How One Podcast App is Using GPT-3 To Improve Your Podcast Listening Experience

Open AI’s GPT-3 has been in the news a lot recently for a number of reasons. Developers have used this new API to do things from writing blog posts that make it to the top of Hacker News, to creating search engines and more.

Today, we’re happy to announce you can add podcasts to the list of things GPT-3 is beginning to revolutionize!

The HyperCatcher app has now applied this exciting technology to our premium podcasts. Using GPT-3 and a little other magic you can now view a summary of your favorite podcasts before you listen. You can even jump past topics you’re not interested in (or commercials) and get to the good stuff more easily than ever!


How Did We Do It?

GPT-3 is a cutting edge machine learning model that’s been trained on content across the internet. This makes GPT-3 the perfect model for creative text generation. GPT-3 or a subsequent Machine Learning model from OpenAI might be great at creating podcasts from scratch one day given a topic, but we’re not quite there yet. One step removed from whole cloth podcast creation is summarizing existing podcasts and that’s exactly what we’re using GPT-3 for!

With GPT-3 the hard part is coming up with the right question and the right format for the question that will prompt the AI to fill in the answer with the right mix of creativity and understanding of the given question.

For example, if you sign up for Open AI’s beta you’ll get access to the GPT-3 playground that comes loaded with some preset prompts. One of these prompts is:

My second grader asked me what this passage means:

From here you can fill in the paragraph with a sample passage. For example, we could use a paragraph of text from a podcast like so:

My second grader asked me what this passage means:

“they age they saved money for a rainy day but this is four months that’s a long time miss you know what it’s not eat but it is to live a normal life with a car payment and a mortgage and a Verizon bill yeah that’s a lot of money but but you know what people shouldn’t have car payments”

Plugging this into the GPT-3 playground we might get a result like this:

I told her that it meant he was going to miss her and be sad because she’s not there. She said, “no, I think he’s saying that he doesn’t have a car payment.”

That’s pretty creative, but it’s not exactly a summary of the given text

This is where some of the parameters given by the API allow us to tweak the creativity and length of the response come in handy. In addition to tweaking parameters we might also give a few examples of what we consider “good” summaries of text that GPT-3 can use as a reference.

After applying those changes GPT-3 will give us something more like this:

“People shouldn’t have car payments.”

Now, we can basically feed a transcript of the entire podcast to GPT-3 in chunks and retrieve summaries of each part of the podcast that can be displayed to users!

If you’d like to check it out yourself tap the button below.


Stay up to date with our news, ideas and updates