Andy Herd, a cartoonist and software extraordinaire, has fed all the scripts of the popular ’90s sitcom into a database. Then, using what’s called a “recurrent neural —i.e. a form of machine learning—it is now outputting scripts for new episodes of the sitcom, reported the Daily Beast. It works by scanning the hundreds of thousands of words in the existing scripts to then predict the next letter after a string of other letters to create dialogue.

“It can generate stuff within minutes, but it’s barely English,” Herd told the Daily Beast.

However, every so often little nuggets of comedic gold pop out of the program, leading to lines like these:

Chandler: (in a muffin) (Runs to the girls to cry) Can I get some presents.”

or

Joey: Seriously give me a clown on the table that’s all.

Herd tweeted about the new project Monday, sharing lines from his program. The initial Tweet has generated over 4,500 retweets and 5.200 likes.

https://twitter.com/_Pandy/status/689209034143084547

He also joked on Twitter that he could sell an improved version to a network, and “all I’d have to do is push a button and BAM, top quality sitcom fodder ready to go.”

Herd used Google’s (GOOG) machine learning library TensorFlow, which went open source in November, to help make his project a reality. Programs like these aren’t all new (though its use to create Friends episodes is amusingly fresh).

It’s much like cognitive computing, and IBM’s (IBM) Watson computer, which uses a much more powerful system to win Jeopardy! or help doctors better find rare diagnoses. Cognitive computing systems are able to simulate human thinking patterns to recognize speech or identify items in pictures or even develop all new recipes for chefs, for example.