Live data from Hacker News

Show HN: Open Source Bot That Summarizes Top Hacker News Stories Using GPT-3

github.com

1–10 of 36 posts

Show HN: Open Source Bot That Summarizes Top Hacker News Stories Using GPT-3

#1
HN Summary is an open source bot which sumarizes top stories on Hacker News and publishes the summaries to a Telegram channel.

Whenever a new story appears on the Hacker News API /topstories.json endpoint, this bot summarizes it (currently using OpenAI GPT-3 text-davinci-002) and sends the story title, summary, and url to the hn_summary channel on Telegram.

The purpose of this project is to help build intuition on the capabilities of the current generation of large language models while making a broader swath of top Hacker News content more easily accessible. It could also serve as a platform for experimentation with other language model capabilites such as semantic search.

Join the HN Summary channel on Telegram to see the bot in action and enjoy the story summaries.

https://t.me/hn_summary

There are a number of potential directions I am interested in exploring here, such as making the bot interactive to allow features like bookmarking urls, semantic search, and semantic filtering.

I am also thinking about interfaces other than telegram. I started with telegram since I had recently used it on another project and it seemed like the easiest way to start playing with it. What other interfaces would make sense here?

Show HN: Open Source Bot That Summarizes Top Hacker News Stories Using GPT-3
github.com

Re: Show HN: Open Source Bot That Summarizes Top Hacker News Stories Using GPT-3

#4
FYI, here is the current prompt prefix we prepend to the story text before sending to GPT-3. Any suggested alternatives? Will try them based on points here.

"Provide a detailed summary of the following web page, including what type of content it is (e.g. news article, essay, technical report, blog post, product documentation, content marketing, etc). If there is anything controversial please highlight the controversy. If there is something unique or clever, please highlight that as well:"

Re: Show HN: Open Source Bot That Summarizes Top Hacker News Stories Using GPT-3

#6

If only n-gate were still publishing their own weekly HN summaries still. The humor was a much needed filter to keep me from getting too serious about things.

what was n-gate? I have a pet theory that humor will be the actual litmus test for AGI.

Re: Show HN: Open Source Bot That Summarizes Top Hacker News Stories Using GPT-3

#7
post #4

FYI, here is the current prompt prefix we prepend to the story text before sending to GPT-3. Any suggested alternatives? Will try them based on points here. "Provide a detailed summary of the following web page, including what type of content it is (e.g. news article, essay, technical report, blog post, product documentation, content marketing, etc). If there is anything controversial please highlight the controversy…

I'm curious: is the "please" you added purely a communication habit or does it lead the model to provide better results?

Re: Show HN: Open Source Bot That Summarizes Top Hacker News Stories Using GPT-3

#9
post #7
post #4

FYI, here is the current prompt prefix we prepend to the story text before sending to GPT-3. Any suggested alternatives? Will try them based on points here. "Provide a detailed summary of the following web page, including what type of content it is (e.g. news article, essay, technical report, blog post, product documentation, content marketing, etc). If there is anything controversial please highlight the controversy…

I'm curious: is the "please" you added purely a communication habit or does it lead the model to provide better results?

It is perhaps unnecessary, and I don't have any data that would support that it makes things better, but I am basically guessing that in the training set of {mostly the entire internet} the responses to questions that include "please" might be ever so slightly higher quality! Perhaps it is just a superstition...

Re: Show HN: Open Source Bot That Summarizes Top Hacker News Stories Using GPT-3

#10
post #8

Doesn't text-davinci-002 have a text input limit , what happens to stories that exceed that limit ?

We use the GPT2 tokenizer (which is apparently the same as the GPT3 tokenizer) to count the number of tokens in the extracted input text and brutally truncate the input text to keep it under the limit. There is a lot of room for more finesse in this area if anyone wants to help out. See open issues in github.
Post reply on HN