Imagine you tell your dad something about a negotiator, or a car. Just enough material for him to land an appalling pun:
That’s because [the] negotiator got my car back to me in one peace.
But this is the work of no dad. Instead, it was created by a pun-generating artificial intelligence, created by Stanford researchers to prove that a neural network can have a sense of humor, too. The results were hilarious, but not always in the way you'd expect. Some of these knee-slappers were barely coherent — but they might provide a preview of what happens when AI starts to become funny.
A preprint of the research, published in the journal arXiv last month, describes how the researchers fed pairs of homophones — like "two" and "too" — to a neural network.
The results often had an almost-there quality that'll be familiar to close watchers of AI-generated text. Here's an example: Even from the outside, I could tell that he’d already lost some wait.
There are plenty of challenges that remain.
Neural networks "are rule-abiding to a fault, and that makes them terrible jokers," writes Wired.
In an organized pun-off against competing human jokesters, the machine was rated to be funnier than a human just ten percent of the time.
"While we improve beyond current techniques, we are still far from human-generated puns," concludes the prewrite. Neural networks tend to conflate creativity with nonsense — which remains a "big challenge in humor."
READ MORE: The Comedian Is in the Machine. AI Is Now Learning Puns [Wired]
More on machine learning: Machine Learning Is Giving Retro Games Cutting-Edge Graphics