OpenAI's language model GPT-3, once used to fool people, is now in the mood for love. Artificial intelligence researcher and writer behind the AI Weirdness blog Janelle Shane decided to train the model to create its own pickup lines like an AI Cupid. Shane trained the neural net on existing sets of "awful" pickup attempts by humans. Thankfully, the results were mostly just odd instead of being cheesy and creepy like the ones made by people. Seriously, we need to improve our game.
Examples of the model's pickup lines include somewhat innocent and cute attempts like "I love you. I don't care if you're a doggo in a trenchcoat." That makes no sense and yet it's still way better than being asked "Are you lost, miss? Because heaven is a long way from here." In 2017, Shane ran another model on the same subject, pickup lines, which resulted in lines like "You are so beautiful that you know what I mean." Look, love isn’t perfect.
Say hello to DaVinci, Ada, Babbage, and Curie — Shane worked with four GPT-3 variants. The biggest one, DaVinci, was the "most competent" one, meaning that it created arguably meaningful sentences, according to Shane. DaVinci generated pickup lines like "You have a lovely face. Can I put it on an air freshener? I want to keep your smell close to me always."
Other gems include "You know what I like about you? Your... long... legs..." and "I once worked with a guy that looked just like you. He was a normal human with a family. Are you a normal human with a family?" plus "Do you like... pancakes?" Fair questions. My personal DaVinci favorite was "You look like Jesus if he were a butler in a Russian mansion." This sounds, to me, like a compliment, though tastes are subjective.
The second GPT-3 variant was Curie. This one was smaller than DaVinci and popped out a few curious pickup lines like "You have the best French Toast I ever had!" and "Picked up some pretty flowers. Wanna smell them? Here, try to take my hand off." Other smaller variants titled Ada and Babbage struggled to create coherent lines, per Shane. Ada especially came out with outlandish results such as "embroidery tags" and "body softening pads.”
In a frank admission, Shane wrote that she has avoided training more effective GPT-3 variants on pickup line samples for a rather hilarious reason. “I’ve resisted trying neural net pickup lines again, because more competent means more human-like, which in this case means worse,” Shane wrote. “Or the new neural nets might even copy existing pickup lines from internet lists, which would also be terrible.”
“Human-written pickup lines,” she added, “are that bad.” Word.