Launched: AI-powered summarization for RSS and Reddit

We launched ChatGPT-powered summarization for Substack a couple weeks ago, to help users see through click-bait. Now you can do the same with items from any RSS feed or subreddit. Check it out!

New Feature
Artificial Intelligence
Author

Brian Kent

Published

April 24, 2023

Apricot’s ChatGPT-powered summarization for Substack newsletters has revolutionized content discovery. Apricot’s article summaries let users see past clickbait headlines so they can make better decisions about what content to spend time on. Following the success of Substack summarization, I’m thrilled to announce that AI-powered summaries are now enabled for all RSS and subreddit feeds on Apricot!

How to generate summaries

Any item in Apricot that has the    icon in the top right can be summarized. Click or press the icon to generate the summary. Click it again to make the summary go away. It’s easier to just show you—check out this quick video.

Why bother? Some examples

Let’s look at a couple examples to get a sense for how Apricot’s summarization works and how it helps. First, an example from the Hacker News front page.1

The first item in this screenshot is Nearly all apps of generative AI are a net negative for society. This is a “hot-take” title whose goal is to persuade you to click by arousing strong emotion without revealing the true take-home message.2

Apricot’s summary shows the author’s true fear is “fake content”, which is a much narrower concern than what the title suggests. Don’t know about you, but I’m not gonna bother clicking that one.

The second item in this screenshot does not have the    icon, which means Apricot can’t generate a summary. This happens when an item simply doesn’t have enough content, which is common with link- or image-only posts on Reddit.

Here’s another screenshot from a Reddit feed, namely r/movies.

The second item—Girl With the Dragon Tattoo, did you notice….—is pure clickbait. Apricot’s summary tells us what the content is really about: the user discovered a hidden scream in the movie. I’m actually intrigued by that—more than I am by the author’s own title—so I’ll click and learn more.

Limitations

Summarization is super powerful but there are some many caveats.

All about the content

As I mentioned above, many RSS and subreddit items don’t have enough content to justify summarization. Apricot does not crawl underlying links or comments, so often there’s just nothing.

Balancing values

Summarization creates tension between two of our core values. On one hand, we want to see through clickbait, but on the other hand, a perfect summary would mean you never read original content, which is bad for multiple reasons. Apricot’s compromise is to limit summaries to one sentence. The summary gives the reader an accurate gist, but they still need to open the original content for details.

Spotify and YouTube are coming

We have not yet enabled summarization for Spotify podcasts or YouTube videos. We’re working on summarizing transcripts of the audio from these sources.

ChatGPT’s not perfect

Lastly and probably most importantly, Apricot’s summaries are auto-generated by ChatGPT, which is a probabilistic tool prone to errors and hallucinations. Please do not rely on Apricot summaries for high-stakes tasks. When in doubt, just click through to the original source.

Testing in prod

RSS feeds are total chaos compared to the predictable schema and formatting of Substack newsletters. In my internal testing, I’ve been happy with Apricot and ChatGPT’s ability to handle all sorts of inputs, but I can’t possibly cover all corner cases myself. So, please don’t hesitate to send feedback if you see anything that looks buggy. I’m particularly interested in the quality of Apricot’s summaries, especially for non-English content.

Footnotes

  1. To add this feed to Apricot, use this link: https://hnrss.org/frontpage↩︎

  2. Is there a proper taxonomy of the types of clickbait?↩︎