Robots dream up hilarious cooking recipes

With Thanksgiving just weeks away, you might be thinking about how you can make your holiday meal unique. Those wanting a futuristic menu might want to take a look at a new program on Github that uses artificial intelligence to auto-generate recipes: the Robo-Meal Master.

Here is one of its meal suggestions: chicken beans muffins. This is how you make them:

If you read that closely, you may be rethinking your futuristic Thanksgiving. Chicken beans muffins don’t sound very appetizing, to start. You also likely did a double-take at “chopped beer.” And what on earth are “whipped cream ends of honey”!?

Coder Tom Brewe trained the Robo-Meal Master from a public database of about 160,000 recipes. It’s come up with barbecue ribs, caramel corn garlic beef, country dip cookies, hot garlic casserole, caramelled and cinnamon rolls, and grilled cheesecake, which might be something you’d find in a hipster dessert bar in San Francisco.

It’s built atop a program called char-RNN developed by Stanford University computer scientist, Andrej Karpathy. After reading through a lot of material, it simply learns the patterns of letters and words in a certain kind of text and uses that to craft its own, one letter at a time. I used this same program earlier this year to create a bot that wrote its own erotica. Its musing were more comedic than they were sexy, just as this program’s recipes are more funny than edible.

One recipe calls for one pound of “lean bag in microwave.” Another calls for one teaspoon of “poppy strip powder.” Still another calls for one tablespoon of “fat flour.” Rarely do the instructions match the ingredients, nor do they make much sense. (“Remove and canned the cookies, and cook till the flavors.)

This isn’t the first time a computer has been put to work as a sous chef.

Earlier this year, computer scientists taught robots how to make pancakes and pizza after reading recipes on the web. IBM released an app with food magazine Bon Appetite that’s meant to help humans come up with new, exciting and surprising plates. The thing has “thought up” recipes for meals like Peruvian potato poutine, Indonesian rice chili con carne and Belgian bacon pudding. Watson is trained on thousands of recipes from Bon Appetite, plus data on food chemistry and human taste preferences. From this data, it develops an understanding of how different ingredients work together to create grub that people will actually enjoy, and so it’s able to give you suggestions of things that are actually edible.

This program isn’t nearly as good at Chef Watson, but at least you’ll have a good laugh reading through some of the concoctions it’s dreamed up. While the names of some of the dishes sound tasty, once you read down into the ingredients and instructions, the recipes quickly devolve into computer gibberish. They’re better fit for an absurdist smorgasbord (or an episode of Jackass) than a gourmet foodie fest.

One of my faves is the Bar Rest Layer Sauce. At first glance, it sounded like a fabulous cure for a holiday hangover. Instead it’s a weird, non-sensical concoction of sugar, flour, butter, baking powder, lemon juice and garlic.

As you may have noticed, the recipe is filed under “poultry, beef”, but there’s no mention of either type of protein in the recipe. That’s because the bot that’s writing these cooking instructions doesn’t actually understand what it’s doing. Its brain is really primitive.

Still, it’s kind of impressive that it’s managed to learn the general “look” of a recipe: title, serving size, list of ingredients, followed by the actual instructions. In some cases, it tells the cook to preheat the oven to 350 degrees Fahrenheit, just like a real recipe might, and to “bake in a preheated 325 degree oven for about 5 minutes. As with our Erotibot, this program has glimpses of brilliance. Unfortunately, they give out pretty quickly.

We’ve picked out a couple of its more off-beat recipes. Enjoy, and attempt at your own risk!

For now, char-RNN, the program upon which robo-recipe generator is based, is “mostly a toy,” its creator Andrej Karpathy told me recently. After all, it can only predict the next character in a sequence. But, it is actually a type of AI, called a deep recurrent neural network, that’s being used at places like Google and Facebook for image recognition and language understanding. These networks have a tiny built-in memory that makes them especially good at these tasks.

char-RNN’s “neural network guts,” he says “could be easily repurposed for a huge variety of real and important problems in society.”

Daniela Hernandez is a senior writer at Fusion. She likes science, robots, pugs, and coffee.