Generating Recommendations with GPT-3

Serendipity is a natural language recommendation engine powered by GPT-3. As opposed to traditional recommendation algorithm that generate recommendations based on implicit or explicit user feedback (such as ratings or items clicked), Serendipity works through a natural language search interface: you just tell it what you want. There are pros and cons to a system like this:


  • Recommendations are context aware: you may be in the mood for a thriller one night but a comedy the next

  • We are able to avoid the cold start problem -- recommendations can be generated even for users with no data

  • You can use this system to find very specific items based on somewhat esoteric clues (examples below)


  • Recommendations will be poor if the query is poorly worded

  • It's a challenge to augment this using traditional ratings or engagement data

In this blog post, we'll showcase some interesting results GPT-3 is able to find.

Very specific queries

Foreign language queries

Soon after I create the search interface, some people from Japan found this website and started using it. Although I didn't code any special logic for foreign languages, it kind of just works. Here are some example queries:

I tried a couple other languages and the results seem reasonable:

Movies/Books where you forgot the name but it's on your mind