpart (professional chefs supplied by IBM’s partner, the Institute of Culinary Education) to handle the actual cooking. Part of the reason is purely logistical–building a robotic chef would be a serious undertaking unto itself–but it’s also a symbolic gesture. IBM was quick to stress that their flavorbot is designed to be a collaborative system. In other words, it’s a far cry from Deep Blue, IBM’s highly secretive supercomputer crafted with the sole goal of publicly trouncing the greatest chess player of all time (and then refusing him a rematch). It’s not even Watson, the well-documented, borderline likeable computer that led Jennings to quip, “I for one welcome our new computer overlords.”
Indeed, the rhetoric around IBM’s culinary creation is not that computers will one day triumph over mankind, but that they just might save it. Varshney imagines his system as a powerful counterforce to the obesity epidemic, capable of developing some of the tastiest, healthiest, and least processed recipes that are theoretically possible–and even doing so on a personalized level to please picky eaters. It’s what I imagine as one part Deep Blue, one part Michelin chef, and one part helpful hippie checkout clerk from your local co-op.
Now, as for how these dishes actually taste, well, don’t knock bear meat until you try it.
“Nothing is really crazily bad, though there are certainly things we’ve tried making that weren’t spectacular, like a mideastern mushroom stroganoff,” Varshney insists when I ask. “Out of the 20 most recent dishes, 17 or 18 have been really good.”
And if you’d like to taste the flavorbot’s food yourself, see the “recipe” for Indian Turmeric Paella above.