These Creepy Halloween Jokes Were Written By AI! 🎃

Halloween is a time for scares, and there's nothing more terrifying than a bad joke. So, I decided to train a the GPT-3 AI language model to write jokes about halloween and digital marketing.

What is GPT-3?

GPT-3 is a large-scale artificial intelligence model released in 2020 by the research lab OpenAI. The model was trained on a dataset of over 45TB of text data and can generate text that is indistinguishable from the human-written text.

In other words, GPT-3 is a machine that can write like a human... including jokes (possibly).

How do large-language AI models work?

Large-language AI models, such as OpenAI's GPT-3, are trained using data and deep learning. Deep learning is a neural network that can learn to recognize patterns in data. The data is fed into the neural network, and the network learns to recognise the patterns.

The training data for these models is typically a large corpus of text. The text is fed into the neural network, and the network learns to recognise the patterns in the text. The more data the network is trained on, the more accurate it becomes.

These models are used for tasks such as machine translation, text classification, and question answering. The model’s can also generate text, and it does this by predicting the next word in a sequence.

How to train generative A.I. to write jokes

Generative A.I. technologies such as OpenAI's GPT-3 can be trained quite easily to perform tasks. Some models, such as GPT-3's Da Vinci, can simply be instructed to perform a command. For example, when I instruct it to:

Write a joke about Halloween and corporate life

It replies:

Why don't corporate types like Halloween? Because it's too much of a hassle to change out of their work clothes!

This is what is known as a zero-shot prompt, as I am providing zero examples in my instruction.

Alternatively, you can provide some example data which the model will use to learn from. For example, in a single-shot prompt, I could instruct it to write a joke but also provide an example joke, such as:

Write a joke about Halloween and corporate life:

Why did the headless horseman go into business?

He wanted to get ahead in life.

###

Note, the use of the ### to indicate to the model that the example has ended (this is called the stop sequence).

So, what did the A.I. generate after this example:

Write a joke about Halloween and corporate life:

Why did the headless horseman go into business?

He wanted to get ahead in life.

###

And it replies:

Why did the vampire open a business?

Because he wanted to make a killing.

A few-shot prompt is where multiple examples are given to the model. For example, I could provide two different jokes about Halloween and corporate life:

Write a joke about Halloween and corporate life:

Why did the headless horseman go into business?

He wanted to get ahead in life.

###

Why did the vampire open a business?

Because he wanted to make a killing.

###

And it replies:

What do you call a group of zombies working in a business?

A brain trust.

###

To learn more about writing prompts, check out this guide to creating prompts from Riku.ai.

20 Halloween themed jokes written by A.I.

Can A.I. really write good jokes?

I'll let you be the judge. The following jokes were all written by GPT-3, with some light editing to make them flow slightly better. The only joke on the list that wasn't written by AI was the first joke about Freddie Kruger, which I wrote to provide a one-off example in my prompt.

All of the accompanying images were created by entering the joke into the A.I. image generator on Riku.ai

Read on to see just how good jokes written by A.I. about Halloween and digital marketing really are:

The training data

In order to give the AI something to learn from, I entered a single-shot prompt with an example joke after the instruction “write a joke about digital marketing and halloween”.

This is the example joke I entered:

What happened to the marketing budgets when Freddy Kruger joined the company?