At Shuffle, we strive to help podcast fans chat and discuss episode content with each other. After seeing how often podcast fans bring up inside jokes and common podcaster quotes, we thought: what if we can bring podcasters into the conversation?
Using podcasters' transcripts and quotes, we seemed to have directed the completion model to finish sentences in a way that reflects the podcaster's style and past quotes.
Me: What do you think of Mark Zuckerburg? Kara Swisher (bot): I like Zuckerberg, he has grown up on me. I don't know him. He seems very cocksure - as he should be - at the age of 25 he built one of the most successful companies in history.
But the harder you push towards reality, the harder the completion model fails:
Me: Who is better, Reggie Miller or Michael Jordan? Bill Simmons (bot): Funnily enough I just did a piece on this. The short answer is that in terms of career Michael Jordan has been slightly better. But on a per 36 minute basis for his career.....Reggie Miller.
Me: How tall are you? Joe Rogan: I'm six feet and two-tenths of an inch tall.
Given how few samples and quotes we fed the API, we were pleasantly surprised by how well GPT-3 picked up on the styles and personalities we wanted. At the same time, its inability to answer factual questions in a consistent manner suggests that it has little understanding of entities and real world relationships.
We've shared this with a few fan subreddits[1] and have gotten over 14k messages within 24 hours. Play around with it and see what you find. Would love to hear your insights!
[1] example: https://www.reddit.com/r/MBMBAM/comments/os4ndz/got_sent_thi...