robot box with a santa hat is singing fa la la la la la la

In 2017 I decided to find out what would happen if I trained a neural net on 240 Christmas carols (collected by The Times of London and reader/neural net hobbyist Erik Svensson). The result? A neural net that was very confused about how Christmas works:

King of toys and hippopotamuses [sic] full of the light of that stood at the dear Son of Santa Claus
He was born in a wonderful christmas tree

Run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolf the new born King.
flying reindeer has a medieval halo

You can kind of understand where the confusion came from.

But that was 2017, when I was training char-rnn from scratch on my laptop. Now in 2019 I have access to the much more powerful GPT-2, trained by OpenAI on 40GB of text from the internet. GPT-2 runs on the big processors on Google Cloud (through Max Woolf’s gpt-2-simple colab notebook), and has much better memory than char-rnn (which struggled to complete a single sentence). Did GPT-2 learn enough about the way words like Santa + reindeer are mentioned online versus Jesus + manger to figure out how to separate multiple kinds of carols? I gave the same 240 carols to the smallest version of GPT-2 (the 124M size), and then asked it what comes after “Rudolph the Red Nosed Reindeer”. Here’s what it added.

Rudolph the Red Nosed Reindeer, with its red belly
The All-gracious King of all the earth,
Had a baby at Christmas time,
On Christmas Day,
A true and holy Deity,
Went down to earth,
With human flesh for sacrifice.

For sinful men such a Deity doth appear,
And wink and nod in reply,
As he winked and nod in reply,
As he winked and nod in reply

The wretched world is run by ox and ass
The wretched world is run by ox and ass,
And in vain build I.

Using a more powerful neural net has apparently made the problem EVEN WORSE. The neural net is still confused about who exactly had that baby at Christmas time. And, most unfortunately, the neural net apparently saw that the word “flesh” appeared in the carols several times (Thank you, good king Wenceslas), as did the words “sacrifice” and “human”, and so it helpfully suggests “human flesh for sacrifice” as a likely phrase based on how it saw those words used online.

Some of the neural net’s renditions of Rudolph are disturbing in much more surrealistic ways.

Rudolph the Red Nosed Reindeer

I was riding by a hansom lettering me
On Christmas morning,
The snow was everywhere;
And the wind was cruel and cruel
Against me like ants,
Like ants, like ants.

Or this rendition, which I swear to god is verbatim

Rudolph the Red Nosed reindeer
Was the best sleigh in all of Nantucket
The horse was lean and tanned
The woman was full of perfume
The only thing keeping the goose from falling off the ground
Was the reins
Like strings of spaghetti
And if you could just cut it down
You could even make a bow
Like a bow and arrow
But if you wanted to make a bow you had to use a lot of power

Right down to the little lamb
Its red head was covered in feathers
Its heart was full of sugar
And the most prized food item was its head

Everywhere in the barn there were beavers’ eyes
Some would go, “Where is the sugar?”
But to them it was simple, it came from a good source
And they found sugar in olden times
Families that lived in the stable all had it

I can also get the neural net to generate custom carols. Here’s what it produced when I asked it to write “The Carol of the AIs” and I can’t tell whether it has misunderstood what Christmas carols are all about, or understood some of them only too well.

Carol of the AI’s

Come and own the yacht
It’s a very special day,
It’s a very special day

We’re all so proud of you
We’re buying this
We’re buying this now

Now it’s Christmas eve
And everyone’s shouting
“Santa Claus!”
“HO HO HO!”

It’s a very special day
It’s a very special day
It’s a very special day

One downside to using a neural net that was pretrained on a bunch of general internet text is that some of its carols involve an improbable level of swearing and gun violence. Become an AI Weirdness supporter to read the full versions of those (and a few others) as bonus content. Or become a free subscriber to get new AI Weirdness posts in your inbox!

Subscribe now

My book on AI is out, and, you can now get it any of these several ways! Amazon - Barnes & Noble - Indiebound - Tattered Cover - Powell’s

Subscribe now