Herd loaded every single Friends script into a recurrent neural network (RNN), which he briefly defined in an interview with The Daily Beast. “It works by predicting the next letter to follow a given sequence of letters, and the predictions are determined by what it learned about language from the Friends dialogue provided.” Basically, it learned to write an episode of Friends using just the information provided, which in this case, was episodes of Friends.
The “recurrent” part of an RNN defines the basic function of the process. To predict the next letter or word, the system relies on what came before it. It builds on what it learned in the previous few steps, and continues this pattern throughout the sequence. An RNN can also be “trained” for more predictable and fitting results, but is also limited by long spaces between dependencies, among other factors. Training is not easy and can include teaching the program to scan deeper, and also make predictions for future sentences, rather than just relying on current and previous data.
Herd’s computer initially spouted nonsense, but eventually some gold was struck. After running his program for 12 hours, Herd’s computer produced scenes that shows an inkling of what an RNN process is capable of in this type of application. Herd tweeted some screenshots of this bizarro Friends episode, and we’ve included some of our favorite lines below:
“Joey: Seriously give me a clown on the table that’s all.”
“Chandler: (in a muffin) (Runs to the girls to cry) Can I get some presents.”
“Monica: Happy Gandolf.”
Herd used Google’s artificial intelligence engine TensorFlow, which became open source in November. “I’ve been reading up about machine learning in my spare time lately because it’s really fascinating,” said Herd. “I honestly don’t understand a lot of the underlying mathematics, but I’m trying! Also, Satan helped a bit.”
Herd plans to further develop his Friends bot to make more sense, after which he may release it to the public. He hopes that his efforts can get people interested in learning new things.
