Fashion designer Agnes Cameron poses in a jumpsuit inspired by an AI-generated design. Photo: Pinar Yanardag
Shrimp and jam pizza: a concoction so absurd no human could have dreamed it up. Because no one did. The recipe was created by a computer that read hundreds of artisanal pizza recipes, and was realized in edible form by a Boston pizza chef.
What’s going on: A group of MIT students are training artificial intelligence systems to come up with crazy new ideas in fashion, food, art, cocktails, and dance — and then bring them to life.
The project was started by students who were disappointed by news articles that cast AI as a fast-approaching threat. Having taken a flagship MIT course called "How to make almost everything," they saw a different role for AI.
- "We want to show the public that with AI you can actually create whatever you want to create," said Pinar Yanardag, an MIT post-doctoral researcher and the project’s founder.
Like the AI systems that drew the image we reported on last week, the ones Yanardag's group trained are not autonomously creative — they are useful only when paired with humans.
- Having ingested tons of examples of previous creations, the system cobbles together odd combinations that a person may not have thought of on their own.
- Despite the extreme variety in their projects — from fashion design and poetry to graffiti and board-game design — the group alters the programs as little as possible between uses. Most of their time is spent curating and cleaning a training dataset.
- "We don't want to give the impression that only people who are AI experts can tailor algorithms," Yanardag said.
Once AI draws up plans for the project at hand, the MIT team and friends execute them. Unbound by convention, the results are often wacky.
- The shrimp and jam pizza, for example, combined elements of seafood and dessert pizzas to make something that "sounds really weird but actually tastes really good," Yanardag says. The chef who cooked the oddity is planning to add it to the menu, she told me.
- Recently, the team made AI-invented chocolate truffles. Most of the results were pleasant — pumpkin and matcha, rosemary and peppermint — but one stood apart: a gingersnap and ground beef truffle.
- "Some people didn't want to bring this monster into reality," said Yanardag. But they did — for science — and it wasn't totally awful. After the truffle-making party, somebody took home all the meat chocolates.
Beyond a side project for MIT students, the AI-maker is a reminder of both the promise and limitations of today's artificial intelligence.
- The data that AI is trained on can introduce biases. This is how automated systems used for hiring, accustomed to seeing resumes from men, end up selecting against female candidates.
- Yanardag's team used bias to their advantage to cook creative pizzas: They trained their AI on artisanal recipes rather than Papa Johns', so the output would be novel.
- The team's programs, which use deep learning, also suffer from what's sometimes called the black-box problem: The programmers can look into the database to see where shrimp and jam came from, but they can't ask their AI to explain why it put them on the same pizza.
What's next: "I see a future where humans have to work with AI to boost their creativity," says Yanardag. AI may not be creative on its own, she said, but it's able to come up with "crazy combinations that humans can interpret and boost their own creativity."