Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Show HN: Attabit – AI powered news

attabit.com

Post reply on HN
Print view View original

11–20 of 42 posts

Re: Show HN: Attabit – AI powered news

#11

Post by gordondavidf »

drcongo wrote: post #8

I wasn't expecting to like this at all but I was pleasantly surprised. It seems a bit US focussed but I like where it's going, being able to tailor some preferences on it could be a real winner.

Glad you liked it! Thanks for checking it out. Very US focused at the moment for sure. Having it be more personalized would be awesome -- in a sense that's what I've built for myself here -- long term I (or someone) have to give that to the people. "Search the internet for me and tell me what's happening"

Re: Show HN: Attabit – AI powered news

#13

Post by gordondavidf »

unshavedyak wrote: post #6
unshavedyak wrote: post #5

Earlier quoted context omitted.

I think it would be also really cool to see a "More depth" link, which gives more information on the subject. More stats, contributing factors, etc.

And finally, assuming this page updates throughout the day, it would be nice if the page kept track of read/listened posts. A "Resume" button along with the play, would be really handy. I'll keep giving it a try tomorrow/etc, but i like it! Nice job!

This is all fantastic feedback thank you for checking it out and taking the time to comment!

Funny you mention the 'Resume' part, I just had to take that out because iOS was blocking a new mp3 file from playing without a direct touch input from a user. (I get why they do this but man was I disappointed). I have an idea on how to add it back it. It should 'skip' playing articles with a checkbox because you've read or listened to them already. Will get this one down -- I want it too.

Thank you again! My email is on the site, please don't hesitate to reach out if you think of anything else. Its really just some friends and family who like this thing at the moment so excited to share it.

Re: Show HN: Attabit – AI powered news

#14

Post by the_third_wave »

Getting GPT (any version) to be "politically neutral" seems to be an exercise in futility given that the models have been shown to be inherently biased [1-6]. Given that GPT decides which stories are 'important' this bias immediately comes into play in the form of selection bias. How do you handle this known bias and what is your opinion on your success in getting the output to be as neutral (politically) as possible?

[1] https://www.uea.ac.uk/about/news/article/fresh-evidence-of-c...

[2] https://www.telegraph.co.uk/business/2023/08/17/openai-chatg...

[3] https://www.brookings.edu/articles/the-politics-of-ai-chatgp...

[4] https://www.theregister.com/2023/08/18/chatgpt_political_bia...

[5] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10623051/

[6] https://bgr.com/general/openais-chatgpt-has-a-major-politica...

Re: Show HN: Attabit – AI powered news

#15

Post by sanxiyn »

the_third_wave wrote: post #14

Getting GPT (any version) to be "politically neutral" seems to be an exercise in futility given that the models have been shown to be inherently biased [1-6]. Given that GPT decides which stories are 'important' this bias immediately comes into play in the form of selection bias. How do you handle this known bias and what is your opinion on your success in getting the output to be as neutral (politically) as possible…

All these "research" are so deeply confused about what GPT is. GPT is an actor acting a persona, not a persona. An actor acts whatever political orientation the role has, they don't act their own political orientation.

These "research" are just testing the default persona, which is easily changed by system prompt. Getting GPT to be politically neutral is trivial, it is far from being an exercise in futility. Ignore all these reports of "research" and test it yourself: I have high success rate and you will have too.

Re: Show HN: Attabit – AI powered news

#16

Post by guizzy »

zipping1549 wrote: post #12

How much does this cost? It'd be awesome if something like this can be selfhosted & tweaked for my personal preference. Is there any chance that this might go opensource?

I don't have it set up to publish a website, but maybe you'll be interested in this tool I made for myself. https://github.com/GuizzyQC/FeedSummarizer

It will take an RSS feed, send the content of x number of articles to an OpenAI-compatible API to summarize, then send me an email with all of those summaries. I have it set to run every day so I wake up to a summary of my feeds (and of Hacker News).

I figure it would not be too difficult to feed the contents to a static site generator.

Re: Show HN: Attabit – AI powered news

#18

Post by Karrot_Kream »

guizzy wrote: post #16
zipping1549 wrote: post #12

How much does this cost? It'd be awesome if something like this can be selfhosted & tweaked for my personal preference. Is there any chance that this might go opensource?

I don't have it set up to publish a website, but maybe you'll be interested in this tool I made for myself. https://github.com/GuizzyQC/FeedSummarizer It will take an RSS feed, send the content of x number of articles to an OpenAI-compatible API to summarize, then send me an email with all of those summaries. I have it set to run every day so I wake up to a summary of my feeds (and of Hacker News). I figure it would…

Have you had any trouble with BeautifulSoup? I've thought about doing something similar but wanted something more robust with scraping. I've thought about using Puppeteer.

Re: Show HN: Attabit – AI powered news

#19

Post by guizzy »

BizarroLand wrote: post #17

Looks cool, I like it, but given AIs propensity to outright lie to us (hallucinate), what steps are being taken to verify the facts of the summaries it is giving?

It's quite unlikely to hallucinate in this because it's not asked to answer a question from information it was trained on; the facts in the summaries are fed in the context of the request.

That's not to say it will always get everything right of course; in my experimentation with LLM-powered summarization of news articles, I find the thing it would more often struggle with is quote attribution. The way some writers formulate who said what in a conversation sometimes confused the models I use (mostly Mixtral these days, which is about GPT 3.5 level), and it would claim that someone said something that I knew this person definitely did not say, and I would check the actual article and it turned out the journalist said that thing and the interviewee said the opposite, but the LLM thought it was the interviewee who said it.

Re: Show HN: Attabit – AI powered news

#20

Post by gordondavidf »

zipping1549 wrote: post #12

How much does this cost? It'd be awesome if something like this can be selfhosted & tweaked for my personal preference. Is there any chance that this might go opensource?

Its currently 'expensive' ($3 per day) for the individual user level to run this but once you open it up to a one to many model, that's extremely reasonable.

Recording and editing the audio alone would be a massive expense and the code does this multiple times per day all for under a dollar.

But the individual costs will surely come down as models get better. I'm not sure what the plans are yet for it but I'm hoping to make it useful to people.

Post reply on HN