Research scientist Janelle Shane used GPT-3 to create new Christmas carols. My favorite is "Mild is Rudolph"
Now in 2020 we have GPT-3, trained on so much of the internet that my challenge is now to stop it from simply repeating carols that it has memorized. Give it just three example carols and it will respond with another from the internet, which is technically exactly what it was rewarded for during training. To get it to generate new carols, I discovered I could give it three existing carols (I chose Hark! The Herald Angels Sing ("bruise in us the serpent's head"), All My Heart This Night Rejoices ("sin away and for aye full atonement maketh"), and Run Run Rudolph ("and away went Rudolph a whizzing like a merry-go-round")), plus just the title of a nonexistent carol. It would generate, for example:
Mild is Rudolph
Mild is Rudolph's image in the snow
He has a shiny nose and knows how to cozy-up-nose
He has a little red nose and tiny little hooves
He persuaded a mischievous infant to go straight to sleepGoodness, how downy is Rudolph's hide
He has, oh his forehead, so plump and smooth and divine
He has a little black snout and tiny little hooves
Benevolence makes foolish young children go straight to sleepShy is Rudolph's figure when he's near
He has a twinkle, oh in his eye, and he's kind of lean
He has a bright shiny nose, and tiny little hooves
He can make any child go straight to sleep
