Artificial intelligence

Can AI Generate Headlines That Will Make You Click?

Using the few-shot capabilities of GPT-3 for headline generation.

Image generated by Jacob Ferus using OpenAI DALL·E 2.

One of the powerful features of big language models, such as GPT-3, comes from its ability to learn after it has been trained and without being updated. You simply feed it a few examples of input and output pairs and then end the prompt with only input to let it complete the output. For example:

3 => 9
4 => 16
5 => 25
6 =>

and 36 will be completed. This is called few-shot learning. With this, we can quickly assign the model tasks without much work and effort.

In this article, I will use this feature to let GPT-3 predict:

• What headlines my future articles could potentially have and their view count
• What headlines will achieve a large view count

This will be based on my historical articles, their headlines and popularity.

First attempt

For the prompt I used the following framing:

The article “<title>” had a view count of <view-count>